do not map the b_pages pages into buffer_map KVA. The use of the
unmapped buffers eliminate the need to perform TLB shootdown for
mapping on the buffer creation and reuse, greatly reducing the amount
of IPIs for shootdown on big-SMP machines and eliminating up to 25-30%
of the system time on i/o intensive workloads.
The unmapped buffer should be explicitely requested by the GB_UNMAPPED
flag by the consumer. For unmapped buffer, no KVA reservation is
performed at all. The consumer might request unmapped buffer which
does have a KVA reserve, to manually map it without recursing into
buffer cache and blocking, with the GB_KVAALLOC flag.
When the mapped buffer is requested and unmapped buffer already
exists, the cache performs an upgrade, possibly reusing the KVA
reservation.
Unmapped buffer is translated into unmapped bio in g_vfs_strategy().
Unmapped bio carry a pointer to the vm_page_t array, offset and length
instead of the data pointer. The provider which processes the bio
should explicitely specify a readiness to accept unmapped bio,
otherwise g_down geom thread performs the transient upgrade of the bio
request by mapping the pages into the new bio_transient_map KVA
submap.
The bio_transient_map submap claims up to 10% of the buffer map, and
the total buffer_map + bio_transient_map KVA usage stays the
same. Still, it could be manually tuned by kern.bio_transient_maxcnt
tunable, in the units of the transient mappings. Eventually, the
bio_transient_map could be removed after all geom classes and drivers
can accept unmapped i/o requests.
Unmapped support can be turned off by the vfs.unmapped_buf_allowed
tunable, disabling which makes the buffer (or cluster) creation
requests to ignore GB_UNMAPPED and GB_KVAALLOC flags. Unmapped
buffers are only enabled by default on the architectures where
pmap_copy_page() was implemented and tested.
In the rework, filesystem metadata is not the subject to maxbufspace
limit anymore. Since the metadata buffers are always mapped, the
buffers still have to fit into the buffer map, which provides a
reasonable (but practically unreachable) upper bound on it. The
non-metadata buffer allocations, both mapped and unmapped, is
accounted against maxbufspace, as before. Effectively, this means that
the maxbufspace is forced on mapped and unmapped buffers separately.
The pre-patch bufspace limiting code did not worked, because
buffer_map fragmentation does not allow the limit to be reached.
By Jeff Roberson request, the getnewbuf() function was split into
smaller single-purpose functions.
Sponsored by: The FreeBSD Foundation
Discussed with: jeff (previous version)
Tested by: pho, scottl (previous version), jhb, bf
MFC after: 2 weeks
Replace the per-object resident and cached pages splay tree with a
path-compressed multi-digit radix trie.
Along with this, switch also the x86-specific handling of idle page
tables to using the radix trie.
This change is supposed to do the following:
- Allowing the acquisition of read locking for lookup operations of the
resident/cached pages collections as the per-vm_page_t splay iterators
are now removed.
- Increase the scalability of the operations on the page collections.
The radix trie does rely on the consumers locking to ensure atomicity of
its operations. In order to avoid deadlocks the bisection nodes are
pre-allocated in the UMA zone. This can be done safely because the
algorithm needs at maximum one new node per insert which means the
maximum number of the desired nodes is the number of available physical
frames themselves. However, not all the times a new bisection node is
really needed.
The radix trie implements path-compression because UFS indirect blocks
can lead to several objects with a very sparse trie, increasing the number
of levels to usually scan. It also helps in the nodes pre-fetching by
introducing the single node per-insert property.
This code is not generalized (yet) because of the possible loss of
performance by having much of the sizes in play configurable.
However, efforts to make this code more general and then reusable in
further different consumers might be really done.
The only KPI change is the removal of the function vm_page_splay() which
is now reaped.
The only KBI change, instead, is the removal of the left/right iterators
from struct vm_page, which are now reaped.
Further technical notes broken into mealpieces can be retrieved from the
svn branch:
http://svn.freebsd.org/base/user/attilio/vmcontention/
Sponsored by: EMC / Isilon storage division
In collaboration with: alc, jeff
Tested by: flo, pho, jhb, davide
Tested by: ian (arm)
Tested by: andreast (powerpc)
pages around, taking array of vm_page_t both for source and
destination. Starting offsets and total transfer size are specified.
The function implements optimal algorithm for copying using the
platform-specific optimizations. For instance, on the architectures
were the direct map is available, no transient mappings are created,
for i386 the per-cpu ephemeral page frame is used. The code was
typically borrowed from the pmap_copy_page() for the same
architecture.
Only i386/amd64, powerpc aim and arm/arm-v6 implementations were
tested at the time of commit. High-level code, not committed yet to
the tree, ensures that the use of the function is only allowed after
explicit enablement.
For sparc64, the existing code has known issues and a stab is added
instead, to allow the kernel linking.
Sponsored by: The FreeBSD Foundation
Tested by: pho (i386, amd64), scottl (amd64), ian (arm and arm-v6)
MFC after: 2 weeks
future further optimizations where the vm_object lock will be held
in read mode most of the time the page cache resident pool of pages
are accessed for reading purposes.
The change is mostly mechanical but few notes are reported:
* The KPI changes as follow:
- VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK()
- VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK()
- VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK()
- VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED()
(in order to avoid visibility of implementation details)
- The read-mode operations are added:
VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(),
VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED()
* The vm/vm_pager.h namespace pollution avoidance (forcing requiring
sys/mutex.h in consumers directly to cater its inlining functions
using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h
consumers now must include also sys/rwlock.h.
* zfs requires a quite convoluted fix to include FreeBSD rwlocks into
the compat layer because the name clash between FreeBSD and solaris
versions must be avoided.
At this purpose zfs redefines the vm_object locking functions
directly, isolating the FreeBSD components in specific compat stubs.
The KPI results heavilly broken by this commit. Thirdy part ports must
be updated accordingly (I can think off-hand of VirtualBox, for example).
Sponsored by: EMC / Isilon storage division
Reviewed by: jeff
Reviewed by: pjd (ZFS specific review)
Discussed with: alc
Tested by: pho
Switch eventtimers(9) from using struct bintime to sbintime_t.
Even before this not a single driver really supported full dynamic range of
struct bintime even in theory, not speaking about practical inexpediency.
This change legitimates the status quo and cleans up the code.
VM_OBJECT_LOCKED() macro is only used to implement a custom version
of lock assertions right now (which likely spread out thanks to
copy and paste).
Remove it and implement actual assertions.
Sponsored by: EMC / Isilon storage division
Reviewed by: alc
Tested by: pho
upon pmap_enter() to create a mapping within a different address space,
i.e., not the thread's own address space. On i386, this entails the
creation of a temporary mapping to the affected page table page (PTP). In
general, pmap_enter() will read from this PTP, allocate a PV entry, and
write to this PTP. The trouble comes when the system is short of memory.
In order to allocate a new PV entry, an older PV entry has to be
reclaimed. Reclaiming a PV entry involves destroying a mapping, which
requires access to the affected PTP. Thus, the PTP mapped at the
beginning of pmap_enter() is no longer mapped at the end of pmap_enter(),
which leads to pmap_enter() modifying the wrong PTP. To address this
problem, pmap_pv_reclaim() is changed to use an alternate method of
mapping PTPs.
Update a related comment.
Reported by: pho
Diagnosed by: kib
MFC after: 5 days
pmap_unmapdev()'s own direct efforts to destroy the page table entries are
redundant, so eliminate them.
Don't set PTE_W on the page table entry in pmap_kenter{,_attr}() on MIPS.
Setting PTE_W on MIPS is inconsistent with the implementation of this
function on other architectures. Moreover, PTE_W should not be set, unless
the pmap's wired mapping count is incremented, which pmap_kenter{,_attr}()
doesn't do.
MFC after: 10 days
comment describing them. Both the function names and the comment had grown
stale. Quite some time has passed since these pmap implementations last
used the page's hold count to track the number of valid mapping within a
page table page. Also, returning TRUE from pmap_unwire_ptp() rather than
_pmap_unwire_ptp() eliminates a few instructions from callers like
pmap_enter_quick_locked() where pmap_unwire_ptp()'s return value is used
directly by a conditional statement.
from pmap_pte(). PT_SET_MA() is not a queued mapping update, but instead
an immediate mapping update, so the page queues lock is not required here.
Reviewed by: cperciva
Constify pc_freemask[].
pmap_pv_reclaim()
Eliminate "freemask" because it was a pessimization. Add a comment about
the resident count adjustment.
free_pv_entry() [i386 only]
Merge an optimization from amd64 (r233954).
get_pv_entry()
Eliminate the move to tail of the pv_chunk on the global pv_chunks list.
(The right strategy needs more thought. Moreover, there were unintended
differences between the amd64 and i386 implementation.)
pmap_remove_pages()
Eliminate unnecessary ()'s.
longer uses the active and inactive paging queues. Instead, the pmap now
maintains an LRU-ordered list of pv entry pages, and pmap_pv_reclaim() uses
this list to select pv entries for reclamation.
Note: The old pmap_collect() tried to avoid reclaiming mappings for pages
that have either a hold_count or a busy field that is non-zero. However,
this isn't necessary for correctness, and the locking in pmap_collect() was
insufficient to guarantee that such mappings weren't reclaimed. The new
pmap_pv_reclaim() doesn't even try.
Tested by: sbruno
MFC after: 5 weeks
When r207410 eliminated the acquisition and release of the page queues
lock from pmap_extract_and_hold(), it didn't take into account that
pmap_pte_quick() sometimes requires the page queues lock to be held.
This change reimplements pmap_extract_and_hold() such that it no
longer uses pmap_pte_quick(), and thus never requires the page queues
lock.
Merge r177525 from the native pmap
Prevent the overflow in the calculation of the next page directory.
The overflow causes the wraparound with consequent corruption of the
(almost) whole address space mapping.
Strictly speaking, r177525 is not required by the Xen pmap because the
hypervisor steals the uppermost region of the normal kernel address
space. I am nonetheless merging it in order to reduce the number of
unnecessary differences between the native and Xen pmap implementations.
Tested by: sbruno
paravirtualized pmap implementations for i386. This includes some
style fixes to the native pmap and several bug fixes that were not
previously applied to the paravirtualized pmap.
Tested by: sbruno
MFC after: 3 weeks
initializing structures, like the pv table, that are only used to
implement superpages. In fact, some of the unnecessary code in
pmap_init() was actually doing harm. It was preventing the kernel from
booting on virtual machines with more than 768 MB of memory.
Tested by: sbruno
The SYSCTL_NODE macro defines a list that stores all child-elements of
that node. If there's no SYSCTL_DECL macro anywhere else, there's no
reason why it shouldn't be static.
flags field. Updates to the atomic flags are performed using the atomic
ops on the containing word, do not require any vm lock to be held, and
are non-blocking. The vm_page_aflag_set(9) and vm_page_aflag_clear(9)
functions are provided to modify afalgs.
Document the changes to flags field to only require the page lock.
Introduce vm_page_reference(9) function to provide a stable KPI and
KBI for filesystems like tmpfs and zfs which need to mark a page as
referenced.
Reviewed by: alc, attilio
Tested by: marius, flo (sparc64); andreast (powerpc, powerpc64)
Approved by: re (bz)
to VPO_UNMANAGED (and also making the flag protected by the vm object
lock, instead of vm page queue lock).
- Mark the fake pages with both PG_FICTITIOUS (as it is now) and
VPO_UNMANAGED. As a consequence, pmap code now can use use just
VPO_UNMANAGED to decide whether the page is unmanaged.
Reviewed by: alc
Tested by: pho (x86, previous version), marius (sparc64),
marcel (arm, ia64, powerpc), ray (mips)
Sponsored by: The FreeBSD Foundation
Approved by: re (bz)
option to vm_object_page_remove() asserts that the specified range of pages
is not mapped, or more precisely that none of these pages have any managed
mappings. Thus, vm_object_page_remove() need not call pmap_remove_all() on
the pages.
This change not only saves time by eliminating pointless calls to
pmap_remove_all(), but it also eliminates an inconsistency in the use of
pmap_remove_all() versus related functions, like pmap_remove_write(). It
eliminates harmless but pointless calls to pmap_remove_all() that were being
performed on PG_UNMANAGED pages.
Update all of the existing assertions on pmap_remove_all() to reflect this
change.
Reviewed by: kib
Xen timer and time counter to provide one-shot and periodic time events.
On my tests this reduces idle interruts rate down to about 30Hz, and accor-
ding to Xen VM Manager reduces host CPU load by three times comparing to
the previous periodic 100Hz clock. Also now, when needed, it is possible to
increase HZ rate without useless CPU burning during idle periods.
Now only ia64 and some ARMs left not migrated to the new event timers.
cpuset_t objects.
That is going to offer the underlying support for a simple bump of
MAXCPU and then support for number of cpus > 32 (as it is today).
Right now, cpumask_t is an int, 32 bits on all our supported architecture.
cpumask_t on the other side is implemented as an array of longs, and
easilly extendible by definition.
The architectures touched by this commit are the following:
- amd64
- i386
- pc98
- arm
- ia64
- XEN
while the others are still missing.
Userland is believed to be fully converted with the changes contained
here.
Some technical notes:
- This commit may be considered an ABI nop for all the architectures
different from amd64 and ia64 (and sparc64 in the future)
- per-cpu members, which are now converted to cpuset_t, needs to be
accessed avoiding migration, because the size of cpuset_t should be
considered unknown
- size of cpuset_t objects is different from kernel and userland (this is
primirally done in order to leave some more space in userland to cope
with KBI extensions). If you need to access kernel cpuset_t from the
userland please refer to example in this patch on how to do that
correctly (kgdb may be a good source, for example).
- Support for other architectures is going to be added soon
- Only MAXCPU for amd64 is bumped now
The patch has been tested by sbruno and Nicholas Esborn on opteron
4 x 12 pack CPUs. More testing on big SMP is expected to came soon.
pluknet tested the patch with his 8-ways on both amd64 and i386.
Tested by: pluknet, sbruno, gianni, Nicholas Esborn
Reviewed by: jeff, jhb, sbruno
function on the possibility of a thread to not preempt.
As this function is very tied to x86 (interrupts disabled checkings)
it is not intended to be used in MI code.
