mirror/freebsd
1
0
Fork 0
mirror of https://git.FreeBSD.org/src.git synced 2024-12-18 10:35:55 +00:00
Code Issues Releases Activity
0b204f9c17
freebsd/sys/kern/sched_ule.c

2187 lines
54 KiB
C
Raw Normal View History

- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*-
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* Copyright (c) 2002-2007, Jeffrey Roberson <jeff@freebsd.org>
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice unmodified, this list of conditions, and the following
* disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
* NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
* THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
Use __FBSDID().
2003-06-11 00:56:59 +00:00
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
Move HWPMC_HOOKS into its own opt_hwpmc_hooks.h file. It doesn't merit being in opt_global.h and forcing a global recompile when only a few files reference it. Approved by: re
2005-06-24 00:16:57 +00:00
#include "opt_hwpmc_hooks.h"
#include "opt_sched.h"
Turn PREEMPTION into a kernel option. Make sure that it's defined if FULL_PREEMPTION is defined. Add a runtime warning to ULE if PREEMPTION is enabled (code inspired by the PREEMPTION warning in kern_switch.c). This is a possible MT5 candidate.
2004-09-02 18:59:15 +00:00
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
#include <sys/param.h>
#include <sys/systm.h>
Update for the KDB framework: o Call kdb_backtrace() instead of backtrace().
2004-07-10 21:38:22 +00:00
#include <sys/kdb.h>
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
#include <sys/kernel.h>
#include <sys/ktr.h>
#include <sys/lock.h>
#include <sys/mutex.h>
#include <sys/proc.h>
- Add in support for KSEs with 0 slice values on the run queue. If we try to select a KSE with a slice of 0 we will update its slice and insert it onto the next queue. - Pass the KSE instead of the ksegrp into sched_slice(). This more accurately reflects the behavior of the code. Slices are granted to kses. - Add a function kseq_nice_min() which finds the smallest nice value assigned to the kseg of any KSE on the queue. - Rewrite the logic in sched_slice(). Add a large comment describing the new slice selection scheme. To summarize, slices are assigned based on the nice value. Priorities are still calculated based on the nice and interactivity of a process. Slice sizes of 0 may be granted for KSEs whos nice is 20 or futher away from the lowest nice on the run queue. Other nice values are scaled across the range [min, min+20]. This fixes ULEs bad behavior with positively niced processes.
2003-04-02 06:46:43 +00:00
#include <sys/resource.h>
- Add initial support for pinning and binding.
2003-11-04 07:45:41 +00:00
#include <sys/resourcevar.h>
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
#include <sys/sched.h>
#include <sys/smp.h>
#include <sys/sx.h>
#include <sys/sysctl.h>
#include <sys/sysproto.h>
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
#include <sys/turnstile.h>
Add user priority loaning code to support priority propagation for 1:1 threading's POSIX priority mutexes, the code is no-op unless priority-aware umtx code is committed.
2006-08-25 06:12:53 +00:00
#include <sys/umtx.h>
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
#include <sys/vmmeter.h>
#ifdef KTRACE
#include <sys/uio.h>
#include <sys/ktrace.h>
#endif
Bring a working snapshot of hwpmc(4), its associated libraries, userland utilities and documentation into -CURRENT. Bump FreeBSD_version. Reviewed by: alc, jhb (kernel changes)
2005-04-19 04:01:25 +00:00
#ifdef HWPMC_HOOKS
#include <sys/pmckern.h>
#endif
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
#include <machine/cpu.h>
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#include <machine/smp.h>
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
#ifndef PREEMPTION
#error "SCHED_ULE requires options PREEMPTION"
#endif
- In tdq_transfer() always set NEEDRESCHED when necessary regardless of the ipi settings. If NEEDRESCHED is set and an ipi is later delivered it will clear it rather than cause extra context switches. However, if we miss setting it we can have terrible latency. - In sched_bind() correctly implement bind. Also be slightly more tolerant of code which calls bind multiple times. However, we don't change binding if another call is made with a different cpu. This does not presently work with hwpmc which I believe should be changed.
2007-01-20 09:03:43 +00:00
/*
* TODO:
* Pick idle from affinity group or self group first.
* Implement pick_score.
*/
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
#define KTR_ULE 0x0 /* Enable for pickpri debugging. */
- Don't SLOT_USE() in the preempt case, sched_add() has already taken the slot for us. Previously, we would take two slots on every preempt, and setrunqueue() would fix it up for us in the non threaded case. The threaded case was simply broken. - Clean up flags, prototypes, comments.
2005-06-04 09:23:28 +00:00
/*
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* Thread scheduler specific section.
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
*/
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched {
TAILQ_ENTRY(td_sched) ts_procq; /* (j/z) Run queue. */
int ts_flags; /* (j) TSF_* flags. */
struct thread *ts_thread; /* (*) Active associated thread. */
u_char ts_rqindex; /* (j) Run queue index. */
int ts_slptime;
int ts_slice;
struct runq *ts_runq;
u_char ts_cpu; /* CPU that we have affinity for. */
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
/* The following variables are only used for pctcpu calculation */
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
int ts_ltick; /* Last tick that we were running on */
int ts_ftick; /* First tick that we were running on */
int ts_ticks; /* Tick count */
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#ifdef SMP
int ts_rltick; /* Real last tick, for affinity. */
#endif
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
/* originally from kg_sched */
- With a sleep time over 2097 seconds hzticks and slptime could end up negative. Use unsigned integers for sleep and run time so this doesn't disturb sched_interact_score(). This should fix the invalid interactive priority panics reported by several users.
2007-01-24 18:18:43 +00:00
u_int skg_slptime; /* Number of ticks we vol. slept */
u_int skg_runtime; /* Number of ticks we were running */
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
};
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
/* flags kept in ts_flags */
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#define TSF_BOUND 0x0001 /* Thread can not migrate. */
#define TSF_XFERABLE 0x0002 /* Thread was added as transferable. */
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct td_sched td_sched0;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* Cpu percentage computation macros and defines.
*
* SCHED_TICK_SECS: Number of seconds to average the cpu usage across.
* SCHED_TICK_TARG: Number of hz ticks to average the cpu usage across.
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
* SCHED_TICK_MAX: Maximum number of ticks before scaling back.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* SCHED_TICK_SHIFT: Shift factor to avoid rounding away results.
* SCHED_TICK_HZ: Compute the number of hz ticks for a given ticks count.
* SCHED_TICK_TOTAL: Gives the amount of time we've been recording ticks.
*/
#define SCHED_TICK_SECS 10
#define SCHED_TICK_TARG (hz * SCHED_TICK_SECS)
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
#define SCHED_TICK_MAX (SCHED_TICK_TARG + hz)
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_TICK_SHIFT 10
#define SCHED_TICK_HZ(ts) ((ts)->ts_ticks >> SCHED_TICK_SHIFT)
- Don't let SCHED_TICK_TOTAL() return less than hz. This can cause integer divide faults in roundup() later if it is able to return 0. For some reason this bug only shows up on my laptop and not my testboxes.
2007-01-06 12:33:43 +00:00
#define SCHED_TICK_TOTAL(ts) (max((ts)->ts_ltick - (ts)->ts_ftick, hz))
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* These macros determine priorities for non-interactive threads. They are
* assigned a priority based on their recent cpu utilization as expressed
* by the ratio of ticks to the tick total. NHALF priorities at the start
* and end of the MIN to MAX timeshare range are only reachable with negative
* or positive nice respectively.
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* PRI_RANGE: Priority range for utilization dependent priorities.
- Don't allow nice to have such a large effect on priority. This was causing poor interactive performance while unnice processes were running. The new scheme still allows nice to have an effect on priority but it is not as dramatic as the effect of the interactivity score.
2003-06-21 02:22:47 +00:00
* PRI_NRESV: Number of nice values.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* PRI_TICKS: Compute a priority in PRI_RANGE from the ticks count and total.
* PRI_NICE: Determines the part of the priority inherited from nice.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_PRI_NRESV (PRIO_MAX - PRIO_MIN)
- Remove uses of PRIO_TOTAL and replace them with SCHED_PRI_NRESV - SCHED_PRI_NRESV does not have the off by one error in PRIO_TOTAL so we do not have to account for it in the few places that we use it. Requested by: bde
2003-11-02 03:49:32 +00:00
#define SCHED_PRI_NHALF (SCHED_PRI_NRESV / 2)
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_PRI_MIN (PRI_MIN_TIMESHARE + SCHED_PRI_NHALF)
#define SCHED_PRI_MAX (PRI_MAX_TIMESHARE - SCHED_PRI_NHALF)
- Fix an off by one error in sched_pri_range. - In tdq_choose() only assert that a thread does not have too high a priority (low value) for the queue we removed it from. This will catch bugs in priority elevation. It's not a serious error for the thread to have too low a priority as we don't change queues in this case as an optimization. Reported by: kris
2007-06-15 19:33:58 +00:00
#define SCHED_PRI_RANGE (SCHED_PRI_MAX - SCHED_PRI_MIN)
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_PRI_TICKS(ts) \
(SCHED_TICK_HZ((ts)) / \
- Fix the sched_priority() invalid priority bugs. Use roundup() instead of max() when computing the divisor in SCHED_TICK_PRI(). This prevents cases where rounding down would allow the quotient to exceed SCHED_PRI_RANGE. - Garbage collect some unused flags and fields. - Replace TDF_HOLD with sched_pin_td()/sched_unpin_td() since it simply duplicated this functionality. - Re-enable the rebalancer by default and fix the sysctl so it can be modified.
2007-01-06 08:44:13 +00:00
(roundup(SCHED_TICK_TOTAL((ts)), SCHED_PRI_RANGE) / SCHED_PRI_RANGE))
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_PRI_NICE(nice) (nice)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* These determine the interactivity of a process. Interactivity differs from
* cpu utilization in that it expresses the voluntary time slept vs time ran
* while cpu utilization includes all time not running. This more accurately
* models the intent of the thread.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*
- Add a new variable 'kg_runtime' that tracks the amount of time we've run. - Use the ratio of kg_runtime / kg_slptime to determine our dynamic priority. - Scale kg_runtime and kg_slptime back when the sum of the two exceeds SCHED_SLP_RUN_MAX. This allows us to slowly forget old behavior. - Scale back the runtime and slptime in fork so that the new process has the same ratio but much less accumulated time. This causes new behavior to be noticed more quickly.
2003-02-10 14:03:45 +00:00
* SLP_RUN_MAX: Maximum amount of sleep time + run time we'll accumulate
* before throttling back.
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
* SLP_RUN_FORK: Maximum slp+run time to inherit at fork time.
- Fix the maximum slice value. I accidentally checked in a value of '2' which meant no process would run for longer than 20ms. - Slightly redo the interactivity scorer. It follows the same algorithm but in a slightly more correct way. Previously values above half were incorrect. - Lower the interactivity threshold to 20. It seems that in testing non- interactive tasks are hardly ever near there and expensive interactive tasks can sometimes surpass it. This area needs more testing. - Remove an unnecessary KTR. - Fix a case where an idle thread that had an elevated priority due to priority prop. would be placed back on the idle queue. - Delay setting NEEDRESCHED until userret() for threads that haad their priority elevated while in kernel. This gives us the same context switch optimization as SCHED_4BSD. - Limit the child's slice to 1 in sched_fork_kse() so we detect its behavior more quickly. - Inhert some of the run/slp time from the child in sched_exit_ksegrp(). - Redo some of the priority comparisons so they are more clear. - Throttle the frequency of sched_pctcpu_update() so that rounding errors do not make it invalid.
2003-06-15 02:18:29 +00:00
* INTERACT_MAX: Maximum interactivity value. Smaller is better.
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
* INTERACT_THRESH: Threshhold for placement on the current runq.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#define SCHED_SLP_RUN_MAX ((hz * 5) << SCHED_TICK_SHIFT)
#define SCHED_SLP_RUN_FORK ((hz / 2) << SCHED_TICK_SHIFT)
- Fix the maximum slice value. I accidentally checked in a value of '2' which meant no process would run for longer than 20ms. - Slightly redo the interactivity scorer. It follows the same algorithm but in a slightly more correct way. Previously values above half were incorrect. - Lower the interactivity threshold to 20. It seems that in testing non- interactive tasks are hardly ever near there and expensive interactive tasks can sometimes surpass it. This area needs more testing. - Remove an unnecessary KTR. - Fix a case where an idle thread that had an elevated priority due to priority prop. would be placed back on the idle queue. - Delay setting NEEDRESCHED until userret() for threads that haad their priority elevated while in kernel. This gives us the same context switch optimization as SCHED_4BSD. - Limit the child's slice to 1 in sched_fork_kse() so we detect its behavior more quickly. - Inhert some of the run/slp time from the child in sched_exit_ksegrp(). - Redo some of the priority comparisons so they are more clear. - Throttle the frequency of sched_pctcpu_update() so that rounding errors do not make it invalid.
2003-06-15 02:18:29 +00:00
#define SCHED_INTERACT_MAX (100)
#define SCHED_INTERACT_HALF (SCHED_INTERACT_MAX / 2)
- The non iterative algorithm for interact_update was broken due to rounding errors. This was the source of the majority of the interactivity problems. Reintroduce the old algorithm and its XXX. - Up the interactivity threshold to 30. It really could stand to be even a tiny bit higher. - Let the sleep and run time accumulate up to 5 seconds of history rather than two. This helps stop XFree86 from becoming non-interactive during bursts of activity.
2003-10-16 08:17:43 +00:00
#define SCHED_INTERACT_THRESH (30)
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* tickincr: Converts a stathz tick into a hz domain scaled by
* the shift factor. Without the shift the error rate
* due to rounding would be unacceptably high.
* realstathz: stathz is sometimes 0 and run off of hz.
* sched_slice: Runtime of each thread before rescheduling.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static int sched_interact = SCHED_INTERACT_THRESH;
static int realstathz;
static int tickincr;
static int sched_slice;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* tdq - per processor runqs and statistics.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
struct runq tdq_idle; /* Queue of IDLE threads. */
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
struct runq tdq_timeshare; /* timeshare run queue. */
struct runq tdq_realtime; /* real-time run queue. */
- Change types for necent runq additions to u_char rather than int. - Fix these types in ULE as well. This fixes bugs in priority index calculations in certain edge cases. (int)-1 % 64 != (uint)-1 % 64. Reported by: kkenn using pho's stress2.
2007-02-08 01:52:25 +00:00
u_char tdq_idx; /* Current insert index. */
u_char tdq_ridx; /* Current removal index. */
short tdq_flags; /* Thread queue flags */
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
int tdq_load; /* Aggregate load. */
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
#ifdef SMP
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
int tdq_transferable;
LIST_ENTRY(tdq) tdq_siblings; /* Next in tdq group. */
struct tdq_group *tdq_group; /* Our processor group. */
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
#else
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
int tdq_sysload; /* For loadavg, !ITHD load. */
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
#endif
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
};
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#define TDQF_BUSY 0x0001 /* Queue is marked as busy */
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
#ifdef SMP
/*
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* tdq groups are groups of processors which can cheaply share threads. When
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
* one processor in the group goes idle it will check the runqs of the other
* processors in its group prior to halting and waiting for an interrupt.
* These groups are suitable for SMT (Symetric Multi-Threading) and not NUMA.
* In a numa environment we'd want an idle bitmap per group and a two tiered
* load balancer.
*/
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq_group {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
int tdg_cpus; /* Count of CPUs in this tdq group. */
cpumask_t tdg_cpumask; /* Mask of cpus in this group. */
cpumask_t tdg_idlemask; /* Idle cpus in this group. */
cpumask_t tdg_mask; /* Bit mask for first cpu. */
int tdg_load; /* Total load of this group. */
int tdg_transferable; /* Transferable load of this group. */
LIST_HEAD(, tdq) tdg_members; /* Linked list of all members. */
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
};
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#define SCHED_AFFINITY_DEFAULT (hz / 100)
#define SCHED_AFFINITY(ts) ((ts)->ts_rltick > ticks - affinity)
/*
* Run-time tunables.
*/
- Disable the long-term load balancer. I believe that steal_busy works better and gives more predictable results.
2007-01-20 21:24:05 +00:00
static int rebalance = 0;
- Define TDQ_ID() for the !SMP case. - Default pick_pri to off. It is not faster in most cases.
2007-06-05 02:53:51 +00:00
static int pick_pri = 0;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
static int affinity;
static int tryself = 1;
static int tryselfidle = 1;
static int ipi_ast = 0;
static int ipi_preempt = 1;
static int ipi_thresh = PRI_MIN_KERN;
static int steal_htt = 1;
static int steal_busy = 1;
static int busy_thresh = 4;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
static int topology = 0;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
* One thread queue per processor.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
static volatile cpumask_t tdq_idle;
static volatile cpumask_t tdq_busy;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
static int tdg_maxid;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct tdq tdq_cpu[MAXCPU];
static struct tdq_group tdq_groups[MAXCPU];
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
static int bal_tick;
static int gbal_tick;
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
static int balance_groups;
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
#define TDQ_SELF() (&tdq_cpu[PCPU_GET(cpuid)])
#define TDQ_CPU(x) (&tdq_cpu[(x)])
#define TDQ_ID(x) ((x) - tdq_cpu)
#define TDQ_GROUP(x) (&tdq_groups[(x)])
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
#else /* !SMP */
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct tdq tdq_cpu;
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
- Define TDQ_ID() for the !SMP case. - Default pick_pri to off. It is not faster in most cases.
2007-06-05 02:53:51 +00:00
#define TDQ_ID(x) (0)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
#define TDQ_SELF() (&tdq_cpu)
#define TDQ_CPU(x) (&tdq_cpu)
- Use ksq_load as the authoritive count of kses on the pair of kseqs for sched_runnable() et all. - Remove some dead code in sched_clock(). - Define two macros KSEQ_SELF() and KSEQ_CPU() for getting the kseq of the current cpu or some alternate cpu. - Start introducing kseq_() functions, such as kseq_choose() and kseq_setup().
2003-01-29 07:00:51 +00:00
#endif
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
static void sched_priority(struct thread *);
- Don't SLOT_USE() in the preempt case, sched_add() has already taken the slot for us. Previously, we would take two slots on every preempt, and setrunqueue() would fix it up for us in the non threaded case. The threaded case was simply broken. - Clean up flags, prototypes, comments.
2005-06-04 09:23:28 +00:00
static void sched_thread_priority(struct thread *, u_char);
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
static int sched_interact_score(struct thread *);
static void sched_interact_update(struct thread *);
static void sched_interact_fork(struct thread *);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static void sched_pctcpu_update(struct td_sched *);
- Fix the sched_priority() invalid priority bugs. Use roundup() instead of max() when computing the divisor in SCHED_TICK_PRI(). This prevents cases where rounding down would allow the quotient to exceed SCHED_PRI_RANGE. - Garbage collect some unused flags and fields. - Replace TDF_HOLD with sched_pin_td()/sched_unpin_td() since it simply duplicated this functionality. - Re-enable the rebalancer by default and fix the sysctl so it can be modified.
2007-01-06 08:44:13 +00:00
static inline void sched_pin_td(struct thread *td);
static inline void sched_unpin_td(struct thread *td);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
/* Operations on per processor queues */
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct td_sched * tdq_choose(struct tdq *);
static void tdq_setup(struct tdq *);
static void tdq_load_add(struct tdq *, struct td_sched *);
static void tdq_load_rem(struct tdq *, struct td_sched *);
static __inline void tdq_runq_add(struct tdq *, struct td_sched *, int);
static __inline void tdq_runq_rem(struct tdq *, struct td_sched *);
void tdq_print(int cpu);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void runq_print(struct runq *rq);
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
#ifdef SMP
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
static int tdq_pickidle(struct tdq *, struct td_sched *);
static int tdq_pickpri(struct tdq *, struct td_sched *, int);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct td_sched *runq_steal(struct runq *);
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
static void sched_balance(void);
static void sched_balance_groups(void);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static void sched_balance_group(struct tdq_group *);
static void sched_balance_pair(struct tdq *, struct tdq *);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
static void sched_smp_tick(struct thread *);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static void tdq_move(struct tdq *, int);
static int tdq_idled(struct tdq *);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
static void tdq_notify(struct td_sched *);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct td_sched *tdq_steal(struct tdq *, int);
- Fix the sched_priority() invalid priority bugs. Use roundup() instead of max() when computing the divisor in SCHED_TICK_PRI(). This prevents cases where rounding down would allow the quotient to exceed SCHED_PRI_RANGE. - Garbage collect some unused flags and fields. - Replace TDF_HOLD with sched_pin_td()/sched_unpin_td() since it simply duplicated this functionality. - Re-enable the rebalancer by default and fix the sysctl so it can be modified.
2007-01-06 08:44:13 +00:00
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#define THREAD_CAN_MIGRATE(td) ((td)->td_pinned == 0)
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
#endif
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void sched_setup(void *dummy);
SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL)
static void sched_initticks(void *dummy);
SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL)
- Fix the sched_priority() invalid priority bugs. Use roundup() instead of max() when computing the divisor in SCHED_TICK_PRI(). This prevents cases where rounding down would allow the quotient to exceed SCHED_PRI_RANGE. - Garbage collect some unused flags and fields. - Replace TDF_HOLD with sched_pin_td()/sched_unpin_td() since it simply duplicated this functionality. - Re-enable the rebalancer by default and fix the sysctl so it can be modified.
2007-01-06 08:44:13 +00:00
static inline void
sched_pin_td(struct thread *td)
{
td->td_pinned++;
}
static inline void
sched_unpin_td(struct thread *td)
{
td->td_pinned--;
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void
runq_print(struct runq *rq)
{
struct rqhead *rqh;
struct td_sched *ts;
int pri;
int j;
int i;
for (i = 0; i < RQB_LEN; i++) {
printf("\t\trunq bits %d 0x%zx\n",
i, rq->rq_status.rqb_bits[i]);
for (j = 0; j < RQB_BPW; j++)
if (rq->rq_status.rqb_bits[i] & (1ul << j)) {
pri = j + (i << RQB_L2BPW);
rqh = &rq->rq_queues[pri];
TAILQ_FOREACH(ts, rqh, ts_procq) {
printf("\t\t\ttd %p(%s) priority %d rqindex %d pri %d\n",
ts->ts_thread, ts->ts_thread->td_proc->p_comm, ts->ts_thread->td_priority, ts->ts_rqindex, pri);
}
}
}
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_print(int cpu)
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
- Keep seperate statistics and run queues for different scheduling classes. - Treat each class specially in kseq_{choose,add,rem}. Let the rest of the code be less aware of scheduling classes. - Skip the interactivity calculation for non TIMESHARE ksegrps. - Move slice and runq selection into kseq_add(). Uninline it now that it's big.
2003-04-03 00:29:28 +00:00
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq = TDQ_CPU(cpu);
- Keep seperate statistics and run queues for different scheduling classes. - Treat each class specially in kseq_{choose,add,rem}. Let the rest of the code be less aware of scheduling classes. - Skip the interactivity calculation for non TIMESHARE ksegrps. - Move slice and runq selection into kseq_add(). Uninline it now that it's big.
2003-04-03 00:29:28 +00:00
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
printf("tdq:\n");
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
printf("\tload: %d\n", tdq->tdq_load);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
printf("\ttimeshare idx: %d\n", tdq->tdq_idx);
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
printf("\ttimeshare ridx: %d\n", tdq->tdq_ridx);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
printf("\trealtime runq:\n");
runq_print(&tdq->tdq_realtime);
printf("\ttimeshare runq:\n");
runq_print(&tdq->tdq_timeshare);
printf("\tidle runq:\n");
runq_print(&tdq->tdq_idle);
- Remove the ksq_loads[] array. We are only interested in three counts, the total load, the timeshare load, and the number of threads that can be migrated to another cpu. Account for these seperately. - Introduce a KSE_CAN_MIGRATE() macro which determines whether or not a KSE can be migrated to another CPU. Currently, this only checks to see if we're an interrupt handler. Eventually this will also be used to support CPU binding.
2003-11-02 10:56:48 +00:00
#ifdef SMP
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
printf("\tload transferable: %d\n", tdq->tdq_transferable);
- Remove the ksq_loads[] array. We are only interested in three counts, the total load, the timeshare load, and the number of threads that can be migrated to another cpu. Account for these seperately. - Introduce a KSE_CAN_MIGRATE() macro which determines whether or not a KSE can be migrated to another CPU. Currently, this only checks to see if we're an interrupt handler. Eventually this will also be used to support CPU binding.
2003-11-02 10:56:48 +00:00
#endif
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
}
- Keep seperate statistics and run queues for different scheduling classes. - Treat each class specially in kseq_{choose,add,rem}. Let the rest of the code be less aware of scheduling classes. - Skip the interactivity calculation for non TIMESHARE ksegrps. - Move slice and runq selection into kseq_add(). Uninline it now that it's big.
2003-04-03 00:29:28 +00:00
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
static __inline void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_runq_add(struct tdq *tdq, struct td_sched *ts, int flags)
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
{
#ifdef SMP
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (THREAD_CAN_MIGRATE(ts->ts_thread)) {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_transferable++;
tdq->tdq_group->tdg_transferable++;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts->ts_flags |= TSF_XFERABLE;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (tdq->tdq_transferable >= busy_thresh &&
(tdq->tdq_flags & TDQF_BUSY) == 0) {
tdq->tdq_flags |= TDQF_BUSY;
atomic_set_int(&tdq_busy, 1 << TDQ_ID(tdq));
}
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
#endif
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (ts->ts_runq == &tdq->tdq_timeshare) {
- Change types for necent runq additions to u_char rather than int. - Fix these types in ULE as well. This fixes bugs in priority index calculations in certain edge cases. (int)-1 % 64 != (uint)-1 % 64. Reported by: kkenn using pho's stress2.
2007-02-08 01:52:25 +00:00
u_char pri;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
pri = ts->ts_thread->td_priority;
KASSERT(pri <= PRI_MAX_TIMESHARE && pri >= PRI_MIN_TIMESHARE,
("Invalid priority %d on timeshare runq", pri));
/*
* This queue contains only priorities between MIN and MAX
* realtime. Use the whole queue to represent these values.
*/
#define TS_RQ_PPQ (((PRI_MAX_TIMESHARE - PRI_MIN_TIMESHARE) + 1) / RQ_NQS)
if ((flags & SRQ_BORROWING) == 0) {
pri = (pri - PRI_MIN_TIMESHARE) / TS_RQ_PPQ;
pri = (pri + tdq->tdq_idx) % RQ_NQS;
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
/*
* This effectively shortens the queue by one so we
* can have a one slot difference between idx and
* ridx while we wait for threads to drain.
*/
if (tdq->tdq_ridx != tdq->tdq_idx &&
pri == tdq->tdq_ridx)
- Cast the intermediate value in priority computtion back down to unsigned char. Weirdly, casting the 1 constant to u_char still produces a signed integer result that is then used in the % computation. This avoids that mess all together and causes a 0 pri to turn into 255 % 64 as we expect. Reported by: kkenn (about 4 times, thanks)
2007-03-17 18:13:32 +00:00
pri = (unsigned char)(pri - 1) % RQ_NQS;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
} else
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
pri = tdq->tdq_ridx;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
runq_add_pri(ts->ts_runq, ts, pri, flags);
} else
runq_add(ts->ts_runq, ts, flags);
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
}
static __inline void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_runq_rem(struct tdq *tdq, struct td_sched *ts)
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
{
#ifdef SMP
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (ts->ts_flags & TSF_XFERABLE) {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_transferable--;
tdq->tdq_group->tdg_transferable--;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts->ts_flags &= ~TSF_XFERABLE;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (tdq->tdq_transferable < busy_thresh &&
(tdq->tdq_flags & TDQF_BUSY)) {
atomic_clear_int(&tdq_busy, 1 << TDQ_ID(tdq));
tdq->tdq_flags &= ~TDQF_BUSY;
}
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
#endif
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
if (ts->ts_runq == &tdq->tdq_timeshare) {
if (tdq->tdq_idx != tdq->tdq_ridx)
runq_remove_idx(ts->ts_runq, ts, &tdq->tdq_ridx);
else
runq_remove_idx(ts->ts_runq, ts, NULL);
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
/*
* For timeshare threads we update the priority here so
* the priority reflects the time we've been sleeping.
*/
ts->ts_ltick = ticks;
sched_pctcpu_update(ts);
sched_priority(ts->ts_thread);
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
} else
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
runq_remove(ts->ts_runq, ts);
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
static void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_load_add(struct tdq *tdq, struct td_sched *ts)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
{
- Remove the ksq_loads[] array. We are only interested in three counts, the total load, the timeshare load, and the number of threads that can be migrated to another cpu. Account for these seperately. - Introduce a KSE_CAN_MIGRATE() macro which determines whether or not a KSE can be migrated to another CPU. Currently, this only checks to see if we're an interrupt handler. Eventually this will also be used to support CPU binding.
2003-11-02 10:56:48 +00:00
int class;
- When a new thread is added to a kseq the load is incremented prior to adding it to the nice tables. Therefore, in kseq_add_nice, we should keep in mind that the load will be 1 if we are the only thread, and not 0. - Assert that the sched lock is held in all the appropriate places. - Increase the scope of the sched lock in sched_pctcpu_update(). - Hold the sched lock in sched_runnable(). It is not held by the caller.
2003-06-08 00:47:33 +00:00
mtx_assert(&sched_lock, MA_OWNED);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
class = PRI_BASE(ts->ts_thread->td_pri_class);
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_load++;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
CTR2(KTR_SCHED, "cpu %jd load: %d", TDQ_ID(tdq), tdq->tdq_load);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (class != PRI_ITHD &&
(ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
#ifdef SMP
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_group->tdg_load++;
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
#else
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_sysload++;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
#endif
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
- Keep seperate statistics and run queues for different scheduling classes. - Treat each class specially in kseq_{choose,add,rem}. Let the rest of the code be less aware of scheduling classes. - Skip the interactivity calculation for non TIMESHARE ksegrps. - Move slice and runq selection into kseq_add(). Uninline it now that it's big.
2003-04-03 00:29:28 +00:00
static void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_load_rem(struct tdq *tdq, struct td_sched *ts)
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
{
- Remove the ksq_loads[] array. We are only interested in three counts, the total load, the timeshare load, and the number of threads that can be migrated to another cpu. Account for these seperately. - Introduce a KSE_CAN_MIGRATE() macro which determines whether or not a KSE can be migrated to another CPU. Currently, this only checks to see if we're an interrupt handler. Eventually this will also be used to support CPU binding.
2003-11-02 10:56:48 +00:00
int class;
- When a new thread is added to a kseq the load is incremented prior to adding it to the nice tables. Therefore, in kseq_add_nice, we should keep in mind that the load will be 1 if we are the only thread, and not 0. - Assert that the sched lock is held in all the appropriate places. - Increase the scope of the sched lock in sched_pctcpu_update(). - Hold the sched lock in sched_runnable(). It is not held by the caller.
2003-06-08 00:47:33 +00:00
mtx_assert(&sched_lock, MA_OWNED);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
class = PRI_BASE(ts->ts_thread->td_pri_class);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (class != PRI_ITHD &&
(ts->ts_thread->td_proc->p_flag & P_NOLOAD) == 0)
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
#ifdef SMP
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_group->tdg_load--;
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
#else
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_sysload--;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
#endif
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdq->tdq_load--;
CTR1(KTR_SCHED, "load: %d", tdq->tdq_load);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts->ts_runq = NULL;
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
#ifdef SMP
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
static void
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
sched_smp_tick(struct thread *td)
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
{
struct tdq *tdq;
tdq = TDQ_SELF();
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (rebalance) {
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
if (ticks >= bal_tick)
sched_balance();
if (ticks >= gbal_tick && balance_groups)
sched_balance_groups();
}
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
td->td_sched->ts_rltick = ticks;
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
}
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
/*
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
* sched_balance is a simple CPU load balancing algorithm. It operates by
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
* finding the least loaded and most loaded cpu and equalizing their load
* by migrating some processes.
*
* Dealing only with two CPUs at a time has two advantages. Firstly, most
* installations will only have 2 cpus. Secondly, load balancing too much at
* once can have an unpleasant effect on the system. The scheduler rarely has
* enough information to make perfect decisions. So this algorithm chooses
* algorithm simplicity and more gradual effects on load in larger systems.
*
* It could be improved by considering the priorities and slices assigned to
* each task prior to balancing them. There are many pathological cases with
* any approach and so the semi random algorithm below may work as well as any.
*
*/
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
static void
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
sched_balance(void)
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq_group *high;
struct tdq_group *low;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
struct tdq_group *tdg;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
int cnt;
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
int i;
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
bal_tick = ticks + (random() % (hz * 2));
- If smp is not started yet don't try to load balance or we'll put threads on cpus that aren't running yet.
2003-06-28 08:24:42 +00:00
if (smp_started == 0)
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
return;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
low = high = NULL;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
i = random() % (tdg_maxid + 1);
for (cnt = 0; cnt <= tdg_maxid; cnt++) {
tdg = TDQ_GROUP(i);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
* Find the CPU with the highest load that has some
* threads to transfer.
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
*/
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if ((high == NULL || tdg->tdg_load > high->tdg_load)
&& tdg->tdg_transferable)
high = tdg;
if (low == NULL || tdg->tdg_load < low->tdg_load)
low = tdg;
if (++i > tdg_maxid)
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
i = 0;
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
}
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
if (low != NULL && high != NULL && high != low)
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
sched_balance_pair(LIST_FIRST(&high->tdg_members),
LIST_FIRST(&low->tdg_members));
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
}
static void
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
sched_balance_groups(void)
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
{
int i;
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
gbal_tick = ticks + (random() % (hz * 2));
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
mtx_assert(&sched_lock, MA_OWNED);
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
if (smp_started)
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
for (i = 0; i <= tdg_maxid; i++)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
sched_balance_group(TDQ_GROUP(i));
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
}
static void
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
sched_balance_group(struct tdq_group *tdg)
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
struct tdq *high;
struct tdq *low;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
int load;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (tdg->tdg_transferable == 0)
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
return;
low = NULL;
high = NULL;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
load = tdq->tdq_load;
if (high == NULL || load > high->tdq_load)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
high = tdq;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (low == NULL || load < low->tdq_load)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
low = tdq;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
}
if (high != NULL && low != NULL && high != low)
sched_balance_pair(high, low);
}
static void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
sched_balance_pair(struct tdq *high, struct tdq *low)
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
{
int transferable;
int high_load;
int low_load;
int move;
int diff;
int i;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
* If we're transfering within a group we have to use this specific
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* tdq's transferable count, otherwise we can steal from other members
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
* of the group.
*/
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (high->tdq_group == low->tdq_group) {
transferable = high->tdq_transferable;
high_load = high->tdq_load;
low_load = low->tdq_load;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
} else {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
transferable = high->tdq_group->tdg_transferable;
high_load = high->tdq_group->tdg_load;
low_load = low->tdq_group->tdg_load;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
}
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
if (transferable == 0)
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
return;
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
/*
* Determine what the imbalance is and then adjust that to how many
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
* threads we actually have to give up (transferable).
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
*/
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
diff = high_load - low_load;
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
move = diff / 2;
if (diff & 0x1)
move++;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
move = min(move, transferable);
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
for (i = 0; i < move; i++)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_move(high, TDQ_ID(low));
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
return;
}
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
static void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_move(struct tdq *from, int cpu)
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
struct tdq *to;
struct td_sched *ts;
tdq = from;
to = TDQ_CPU(cpu);
ts = tdq_steal(tdq, 1);
if (ts == NULL) {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
struct tdq_group *tdg;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg = tdq->tdq_group;
LIST_FOREACH(tdq, &tdg->tdg_members, tdq_siblings) {
if (tdq == from || tdq->tdq_transferable == 0)
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
continue;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = tdq_steal(tdq, 1);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
break;
}
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (ts == NULL)
panic("tdq_move: No threads available with a "
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
"transferable count of %d\n",
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg->tdg_transferable);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (tdq == to)
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
return;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
sched_rem(ts->ts_thread);
ts->ts_cpu = cpu;
sched_pin_td(ts->ts_thread);
sched_add(ts->ts_thread, SRQ_YIELDING);
sched_unpin_td(ts->ts_thread);
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
}
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
static int
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_idled(struct tdq *tdq)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
{
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
struct tdq_group *tdg;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *steal;
struct td_sched *ts;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg = tdq->tdq_group;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
* If we're in a cpu group, try and steal threads from another cpu in
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
* the group before idling.
*/
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (steal_htt && tdg->tdg_cpus > 1 && tdg->tdg_transferable) {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
LIST_FOREACH(steal, &tdg->tdg_members, tdq_siblings) {
if (steal == tdq || steal->tdq_transferable == 0)
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
continue;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = tdq_steal(steal, 0);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (ts)
goto steal;
}
}
if (steal_busy) {
while (tdq_busy) {
int cpu;
cpu = ffs(tdq_busy);
if (cpu == 0)
break;
cpu--;
steal = TDQ_CPU(cpu);
if (steal->tdq_transferable == 0)
continue;
ts = tdq_steal(steal, 1);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (ts == NULL)
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
continue;
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR5(KTR_ULE,
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
"tdq_idled: stealing td %p(%s) pri %d from %d busy 0x%X",
ts->ts_thread, ts->ts_thread->td_proc->p_comm,
ts->ts_thread->td_priority, cpu, tdq_busy);
goto steal;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
}
/*
* We only set the idled bit when all of the cpus in the group are
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* idle. Otherwise we could get into a situation where a thread bounces
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
* back and forth between two idle cores on seperate physical CPUs.
*/
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg->tdg_idlemask |= PCPU_GET(cpumask);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (tdg->tdg_idlemask == tdg->tdg_cpumask)
atomic_set_int(&tdq_idle, tdg->tdg_mask);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
return (1);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
steal:
sched_rem(ts->ts_thread);
ts->ts_cpu = PCPU_GET(cpuid);
sched_pin_td(ts->ts_thread);
sched_add(ts->ts_thread, SRQ_YIELDING);
sched_unpin_td(ts->ts_thread);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
return (0);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
static void
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
tdq_notify(struct td_sched *ts)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
{
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
struct thread *ctd;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
struct pcpu *pcpu;
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
int cpri;
int pri;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
int cpu;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
cpu = ts->ts_cpu;
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
pri = ts->ts_thread->td_priority;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
pcpu = pcpu_find(cpu);
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
ctd = pcpu->pc_curthread;
cpri = ctd->td_priority;
- In tdq_transfer() always set NEEDRESCHED when necessary regardless of the ipi settings. If NEEDRESCHED is set and an ipi is later delivered it will clear it rather than cause extra context switches. However, if we miss setting it we can have terrible latency. - In sched_bind() correctly implement bind. Also be slightly more tolerant of code which calls bind multiple times. However, we don't change binding if another call is made with a different cpu. This does not presently work with hwpmc which I believe should be changed.
2007-01-20 09:03:43 +00:00
/*
* If our priority is not better than the current priority there is
* nothing to do.
*/
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
if (pri > cpri)
- In tdq_transfer() always set NEEDRESCHED when necessary regardless of the ipi settings. If NEEDRESCHED is set and an ipi is later delivered it will clear it rather than cause extra context switches. However, if we miss setting it we can have terrible latency. - In sched_bind() correctly implement bind. Also be slightly more tolerant of code which calls bind multiple times. However, we don't change binding if another call is made with a different cpu. This does not presently work with hwpmc which I believe should be changed.
2007-01-20 09:03:43 +00:00
return;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
/*
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
* Always IPI idle.
*/
if (cpri > PRI_MIN_IDLE)
goto sendipi;
/*
* If we're realtime or better and there is timeshare or worse running
* send an IPI.
*/
if (pri < PRI_MAX_REALTIME && cpri > PRI_MAX_REALTIME)
goto sendipi;
/*
* Otherwise only IPI if we exceed the threshold.
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
*/
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
if (pri > ipi_thresh)
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
return;
- Implement much more intelligent ipi sending. This algorithm tries to minimize IPIs and rescheduling when scheduling like tasks while keeping latency low for important threads. 1) An idle thread is running. 2) The current thread is worse than realtime and the new thread is better than realtime. Realtime to realtime doesn't preempt. 3) The new thread's priority is less than the threshold.
2007-01-25 23:51:59 +00:00
sendipi:
ctd->td_flags |= TDF_NEEDRESCHED;
if (cpri < PRI_MIN_IDLE) {
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
if (ipi_ast)
ipi_selected(1 << cpu, IPI_AST);
else if (ipi_preempt)
ipi_selected(1 << cpu, IPI_PREEMPT);
} else
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
ipi_selected(1 << cpu, IPI_PREEMPT);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct td_sched *
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
runq_steal(struct runq *rq)
{
struct rqhead *rqh;
struct rqbits *rqb;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched *ts;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
int word;
int bit;
mtx_assert(&sched_lock, MA_OWNED);
rqb = &rq->rq_status;
for (word = 0; word < RQB_LEN; word++) {
if (rqb->rqb_bits[word] == 0)
continue;
for (bit = 0; bit < RQB_BPW; bit++) {
rqb_bits[] may be an int64_t (eg: on alpha, and recently on amd64). Be sure to shift (long)1 << 33 and higher, not (int)1. Otherwise bad things happen(TM). This is why beast.freebsd.org paniced with ULE. Reviewed by: jeff
2003-12-07 09:57:51 +00:00
if ((rqb->rqb_bits[word] & (1ul << bit)) == 0)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
continue;
rqh = &rq->rq_queues[bit + (word << RQB_L2BPW)];
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
TAILQ_FOREACH(ts, rqh, ts_procq) {
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (THREAD_CAN_MIGRATE(ts->ts_thread))
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
return (ts);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
}
}
return (NULL);
}
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct td_sched *
tdq_steal(struct tdq *tdq, int stealidle)
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched *ts;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
* Steal from next first to try to get a non-interactive task that
* may not have run for a while.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* XXX Need to effect steal order for timeshare threads.
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if ((ts = runq_steal(&tdq->tdq_realtime)) != NULL)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
return (ts);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if ((ts = runq_steal(&tdq->tdq_timeshare)) != NULL)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
return (ts);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
if (stealidle)
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
return (runq_steal(&tdq->tdq_idle));
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
return (NULL);
}
int
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
tdq_pickidle(struct tdq *tdq, struct td_sched *ts)
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
{
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
struct tdq_group *tdg;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
int self;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
int cpu;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
self = PCPU_GET(cpuid);
- In kseq_transfer() return if smp has not been started. - In sched_add(), do the idle check prior to the transfer check so that we don't try to transfer load from an idle cpu. This fixes panics caused by IPIs on UP machines running SMP kernels. Reported/Debugged by: seanc
2003-12-20 14:03:14 +00:00
if (smp_started == 0)
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
return (self);
- In kseq_transfer() return if smp has not been started. - In sched_add(), do the idle check prior to the transfer check so that we don't try to transfer load from an idle cpu. This fixes panics caused by IPIs on UP machines running SMP kernels. Reported/Debugged by: seanc
2003-12-20 14:03:14 +00:00
/*
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
* If the current CPU has idled, just run it here.
- In kseq_transfer() return if smp has not been started. - In sched_add(), do the idle check prior to the transfer check so that we don't try to transfer load from an idle cpu. This fixes panics caused by IPIs on UP machines running SMP kernels. Reported/Debugged by: seanc
2003-12-20 14:03:14 +00:00
*/
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if ((tdq->tdq_group->tdg_idlemask & PCPU_GET(cpumask)) != 0)
return (self);
/*
* Try the last group we ran on.
*/
tdg = TDQ_CPU(ts->ts_cpu)->tdq_group;
cpu = ffs(tdg->tdg_idlemask);
if (cpu)
return (cpu - 1);
/*
* Search for an idle group.
*/
cpu = ffs(tdq_idle);
if (cpu)
return (cpu - 1);
/*
* XXX If there are no idle groups, check for an idle core.
*/
/*
* No idle CPUs?
*/
return (self);
}
static int
tdq_pickpri(struct tdq *tdq, struct td_sched *ts, int flags)
{
struct pcpu *pcpu;
int lowpri;
int lowcpu;
int lowload;
int load;
int self;
int pri;
int cpu;
self = PCPU_GET(cpuid);
if (smp_started == 0)
return (self);
pri = ts->ts_thread->td_priority;
/*
* Regardless of affinity, if the last cpu is idle send it there.
*/
pcpu = pcpu_find(ts->ts_cpu);
if (pcpu->pc_curthread->td_priority > PRI_MIN_IDLE) {
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR5(KTR_ULE,
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
"ts_cpu %d idle, ltick %d ticks %d pri %d curthread %d",
ts->ts_cpu, ts->ts_rltick, ticks, pri,
pcpu->pc_curthread->td_priority);
return (ts->ts_cpu);
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
}
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
/*
* If we have affinity, try to place it on the cpu we last ran on.
*/
if (SCHED_AFFINITY(ts) && pcpu->pc_curthread->td_priority > pri) {
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR5(KTR_ULE,
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
"affinity for %d, ltick %d ticks %d pri %d curthread %d",
ts->ts_cpu, ts->ts_rltick, ticks, pri,
pcpu->pc_curthread->td_priority);
return (ts->ts_cpu);
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
}
/*
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
* Try ourself first; If we're running something lower priority this
* may have some locality with the waking thread and execute faster
* here.
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
*/
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (tryself) {
/*
* If we're being awoken by an interrupt thread or the waker
* is going right to sleep run here as well.
*/
if ((TDQ_SELF()->tdq_load == 1) && (flags & SRQ_YIELDING ||
curthread->td_pri_class == PRI_ITHD)) {
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR2(KTR_ULE, "tryself load %d flags %d",
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
TDQ_SELF()->tdq_load, flags);
return (self);
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
}
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
/*
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
* Look for an idle group.
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
*/
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR1(KTR_ULE, "tdq_idle %X", tdq_idle);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
cpu = ffs(tdq_idle);
if (cpu)
return (cpu - 1);
if (tryselfidle && pri < curthread->td_priority) {
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR1(KTR_ULE, "tryself %d",
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
curthread->td_priority);
return (self);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
/*
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
* Now search for the cpu running the lowest priority thread with
* the least load.
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
*/
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
lowload = 0;
lowpri = lowcpu = 0;
for (cpu = 0; cpu <= mp_maxid; cpu++) {
if (CPU_ABSENT(cpu))
continue;
pcpu = pcpu_find(cpu);
pri = pcpu->pc_curthread->td_priority;
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR4(KTR_ULE,
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
"cpu %d pri %d lowcpu %d lowpri %d",
cpu, pri, lowcpu, lowpri);
if (pri < lowpri)
continue;
load = TDQ_CPU(cpu)->tdq_load;
if (lowpri && lowpri == pri && load > lowload)
continue;
lowpri = pri;
lowcpu = cpu;
lowload = load;
}
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
return (lowcpu);
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
}
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#endif /* SMP */
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
- When stealing a kse in kseq_move() ignore the current kseq's min nice value. We want to steal any thread, even one that is not given a slice on its current queue.
2003-07-08 06:19:40 +00:00
/*
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
* Pick the highest priority task we have and return it.
- When stealing a kse in kseq_move() ignore the current kseq's min nice value. We want to steal any thread, even one that is not given a slice on its current queue.
2003-07-08 06:19:40 +00:00
*/
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
static struct td_sched *
tdq_choose(struct tdq *tdq)
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched *ts;
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
- When a new thread is added to a kseq the load is incremented prior to adding it to the nice tables. Therefore, in kseq_add_nice, we should keep in mind that the load will be 1 if we are the only thread, and not 0. - Assert that the sched lock is held in all the appropriate places. - Increase the scope of the sched lock in sched_pctcpu_update(). - Hold the sched lock in sched_runnable(). It is not held by the caller.
2003-06-08 00:47:33 +00:00
mtx_assert(&sched_lock, MA_OWNED);
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
ts = runq_choose(&tdq->tdq_realtime);
- Fix an off by one error in sched_pri_range. - In tdq_choose() only assert that a thread does not have too high a priority (low value) for the queue we removed it from. This will catch bugs in priority elevation. It's not a serious error for the thread to have too low a priority as we don't change queues in this case as an optimization. Reported by: kris
2007-06-15 19:33:58 +00:00
if (ts != NULL)
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
return (ts);
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
ts = runq_choose_from(&tdq->tdq_timeshare, tdq->tdq_ridx);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (ts != NULL) {
- Fix an off by one error in sched_pri_range. - In tdq_choose() only assert that a thread does not have too high a priority (low value) for the queue we removed it from. This will catch bugs in priority elevation. It's not a serious error for the thread to have too low a priority as we don't change queues in this case as an optimization. Reported by: kris
2007-06-15 19:33:58 +00:00
KASSERT(ts->ts_thread->td_priority >= PRI_MIN_TIMESHARE,
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
("tdq_choose: Invalid priority on timeshare queue %d",
ts->ts_thread->td_priority));
return (ts);
}
ts = runq_choose(&tdq->tdq_idle);
if (ts != NULL) {
KASSERT(ts->ts_thread->td_priority >= PRI_MIN_IDLE,
("tdq_choose: Invalid priority on idle queue %d",
ts->ts_thread->td_priority));
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
return (ts);
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
return (NULL);
- Add in support for KSEs with 0 slice values on the run queue. If we try to select a KSE with a slice of 0 we will update its slice and insert it onto the next queue. - Pass the KSE instead of the ksegrp into sched_slice(). This more accurately reflects the behavior of the code. Slices are granted to kses. - Add a function kseq_nice_min() which finds the smallest nice value assigned to the kseg of any KSE on the queue. - Rewrite the logic in sched_slice(). Add a large comment describing the new slice selection scheme. To summarize, slices are assigned based on the nice value. Priorities are still calculated based on the nice and interactivity of a process. Slice sizes of 0 may be granted for KSEs whos nice is 20 or futher away from the lowest nice on the run queue. Other nice values are scaled across the range [min, min+20]. This fixes ULEs bad behavior with positively niced processes.
2003-04-02 06:46:43 +00:00
}
- Use ksq_load as the authoritive count of kses on the pair of kseqs for sched_runnable() et all. - Remove some dead code in sched_clock(). - Define two macros KSEQ_SELF() and KSEQ_CPU() for getting the kseq of the current cpu or some alternate cpu. - Start introducing kseq_() functions, such as kseq_choose() and kseq_setup().
2003-01-29 07:00:51 +00:00
static void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_setup(struct tdq *tdq)
- Use ksq_load as the authoritive count of kses on the pair of kseqs for sched_runnable() et all. - Remove some dead code in sched_clock(). - Define two macros KSEQ_SELF() and KSEQ_CPU() for getting the kseq of the current cpu or some alternate cpu. - Start introducing kseq_() functions, such as kseq_choose() and kseq_setup().
2003-01-29 07:00:51 +00:00
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
runq_init(&tdq->tdq_realtime);
runq_init(&tdq->tdq_timeshare);
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
runq_init(&tdq->tdq_idle);
tdq->tdq_load = 0;
- Use ksq_load as the authoritive count of kses on the pair of kseqs for sched_runnable() et all. - Remove some dead code in sched_clock(). - Define two macros KSEQ_SELF() and KSEQ_CPU() for getting the kseq of the current cpu or some alternate cpu. - Start introducing kseq_() functions, such as kseq_choose() and kseq_setup().
2003-01-29 07:00:51 +00:00
}
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
static void
sched_setup(void *dummy)
{
- Clean up an unused variable. Submitted by: Steve Kargl <skg@routmask.apl.washington.edu>
2003-07-07 21:08:28 +00:00
#ifdef SMP
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
int i;
- Clean up an unused variable. Submitted by: Steve Kargl <skg@routmask.apl.washington.edu>
2003-07-07 21:08:28 +00:00
#endif
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Fix a bug in slice calculation code, current code uses hz but sched_clock() is called by state clock. Submitted by: taku at tackymt dot homeip dot net
2005-12-19 08:26:09 +00:00
/*
* To avoid divide-by-zero, we set realstathz a dummy value
* in case which sched_clock() called before sched_initticks().
*/
realstathz = hz;
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
sched_slice = (realstathz/10); /* ~100ms */
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
tickincr = 1 << SCHED_TICK_SHIFT;
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
#ifdef SMP
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
balance_groups = 0;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* Initialize the tdqs.
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
*/
for (i = 0; i < MAXCPU; i++) {
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
struct tdq *tdq;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
tdq = &tdq_cpu[i];
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_setup(&tdq_cpu[i]);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
if (smp_topology == NULL) {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
struct tdq_group *tdg;
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
struct tdq *tdq;
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
int cpus;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
for (cpus = 0, i = 0; i < MAXCPU; i++) {
if (CPU_ABSENT(i))
continue;
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
tdq = &tdq_cpu[i];
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg = &tdq_groups[cpus];
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* Setup a tdq group with one member.
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
*/
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
tdq->tdq_transferable = 0;
tdq->tdq_group = tdg;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg->tdg_cpus = 1;
tdg->tdg_idlemask = 0;
tdg->tdg_cpumask = tdg->tdg_mask = 1 << i;
tdg->tdg_load = 0;
tdg->tdg_transferable = 0;
LIST_INIT(&tdg->tdg_members);
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
LIST_INSERT_HEAD(&tdg->tdg_members, tdq, tdq_siblings);
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
cpus++;
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
}
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg_maxid = cpus - 1;
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
} else {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
struct tdq_group *tdg;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
struct cpu_group *cg;
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
int j;
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
topology = 1;
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
for (i = 0; i < smp_topology->ct_count; i++) {
cg = &smp_topology->ct_group[i];
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg = &tdq_groups[i];
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
* Initialize the group.
*/
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg->tdg_idlemask = 0;
tdg->tdg_load = 0;
tdg->tdg_transferable = 0;
tdg->tdg_cpus = cg->cg_count;
tdg->tdg_cpumask = cg->cg_mask;
LIST_INIT(&tdg->tdg_members);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
* Find all of the group members and add them.
*/
for (j = 0; j < MAXCPU; j++) {
if ((cg->cg_mask & (1 << j)) != 0) {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (tdg->tdg_mask == 0)
tdg->tdg_mask = 1 << j;
tdq_cpu[j].tdq_transferable = 0;
tdq_cpu[j].tdq_group = tdg;
LIST_INSERT_HEAD(&tdg->tdg_members,
&tdq_cpu[j], tdq_siblings);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
}
}
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (tdg->tdg_cpus > 1)
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
balance_groups = 1;
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
}
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
tdg_maxid = smp_topology->ct_count - 1;
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
}
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
/*
* Stagger the group and global load balancer so they do not
* interfere with each other.
*/
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
bal_tick = ticks + hz;
- Now that we have kseq groups, balance them seperately. - The new sched_balance_groups() function does intra-group balancing while sched_balance() balances the available groups. - Pick a random time between 0 ticks and hz * 2 ticks to restart each balancing process. Each balancer has its own timeout. - Pick a random place in the list of groups to start the search for lowest and highest group loads. This prevents us from prefering a group based on numeric position. - Use a nasty hack to stop us from preferring cpu 0. The problem is that softclock always runs on cpu 0, so it always has a little extra load. We ignore this load in the balancer for now. In the future softclock should run on a random cpu and these hacks can go away.
2003-12-12 07:33:51 +00:00
if (balance_groups)
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
gbal_tick = ticks + (hz / 2);
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
#else
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_setup(TDQ_SELF());
- Add a simple CPU load balancing algorithm. This works by executing once a second and equalizing the load between the two most imbalanced CPU. This is intended to clear up long term load imbalances that would not be handled by the 'pull' method in sched_choose(). - Pull out some bits of sched_choose() into a kseq_move() function that moves an arbitrary thread from one kseq to another.
2003-06-09 00:39:09 +00:00
#endif
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
mtx_lock_spin(&sched_lock);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_load_add(TDQ_SELF(), &td_sched0);
- Parse the cpu topology map in sched_setup(). - Associate logical CPUs on the same physical core with the same kseq. - Adjust code that assumed there would only be one running thread in any kseq. - Wrap the HTT code with a ULE_HTT_EXPERIMENTAL ifdef. This is a start towards HyperThreading support but it isn't quite there yet.
2003-07-04 19:59:00 +00:00
mtx_unlock_spin(&sched_lock);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
Fix a bug in slice calculation code, current code uses hz but sched_clock() is called by state clock. Submitted by: taku at tackymt dot homeip dot net
2005-12-19 08:26:09 +00:00
/* ARGSUSED */
static void
sched_initticks(void *dummy)
{
mtx_lock_spin(&sched_lock);
realstathz = stathz ? stathz : hz;
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
sched_slice = (realstathz/10); /* ~100ms */
Fix a bug in slice calculation code, current code uses hz but sched_clock() is called by state clock. Submitted by: taku at tackymt dot homeip dot net
2005-12-19 08:26:09 +00:00
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* tickincr is shifted out by 10 to avoid rounding errors due to
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
* hz not being evenly divisible by stathz on all platforms.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
*/
tickincr = (hz << SCHED_TICK_SHIFT) / realstathz;
/*
* This does not work for values of stathz that are more than
* 1 << SCHED_TICK_SHIFT * hz. In practice this does not happen.
Fix a bug in slice calculation code, current code uses hz but sched_clock() is called by state clock. Submitted by: taku at tackymt dot homeip dot net
2005-12-19 08:26:09 +00:00
*/
if (tickincr == 0)
tickincr = 1;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#ifdef SMP
affinity = SCHED_AFFINITY_DEFAULT;
#endif
Fix a bug in slice calculation code, current code uses hz but sched_clock() is called by state clock. Submitted by: taku at tackymt dot homeip dot net
2005-12-19 08:26:09 +00:00
mtx_unlock_spin(&sched_lock);
}
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
* Scale the scheduling priority according to the "interactivity" of this
* process.
*/
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
static void
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_priority(struct thread *td)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
int score;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
int pri;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
if (td->td_pri_class != PRI_TIMESHARE)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
return;
- Add in support for KSEs with 0 slice values on the run queue. If we try to select a KSE with a slice of 0 we will update its slice and insert it onto the next queue. - Pass the KSE instead of the ksegrp into sched_slice(). This more accurately reflects the behavior of the code. Slices are granted to kses. - Add a function kseq_nice_min() which finds the smallest nice value assigned to the kseg of any KSE on the queue. - Rewrite the logic in sched_slice(). Add a large comment describing the new slice selection scheme. To summarize, slices are assigned based on the nice value. Priorities are still calculated based on the nice and interactivity of a process. Slice sizes of 0 may be granted for KSEs whos nice is 20 or futher away from the lowest nice on the run queue. Other nice values are scaled across the range [min, min+20]. This fixes ULEs bad behavior with positively niced processes.
2003-04-02 06:46:43 +00:00
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* If the score is interactive we place the thread in the realtime
* queue with a priority that is less than kernel and interrupt
* priorities. These threads are not subject to nice restrictions.
- Add in support for KSEs with 0 slice values on the run queue. If we try to select a KSE with a slice of 0 we will update its slice and insert it onto the next queue. - Pass the KSE instead of the ksegrp into sched_slice(). This more accurately reflects the behavior of the code. Slices are granted to kses. - Add a function kseq_nice_min() which finds the smallest nice value assigned to the kseg of any KSE on the queue. - Rewrite the logic in sched_slice(). Add a large comment describing the new slice selection scheme. To summarize, slices are assigned based on the nice value. Priorities are still calculated based on the nice and interactivity of a process. Slice sizes of 0 may be granted for KSEs whos nice is 20 or futher away from the lowest nice on the run queue. Other nice values are scaled across the range [min, min+20]. This fixes ULEs bad behavior with positively niced processes.
2003-04-02 06:46:43 +00:00
*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* Scores greater than this are placed on the normal realtime queue
* where the priority is partially decided by the most recent cpu
* utilization and the rest is decided by nice value.
- Add in support for KSEs with 0 slice values on the run queue. If we try to select a KSE with a slice of 0 we will update its slice and insert it onto the next queue. - Pass the KSE instead of the ksegrp into sched_slice(). This more accurately reflects the behavior of the code. Slices are granted to kses. - Add a function kseq_nice_min() which finds the smallest nice value assigned to the kseg of any KSE on the queue. - Rewrite the logic in sched_slice(). Add a large comment describing the new slice selection scheme. To summarize, slices are assigned based on the nice value. Priorities are still calculated based on the nice and interactivity of a process. Slice sizes of 0 may be granted for KSEs whos nice is 20 or futher away from the lowest nice on the run queue. Other nice values are scaled across the range [min, min+20]. This fixes ULEs bad behavior with positively niced processes.
2003-04-02 06:46:43 +00:00
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
score = sched_interact_score(td);
if (score < sched_interact) {
pri = PRI_MIN_REALTIME;
pri += ((PRI_MAX_REALTIME - PRI_MIN_REALTIME) / sched_interact)
* score;
KASSERT(pri >= PRI_MIN_REALTIME && pri <= PRI_MAX_REALTIME,
- With a sleep time over 2097 seconds hzticks and slptime could end up negative. Use unsigned integers for sleep and run time so this doesn't disturb sched_interact_score(). This should fix the invalid interactive priority panics reported by several users.
2007-01-24 18:18:43 +00:00
("sched_priority: invalid interactive priority %d score %d",
pri, score));
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
} else {
pri = SCHED_PRI_MIN;
if (td->td_sched->ts_ticks)
pri += SCHED_PRI_TICKS(td->td_sched);
pri += SCHED_PRI_NICE(td->td_proc->p_nice);
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
if (!(pri >= PRI_MIN_TIMESHARE && pri <= PRI_MAX_TIMESHARE)) {
static int once = 1;
if (once) {
printf("sched_priority: invalid priority %d",
pri);
printf("nice %d, ticks %d ftick %d ltick %d tick pri %d\n",
td->td_proc->p_nice,
td->td_sched->ts_ticks,
td->td_sched->ts_ftick,
td->td_sched->ts_ltick,
SCHED_PRI_TICKS(td->td_sched));
once = 0;
}
pri = min(max(pri, PRI_MIN_TIMESHARE),
PRI_MAX_TIMESHARE);
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
}
sched_user_prio(td, pri);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
- Add in support for KSEs with 0 slice values on the run queue. If we try to select a KSE with a slice of 0 we will update its slice and insert it onto the next queue. - Pass the KSE instead of the ksegrp into sched_slice(). This more accurately reflects the behavior of the code. Slices are granted to kses. - Add a function kseq_nice_min() which finds the smallest nice value assigned to the kseg of any KSE on the queue. - Rewrite the logic in sched_slice(). Add a large comment describing the new slice selection scheme. To summarize, slices are assigned based on the nice value. Priorities are still calculated based on the nice and interactivity of a process. Slice sizes of 0 may be granted for KSEs whos nice is 20 or futher away from the lowest nice on the run queue. Other nice values are scaled across the range [min, min+20]. This fixes ULEs bad behavior with positively niced processes.
2003-04-02 06:46:43 +00:00
return;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
/*
* This routine enforces a maximum limit on the amount of scheduling history
* kept. It is called after either the slptime or runtime is adjusted.
*/
- Add a new function "sched_interact_update()" that scales back the sleep and run time. - Scale the sleep and run time back via sched_interact_update() in more places. This is to keep the statistic more accurate. - Charge a parent one tick for forking a child. - Add only the run time and not the sleep time to the parents kg when a thread exits. This allows us to give a penalty for having an expensive thread exit but does not give a bonus for having an interactive thread exit. - Change the SLP_RUN_THROTTLE to limit us to 4/5th and not 1/2. - Change the SLP_RUN_MAX to two seconds. This keeps bursty interactive applications like mozilla and openoffice in the interactive range even through expensive tasks. - Recalculate the slice after every sleep. This ensures that once a task has been marked interactive it only has a slice of 1 at the risk of giving tasks that sleep for a very brief period a longer time slice.
2003-06-17 06:39:51 +00:00
static void
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_interact_update(struct thread *td)
- Add a new function "sched_interact_update()" that scales back the sleep and run time. - Scale the sleep and run time back via sched_interact_update() in more places. This is to keep the statistic more accurate. - Charge a parent one tick for forking a child. - Add only the run time and not the sleep time to the parents kg when a thread exits. This allows us to give a penalty for having an expensive thread exit but does not give a bonus for having an interactive thread exit. - Change the SLP_RUN_THROTTLE to limit us to 4/5th and not 1/2. - Change the SLP_RUN_MAX to two seconds. This keeps bursty interactive applications like mozilla and openoffice in the interactive range even through expensive tasks. - Recalculate the slice after every sleep. This ensures that once a task has been marked interactive it only has a slice of 1 at the risk of giving tasks that sleep for a very brief period a longer time slice.
2003-06-17 06:39:51 +00:00
{
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
struct td_sched *ts;
- With a sleep time over 2097 seconds hzticks and slptime could end up negative. Use unsigned integers for sleep and run time so this doesn't disturb sched_interact_score(). This should fix the invalid interactive priority panics reported by several users.
2007-01-24 18:18:43 +00:00
u_int sum;
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
ts = td->td_sched;
sum = ts->skg_runtime + ts->skg_slptime;
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
if (sum < SCHED_SLP_RUN_MAX)
return;
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
/*
* This only happens from two places:
* 1) We have added an unusual amount of run time from fork_exit.
* 2) We have added an unusual amount of sleep time from sched_sleep().
*/
if (sum > SCHED_SLP_RUN_MAX * 2) {
if (ts->skg_runtime > ts->skg_slptime) {
ts->skg_runtime = SCHED_SLP_RUN_MAX;
ts->skg_slptime = 1;
} else {
ts->skg_slptime = SCHED_SLP_RUN_MAX;
ts->skg_runtime = 1;
}
return;
}
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
/*
* If we have exceeded by more than 1/5th then the algorithm below
* will not bring us back into range. Dividing by two here forces
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
* us into the range of [4/5 * SCHED_INTERACT_MAX, SCHED_INTERACT_MAX]
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
*/
- Use the proper constant in sched_interact_update(). Previously, SCHED_INTERACT_MAX was used where SCHED_SLP_RUN_MAX was needed. This was causing the interactivity scaler to lose history at a more dramatic rate than intended.
2004-04-04 19:12:56 +00:00
if (sum > (SCHED_SLP_RUN_MAX / 5) * 6) {
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
ts->skg_runtime /= 2;
ts->skg_slptime /= 2;
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
return;
}
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
ts->skg_runtime = (ts->skg_runtime / 5) * 4;
ts->skg_slptime = (ts->skg_slptime / 5) * 4;
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
}
- Use a better algorithm in sched_pctcpu_update() Contributed by: Thomaswuerfl@gmx.de - In sched_prio(), adjust the run queue for threads which may need to move to the current queue due to priority propagation . - In sched_switch(), fix style bug introduced when the KSE support went in. Columns are 80 chars wide, not 90. - In sched_switch(), Fix the comparison in the idle case and explicitly re-initialize the runq in the not propagated case. - Remove dead code in sched_clock(). - In sched_clock(), If we're an IDLE class td set NEEDRESCHED so that threads that have become runnable will get a chance to. - In sched_runnable(), if we're not the IDLETD, we should not consider curthread when examining the load. This mimics the 4BSD behavior of returning 0 when the only runnable thread is running. - In sched_userret(), remove the code for setting NEEDRESCHED entirely. This is not necessary and is not implemented in 4BSD. - Use the correct comparison in sched_add() when checking to see if an idle prio task has had it's priority temporarily elevated.
2003-10-27 06:47:05 +00:00
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
static void
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_interact_fork(struct thread *td)
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
{
int ratio;
int sum;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sum = td->td_sched->skg_runtime + td->td_sched->skg_slptime;
- Change sched_interact_update() to only accept slp+runtime values between 0 and SCHED_SLP_RUN_MAX * 2. This allows us to simplify the algorithm quite a bit. Before, it dealt with arbitrary values which required us to do nasty integer division tricks that didn't quite work out correctly. - Chnage sched_wakeup() to detect conditions where the slp+runtime could exceed SCHED_SLP_RUN_MAX * 2. This can happen if we go to sleep for longer than 6 seconds. In this case, we'll just clear the runtime and set the sleep time to the max. - Define a new function, sched_interact_fork() which updates the slp+runtime of a newly forked thread. We want to limit the amount of history retained from the parent so that we learn the child's behavior quickly. We don't, however want to decay it to nothing. Previously, we would simply divide each parameter by 100 whenever we forked. After a few forks the values would reach 0 and tasks would not be considered interactive. - Add another KTR entry, cleanup some existing entries. - Remove a useless sched_interact_update() from sched_priority(). This is already done by the callers that require it.
2003-11-02 03:36:33 +00:00
if (sum > SCHED_SLP_RUN_FORK) {
ratio = sum / SCHED_SLP_RUN_FORK;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
td->td_sched->skg_runtime /= ratio;
td->td_sched->skg_slptime /= ratio;
- Add a new function "sched_interact_update()" that scales back the sleep and run time. - Scale the sleep and run time back via sched_interact_update() in more places. This is to keep the statistic more accurate. - Charge a parent one tick for forking a child. - Add only the run time and not the sleep time to the parents kg when a thread exits. This allows us to give a penalty for having an expensive thread exit but does not give a bonus for having an interactive thread exit. - Change the SLP_RUN_THROTTLE to limit us to 4/5th and not 1/2. - Change the SLP_RUN_MAX to two seconds. This keeps bursty interactive applications like mozilla and openoffice in the interactive range even through expensive tasks. - Recalculate the slice after every sleep. This ensures that once a task has been marked interactive it only has a slice of 1 at the risk of giving tasks that sleep for a very brief period a longer time slice.
2003-06-17 06:39:51 +00:00
}
}
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
static int
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_interact_score(struct thread *td)
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
{
- Fix the maximum slice value. I accidentally checked in a value of '2' which meant no process would run for longer than 20ms. - Slightly redo the interactivity scorer. It follows the same algorithm but in a slightly more correct way. Previously values above half were incorrect. - Lower the interactivity threshold to 20. It seems that in testing non- interactive tasks are hardly ever near there and expensive interactive tasks can sometimes surpass it. This area needs more testing. - Remove an unnecessary KTR. - Fix a case where an idle thread that had an elevated priority due to priority prop. would be placed back on the idle queue. - Delay setting NEEDRESCHED until userret() for threads that haad their priority elevated while in kernel. This gives us the same context switch optimization as SCHED_4BSD. - Limit the child's slice to 1 in sched_fork_kse() so we detect its behavior more quickly. - Inhert some of the run/slp time from the child in sched_exit_ksegrp(). - Redo some of the priority comparisons so they are more clear. - Throttle the frequency of sched_pctcpu_update() so that rounding errors do not make it invalid.
2003-06-15 02:18:29 +00:00
int div;
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
if (td->td_sched->skg_runtime > td->td_sched->skg_slptime) {
div = max(1, td->td_sched->skg_runtime / SCHED_INTERACT_HALF);
- Fix the maximum slice value. I accidentally checked in a value of '2' which meant no process would run for longer than 20ms. - Slightly redo the interactivity scorer. It follows the same algorithm but in a slightly more correct way. Previously values above half were incorrect. - Lower the interactivity threshold to 20. It seems that in testing non- interactive tasks are hardly ever near there and expensive interactive tasks can sometimes surpass it. This area needs more testing. - Remove an unnecessary KTR. - Fix a case where an idle thread that had an elevated priority due to priority prop. would be placed back on the idle queue. - Delay setting NEEDRESCHED until userret() for threads that haad their priority elevated while in kernel. This gives us the same context switch optimization as SCHED_4BSD. - Limit the child's slice to 1 in sched_fork_kse() so we detect its behavior more quickly. - Inhert some of the run/slp time from the child in sched_exit_ksegrp(). - Redo some of the priority comparisons so they are more clear. - Throttle the frequency of sched_pctcpu_update() so that rounding errors do not make it invalid.
2003-06-15 02:18:29 +00:00
return (SCHED_INTERACT_HALF +
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
(SCHED_INTERACT_HALF - (td->td_sched->skg_slptime / div)));
- Handle the case where slptime == runtime. Submitted by: Atoine Brodin
2007-03-17 23:32:48 +00:00
}
if (td->td_sched->skg_slptime > td->td_sched->skg_runtime) {
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
div = max(1, td->td_sched->skg_slptime / SCHED_INTERACT_HALF);
return (td->td_sched->skg_runtime / div);
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
}
- Handle the case where slptime == runtime. Submitted by: Atoine Brodin
2007-03-17 23:32:48 +00:00
/* runtime == slptime */
if (td->td_sched->skg_runtime)
return (SCHED_INTERACT_HALF);
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
- Fix the maximum slice value. I accidentally checked in a value of '2' which meant no process would run for longer than 20ms. - Slightly redo the interactivity scorer. It follows the same algorithm but in a slightly more correct way. Previously values above half were incorrect. - Lower the interactivity threshold to 20. It seems that in testing non- interactive tasks are hardly ever near there and expensive interactive tasks can sometimes surpass it. This area needs more testing. - Remove an unnecessary KTR. - Fix a case where an idle thread that had an elevated priority due to priority prop. would be placed back on the idle queue. - Delay setting NEEDRESCHED until userret() for threads that haad their priority elevated while in kernel. This gives us the same context switch optimization as SCHED_4BSD. - Limit the child's slice to 1 in sched_fork_kse() so we detect its behavior more quickly. - Inhert some of the run/slp time from the child in sched_exit_ksegrp(). - Redo some of the priority comparisons so they are more clear. - Throttle the frequency of sched_pctcpu_update() so that rounding errors do not make it invalid.
2003-06-15 02:18:29 +00:00
/*
* This can happen if slptime and runtime are 0.
*/
return (0);
- Create a function sched_interact_score() which decides on the interactivity of a kseg and assigns it a value of 0 through 100. - Use sched_interact_score() to determine the dynamic priority. - Define SCHED_CURR() in terms of sched_interact_score(). - Adjust the maximum slice back down to 100ms. - Remove redundant clearing of ke_runq in sched_wakeup() - Clean up #defines and comment them.
2003-03-04 02:45:59 +00:00
}
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* Called from proc0_init() to bootstrap the scheduler.
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
*/
void
schedinit(void)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
/*
* Set up the scheduler specific parts of proc0.
*/
proc0.p_sched = NULL; /* XXX */
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
thread0.td_sched = &td_sched0;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread0.td_lock = &sched_lock;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
td_sched0.ts_ltick = ticks;
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
td_sched0.ts_ftick = ticks;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
td_sched0.ts_thread = &thread0;
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
}
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
/*
* This is only somewhat accurate since given many processes of the same
* priority they will switch when their slices run out, which will be
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* at most sched_slice stathz ticks.
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
*/
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
int
sched_rr_interval(void)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/* Convert sched_slice to hz */
return (hz/(realstathz/sched_slice));
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
static void
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
sched_pctcpu_update(struct td_sched *ts)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (ts->ts_ticks == 0)
return;
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
if (ticks - (hz / 10) < ts->ts_ltick &&
SCHED_TICK_TOTAL(ts) < SCHED_TICK_MAX)
return;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
* Adjust counters and watermark for pctcpu calc.
- Fix the maximum slice value. I accidentally checked in a value of '2' which meant no process would run for longer than 20ms. - Slightly redo the interactivity scorer. It follows the same algorithm but in a slightly more correct way. Previously values above half were incorrect. - Lower the interactivity threshold to 20. It seems that in testing non- interactive tasks are hardly ever near there and expensive interactive tasks can sometimes surpass it. This area needs more testing. - Remove an unnecessary KTR. - Fix a case where an idle thread that had an elevated priority due to priority prop. would be placed back on the idle queue. - Delay setting NEEDRESCHED until userret() for threads that haad their priority elevated while in kernel. This gives us the same context switch optimization as SCHED_4BSD. - Limit the child's slice to 1 in sched_fork_kse() so we detect its behavior more quickly. - Inhert some of the run/slp time from the child in sched_exit_ksegrp(). - Redo some of the priority comparisons so they are more clear. - Throttle the frequency of sched_pctcpu_update() so that rounding errors do not make it invalid.
2003-06-15 02:18:29 +00:00
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (ts->ts_ltick > ticks - SCHED_TICK_TARG)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts->ts_ticks = (ts->ts_ticks / (ticks - ts->ts_ftick)) *
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
SCHED_TICK_TARG;
else
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts->ts_ticks = 0;
ts->ts_ltick = ticks;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
ts->ts_ftick = ts->ts_ltick - SCHED_TICK_TARG;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static void
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
sched_thread_priority(struct thread *td, u_char prio)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched *ts;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
- Remove earlier KTR_ULE tracepoints. - Define new KTR_SCHED points so that we can graph the operation of the scheduler.
2004-12-26 00:15:33 +00:00
CTR6(KTR_SCHED, "sched_prio: %p(%s) prio %d newprio %d by %p(%s)",
td, td->td_proc->p_comm, td->td_priority, prio, curthread,
curthread->td_proc->p_comm);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = td->td_sched;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
if (td->td_priority == prio)
return;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
if (TD_ON_RUNQ(td) && prio < td->td_priority) {
- Use a better algorithm in sched_pctcpu_update() Contributed by: Thomaswuerfl@gmx.de - In sched_prio(), adjust the run queue for threads which may need to move to the current queue due to priority propagation . - In sched_switch(), fix style bug introduced when the KSE support went in. Columns are 80 chars wide, not 90. - In sched_switch(), Fix the comparison in the idle case and explicitly re-initialize the runq in the not propagated case. - Remove dead code in sched_clock(). - In sched_clock(), If we're an IDLE class td set NEEDRESCHED so that threads that have become runnable will get a chance to. - In sched_runnable(), if we're not the IDLETD, we should not consider curthread when examining the load. This mimics the 4BSD behavior of returning 0 when the only runnable thread is running. - In sched_userret(), remove the code for setting NEEDRESCHED entirely. This is not necessary and is not implemented in 4BSD. - Use the correct comparison in sched_add() when checking to see if an idle prio task has had it's priority temporarily elevated.
2003-10-27 06:47:05 +00:00
/*
* If the priority has been elevated due to priority
* propagation, we may have to move ourselves to a new
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* queue. This could be optimized to not re-add in some
* cases.
- Introduce a new flag KEF_HOLD that prevents sched_add() from doing a migration. Use this in sched_prio() and sched_switch() to stop us from migrating threads that are in short term sleeps or are runnable. These extra migrations were added in the patches to support KSE. - Only set NEEDRESCHED if the thread we're adding in sched_add() is a lower priority and is being placed on the current queue. - Fix some minor whitespace problems.
2004-08-12 07:56:33 +00:00
*/
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
MPASS(td->td_lock == &sched_lock);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
sched_rem(td);
td->td_priority = prio;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
sched_add(td, SRQ_BORROWING|SRQ_OURSELF);
- Use a better algorithm in sched_pctcpu_update() Contributed by: Thomaswuerfl@gmx.de - In sched_prio(), adjust the run queue for threads which may need to move to the current queue due to priority propagation . - In sched_switch(), fix style bug introduced when the KSE support went in. Columns are 80 chars wide, not 90. - In sched_switch(), Fix the comparison in the idle case and explicitly re-initialize the runq in the not propagated case. - Remove dead code in sched_clock(). - In sched_clock(), If we're an IDLE class td set NEEDRESCHED so that threads that have become runnable will get a chance to. - In sched_runnable(), if we're not the IDLETD, we should not consider curthread when examining the load. This mimics the 4BSD behavior of returning 0 when the only runnable thread is running. - In sched_userret(), remove the code for setting NEEDRESCHED entirely. This is not necessary and is not implemented in 4BSD. - Use the correct comparison in sched_add() when checking to see if an idle prio task has had it's priority temporarily elevated.
2003-10-27 06:47:05 +00:00
} else
Let SA process work under ULE scheduler, originally it would panic kernel. Reviewed by: jeff
2003-08-26 11:33:15 +00:00
td->td_priority = prio;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
/*
* Update a thread's priority when it is lent another thread's
* priority.
*/
void
sched_lend_prio(struct thread *td, u_char prio)
{
td->td_flags |= TDF_BORROWING;
sched_thread_priority(td, prio);
}
/*
* Restore a thread's priority when priority propagation is
* over. The prio argument is the minimum priority the thread
* needs to have to satisfy other possible priority lending
* requests. If the thread's regular priority is less
* important than prio, the thread will keep a priority boost
* of prio.
*/
void
sched_unlend_prio(struct thread *td, u_char prio)
{
u_char base_pri;
if (td->td_base_pri >= PRI_MIN_TIMESHARE &&
td->td_base_pri <= PRI_MAX_TIMESHARE)
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
base_pri = td->td_user_pri;
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
else
base_pri = td->td_base_pri;
if (prio >= base_pri) {
Fix a typo and two whitespace nits.
2004-12-30 22:17:00 +00:00
td->td_flags &= ~TDF_BORROWING;
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
sched_thread_priority(td, base_pri);
} else
sched_lend_prio(td, prio);
}
void
sched_prio(struct thread *td, u_char prio)
{
u_char oldprio;
/* First, update the base priority. */
td->td_base_pri = prio;
/*
Fix a typo and two whitespace nits.
2004-12-30 22:17:00 +00:00
* If the thread is borrowing another thread's priority, don't
Rework the interface between priority propagation (lending) and the schedulers a bit to ensure more correct handling of priorities and fewer priority inversions: - Add two functions to the sched(9) API to handle priority lending: sched_lend_prio() and sched_unlend_prio(). The turnstile code uses these functions to ask the scheduler to lend a thread a set priority and to tell the scheduler when it thinks it is ok for a thread to stop borrowing priority. The unlend case is slightly complex in that the turnstile code tells the scheduler what the minimum priority of the thread needs to be to satisfy the requirements of any other threads blocked on locks owned by the thread in question. The scheduler then decides where the thread can go back to normal mode (if it's normal priority is high enough to satisfy the pending lock requests) or it it should continue to use the priority specified to the sched_unlend_prio() call. This involves adding a new per-thread flag TDF_BORROWING that replaces the ULE-only kse flag for priority elevation. - Schedulers now refuse to lower the priority of a thread that is currently borrowing another therad's priority. - If a scheduler changes the priority of a thread that is currently sitting on a turnstile, it will call a new function turnstile_adjust() to inform the turnstile code of the change. This function resorts the thread on the priority list of the turnstile if needed, and if the thread ends up at the head of the list (due to having the highest priority) and its priority was raised, then it will propagate that new priority to the owner of the lock it is blocked on. Some additional fixes specific to the 4BSD scheduler include: - Common code for updating the priority of a thread when the user priority of its associated kse group has been consolidated in a new static function resetpriority_thread(). One change to this function is that it will now only adjust the priority of a thread if it already has a time sharing priority, thus preserving any boosts from a tsleep() until the thread returns to userland. Also, resetpriority() no longer calls maybe_resched() on each thread in the group. Instead, the code calling resetpriority() is responsible for calling resetpriority_thread() on any threads that need to be updated. - schedcpu() now uses resetpriority_thread() instead of just calling sched_prio() directly after it updates a kse group's user priority. - sched_clock() now uses resetpriority_thread() rather than writing directly to td_priority. - sched_nice() now updates all the priorities of the threads after the group priority has been adjusted. Discussed with: bde Reviewed by: ups, jeffr Tested on: 4bsd, ule Tested on: i386, alpha, sparc64
2004-12-30 20:52:44 +00:00
* ever lower the priority.
*/
if (td->td_flags & TDF_BORROWING && td->td_priority < prio)
return;
/* Change the real priority. */
oldprio = td->td_priority;
sched_thread_priority(td, prio);
/*
* If the thread is on a turnstile, then let the turnstile update
* its state.
*/
if (TD_ON_LOCK(td) && oldprio != prio)
turnstile_adjust(td, oldprio);
}
Fix a typo and two whitespace nits.
2004-12-30 22:17:00 +00:00
Add user priority loaning code to support priority propagation for 1:1 threading's POSIX priority mutexes, the code is no-op unless priority-aware umtx code is committed.
2006-08-25 06:12:53 +00:00
void
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_user_prio(struct thread *td, u_char prio)
Add user priority loaning code to support priority propagation for 1:1 threading's POSIX priority mutexes, the code is no-op unless priority-aware umtx code is committed.
2006-08-25 06:12:53 +00:00
{
u_char oldprio;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
td->td_base_user_pri = prio;
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
if (td->td_flags & TDF_UBORROWING && td->td_user_pri <= prio)
return;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
oldprio = td->td_user_pri;
td->td_user_pri = prio;
Add user priority loaning code to support priority propagation for 1:1 threading's POSIX priority mutexes, the code is no-op unless priority-aware umtx code is committed.
2006-08-25 06:12:53 +00:00
if (TD_ON_UPILOCK(td) && oldprio != prio)
umtx_pi_adjust(td, oldprio);
}
void
sched_lend_user_prio(struct thread *td, u_char prio)
{
u_char oldprio;
td->td_flags |= TDF_UBORROWING;
o Fix a couple of obvious typos.
2006-11-08 09:09:07 +00:00
oldprio = td->td_user_pri;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
td->td_user_pri = prio;
Add user priority loaning code to support priority propagation for 1:1 threading's POSIX priority mutexes, the code is no-op unless priority-aware umtx code is committed.
2006-08-25 06:12:53 +00:00
if (TD_ON_UPILOCK(td) && oldprio != prio)
umtx_pi_adjust(td, oldprio);
}
void
sched_unlend_user_prio(struct thread *td, u_char prio)
{
u_char base_pri;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
base_pri = td->td_base_user_pri;
Add user priority loaning code to support priority propagation for 1:1 threading's POSIX priority mutexes, the code is no-op unless priority-aware umtx code is committed.
2006-08-25 06:12:53 +00:00
if (prio >= base_pri) {
td->td_flags &= ~TDF_UBORROWING;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_user_prio(td, base_pri);
Add user priority loaning code to support priority propagation for 1:1 threading's POSIX priority mutexes, the code is no-op unless priority-aware umtx code is committed.
2006-08-25 06:12:53 +00:00
} else
sched_lend_user_prio(td, prio);
}
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
void
Add some code to allow threads to nominat a sibling to run if theyu are going to sleep. MFC after: 1 week
2004-09-10 21:04:38 +00:00
sched_switch(struct thread *td, struct thread *newtd, int flags)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
struct tdq *tdq;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched *ts;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
int preempt;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
preempt = flags & SW_PREEMPT;
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
tdq = TDQ_SELF();
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
ts = td->td_sched;
- Introduce a new flag KEF_HOLD that prevents sched_add() from doing a migration. Use this in sched_prio() and sched_switch() to stop us from migrating threads that are in short term sleeps or are runnable. These extra migrations were added in the patches to support KSE. - Only set NEEDRESCHED if the thread we're adding in sched_add() is a lower priority and is being placed on the current queue. - Fix some minor whitespace problems.
2004-08-12 07:56:33 +00:00
td->td_lastcpu = td->td_oncpu;
Move the _oncpu entry from the KSE to the thread. The entry in the KSE still exists but it's purpose will change a bit when we add the ability to lock a KSE to a cpu.
2003-04-10 17:35:44 +00:00
td->td_oncpu = NOCPU;
- Move TDF_OWEPREEMPT, TDF_OWEUPC, and TDF_USTATCLOCK over to td_pflags since they are only accessed by curthread and thus do not need any locking. - Move pr_addr and pr_ticks out of struct uprof (which is per-process) and directly into struct thread as td_profil_addr and td_profil_ticks as these variables are really per-thread. (They are used to defer an addupc_intr() that was too "hard" until ast()).
2004-07-16 21:04:55 +00:00
td->td_flags &= ~TDF_NEEDRESCHED;
Sprinkle some volatile magic and rearrange things a bit to avoid race conditions in critical_exit now that it no longer blocks interrupts. Reviewed by: jhb
2005-04-08 03:37:53 +00:00
td->td_owepreempt = 0;
- In sched_switch(), if a thread has been assigned, don't touch the runqueues or load. These things have already been taken care of in sched_bind() which should be the only place that we're switching in an assigned thread.
2003-12-11 04:00:49 +00:00
/*
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
* If the thread has been assigned it may be in the process of switching
- In sched_switch(), if a thread has been assigned, don't touch the runqueues or load. These things have already been taken care of in sched_bind() which should be the only place that we're switching in an assigned thread.
2003-12-11 04:00:49 +00:00
* to the new cpu. This is the case in sched_bind().
*/
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
/*
* Switch to the sched lock to fix things up and pick
* a new thread.
*/
if (td->td_lock != &sched_lock) {
mtx_lock_spin(&sched_lock);
thread_unlock(td);
}
Instead of doing comparisons using the pcpu area to see if a thread is an idle thread, just see if it has the IDLETD flag set. That flag will probably move to the pflags word as it's permenent and never chenges for the life of the system so it doesn't need locking.
2007-03-08 06:44:34 +00:00
if (TD_IS_IDLETHREAD(td)) {
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
MPASS(td->td_lock == &sched_lock);
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
TD_SET_CAN_RUN(td);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
} else if (TD_IS_RUNNING(td)) {
/*
* Don't allow the thread to migrate
* from a preemption.
*/
- More search and replace prettying.
2006-12-29 12:55:32 +00:00
tdq_load_rem(tdq, ts);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
if (preempt)
sched_pin_td(td);
sched_add(td, preempt ?
SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED :
SRQ_OURSELF|SRQ_YIELDING);
if (preempt)
sched_unpin_td(td);
} else
tdq_load_rem(tdq, ts);
mtx_assert(&sched_lock, MA_OWNED);
Use some macros to trach available scheduler slots to allow easier debugging. MFC after: 4 days
2004-10-05 21:10:44 +00:00
if (newtd != NULL) {
When preempting a thread, put it back on the HEAD of its run queue. (Only really implemented in 4bsd) MFC after: 4 days
2004-10-05 22:03:10 +00:00
/*
- Fix the case where we're not preempting but there is already a newtd as this happens via thread_switchout(). I don't particularly like the structure of the code here. We twice call out to thread code when a thread is voluntarily switching. Once to thread_switchout() and once to slot_fill(), while sched_4BSD does even more work which is redundant to select another thread to use our remaining slice. This should be simplified in the future, but for now I'm only going to fix the bug not the bad design.
2005-06-07 02:59:16 +00:00
* If we bring in a thread account for it as if it had been
* added to the run queue and then chosen.
When preempting a thread, put it back on the HEAD of its run queue. (Only really implemented in 4bsd) MFC after: 4 days
2004-10-05 22:03:10 +00:00
*/
Fix whitespace botch that only showed up in the commit message diff :-/ MFC after: 4 days
2004-10-05 22:14:02 +00:00
TD_SET_RUNNING(newtd);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_load_add(TDQ_SELF(), newtd->td_sched);
Use some macros to trach available scheduler slots to allow easier debugging. MFC after: 4 days
2004-10-05 21:10:44 +00:00
} else
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
newtd = choosethread();
Bring a working snapshot of hwpmc(4), its associated libraries, userland utilities and documentation into -CURRENT. Bump FreeBSD_version. Reviewed by: alc, jhb (kernel changes)
2005-04-19 04:01:25 +00:00
if (td != newtd) {
#ifdef HWPMC_HOOKS
if (PMC_PROC_IS_USING_PMCS(td->td_proc))
PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT);
#endif
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
- Placing the 'volatile' on the right side of the * in the td_lock declaration removes the need for __DEVOLATILE(). Pointed out by: tegge
2007-06-06 03:40:47 +00:00
cpu_switch(td, newtd, td->td_lock);
Bring a working snapshot of hwpmc(4), its associated libraries, userland utilities and documentation into -CURRENT. Bump FreeBSD_version. Reviewed by: alc, jhb (kernel changes)
2005-04-19 04:01:25 +00:00
#ifdef HWPMC_HOOKS
if (PMC_PROC_IS_USING_PMCS(td->td_proc))
PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN);
#endif
}
- Collapse sched_switchin() and sched_switchout() into sched_switch(). Now mi_switch() calls sched_switch() which calls cpu_switch(). This is actually one less function call than it had been.
2003-10-16 08:53:46 +00:00
sched_lock.mtx_lock = (uintptr_t)td;
Move the _oncpu entry from the KSE to the thread. The entry in the KSE still exists but it's purpose will change a bit when we add the ability to lock a KSE to a cpu.
2003-04-10 17:35:44 +00:00
td->td_oncpu = PCPU_GET(cpuid);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
MPASS(td->td_lock == &sched_lock);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
void
Nice, is a property of a process as a whole.. I mistakenly moved it to the ksegroup when breaking up the process structure. Put it back in the proc structure.
2004-06-16 00:26:31 +00:00
sched_nice(struct proc *p, int nice)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
struct thread *td;
Nice, is a property of a process as a whole.. I mistakenly moved it to the ksegroup when breaking up the process structure. Put it back in the proc structure.
2004-06-16 00:26:31 +00:00
PROC_LOCK_ASSERT(p, MA_OWNED);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
PROC_SLOCK_ASSERT(p, MA_OWNED);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
Nice, is a property of a process as a whole.. I mistakenly moved it to the ksegroup when breaking up the process structure. Put it back in the proc structure.
2004-06-16 00:26:31 +00:00
p->p_nice = nice;
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
FOREACH_THREAD_IN_PROC(p, td) {
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_lock(td);
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_priority(td);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
sched_prio(td, td->td_base_user_pri);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_unlock(td);
Nice, is a property of a process as a whole.. I mistakenly moved it to the ksegroup when breaking up the process structure. Put it back in the proc structure.
2004-06-16 00:26:31 +00:00
}
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
void
Switch the sleep/wakeup and condition variable implementations to use the sleep queue interface: - Sleep queues attempt to merge some of the benefits of both sleep queues and condition variables. Having sleep qeueus in a hash table avoids having to allocate a queue head for each wait channel. Thus, struct cv has shrunk down to just a single char * pointer now. However, the hash table does not hold threads directly, but queue heads. This means that once you have located a queue in the hash bucket, you no longer have to walk the rest of the hash chain looking for threads. Instead, you have a list of all the threads sleeping on that wait channel. - Outside of the sleepq code and the sleep/cv code the kernel no longer differentiates between cv's and sleep/wakeup. For example, calls to abortsleep() and cv_abort() are replaced with a call to sleepq_abort(). Thus, the TDF_CVWAITQ flag is removed. Also, calls to unsleep() and cv_waitq_remove() have been replaced with calls to sleepq_remove(). - The sched_sleep() function no longer accepts a priority argument as sleep's no longer inherently bump the priority. Instead, this is soley a propery of msleep() which explicitly calls sched_prio() before blocking. - The TDF_ONSLEEPQ flag has been dropped as it was never used. The associated TDF_SET_ONSLEEPQ and TDF_CLR_ON_SLEEPQ macros have also been dropped and replaced with a single explicit clearing of td_wchan. TD_SET_ONSLEEPQ() would really have only made sense if it had taken the wait channel and message as arguments anyway. Now that that only happens in one place, a macro would be overkill.
2004-02-27 18:52:44 +00:00
sched_sleep(struct thread *td)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
td->td_sched->ts_slptime = ticks;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
void
sched_wakeup(struct thread *td)
{
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
struct td_sched *ts;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
int slptime;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
ts = td->td_sched;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
* If we slept for more than a tick update our interactivity and
* priority.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
slptime = ts->ts_slptime;
ts->ts_slptime = 0;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if (slptime && slptime != ticks) {
- With a sleep time over 2097 seconds hzticks and slptime could end up negative. Use unsigned integers for sleep and run time so this doesn't disturb sched_interact_score(). This should fix the invalid interactive priority panics reported by several users.
2007-01-24 18:18:43 +00:00
u_int hzticks;
- Refrain from setting the td_priority in sched_wakeup(). It will be reset before we return to user space.
2003-03-03 04:11:40 +00:00
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
hzticks = (ticks - slptime) << SCHED_TICK_SHIFT;
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
ts->skg_slptime += hzticks;
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
sched_interact_update(td);
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
sched_pctcpu_update(ts);
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_priority(td);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
/* Reset the slice value after we sleep. */
ts->ts_slice = sched_slice;
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
sched_add(td, SRQ_BORING);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
/*
* Penalize the parent for creating a new child and initialize the child's
* priority.
*/
void
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_fork(struct thread *td, struct thread *child)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
sched_fork_thread(td, child);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* Penalize the parent and child for forking.
*/
sched_interact_fork(child);
sched_priority(child);
td->td_sched->skg_runtime += tickincr;
sched_interact_update(td);
sched_priority(td);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
}
void
sched_fork_thread(struct thread *td, struct thread *child)
{
struct td_sched *ts;
struct td_sched *ts2;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* Initialize child.
*/
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
sched_newthread(child);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
child->td_lock = &sched_lock;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = td->td_sched;
ts2 = child->td_sched;
ts2->ts_cpu = ts->ts_cpu;
ts2->ts_runq = NULL;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
* Grab our parents cpu estimation information and priority.
*/
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts2->ts_ticks = ts->ts_ticks;
ts2->ts_ltick = ts->ts_ltick;
ts2->ts_ftick = ts->ts_ftick;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
child->td_user_pri = td->td_user_pri;
child->td_base_user_pri = td->td_base_user_pri;
/*
* And update interactivity score.
*/
ts2->skg_slptime = ts->skg_slptime;
ts2->skg_runtime = ts->skg_runtime;
ts2->ts_slice = 1; /* Attempt to quickly learn interactivity. */
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
}
void
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_class(struct thread *td, int class)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
{
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
if (td->td_pri_class == class)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
return;
- Remove the ksq_loads[] array. We are only interested in three counts, the total load, the timeshare load, and the number of threads that can be migrated to another cpu. Account for these seperately. - Introduce a KSE_CAN_MIGRATE() macro which determines whether or not a KSE can be migrated to another CPU. Currently, this only checks to see if we're an interrupt handler. Eventually this will also be used to support CPU binding.
2003-11-02 10:56:48 +00:00
#ifdef SMP
- Fix the sched_priority() invalid priority bugs. Use roundup() instead of max() when computing the divisor in SCHED_TICK_PRI(). This prevents cases where rounding down would allow the quotient to exceed SCHED_PRI_RANGE. - Garbage collect some unused flags and fields. - Replace TDF_HOLD with sched_pin_td()/sched_unpin_td() since it simply duplicated this functionality. - Re-enable the rebalancer by default and fix the sysctl so it can be modified.
2007-01-06 08:44:13 +00:00
/*
* On SMP if we're on the RUNQ we must adjust the transferable
* count because could be changing to or from an interrupt
* class.
*/
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
if (TD_ON_RUNQ(td)) {
- Fix the sched_priority() invalid priority bugs. Use roundup() instead of max() when computing the divisor in SCHED_TICK_PRI(). This prevents cases where rounding down would allow the quotient to exceed SCHED_PRI_RANGE. - Garbage collect some unused flags and fields. - Replace TDF_HOLD with sched_pin_td()/sched_unpin_td() since it simply duplicated this functionality. - Re-enable the rebalancer by default and fix the sysctl so it can be modified.
2007-01-06 08:44:13 +00:00
struct tdq *tdq;
tdq = TDQ_CPU(td->td_sched->ts_cpu);
if (THREAD_CAN_MIGRATE(td)) {
tdq->tdq_transferable--;
tdq->tdq_group->tdg_transferable--;
}
td->td_pri_class = class;
if (THREAD_CAN_MIGRATE(td)) {
tdq->tdq_transferable++;
tdq->tdq_group->tdg_transferable++;
- Introduce kseq_runq_{add,rem}() which are used to insert and remove kses from the run queues. Also, on SMP, we track the transferable count here. Threads are transferable only as long as they are on the run queue. - Previously, we adjusted our load balancing based on the transferable count minus the number of actual cpus. This was done to account for the threads which were likely to be running. All of this logic is simpler now that transferable accounts for only those threads which can actually be taken. Updated various places in sched_add() and kseq_balance() to account for this. - Rename kseq_{add,rem} to kseq_load_{add,rem} to reflect what they're really doing. The load is accounted for seperately from the runq because the load is accounted for even as the thread is running. - Fix a bug in sched_class() where we weren't properly using the PRI_BASE() version of the kg_pri_class. - Add a large comment that describes the impact of a seemingly simple conditional in sched_add(). - Also in sched_add() check the transferable count and KSE_CAN_MIGRATE() prior to checking kseq_idle. This reduces the frequency of access for kseq_idle which is a shared resource.
2003-11-15 07:32:07 +00:00
}
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
}
- Fix the sched_priority() invalid priority bugs. Use roundup() instead of max() when computing the divisor in SCHED_TICK_PRI(). This prevents cases where rounding down would allow the quotient to exceed SCHED_PRI_RANGE. - Garbage collect some unused flags and fields. - Replace TDF_HOLD with sched_pin_td()/sched_unpin_td() since it simply duplicated this functionality. - Re-enable the rebalancer by default and fix the sysctl so it can be modified.
2007-01-06 08:44:13 +00:00
#endif
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
td->td_pri_class = class;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
/*
* Return some of the child's priority and interactivity to the parent.
*/
void
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
sched_exit(struct proc *p, struct thread *child)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
struct thread *td;
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
CTR3(KTR_SCHED, "sched_exit: %p(%s) prio %d",
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
child, child->td_proc->p_comm, child->td_priority);
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
PROC_SLOCK_ASSERT(p, MA_OWNED);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
td = FIRST_THREAD_IN_PROC(p);
sched_exit_thread(td, child);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
}
void
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
sched_exit_thread(struct thread *td, struct thread *child)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
CTR3(KTR_SCHED, "sched_exit_thread: %p(%s) prio %d",
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
child, child->td_proc->p_comm, child->td_priority);
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_lock(child);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
tdq_load_rem(TDQ_CPU(child->td_sched->ts_cpu), child->td_sched);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_unlock(child);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
#ifdef KSE
/*
* KSE forks and exits so often that this penalty causes short-lived
* threads to always be non-interactive. This causes mozilla to
* crawl under load.
*/
if ((td->td_pflags & TDP_SA) && td->td_proc == child->td_proc)
return;
#endif
/*
* Give the child's runtime to the parent without returning the
* sleep time as a penalty to the parent. This causes shells that
* launch expensive things to mark their children as expensive.
*/
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_lock(td);
Changes to try fix sched_ule.c courtesy of David Xu.
2006-12-06 06:55:59 +00:00
td->td_sched->skg_runtime += child->td_sched->skg_runtime;
sched_interact_update(td);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
sched_priority(td);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_unlock(td);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
}
void
sched_userret(struct thread *td)
{
/*
* XXX we cheat slightly on the locking here to avoid locking in
* the usual case. Setting td_priority here is essentially an
* incomplete workaround for not setting it properly elsewhere.
* Now that some interrupt handlers are threads, not setting it
* properly elsewhere can clobber it in the window between setting
* it here and returning to user mode, so don't waste time setting
* it perfectly here.
*/
KASSERT((td->td_flags & TDF_BORROWING) == 0,
("thread with borrowed priority returning to userland"));
if (td->td_priority != td->td_user_pri) {
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_lock(td);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
td->td_priority = td->td_user_pri;
td->td_base_pri = td->td_user_pri;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_unlock(td);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
}
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
void
- Update the sched api. sched_{add,rem,clock,pctcpu} now all accept a td argument rather than a kse.
2003-10-16 08:39:15 +00:00
sched_clock(struct thread *td)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
struct td_sched *ts;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
- Run sched_balance() and sched_balance_groups() from hardclock via sched_clock() rather than using callouts. This means we no longer have to take the load of the callout thread into consideration while balancing and should make the balancing decisions simpler and more accurate. Tested on: x86/UP, amd64/SMP
2004-06-02 05:46:48 +00:00
mtx_assert(&sched_lock, MA_OWNED);
#ifdef SMP
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
sched_smp_tick(td);
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
#endif
tdq = TDQ_SELF();
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
/*
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
* Advance the insert index once for each tick to ensure that all
* threads get a chance to run.
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
*/
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
if (tdq->tdq_idx == tdq->tdq_ridx) {
tdq->tdq_idx = (tdq->tdq_idx + 1) % RQ_NQS;
if (TAILQ_EMPTY(&tdq->tdq_timeshare.rq_queues[tdq->tdq_ridx]))
tdq->tdq_ridx = tdq->tdq_idx;
}
ts = td->td_sched;
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
/*
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
* We only do slicing code for TIMESHARE threads.
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
*/
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
if (td->td_pri_class != PRI_TIMESHARE)
- Use a better algorithm in sched_pctcpu_update() Contributed by: Thomaswuerfl@gmx.de - In sched_prio(), adjust the run queue for threads which may need to move to the current queue due to priority propagation . - In sched_switch(), fix style bug introduced when the KSE support went in. Columns are 80 chars wide, not 90. - In sched_switch(), Fix the comparison in the idle case and explicitly re-initialize the runq in the not propagated case. - Remove dead code in sched_clock(). - In sched_clock(), If we're an IDLE class td set NEEDRESCHED so that threads that have become runnable will get a chance to. - In sched_runnable(), if we're not the IDLETD, we should not consider curthread when examining the load. This mimics the 4BSD behavior of returning 0 when the only runnable thread is running. - In sched_userret(), remove the code for setting NEEDRESCHED entirely. This is not necessary and is not implemented in 4BSD. - Use the correct comparison in sched_add() when checking to see if an idle prio task has had it's priority temporarily elevated.
2003-10-27 06:47:05 +00:00
return;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
- Only allow the tdq_idx to increase by one each tick rather than up to the most recently chosen index. This significantly improves nice behavior. This allows a lower priority thread to run some multiple of times before the higher priority thread makes it to the front of the queue. A nice +20 cpu hog now only gets ~5% of the cpu when running with a nice 0 cpu hog and about 1.5% with a nice -20 hog. A nice difference of 1 makes a 4% difference in cpu usage between two hogs. - Track a seperate insert and removal index. When the removal index is empty it is updated to point at the current insert index. - Don't remove and re-add a thread to the runq when it is being adjusted down in priority. - Pull some conditional code out of sched_tick(). It's looking a bit large now.
2007-01-04 12:16:19 +00:00
* We used a tick; charge it to the thread so that we can compute our
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
* interactivity.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
td->td_sched->skg_runtime += tickincr;
sched_interact_update(td);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
* We used up one time slice.
*/
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (--ts->ts_slice > 0)
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
return;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/*
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
* We're out of time, recompute priorities and requeue.
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
*/
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
sched_priority(td);
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
td->td_flags |= TDF_NEEDRESCHED;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
int
sched_runnable(void)
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
- When a new thread is added to a kseq the load is incremented prior to adding it to the nice tables. Therefore, in kseq_add_nice, we should keep in mind that the load will be 1 if we are the only thread, and not 0. - Assert that the sched lock is held in all the appropriate places. - Increase the scope of the sched lock in sched_pctcpu_update(). - Hold the sched lock in sched_runnable(). It is not held by the caller.
2003-06-08 00:47:33 +00:00
int load;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
- When a new thread is added to a kseq the load is incremented prior to adding it to the nice tables. Therefore, in kseq_add_nice, we should keep in mind that the load will be 1 if we are the only thread, and not 0. - Assert that the sched lock is held in all the appropriate places. - Increase the scope of the sched lock in sched_pctcpu_update(). - Hold the sched lock in sched_runnable(). It is not held by the caller.
2003-06-08 00:47:33 +00:00
load = 1;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq = TDQ_SELF();
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#ifdef SMP
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (tdq_busy)
goto out;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#endif
- Use a better algorithm in sched_pctcpu_update() Contributed by: Thomaswuerfl@gmx.de - In sched_prio(), adjust the run queue for threads which may need to move to the current queue due to priority propagation . - In sched_switch(), fix style bug introduced when the KSE support went in. Columns are 80 chars wide, not 90. - In sched_switch(), Fix the comparison in the idle case and explicitly re-initialize the runq in the not propagated case. - Remove dead code in sched_clock(). - In sched_clock(), If we're an IDLE class td set NEEDRESCHED so that threads that have become runnable will get a chance to. - In sched_runnable(), if we're not the IDLETD, we should not consider curthread when examining the load. This mimics the 4BSD behavior of returning 0 when the only runnable thread is running. - In sched_userret(), remove the code for setting NEEDRESCHED entirely. This is not necessary and is not implemented in 4BSD. - Use the correct comparison in sched_add() when checking to see if an idle prio task has had it's priority temporarily elevated.
2003-10-27 06:47:05 +00:00
if ((curthread->td_flags & TDF_IDLETD) != 0) {
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (tdq->tdq_load > 0)
- Use a better algorithm in sched_pctcpu_update() Contributed by: Thomaswuerfl@gmx.de - In sched_prio(), adjust the run queue for threads which may need to move to the current queue due to priority propagation . - In sched_switch(), fix style bug introduced when the KSE support went in. Columns are 80 chars wide, not 90. - In sched_switch(), Fix the comparison in the idle case and explicitly re-initialize the runq in the not propagated case. - Remove dead code in sched_clock(). - In sched_clock(), If we're an IDLE class td set NEEDRESCHED so that threads that have become runnable will get a chance to. - In sched_runnable(), if we're not the IDLETD, we should not consider curthread when examining the load. This mimics the 4BSD behavior of returning 0 when the only runnable thread is running. - In sched_userret(), remove the code for setting NEEDRESCHED entirely. This is not necessary and is not implemented in 4BSD. - Use the correct comparison in sched_add() when checking to see if an idle prio task has had it's priority temporarily elevated.
2003-10-27 06:47:05 +00:00
goto out;
} else
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (tdq->tdq_load - 1 > 0)
- Use a better algorithm in sched_pctcpu_update() Contributed by: Thomaswuerfl@gmx.de - In sched_prio(), adjust the run queue for threads which may need to move to the current queue due to priority propagation . - In sched_switch(), fix style bug introduced when the KSE support went in. Columns are 80 chars wide, not 90. - In sched_switch(), Fix the comparison in the idle case and explicitly re-initialize the runq in the not propagated case. - Remove dead code in sched_clock(). - In sched_clock(), If we're an IDLE class td set NEEDRESCHED so that threads that have become runnable will get a chance to. - In sched_runnable(), if we're not the IDLETD, we should not consider curthread when examining the load. This mimics the 4BSD behavior of returning 0 when the only runnable thread is running. - In sched_userret(), remove the code for setting NEEDRESCHED entirely. This is not necessary and is not implemented in 4BSD. - Use the correct comparison in sched_add() when checking to see if an idle prio task has had it's priority temporarily elevated.
2003-10-27 06:47:05 +00:00
goto out;
- When a new thread is added to a kseq the load is incremented prior to adding it to the nice tables. Therefore, in kseq_add_nice, we should keep in mind that the load will be 1 if we are the only thread, and not 0. - Assert that the sched lock is held in all the appropriate places. - Increase the scope of the sched lock in sched_pctcpu_update(). - Hold the sched lock in sched_runnable(). It is not held by the caller.
2003-06-08 00:47:33 +00:00
load = 0;
out:
return (load);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
struct thread *
- Fix the ksq_load calculation. It now reflects the number of entries on the run queue for each cpu. - Introduce kse stealing into the sched_choose() code. This helps balance cpus better in cases where process turnover is high. This implementation is fairly trivial and will likely be only a temporary measure until something more sophisticated has been written.
2003-01-28 09:28:20 +00:00
sched_choose(void)
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
struct td_sched *ts;
- Fix the ksq_load calculation. It now reflects the number of entries on the run queue for each cpu. - Introduce kse stealing into the sched_choose() code. This helps balance cpus better in cases where process turnover is high. This implementation is fairly trivial and will likely be only a temporary measure until something more sophisticated has been written.
2003-01-28 09:28:20 +00:00
- When a new thread is added to a kseq the load is incremented prior to adding it to the nice tables. Therefore, in kseq_add_nice, we should keep in mind that the load will be 1 if we are the only thread, and not 0. - Assert that the sched lock is held in all the appropriate places. - Increase the scope of the sched lock in sched_pctcpu_update(). - Hold the sched lock in sched_runnable(). It is not held by the caller.
2003-06-08 00:47:33 +00:00
mtx_assert(&sched_lock, MA_OWNED);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq = TDQ_SELF();
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
#ifdef SMP
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
restart:
- Add a SYSCTL node for the ule scheduler. - Allow user adjustable min and max time slices (suggested by hiten). - Change the SLP_RUN_MAX to 100ms from 2 seconds so that we learn whether a process is interactive or not much more quickly. - Place a process on the current run queue if it is interactive or if it is running at an interrupt thread priority due to priority prop. - Use the 'current' timeshare queue for interrupt threads, realtime threads, and idle threads that are running at higher priority due to priority prop. This fixes problems where priorities would have been elevated but we would not check the timeshare run queue until other lower priority tasks were no longer runnable. - Keep an array of loads indexed by the priority class as well as a global load. - Keep an bucket of nice values with a count of the number of kses currently runnable with that nice value. - Keep track of the minimum nice value of any running thread. - Remove the unused short term sleep accounting. I was attempting to use this for load balancing but it didn't work out. - Define a kseq_print() for use with debugging. - Add KTR debugging at useful places so we can easily debug slice and priority assignment. - Decouple the runq assignment from the kseq assignment. kseq_add now keeps track of statistics. This is done so that the nice and load is still tracked for the currently running process. Previously if a niced process was added while a non nice process was running the niced process would still get a slice since it was not aware of the unnice process. - Make adjustments for the sched api changes.
2003-04-11 03:47:14 +00:00
#endif
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = tdq_choose(tdq);
if (ts) {
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#ifdef SMP
- Fix a comparison in sched_choose() that caused cpus to be constantly marked idle, thus breaking cpu load balancing. - Change sched_interact_update() to fix cases where the stored history has expanded significantly rather than handling them in the callers. This fixes a case where sched_priority() could compute a bad value. - Add a sysctl to disable the global load balancer for experimentation.
2007-01-05 23:45:38 +00:00
if (ts->ts_thread->td_priority > PRI_MIN_IDLE)
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (tdq_idled(tdq) == 0)
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
goto restart;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#endif
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_runq_rem(tdq, ts);
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
return (ts->ts_thread);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
- Fix the ksq_load calculation. It now reflects the number of entries on the run queue for each cpu. - Introduce kse stealing into the sched_choose() code. This helps balance cpus better in cases where process turnover is high. This implementation is fairly trivial and will likely be only a temporary measure until something more sophisticated has been written.
2003-01-28 09:28:20 +00:00
#ifdef SMP
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (tdq_idled(tdq) == 0)
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
goto restart;
- Fix the ksq_load calculation. It now reflects the number of entries on the run queue for each cpu. - Introduce kse stealing into the sched_choose() code. This helps balance cpus better in cases where process turnover is high. This implementation is fairly trivial and will likely be only a temporary measure until something more sophisticated has been written.
2003-01-28 09:28:20 +00:00
#endif
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
return (PCPU_GET(idlethread));
}
static int
sched_preempt(struct thread *td)
{
struct thread *ctd;
int cpri;
int pri;
ctd = curthread;
pri = td->td_priority;
cpri = ctd->td_priority;
if (panicstr != NULL || pri >= cpri || cold || TD_IS_INHIBITED(ctd))
return (0);
/*
* Always preempt IDLE threads. Otherwise only if the preempting
* thread is an ithread.
*/
if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE)
return (0);
if (ctd->td_critnest > 1) {
CTR1(KTR_PROC, "sched_preempt: in critical section %d",
ctd->td_critnest);
ctd->td_owepreempt = 1;
return (0);
}
/*
* Thread is runnable but not yet put on system run queue.
*/
MPASS(TD_ON_RUNQ(td));
TD_SET_RUNNING(td);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
MPASS(ctd->td_lock == &sched_lock);
MPASS(td->td_lock == &sched_lock);
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
CTR3(KTR_PROC, "preempting to thread %p (pid %d, %s)\n", td,
td->td_proc->p_pid, td->td_proc->p_comm);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
/*
* We enter the switch with two runnable threads that both have
* the same lock. When we return td may be sleeping so we need
* to switch locks to make sure he's locked correctly.
*/
SCHED_STAT_INC(switch_preempt);
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
mi_switch(SW_INVOL|SW_PREEMPT, td);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
spinlock_enter();
thread_unlock(ctd);
thread_lock(td);
spinlock_exit();
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
return (1);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
void
Give setrunqueue() and sched_add() more of a clue as to where they are coming from and what is expected from them. MFC after: 2 days
2004-09-01 02:11:28 +00:00
sched_add(struct thread *td, int flags)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
struct td_sched *ts;
- Fix a long standing problem where an ithread would not honor sched_pin(). - Remove the sched_add wrapper that used sched_add_internal() as a backend. Its only purpose was to interpret one flag and turn it into an int. Do the right thing and interpret the flag in sched_add() instead. - Pass the flag argument to sched_add() to kseq_runq_add() so that we can get the SRQ_PREEMPT optimization too. - Add a KEF_INTERNAL flag. If KEF_INTERNAL is set we don't adjust the SLOT counts, otherwise the slot counts are adjusted as soon as we enter sched_add() or sched_rem() rather than when the thread is actually placed on the run queue. This greatly simplifies the handling of slots. - Remove the explicit prevention of migration for ithreads on non-x86 platforms. This was never shown to have any real benefit. - Remove the unused class argument to KSE_CAN_MIGRATE(). - Add ktr points for thread migration events. - Fix a long standing bug on platforms which don't initialize the cpu topology. The ksg_maxid variable was never correctly set on these platforms which caused the long term load balancer to never inspect more than the first group or processor. - Fix another bug which prevented the long term load balancer from working properly. If stathz != hz we can't expect sched_clock() to be called on the exact tick count that we're anticipating. - Rearrange sched_switch() a bit to reduce indentation levels.
2004-12-26 22:56:08 +00:00
int preemptive;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
int class;
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#ifdef SMP
int cpuid;
int cpumask;
#endif
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
ts = td->td_sched;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
- Remove earlier KTR_ULE tracepoints. - Define new KTR_SCHED points so that we can graph the operation of the scheduler.
2004-12-26 00:15:33 +00:00
CTR5(KTR_SCHED, "sched_add: %p(%s) prio %d by %p(%s)",
td, td->td_proc->p_comm, td->td_priority, curthread,
curthread->td_proc->p_comm);
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
KASSERT((td->td_inhibitors == 0),
("sched_add: trying to run inhibited thread"));
KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)),
("sched_add: bad thread state"));
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
KASSERT(td->td_proc->p_sflag & PS_INMEM,
- Make some context switches conditional on SCHED_STRICT_RESCHED. This may have some negative effect on interactivity but it yields great perf. gains. This also brings the conditions under which ULE context switches inline with SCHED_4BSD. - Define some new kseq_* functions for manipulating the run queue. - Add a new kseq member ksq_rslices and ksq_bload. rslices is the sum of the slices of runnable kses. This will be used for push load balance decisions. bload is the number of threads blocked waiting on IO.
2003-02-03 05:30:07 +00:00
("sched_add: process swapped out"));
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
/*
* Now that the thread is moving to the run-queue, set the lock
* to the scheduler's lock.
*/
if (td->td_lock != &sched_lock) {
mtx_lock_spin(&sched_lock);
thread_lock_set(td, &sched_lock);
}
mtx_assert(&sched_lock, MA_OWNED);
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
TD_SET_RUNQ(td);
tdq = TDQ_SELF();
class = PRI_BASE(td->td_pri_class);
preemptive = !(flags & SRQ_YIELDING);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/*
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
* Recalculate the priority before we select the target cpu or
* run-queue.
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
*/
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
if (class == PRI_TIMESHARE)
sched_priority(td);
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
if (ts->ts_slice == 0)
ts->ts_slice = sched_slice;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#ifdef SMP
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
cpuid = PCPU_GET(cpuid);
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
/*
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
* Pick the destination cpu and if it isn't ours transfer to the
* target cpu.
- Use a new flag, KEF_XFERABLE, to record with certainty that this kse had contributed to the transferable load count. This prevents any potential problems with sched_pin() being used around calls to setrunqueue(). - Change the sched_add() load balancing algorithm to try to migrate on wakeup. This attempts to place threads that communicate with each other on the same CPU. - Don't clear the idle counts in kseq_transfer(), let the cpus do that when they call sched_add() from kseq_assign(). - Correct a few out of date comments. - Make sure the ke_cpu field is correct when we preempt. - Call kseq_assign() from sched_clock() to catch any assignments that were done without IPI. Presently all assignments are done with an IPI, but I'm trying a patch that limits that. - Don't migrate a thread if it is still runnable in sched_add(). Previously, this could only happen for KSE threads, but due to changes to sched_switch() all threads went through this path. - Remove some code that was added with preemption but is not necessary.
2004-08-10 07:52:21 +00:00
*/
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (THREAD_CAN_MIGRATE(td)) {
if (td->td_priority <= PRI_MAX_ITHD) {
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR2(KTR_ULE, "ithd %d < %d",
td->td_priority, PRI_MAX_ITHD);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
ts->ts_cpu = cpuid;
Schedule the ithread on the same cpu as the interrupt Tested by: kmacy Submitted by: jeffr
2007-04-20 05:45:46 +00:00
} else if (pick_pri)
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
ts->ts_cpu = tdq_pickpri(tdq, ts, flags);
else
ts->ts_cpu = tdq_pickidle(tdq, ts);
} else
- Get rid of the unused DIDRUN flag. This was really only present to support sched_4bsd. - Rename the KTR level for non schedgraph parsed events. They take event space from things we'd like to graph. - Reset our slice value after we sleep. The slice is simply there to prevent starvation among equal priorities. A thread which had almost exhausted it's slice and then slept doesn't need to be rescheduled a tick after it wakes up. - Set the maximum slice value to a more conservative 100ms now that it is more accurately enforced.
2007-01-25 19:14:11 +00:00
CTR1(KTR_ULE, "pinned %d", td->td_pinned);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
if (ts->ts_cpu != cpuid)
preemptive = 0;
tdq = TDQ_CPU(ts->ts_cpu);
cpumask = 1 << ts->ts_cpu;
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
/*
- In kseq_transfer() return if smp has not been started. - In sched_add(), do the idle check prior to the transfer check so that we don't try to transfer load from an idle cpu. This fixes panics caused by IPIs on UP machines running SMP kernels. Reported/Debugged by: seanc
2003-12-20 14:03:14 +00:00
* If we had been idle, clear our bit in the group and potentially
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
* the global bitmap.
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
*/
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
if ((class != PRI_IDLE && class != PRI_ITHD) &&
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
(tdq->tdq_group->tdg_idlemask & cpumask) != 0) {
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
/*
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
* Check to see if our group is unidling, and if so, remove it
* from the global idle mask.
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
*/
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
if (tdq->tdq_group->tdg_idlemask ==
tdq->tdq_group->tdg_cpumask)
atomic_clear_int(&tdq_idle, tdq->tdq_group->tdg_mask);
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
/*
* Now remove ourselves from the group specific idle mask.
*/
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
tdq->tdq_group->tdg_idlemask &= ~cpumask;
}
- Add static to local functions and data where it was missing. - Add an IPI based mechanism for migrating kses. This mechanism is broken down into several components. This is intended to reduce cache thrashing by eliminating most cases where one cpu touches another's run queues. - kseq_notify() appends a kse to a lockless singly linked list and conditionally sends an IPI to the target processor. Right now this is protected by sched_lock but at some point I'd like to get rid of the global lock. This is why I used something more complicated than a standard queue. - kseq_assign() processes our list of kses that have been assigned to us by other processors. This simply calls sched_add() for each item on the list after clearing the new KEF_ASSIGNED flag. This flag is used to indicate that we have been appeneded to the assigned queue but not added to the run queue yet. - In sched_add(), instead of adding a KSE to another processor's queue we use kse_notify() so that we don't touch their queue. Also in sched_add(), if KEF_ASSIGNED is already set return immediately. This can happen if a thread is removed and readded so that the priority is recorded properly. - In sched_rem() return immediately if KEF_ASSIGNED is set. All callers immediately readd simply to adjust priorites etc. - In sched_choose(), if we're running an IDLE task or the per cpu idle thread set our cpumask bit in 'kseq_idle' so that other processors may know that we are idle. Before this, make a single pass through the run queues of other processors so that we may find work more immediately if it is available. - In sched_runnable(), don't scan each processor's run queue, they will IPI us if they have work for us to do. - In sched_add(), if we're adding a thread that can be migrated and we have plenty of work to do, try to migrate the thread to an idle kseq. - Simplify the logic in sched_prio() and take the KEF_ASSIGNED flag into consideration. - No longer use kseq_choose() to steal threads, it can lose it's last argument. - Create a new function runq_steal() which operates like runq_choose() but skips threads based on some criteria. Currently it will not steal PRI_ITHD threads. In the future this will be used for CPU binding. - Create a kseq_steal() that checks each run queue with runq_steal(), use kseq_steal() in the places where we used kseq_choose() to steal with before.
2003-10-31 11:16:04 +00:00
#endif
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
/*
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
* Pick the run queue based on priority.
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
*/
if (td->td_priority <= PRI_MAX_REALTIME)
ts->ts_runq = &tdq->tdq_realtime;
else if (td->td_priority <= PRI_MAX_TIMESHARE)
ts->ts_runq = &tdq->tdq_timeshare;
else
ts->ts_runq = &tdq->tdq_idle;
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
if (preemptive && sched_preempt(td))
Implement preemption of kernel threads natively in the scheduler rather than as one-off hacks in various other parts of the kernel: - Add a function maybe_preempt() that is called from sched_add() to determine if a thread about to be added to a run queue should be preempted to directly. If it is not safe to preempt or if the new thread does not have a high enough priority, then the function returns false and sched_add() adds the thread to the run queue. If the thread should be preempted to but the current thread is in a nested critical section, then the flag TDF_OWEPREEMPT is set and the thread is added to the run queue. Otherwise, mi_switch() is called immediately and the thread is never added to the run queue since it is switch to directly. When exiting an outermost critical section, if TDF_OWEPREEMPT is set, then clear it and call mi_switch() to perform the deferred preemption. - Remove explicit preemption from ithread_schedule() as calling setrunqueue() now does all the correct work. This also removes the do_switch argument from ithread_schedule(). - Do not use the manual preemption code in mtx_unlock if the architecture supports native preemption. - Don't call mi_switch() in a loop during shutdown to give ithreads a chance to run if the architecture supports native preemption since the ithreads will just preempt DELAY(). - Don't call mi_switch() from the page zeroing idle thread for architectures that support native preemption as it is unnecessary. - Native preemption is enabled on the same archs that supported ithread preemption, namely alpha, i386, and amd64. This change should largely be a NOP for the default case as committed except that we will do fewer context switches in a few cases and will avoid the run queues completely when preempting. Approved by: scottl (with his re@ hat)
2004-07-02 20:21:44 +00:00
return;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq_runq_add(tdq, ts, flags);
tdq_load_add(tdq, ts);
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#ifdef SMP
if (ts->ts_cpu != cpuid) {
tdq_notify(ts);
return;
}
#endif
if (td->td_priority < curthread->td_priority)
curthread->td_flags |= TDF_NEEDRESCHED;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
void
- Update the sched api. sched_{add,rem,clock,pctcpu} now all accept a td argument rather than a kse.
2003-10-16 08:39:15 +00:00
sched_rem(struct thread *td)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct tdq *tdq;
struct td_sched *ts;
- Update the sched api. sched_{add,rem,clock,pctcpu} now all accept a td argument rather than a kse.
2003-10-16 08:39:15 +00:00
- Remove earlier KTR_ULE tracepoints. - Define new KTR_SCHED points so that we can graph the operation of the scheduler.
2004-12-26 00:15:33 +00:00
CTR5(KTR_SCHED, "sched_rem: %p(%s) prio %d by %p(%s)",
td, td->td_proc->p_comm, td->td_priority, curthread,
curthread->td_proc->p_comm);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = td->td_sched;
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
KASSERT(TD_ON_RUNQ(td),
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
("sched_rem: thread not on run queue"));
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
tdq = TDQ_CPU(ts->ts_cpu);
tdq_runq_rem(tdq, ts);
tdq_load_rem(tdq, ts);
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
TD_SET_CAN_RUN(td);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
fixpt_t
- Update the sched api. sched_{add,rem,clock,pctcpu} now all accept a td argument rather than a kse.
2003-10-16 08:39:15 +00:00
sched_pctcpu(struct thread *td)
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
{
fixpt_t pctcpu;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched *ts;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
pctcpu = 0;
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = td->td_sched;
if (ts == NULL)
- If a thread is not bound to a kse return 0 from sched_pctcpu(). Reported by: pawel.worach@nordea.com
2003-10-20 19:55:21 +00:00
return (0);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_lock(td);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
if (ts->ts_ticks) {
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
int rtick;
- ftick was initialized to -1 for init and any of it's children. Fix this by setting ftick = ltick = ticks in schedinit(). - Update the priority when we are pulled off of the run queue and when we are inserted onto the run queue so that it more accurately reflects our present status. This is important for efficient priority propagation functioning. - Move the frequency test into sched_pctcpu_update() so we don't repeat it each time we'd like to call it. - Put some temporary work-around code in sched_priority() in case the tick mechanism produces a bad priority. Eventually this should revert to an assert again.
2007-01-05 08:50:38 +00:00
sched_pctcpu_update(ts);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
/* How many rtick per second ? */
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
rtick = min(SCHED_TICK_HZ(ts) / SCHED_TICK_SECS, hz);
pctcpu = (FSCALE * ((FSCALE * rtick)/hz)) >> FSHIFT;
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
}
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
td->td_proc->p_swtime = ts->ts_ltick - ts->ts_ftick;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_unlock(td);
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
return (pctcpu);
}
- Add initial support for pinning and binding.
2003-11-04 07:45:41 +00:00
void
sched_bind(struct thread *td, int cpu)
{
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
struct td_sched *ts;
- Add initial support for pinning and binding.
2003-11-04 07:45:41 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts = td->td_sched;
- In tdq_transfer() always set NEEDRESCHED when necessary regardless of the ipi settings. If NEEDRESCHED is set and an ipi is later delivered it will clear it rather than cause extra context switches. However, if we miss setting it we can have terrible latency. - In sched_bind() correctly implement bind. Also be slightly more tolerant of code which calls bind multiple times. However, we don't change binding if another call is made with a different cpu. This does not presently work with hwpmc which I believe should be changed.
2007-01-20 09:03:43 +00:00
if (ts->ts_flags & TSF_BOUND)
- We do need to IPI the idlethread on some systems. It may be stuck in a power saving mode otherwise. - If the thread is already bound in sched_bind() unbind it before re-binding it to a new cpu. I don't like these semantics but they are expected by some code in the tree. Patch by jkoshy.
2007-01-20 17:03:33 +00:00
sched_unbind(td);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
ts->ts_flags |= TSF_BOUND;
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
#ifdef SMP
- In tdq_transfer() always set NEEDRESCHED when necessary regardless of the ipi settings. If NEEDRESCHED is set and an ipi is later delivered it will clear it rather than cause extra context switches. However, if we miss setting it we can have terrible latency. - In sched_bind() correctly implement bind. Also be slightly more tolerant of code which calls bind multiple times. However, we don't change binding if another call is made with a different cpu. This does not presently work with hwpmc which I believe should be changed.
2007-01-20 09:03:43 +00:00
sched_pin();
- Add support for CPU groups to ule. All SMT cores on the same physical cpu are added to a group. - Don't place a cpu into the kseq_idle bitmask until all cpus in that group have idled. - Prefer idle groups over idle group members in the new kseq_transfer() function. In this way we will prefer to balance load across full cores rather than add further load a partial core. - Before a cpu goes idle, check the other group members for threads. Since SMT cpus may freely share threads, this is cheap. - SMT cores may be individually pinned and bound to now. This contrasts the old mechanism where binding or pinning would have allowed a thread to run on any available cpu. - Remove some unnecessary logic from sched_switch(). Priority propagation should be properly taken care of in sched_prio() now.
2003-12-11 03:57:10 +00:00
if (PCPU_GET(cpuid) == cpu)
- Add initial support for pinning and binding.
2003-11-04 07:45:41 +00:00
return;
- In tdq_transfer() always set NEEDRESCHED when necessary regardless of the ipi settings. If NEEDRESCHED is set and an ipi is later delivered it will clear it rather than cause extra context switches. However, if we miss setting it we can have terrible latency. - In sched_bind() correctly implement bind. Also be slightly more tolerant of code which calls bind multiple times. However, we don't change binding if another call is made with a different cpu. This does not presently work with hwpmc which I believe should be changed.
2007-01-20 09:03:43 +00:00
ts->ts_cpu = cpu;
- Add initial support for pinning and binding.
2003-11-04 07:45:41 +00:00
/* When we return from mi_switch we'll be on the correct cpu. */
Add NULL arg to mi_switch() call to stop kernel compiles from breaking.
2004-07-03 16:57:51 +00:00
mi_switch(SW_VOL, NULL);
- Add initial support for pinning and binding.
2003-11-04 07:45:41 +00:00
#endif
}
void
sched_unbind(struct thread *td)
{
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
struct td_sched *ts;
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
ts = td->td_sched;
- In tdq_transfer() always set NEEDRESCHED when necessary regardless of the ipi settings. If NEEDRESCHED is set and an ipi is later delivered it will clear it rather than cause extra context switches. However, if we miss setting it we can have terrible latency. - In sched_bind() correctly implement bind. Also be slightly more tolerant of code which calls bind multiple times. However, we don't change binding if another call is made with a different cpu. This does not presently work with hwpmc which I believe should be changed.
2007-01-20 09:03:43 +00:00
if ((ts->ts_flags & TSF_BOUND) == 0)
return;
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
ts->ts_flags &= ~TSF_BOUND;
#ifdef SMP
sched_unpin();
#endif
- Add initial support for pinning and binding.
2003-11-04 07:45:41 +00:00
}
Bring a working snapshot of hwpmc(4), its associated libraries, userland utilities and documentation into -CURRENT. Bump FreeBSD_version. Reviewed by: alc, jhb (kernel changes)
2005-04-19 04:01:25 +00:00
int
sched_is_bound(struct thread *td)
{
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
THREAD_LOCK_ASSERT(td, MA_OWNED);
Threading cleanup.. part 2 of several. Make part of John Birrell's KSE patch permanent.. Specifically, remove: Any reference of the ksegrp structure. This feature was never fully utilised and made things overly complicated. All code in the scheduler that tried to make threaded programs fair to unthreaded programs. Libpthread processes will already do this to some extent and libthr processes already disable it. Also: Since this makes such a big change to the scheduler(s), take the opportunity to rename some structures and elements that had to be moved anyhow. This makes the code a lot more readable. The ULE scheduler compiles again but I have no idea if it works. The 4bsd scheduler still reqires a little cleaning and some functions that now do ALMOST nothing will go away, but I thought I'd do that as a separate commit. Tested by David Xu, and Dan Eischen using libthr and libpthread.
2006-12-06 06:34:57 +00:00
return (td->td_sched->ts_flags & TSF_BOUND);
Bring a working snapshot of hwpmc(4), its associated libraries, userland utilities and documentation into -CURRENT. Bump FreeBSD_version. Reviewed by: alc, jhb (kernel changes)
2005-04-19 04:01:25 +00:00
}
Add scheduler API sched_relinquish(), the API is used to implement yield() and sched_yield() syscalls. Every scheduler has its own way to relinquish cpu, the ULE and CORE schedulers have two internal run- queues, a timesharing thread which calls yield() syscall should be moved to inactive queue.
2006-06-15 06:37:39 +00:00
void
sched_relinquish(struct thread *td)
{
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_lock(td);
Make KSE a kernel option, turned on by default in all GENERIC kernel configs except sun4v (which doesn't process signals properly with KSE). Reviewed by: davidxu@
2006-10-26 21:42:22 +00:00
if (td->td_pri_class == PRI_TIMESHARE)
Add scheduler API sched_relinquish(), the API is used to implement yield() and sched_yield() syscalls. Every scheduler has its own way to relinquish cpu, the ULE and CORE schedulers have two internal run- queues, a timesharing thread which calls yield() syscall should be moved to inactive queue.
2006-06-15 06:37:39 +00:00
sched_prio(td, PRI_MAX_TIMESHARE);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
SCHED_STAT_INC(switch_relinquish);
Add scheduler API sched_relinquish(), the API is used to implement yield() and sched_yield() syscalls. Every scheduler has its own way to relinquish cpu, the ULE and CORE schedulers have two internal run- queues, a timesharing thread which calls yield() syscall should be moved to inactive queue.
2006-06-15 06:37:39 +00:00
mi_switch(SW_VOL, NULL);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
thread_unlock(td);
Add scheduler API sched_relinquish(), the API is used to implement yield() and sched_yield() syscalls. Every scheduler has its own way to relinquish cpu, the ULE and CORE schedulers have two internal run- queues, a timesharing thread which calls yield() syscall should be moved to inactive queue.
2006-06-15 06:37:39 +00:00
}
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
int
sched_load(void)
{
#ifdef SMP
int total;
int i;
total = 0;
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
for (i = 0; i <= tdg_maxid; i++)
total += TDQ_GROUP(i)->tdg_load;
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
return (total);
#else
- Clean up a bit after the most recent KSE restructuring.
2006-12-29 10:37:07 +00:00
return (TDQ_SELF()->tdq_sysload);
- Add a new member to struct kseq called ksq_sysload. This is intended to track the load for the sched_load() function. In the SMP case this member is not defined because it would be redundant with the ksg_load member which already tracks the non ithd load. - For sched_load() in the UP case simply return ksq_sysload. In the SMP case traverse the list of kseq groups and sum up their ksg_load fields.
2004-02-01 02:48:36 +00:00
#endif
}
- Add the ule scheduler. This is intended to be a general purpose process scheduler with many SMP benefits. It is still very experimental and should be used only in test environments.
2003-01-26 05:23:15 +00:00
int
sched_sizeof_proc(void)
{
return (sizeof(struct proc));
}
int
sched_sizeof_thread(void)
{
return (sizeof(struct thread) + sizeof(struct td_sched));
}
Add scheduler CORE, the work I have done half a year ago, recent, I picked it up again. The scheduler is forked from ULE, but the algorithm to detect an interactive process is almost completely different with ULE, it comes from Linux paper "Understanding the Linux 2.6.8.1 CPU Scheduler", although I still use same word "score" as a priority boost in ULE scheduler. Briefly, the scheduler has following characteristic: 1. Timesharing process's nice value is seriously respected, timeslice and interaction detecting algorithm are based on nice value. 2. per-cpu scheduling queue and load balancing. 3. O(1) scheduling. 4. Some cpu affinity code in wakeup path. 5. Support POSIX SCHED_FIFO and SCHED_RR. Unlike scheduler 4BSD and ULE which using fuzzy RQ_PPQ, the scheduler uses 256 priority queues. Unlike ULE which using pull and push, the scheduelr uses pull method, the main reason is to let relative idle cpu do the work, but current the whole scheduler is protected by the big sched_lock, so the benefit is not visible, it really can be worse than nothing because all other cpu are locked out when we are doing balancing work, which the 4BSD scheduelr does not have this problem. The scheduler does not support hyperthreading very well, in fact, the scheduler does not make the difference between physical CPU and logical CPU, this should be improved in feature. The scheduler has priority inversion problem on MP machine, it is not good for realtime scheduling, it can cause realtime process starving. As a result, it seems the MySQL super-smack runs better on my Pentium-D machine when using libthr, despite on UP or SMP kernel.
2006-06-13 13:12:56 +00:00
void
sched_tick(void)
{
- Catch up to setrunqueue/choosethread/etc. api changes. - Define our own maybe_preempt() as sched_preempt(). We want to be able to preempt idlethread in all cases. - Define our idlethread to require preemption to exit. - Get the cpu estimation tick from sched_tick() so we don't have to worry about errors from a sampling interval that differs from the time domain. This was the source of sched_priority prints/panics and inaccurate pctcpu display in top.
2007-01-23 08:50:34 +00:00
struct td_sched *ts;
ts = curthread->td_sched;
/* Adjust ticks for pctcpu */
ts->ts_ticks += 1 << SCHED_TICK_SHIFT;
ts->ts_ltick = ticks;
/*
* Update if we've exceeded our desired tick threshhold by over one
* second.
*/
if (ts->ts_ftick + SCHED_TICK_MAX < ts->ts_ltick)
sched_pctcpu_update(ts);
}
/*
* The actual idle process.
*/
void
sched_idletd(void *dummy)
{
struct proc *p;
struct thread *td;
td = curthread;
p = td->td_proc;
mtx_assert(&Giant, MA_NOTOWNED);
/* ULE Relies on preemption for idle interruption. */
for (;;)
cpu_idle();
Add scheduler CORE, the work I have done half a year ago, recent, I picked it up again. The scheduler is forked from ULE, but the algorithm to detect an interactive process is almost completely different with ULE, it comes from Linux paper "Understanding the Linux 2.6.8.1 CPU Scheduler", although I still use same word "score" as a priority boost in ULE scheduler. Briefly, the scheduler has following characteristic: 1. Timesharing process's nice value is seriously respected, timeslice and interaction detecting algorithm are based on nice value. 2. per-cpu scheduling queue and load balancing. 3. O(1) scheduling. 4. Some cpu affinity code in wakeup path. 5. Support POSIX SCHED_FIFO and SCHED_RR. Unlike scheduler 4BSD and ULE which using fuzzy RQ_PPQ, the scheduler uses 256 priority queues. Unlike ULE which using pull and push, the scheduelr uses pull method, the main reason is to let relative idle cpu do the work, but current the whole scheduler is protected by the big sched_lock, so the benefit is not visible, it really can be worse than nothing because all other cpu are locked out when we are doing balancing work, which the 4BSD scheduelr does not have this problem. The scheduler does not support hyperthreading very well, in fact, the scheduler does not make the difference between physical CPU and logical CPU, this should be improved in feature. The scheduler has priority inversion problem on MP machine, it is not good for realtime scheduling, it can cause realtime process starving. As a result, it seems the MySQL super-smack runs better on my Pentium-D machine when using libthr, despite on UP or SMP kernel.
2006-06-13 13:12:56 +00:00
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
/*
* A CPU is entering for the first time or a thread is exiting.
*/
void
sched_throw(struct thread *td)
{
/*
* Correct spinlock nesting. The idle thread context that we are
* borrowing was created so that it would start out with a single
* spin lock (sched_lock) held in fork_trampoline(). Since we've
* explicitly acquired locks in this function, the nesting count
* is now 2 rather than 1. Since we are nested, calling
* spinlock_exit() will simply adjust the counts without allowing
* spin lock using code to interrupt us.
*/
if (td == NULL) {
mtx_lock_spin(&sched_lock);
spinlock_exit();
} else {
MPASS(td->td_lock == &sched_lock);
}
mtx_assert(&sched_lock, MA_OWNED);
KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count"));
PCPU_SET(switchtime, cpu_ticks());
PCPU_SET(switchticks, ticks);
cpu_throw(td, choosethread()); /* doesn't return */
}
void
- Move some common code out of sched_fork_exit() and back into fork_exit().
2007-06-12 07:47:09 +00:00
sched_fork_exit(struct thread *td)
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
{
/*
* Finish setting up thread glue so that it begins execution in a
* non-nested critical section with sched_lock held but not recursed.
*/
- Move some common code out of sched_fork_exit() and back into fork_exit().
2007-06-12 07:47:09 +00:00
td->td_oncpu = PCPU_GET(cpuid);
sched_lock.mtx_lock = (uintptr_t)td;
THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED);
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
}
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
static SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RW, 0, "Scheduler");
SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "ule", 0,
"Scheduler name");
SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, interact, CTLFLAG_RW, &sched_interact, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, tickincr, CTLFLAG_RD, &tickincr, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, realstathz, CTLFLAG_RD, &realstathz, 0, "");
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#ifdef SMP
SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri, CTLFLAG_RW, &pick_pri, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_affinity, CTLFLAG_RW,
&affinity, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_tryself, CTLFLAG_RW,
&tryself, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, pick_pri_tryselfidle, CTLFLAG_RW,
&tryselfidle, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, ipi_preempt, CTLFLAG_RW, &ipi_preempt, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, ipi_ast, CTLFLAG_RW, &ipi_ast, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, ipi_thresh, CTLFLAG_RW, &ipi_thresh, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, steal_htt, CTLFLAG_RW, &steal_htt, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, steal_busy, CTLFLAG_RW, &steal_busy, 0, "");
SYSCTL_INT(_kern_sched, OID_AUTO, busy_thresh, CTLFLAG_RW, &busy_thresh, 0, "");
Commit 1/14 of sched_lock decomposition. - Move all scheduler locking into the schedulers utilizing a technique similar to solaris's container locking. - A per-process spinlock is now used to protect the queue of threads, thread count, suspension count, p_sflags, and other process related scheduling fields. - The new thread lock is actually a pointer to a spinlock for the container that the thread is currently owned by. The container may be a turnstile, sleepqueue, or run queue. - thread_lock() is now used to protect access to thread related scheduling fields. thread_unlock() unlocks the lock and thread_set_lock() implements the transition from one lock to another. - A new "blocked_lock" is used in cases where it is not safe to hold the actual thread's lock yet we must prevent access to the thread. - sched_throw() and sched_fork_exit() are introduced to allow the schedulers to fix-up locking at these points. - Add some minor infrastructure for optionally exporting scheduler statistics that were invaluable in solving performance problems with this patch. Generally these statistics allow you to differentiate between different causes of context switches. Tested by: kris, current@ Tested on: i386, amd64, ULE, 4BSD, libthr, libkse, PREEMPTION, etc. Discussed with: kris, attilio, kmacy, jhb, julian, bde (small parts each)
2007-06-04 23:50:30 +00:00
SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0, "");
Major revamp of ULE's cpu load balancing: - Switch back to direct modification of remote CPU run queues. This added a lot of complexity with questionable gain. It's easy enough to reimplement if it's shown to help on huge machines. - Re-implement the old tdq_transfer() call as tdq_pickidle(). Change sched_add() so we have selectable cpu choosers and simplify the logic a bit here. - Implement tdq_pickpri() as the new default cpu chooser. This algorithm is similar to Solaris in that it tries to always run the threads with the best priorities. It is actually slightly more complex than solaris's algorithm because we also tend to favor the local cpu over other cpus which has a boost in latency but also potentially enables cache sharing between the waking thread and the woken thread. - Add a bunch of tunables that can be used to measure effects of different load balancing strategies. Most of these will go away once the algorithm is more definite. - Add a new mechanism to steal threads from busy cpus when we idle. This is enabled with kern.sched.steal_busy and kern.sched.busy_thresh. The threshold is the required length of a tdq's run queue before another cpu will be able to steal runnable threads. This prevents most queue imbalances that contribute the long latencies.
2007-01-19 21:56:08 +00:00
#endif
ULE 2.0: - Remove the double queue mechanism for timeshare threads. It was slow due to excess cache lines in play, caused suboptimal scheduling behavior with niced and other non-interactive processes, complicated priority lending, etc. - Use a circular queue with a floating starting index for timeshare threads. Enforces fairness by moving the insertion point closer to threads with worse priorities over time. - Give interactive timeshare threads real-time user-space priorities and place them on the realtime/ithd queue. - Select non-interactive timeshare thread priorities based on their cpu utilization over the last 10 seconds combined with the nice value. This gives us more sane priorities and behavior in a loaded system as compared to the old method of using the interactivity score. The interactive score quickly hit a ceiling if threads were non-interactive and penalized new hog threads. - Use one slice size for all threads. The slice is not currently dynamically set to adjust scheduling behavior of different threads. - Add some new sysctls for scheduling parameters. Bug fixes/Clean up: - Fix zeroing of td_sched after initialization in sched_fork_thread() caused by recent ksegrp removal. - Fix KSE interactivity issues related to frequent forking and exiting of kse threads. We simply disable the penalty for thread creation and exit for kse threads. - Cleanup the cpu estimator by using tickincr here as well. Keep ticks and ltick/ftick in the same frequency. Previously ticks were stathz and others were hz. - Lots of new and updated comments. - Many many others. Tested on: up x86/amd64, 8way amd64.
2007-01-04 08:56:25 +00:00
/* ps compat */
static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "");
Refactor a bunch of scheduler code to give basically the same behaviour but with slightly cleaned up interfaces. The KSE structure has become the same as the "per thread scheduler private data" structure. In order to not make the diffs too great one is #defined as the other at this time. The KSE (or td_sched) structure is now allocated per thread and has no allocation code of its own. Concurrency for a KSEGRP is now kept track of via a simple pair of counters rather than using KSE structures as tokens. Since the KSE structure is different in each scheduler, kern_switch.c is now included at the end of each scheduler. Nothing outside the scheduler knows the contents of the KSE (aka td_sched) structure. The fields in the ksegrp structure that are to do with the scheduler's queueing mechanisms are now moved to the kg_sched structure. (per ksegrp scheduler private data structure). In other words how the scheduler queues and keeps track of threads is no-one's business except the scheduler's. This should allow people to write experimental schedulers with completely different internal structuring. A scheduler call sched_set_concurrency(kg, N) has been added that notifies teh scheduler that no more than N threads from that ksegrp should be allowed to be on concurrently scheduled. This is also used to enforce 'fainess' at this time so that a ksegrp with 10000 threads can not swamp a the run queue and force out a process with 1 thread, since the current code will not set the concurrency above NCPU, and both schedulers will not allow more than that many onto the system run queue at a time. Each scheduler should eventualy develop their own methods to do this now that they are effectively separated. Rejig libthr's kernel interface to follow the same code paths as linkse for scope system threads. This has slightly hurt libthr's performance but I will work to recover as much of it as I can. Thread exit code has been cleaned up greatly. exit and exec code now transitions a process back to 'standard non-threaded mode' before taking the next step. Reviewed by: scottl, peter MFC after: 1 week
2004-09-05 02:09:54 +00:00
#define KERN_SWITCH_INCLUDE 1
#include "kern/kern_switch.c"
Reference in New Issue Copy Permalink