to get the semantics when setting the PMAP right. Prior to r251782, the
latter already used implicit acquire semantics, which - currently - means
to not employ additional explicit memory barriers under the hood (see also
r225889).
Issues were noted by Bruce Evans and are present on all architectures.
On i386, a counter fetch should use atomic read of 64bit value,
otherwise carry from the increment on other CPU could be lost for the
given fetch, making error of 2^32. If 64bit read (cmpxchg8b) is not
available on the machine, it cannot be SMP and it is enough to disable
preemption around read to avoid the split read.
On x86 the counter increment is not atomic on purpose, which makes it
possible for the store of the incremented result to override just
zeroed per-cpu slot. The effect would be a counter going off by
arbitrary value after zeroing. Perform the counter zeroing on the
same processor which does the increments, making the operations
mutually exclusive. On i386, same as for the fetching, if the
cmpxchg8b is not available, machine is not SMP and we disable
preemption for zeroing.
PowerPC64 is treated the same as amd64.
For other architectures, the changes made to allow the compilation to
succeed, without fixing the issues with zeroing or fetching. It
should be possible to handle them by using the 64bit loads and stores
atomic WRT preemption (assuming the architectures also converted from
using critical sections to proper asm). If architecture does not
provide the facility, using global (spin) mutex would be non-optimal
but working solution.
Noted by: bde
Sponsored by: The FreeBSD Foundation
The atomic_load() and atomic_store() macros conflict with the equally
named macros from <stdatomic.h>. Remove them, as they are only used to
implement functions that are not present on any of the other
architectures.
order to match the MAXCPU concept. The change should also be useful
for consolidation and consistency.
Sponsored by: EMC / Isilon storage division
Obtained from: jeff
Reviewed by: alc
Introduce counter(9) API, that implements fast and raceless counters,
provided (but not limited to) for gathering of statistical data.
See http://lists.freebsd.org/pipermail/freebsd-arch/2013-April/014204.html
for more details.
In collaboration with: kib
Reviewed by: luigi
Tested by: ae, ray
Sponsored by: Nginx, Inc.
every architecture's busdma_machdep.c. It is done by unifying the
bus_dmamap_load_buffer() routines so that they may be called from MI
code. The MD busdma is then given a chance to do any final processing
in the complete() callback.
The cam changes unify the bus_dmamap_load* handling in cam drivers.
The arm and mips implementations are updated to track virtual
addresses for sync(). Previously this was done in a type specific
way. Now it is done in a generic way by recording the list of
virtuals in the map.
Submitted by: jeff (sponsored by EMC/Isilon)
Reviewed by: kan (previous version), scottl,
mjacob (isp(4), no objections for target mode changes)
Discussed with: ian (arm changes)
Tested by: marius (sparc64), mips (jmallet), isci(4) on x86 (jharris),
amd64 (Fabian Keil <freebsd-listen@fabiankeil.de>)
reading registers from other CPUs. As it turns out, the hardware doesn't
really like concurrent IPI'ing causing adverse effects. Also the thought
deadlock when using this spin lock here and the targeted CPU(s) are also
holding or in case of nested locks can't actually happen. This is due to
the fact that on sparc64, spinlock_enter() only raises the PIL but doesn't
disable interrupts completely. Thus direct cross calls as used for the
register reading (and all other MD IPI needs) still will be executed by
the targeted CPU(s) in that case.
MFC after: 3 days
- Implement a function to ensure that all preempted threads have switched
back out at least once. Use this to make sure there are no stale
references to the old ktr_buf or the lock profiling buffers before
updating them.
Reviewed by: marius (sparc64 parts), attilio (earlier patch)
Sponsored by: EMC / Isilon Storage Division
The reason for this is that the SPARC v9 architecture allows nested
interrupts of higher priority/level than that of the current interrupt
to occur (and we can't just entirely bypass this model, also, at least
for tick interrupts, this also wouldn't be wise). However, when a
preemption interrupt interrupts another interrupt of lower priority,
f.e. PIL_ITHREAD, and that one in turn is nested by a third interrupt,
f.e. PIL_TICK, with SCHED_ULE the execution of interrupts higher than
PIL_PREEMPT may be migrated to another CPU. In particular, tl1_ret(),
which is responsible for restoring the state of the CPU prior to entry
to the interrupt based on the (also migrated) trap frame, then is run
on a CPU which actually didn't receive the interrupt in question,
causing an inappropriate processor interrupt level to be "restored".
In turn, this causes interrupts of the first level, i.e. PIL_ITHREAD
in the above scenario, to be blocked on the target of the migration
until the correct PIL happens to be restored again on that CPU again.
Making PIL_PREEMPT the lowest real priority, this effectively prevents
this scenario from happening, as preemption interrupts no longer can
interrupt any other interrupt besides stray ones (which is no issue).
Thanks to attilio@ and especially mav@ for helping me to understand
this problem at the 201208DevSummit.
- Give PIL_STOP (which is also used for IPI_STOP_HARD, given that there's
no real equivalent to NMIs on SPARC v9) the highest possible priority
just below the hardwired PIL_TICK, so it has a chance to interrupt
more things.
MFC after: 1 week
instruction loads/stores at its will.
The macro __compiler_membar() is currently supported for both gcc and
clang, but kernel compilation will fail otherwise.
Reviewed by: bde, kib
Discussed with: dim, theraven
MFC after: 2 weeks
r234247.
Use, instead, the static intializer introduced in r239923 for x86 and
sparc64 intr_cpus, unwinding the code to the initial version.
Reviewed by: marius
other CPUs doesn't require locking so get rid of it. As the latter is used
for the timecounter on certain machine models, using a spin lock in this
case can lead to a deadlock with the upcoming callout(9) rework.
- Merge r134227/r167250 from x86:
Avoid cross-IPI SMP deadlock by using the smp_ipi_mtx spin lock not only
for smp_rendezvous_cpus() but also for the MD cache invalidation and TLB
demapping IPIs.
- Mark some unused function arguments as such.
MFC after: 1 week
This is required for ARM EABI. Section 7.1.1 of the Procedure Call for the
ARM Architecture (AAPCS) defines wchar_t as either an unsigned int or an
unsigned short with the former preferred.
Because of this requirement we need to move the definition of __wchar_t to
a machine dependent header. It also cleans up the macros defining the limits
of wchar_t by defining __WCHAR_MIN and __WCHAR_MAX in the same machine
dependent header then using them to define WCHAR_MIN and WCHAR_MAX
respectively.
Discussed with: bde
usermode, using shared page. The structures and functions have vdso
prefix, to indicate the intended location of the code in some future.
The versioned per-algorithm data is exported in the format of struct
vdso_timehands, which mostly repeats the content of in-kernel struct
timehands. Usermode reading of the structure can be lockless.
Compatibility export for 32bit processes on 64bit host is also
provided. Kernel also provides usermode with indication about
currently used timecounter, so that libc can fall back to syscall if
configured timecounter is unknown to usermode code.
The shared data updates are initiated both from the tc_windup(), where
a fast task is queued to do the update, and from sysctl handlers which
change timecounter. A manual override switch
kern.timecounter.fast_gettime allows to turn off the mechanism.
Only x86 architectures export the real algorithm data, and there, only
for tsc timecounter. HPET counters page could be exported as well, but
I prefer to not further glue the kernel and libc ABI there until
proper vdso-based solution is developed.
Minimal stubs neccessary for non-x86 architectures to still compile
are provided.
Discussed with: bde
Reviewed by: jhb
Tested by: flo
MFC after: 1 month
layer, but it is read directly by the MI VM layer. This change introduces
pmap_page_is_write_mapped() in order to completely encapsulate all direct
access to PGA_WRITEABLE in the pmap layer.
Aesthetics aside, I am making this change because amd64 will likely begin
using an alternative method to track write mappings, and having
pmap_page_is_write_mapped() in place allows me to make such a change
without further modification to the MI VM layer.
As an added bonus, tidy up some nearby comments concerning page flags.
Reviewed by: kib
MFC after: 6 weeks
to this pmap.c. This new r/w lock is used primarily to synchronize access
to the TTE lists. However, it will be used in a somewhat unconventional
way. As finer-grained TTE list locking is added to each of the pmap
functions that acquire this r/w lock, its acquisition will be changed from
write to read, enabling concurrent execution of the pmap functions with
finer-grained locking.
Reviewed by: attilio
Tested by: flo
MFC after: 10 days
in_cksum.h required ip.h to be included for struct ip. To be
able to use some general checksum functions like in_addword()
in a non-IPv4 context, limit the (also exported to user space)
IPv4 specific functions to the times, when the ip.h header is
present and IPVERSION is defined (to 4).
We should consider more general checksum (updating) functions
to also allow easier incremental checksum updates in the L3/4
stack and firewalls, as well as ponder further requirements by
certain NIC drivers needing slightly different pseudo values
in offloading cases. Thinking in terms of a better "library".
Sponsored by: The FreeBSD Foundation
Sponsored by: iXsystems
Reviewed by: gnn (as part of the whole)
MFC After: 3 days
discrepancy between modules and kernel, but deal with SMP differences
within the functions themselves.
As an added bonus this also helps in terms of code readability.
Requested by: gibbs
Reviewed by: jhb, marius
MFC after: 1 week
didn't already have them. This is because the ternary expression will
return int, due to the Usual Arithmetic Conversions. Such casts are not
needed for the 32 and 64 bit variants.
While here, add additional parentheses around the x86 variant, to
protect against unintended consequences.
MFC after: 2 weeks
long for specifying a boundary constraint.
- Change bus_dma tags to use bus_addr_t instead of bus_size_t for boundary
constraints.
These allow boundary constraints to be fully expressed for cases where
sizeof(bus_addr_t) != sizeof(bus_size_t). Specifically, it allows a
driver to properly specify a 4GB boundary in a PAE kernel.
Note that this cannot be safely MFC'd without a lot of compat shims due
to KBI changes, so I do not intend to merge it.
Reviewed by: scottl
helper since r230632, use these for output and panicing during the
early cycles and move cninit() until after the static per-CPU data
has been set up. This solves a couple of issue regarding the non-
availability of the static per-CPU data:
- panic() not working and only making things worse when called,
- having to supply a special DELAY() implementation to the low-level
console drivers,
- curthread accesses of mutex(9) usage in low-level console drivers
that aren't conditional due to compiler optimizations (basically,
this is the problem described in r227537 but in this case for
keyboards attached via uart(4)). [1]
PR: 164123 [1]
implementing a simple OF_panic() that may be used during the early
cycles when panic() isn't available, yet.
- Mark cpu_{exit,shutdown}() as __dead2 as appropriate.
VM_KMEM_SIZE_SCALE to 2, awaiting more insight from alc@. As it turns
out, the VM apparently has problems with machines that have large holes
in the physical address space, causing the kmem_suballoc() call in
kmeminit() to fail with a VM_KMEM_SIZE_SCALE of 1. Using a value of 2
allows these, namely Blade 1500 with 2GB of RAM, to boot.
PR: 164227
no need to additionally add CPU memory barriers to the acquire variants of
atomic(9), these are documented to also include compiler memory barriers.
So add the latter, which were previously included by using membar(), back.
the 16-bit cylinders field of the VTOC8 disk label (at around 502GB). The
geometry chosen for disks above that limit allows to use disks up to 2TB,
which is the limit of the extended VTOC8 format. The geometry used for
disks smaller than the 16-bit cylinders limit stays the same as used by
cam_calc_geometry(9) for extended translation.
Thanks to Hans-Joerg Sirtl for providing hardware for testing this change.
MFC after: 3 days
directly from g7, the pcpu pointer. This guarantees correct behavior
when the thread migrates to a different CPU.
Commit message stolen from r205431. Additional testing by Peter Jeremy.
MFC after: 3 days
implement a deprecated FPU control interface in addition to the
standard one. To make this clearer, further deprecate ieeefp.h
by not declaring the function prototypes except on architectures
that implement them already.
Currently i386 and amd64 implement the ieeefp.h interface for
compatibility, and for fp[gs]etprec(), which doesn't exist on
most other hardware. Powerpc, sparc64, and ia64 partially implement
it and probably shouldn't, and other architectures don't implement it
at all.
and pc_pmap for SMP. This is key to allowing adding support for SCHED_ULE.
Thanks go to Peter Jeremy for additional testing.
- Add support for SCHED_ULE to cpu_switch().
Committed from: 201110DevSummit
- Implement bus_adjust_resource() methods as far as necessary and in non-PCI
bridge drivers as far as feasible without rototilling them.
- As NEW_PCIB does a layering violation by activating resources at layers
above pci(4) without previously bubbling up their allocation there, move
the assignment of bus tags and handles from the bus_alloc_resource() to
the bus_activate_resource() methods like at least the other NEW_PCIB
enabled architectures do. This is somewhat unfortunate as previously
sparc64 (ab)used resource activation to indicate whether SYS_RES_MEMORY
resources should be mapped into KVA, which is only necessary if their
going to be accessed via the pointer returned from rman_get_virtual() but
not for bus_space(9) as the later always uses physical access on sparc64.
Besides wasting KVA if we always map in SYS_RES_MEMORY resources, a driver
also may deliberately not map them in if the firmware already has done so,
possibly in a special way. So in order to still allow a driver to decide
whether a SYS_RES_MEMORY resource should be mapped into KVA we let it
indicate that by calling bus_space_map(9) with BUS_SPACE_MAP_LINEAR as
actually documented in the bus_space(9) page. This is implemented by
allocating a separate bus tag per SYS_RES_MEMORY resource and passing the
resource via the previously unused bus tag cookie so we later on can call
rman_set_virtual() in sparc64_bus_mem_map(). As a side effect this now
also allows to actually indicate that a SYS_RES_MEMORY resource should be
mapped in as cacheable and/or read-only via BUS_SPACE_MAP_CACHEABLE and
BUS_SPACE_MAP_READONLY respectively.
- Do some minor cleanup like taking advantage of rman_init_from_resource(),
factor out the common part of bus tag allocation into a newly added
sparc64_alloc_bus_tag(), hook up some missing newbus methods and replace
some homegrown versions with the generic counterparts etc.
- While at it, let apb_attach() (which can't use the generic NEW_PCIB code
as APB bridges just don't have the base and limit registers implemented)
regarding the config space registers cached in pcib_softc and the SYSCTL
reporting nodes set up.
atomic operations behave as if the were followed by a memory barrier so
there's no need to include ones in the acquire variants of atomic(9).
Removing these results a small performance improvement, specifically this
is sufficient to compensate the performance loss seen in the worldstone
benchmark seen when using SCHED_ULE instead of SCHED_4BSD.
This change is inspired by Linux even more radically doing the equivalent
thing some time ago.
Thanks go to Peter Jeremy for additional testing.
