possible future I-cache coherency operation can succeed. On ARM
for example the L1 cache can be (is) virtually mapped, which
means that any I/O that uses temporary mappings will not see the
I-cache made coherent. On ia64 a similar behaviour has been
observed. By flushing the D-cache, execution of binaries backed
by md(4) and/or NFS work reliably.
For Book-E (powerpc), execution over NFS exhibits SIGILL once in
a while as well, though cpu_flush_dcache() hasn't been implemented
yet.
Doing an explicit D-cache flush as part of the non-DMA based I/O
read operation eliminates the need to do it as part of the
I-cache coherency operation itself and as such avoids pessimizing
the DMA-based I/O read operations for which D-cache are already
flushed/invalidated. It also allows future optimizations whereby
the bcopy() followed by the D-cache flush can be integrated in a
single operation, which could be implemented using on-chips DMA
engines, by-passing the D-cache altogether.
a fair number of static data structures, making this an unlikely
option to try to change without also changing source code. [1]
Change default cache line size on ia64, sparc64, and sun4v to 128
bytes, as this was what rtld-elf was already using on those
platforms. [2]
Suggested by: bde [1], jhb [2]
MFC after: 2 weeks
CACHE_LINE_SIZE constant. These constants are intended to
over-estimate the cache line size, and be used at compile-time
when a run-time tuning alternative isn't appropriate or
available.
Defaults for all architectures are 64 bytes, except powerpc
where it is 128 bytes (used on G5 systems).
MFC after: 2 weeks
Discussed on: arch@
a tag that has BUS_DMA_KEEP_PG_OFFSET set. Otherwise the page could be
reused with a non-zero offset by a tag that doesn't have
BUS_DMA_KEEP_PG_OFFSET leading to data corruption.
Sleuthing by: avg
Reviewed by: scottl
1) Move the new field (brand_note) to the end of the Brandinfo structure.
2) Add a new flag BI_BRAND_NOTE that indicates that the brand_note pointer
is valid.
3) Use the brand_note field if the flag BI_BRAND_NOTE is set and as old
modules won't have the flag set, so the new field brand_note would be
ignored.
Suggested by: jhb
Reviewed by: jhb
Approved by: kib (mentor)
MFC after: 6 days
to the full path of the image that is being executed.
Increase AT_COUNT.
Remove no longer true comment about types used in Linux ELF binaries,
listed types contain FreeBSD-specific entries.
Reviewed by: kan
".note.ABI-tag" section.
The search order of a brand is changed, now first of all the
".note.ABI-tag" is looked through.
Move code which fetch osreldate for ELF binary to check_note() handler.
PR: 118473
Approved by: kib (mentor)
offset. This is needed for the ehci hardware buffer rings that assume
this behavior.
This is an interim solution, and a more general one is being worked
on. This solution doesn't break anything that doesn't ask for it
directly. The mbuf and uio variants with this flag likely don't work
and haven't been tested.
Universe builds with these changes. I don't have a huge-memory
machine to test these changes with, but will be happy to work with
folks that do and hps if this changes turns out not to be sufficient.
Submitted by: alfred@ from Hans Peter Selasky's original
while holding a spin mutex. Instead, it now shoves the machine check
records onto a queue that is later drained to add sysctl nodes for each
record. While a routine to drain the queue is present, it is not currently
called.
Reviewed by: marcel
Sgtty is a programming interface that has been replaced by termios over
the years. In June we already removed <sgtty.h>, which exposes the
ioctl()'s that are implemented by this interface. The importance of this
flag is overrated right now.
of the ABI of the currently executing image. Change some places to test
the flags instead of explicit comparing with address of known sysentvec
structures to determine ABI features.
Discussed with: dchagin, imp, jhb, peter
and ifnet functions
- add memory barriers to <machine/atomic.h>
- update drivers to only conditionally define their own
- add lockless producer / consumer ring buffer
- remove ring buffer implementation from cxgb and update its callers
- add if_transmit(struct ifnet *ifp, struct mbuf *m) to ifnet to
allow drivers to efficiently manage multiple hardware queues
(i.e. not serialize all packets through one ifq)
- expose if_qflush to allow drivers to flush any driver managed queues
This work was supported by Bitgravity Inc. and Chelsio Inc.
all to date and the latter also is only used in ia64 and powerpc
code which no longer serves a real purpose after bring-up and just
can be removed as well. Note that architectures like sun4u also
provide no means of implementing IPI'ing a CPU itself natively
in the first place.
Suggested by: jhb
Reviewed by: arch, grehan, jhb
After I removed all the unit2minor()/minor2unit() calls from the kernel
yesterday, I realised calling minor() everywhere is quite confusing.
Character devices now only have the ability to store a unit number, not
a minor number. Remove the confusion by using dev2unit() everywhere.
This commit could also be considered as a bug fix. A lot of drivers call
minor(), while they should actually be calling dev2unit(). In -CURRENT
this isn't a problem, but it turns out we never had any problem reports
related to that issue in the past. I suspect not many people connect
more than 256 pieces of the same hardware.
Reviewed by: kib
to the C99 style. At least, it is easier to read sysent definitions
that way, and search for the actual instances of sigcode etc.
Explicitely initialize sysentvec.sv_maxssiz that was missed in most
sysvecs.
No objection from: jhb
MFC after: 1 month
and bcmp are not the same thing. 'man bcmp' states that the return is
"non-zero" if the two byte strings are not identical. Where as,
'man memcmp' states that the return is the "difference between the
first two differing bytes (treated as unsigned char values" if the
two byte strings are not identical.
So provide a proper memcmp(9), but it is a C implementation not a tuned
assembly implementation. Therefore bcmp(9) should be preferred over memcmp(9).
The last half year I've been working on a replacement TTY layer for the
FreeBSD kernel. The new TTY layer was designed to improve the following:
- Improved driver model:
The old TTY layer has a driver model that is not abstract enough to
make it friendly to use. A good example is the output path, where the
device drivers directly access the output buffers. This means that an
in-kernel PPP implementation must always convert network buffers into
TTY buffers.
If a PPP implementation would be built on top of the new TTY layer
(still needs a hooks layer, though), it would allow the PPP
implementation to directly hand the data to the TTY driver.
- Improved hotplugging:
With the old TTY layer, it isn't entirely safe to destroy TTY's from
the system. This implementation has a two-step destructing design,
where the driver first abandons the TTY. After all threads have left
the TTY, the TTY layer calls a routine in the driver, which can be
used to free resources (unit numbers, etc).
The pts(4) driver also implements this feature, which means
posix_openpt() will now return PTY's that are created on the fly.
- Improved performance:
One of the major improvements is the per-TTY mutex, which is expected
to improve scalability when compared to the old Giant locking.
Another change is the unbuffered copying to userspace, which is both
used on TTY device nodes and PTY masters.
Upgrading should be quite straightforward. Unlike previous versions,
existing kernel configuration files do not need to be changed, except
when they reference device drivers that are listed in UPDATING.
Obtained from: //depot/projects/mpsafetty/...
Approved by: philip (ex-mentor)
Discussed: on the lists, at BSDCan, at the DevSummit
Sponsored by: Snow B.V., the Netherlands
dcons(4) fixed by: kan
virtualization work done by Marko Zec (zec@).
This is the first in a series of commits over the course
of the next few weeks.
Mark all uses of global variables to be virtualized
with a V_ prefix.
Use macros to map them back to their global names for
now, so this is a NOP change only.
We hope to have caught at least 85-90% of what is needed
so we do not invalidate a lot of outstanding patches again.
Obtained from: //depot/projects/vimage-commit2/...
Reviewed by: brooks, des, ed, mav, julian,
jamie, kris, rwatson, zec, ...
(various people I forgot, different versions)
md5 (with a bit of help)
Sponsored by: NLnet Foundation, The FreeBSD Foundation
X-MFC after: never
V_Commit_Message_Reviewed_By: more people than the patch
Use the new inline function in ia64_invalidate_icache().
While there, add proper synchronization so that we know
the fc.i instructions have taken effect when we return.
Now that st_rdev is being automatically generated by the kernel, there
is no need to define static major/minor numbers for the iodev and
memdev. We still need the minor numbers for the memdev, however, to
distinguish between /dev/mem and /dev/kmem.
Approved by: philip (mentor)
inlining resulted in constant propagation to the extend that cmpval
was known to the compiler to be URWLOCK_WRITE_OWNER (= 0x80000000U).
Unfortunately, instead of zero-extending the unsigned constant, it
was sign-extended. As such, the cmpxchg instruction was comparing
0x0000000080000000LU to 0xffffffff80000000LU and obviously didn't
perform the exchange.
But, since the value returned by cmpxhg equalled cmpval (when zero-
extended), the _thr_rtld_lock_release() function thought the exchange
did happen and as such returned as if having released the lock. This
was not the case. Subsequent locking requests found rw_state non-zero
and the thread in question entered the kernel and block indefinitely.
The work-around is to zero-extend by casting to uint64_t.
hard limit of 512 pending mutexes in the witness code and
we can easily have 1 million bucket mutexes initialized before
witness is up and running. Bumping the limit from 512 to 1M
is not really an option here...
from idle over the next tick.
- Add a new MD routine, cpu_wake_idle() to wakeup idle threads who are
suspended in cpu specific states. This function can fail and cause the
scheduler to fall back to another mechanism (ipi).
- Implement support for mwait in cpu_idle() on i386/amd64 machines that
support it. mwait is a higher performance way to synchronize cpus
as compared to hlt & ipis.
- Allow selecting the idle routine by name via sysctl machdep.idle. This
replaces machdep.cpu_idle_hlt. Only idle routines supported by the
current machine are permitted.
Sponsored by: Nokia
for better structure.
Much of this is related to <sys/clock.h>, which should really have
been called <sys/calendar.h>, but unless and until we need the name,
the repocopy can wait.
In general the kernel does not know about minutes, hours, days,
timezones, daylight savings time, leap-years and such. All that
is theoretically a matter for userland only.
Parts of kernel code does however care: badly designed filesystems
store timestamps in local time and RTC chips almost universally
track time in a YY-MM-DD HH:MM:SS format, and sometimes in local
timezone instead of UTC. For this we have <sys/clock.h>
<sys/time.h> on the other hand, deals with time_t, timeval, timespec
and so on. These know only seconds and fractions thereof.
Move inittodr() and resettodr() prototypes to <sys/time.h>.
Retain the names as it is one of the few surviving PDP/VAX references.
Move startrtclock() to <machine/clock.h> on relevant platforms, it
is a MD call between machdep.c/clock.c. Remove references to it
elsewhere.
Remove a lot of unnecessary <sys/clock.h> includes.
Move the machdep.disable_rtc_set sysctl to subr_rtc.c where it belongs.
XXX: should be kern.disable_rtc_set really, it's not MD.
inittodr() and resettodr(). Have nexus double as the clock device,
because it's the firmware that provides RTC services. We could
create a special (pseudo-) device for it, but that wasn't superior
enough to actually do it. Maybe later...
Requested by: phk
o Implement IPI_PREEMPT,
o Set td_lock for the thread being switched out,
o For ULE & SMP, loop while td_lock points to blocked_lock for
the thread being switched in,
o Enable ULE by default in GENERIC and SKI,
(i.e. fixed delivery) to SAPIC_DELMODE_LOWPRI. While the commit
log doesn't mention the change in behaviour, it is believed to be
deliberate. In the last 5.5 years this hasn't been a problem. Nor
do I think did it make any difference, but who knows. However, I
do know that it break SMP support for Montecito-based machines.
Switch back to fixed-CPU delivery so that SMP works again. This
gives me some time to look more closely at the problem, as well
as make sure the I-cache validation as it's implemented currently
is sufficient in SMP configurations...
lookup hard interrupt events by number. Ignore the irq# for soft intrs.
- Add support to cpuset for binding hardware interrupts. This has the
side effect of binding any ithread associated with the hard interrupt.
As per restrictions imposed by MD code we can only bind interrupts to
a single cpu presently. Interrupts can be 'unbound' by binding them
to all cpus.
Reviewed by: jhb
Sponsored by: Nokia
allows all the INTR_FILTER #ifdef's to be removed from the MD interrupt
code.
- Rename the intr_event 'eoi', 'disable', and 'enable' hooks to
'post_filter', 'pre_ithread', and 'post_ithread' to be less x86-centric.
Also, add a comment describe what the MI code expects them to do.
- On amd64, i386, and powerpc this is effectively a NOP.
- On arm, don't bother masking the interrupt unless the ithread is
scheduled in the non-INTR_FILTER case to match what INTR_FILTER did.
Also, don't bother unmasking the interrupt in the post_filter case if
we never masked it. The INTR_FILTER case had been doing this by having
arm_unmask_irq for the post_filter (formerly 'eoi') hook.
- On ia64, stray interrupts are now masked for the non-INTR_FILTER case.
They were already masked in the INTR_FILTER case.
- On sparc64, use the a NULL pre_ithread hook and use intr_enable_eoi() for
both the 'post_filter' and 'post_ithread' hooks to match what the
non-INTR_FILTER code did.
- On sun4v, retire the ithread wrapper hack by using an appropriate
'post_ithread' hook instead (it's what 'post_ithread'/'enable' was
designed to do even in 5.x).
Glanced at by: piso
Reviewed by: marius
Requested by: marius [1], [5]
Tested on: amd64, i386, arm, sparc64
was a kluge. This implementation matches the behaviour on powerpc
and sparc64.
While on the subject, make sure to invalidate the I-cache after
loading a kernel module.
MFC after: 2 weeks
to detect (or load) kernel NLM support in rpc.lockd. Remove the '-k'
option to rpc.lockd and make kernel NLM the default. A user can still
force the use of the old user NLM by building a kernel without NFSLOCKD
and/or removing the nfslockd.ko module.
these days, so de-generalize the acquire_timer/release_timer api
to just deal with speakers.
The new (optional) MD functions are:
timer_spkr_acquire()
timer_spkr_release()
and
timer_spkr_setfreq()
the last of which configures the timer to generate a tone of a given
frequency, in Hz instead of 1/1193182th of seconds.
Drop entirely timer2 on pc98, it is not used anywhere at all.
Move sysbeep() to kern/tty_cons.c and use the timer_spkr*() if
they exist, and do nothing otherwise.
Remove prototypes and empty acquire-/release-timer() and sysbeep()
functions from the non-beeping archs.
This eliminate the need for the speaker driver to know about
i8254frequency at all. In theory this makes the speaker driver MI,
contingent on the timer_spkr_*() functions existing but the driver
does not know this yet and still attaches to the ISA bus.
Syscons is more tricky, in one function, sc_tone(), it knows the hz
and things are just fine.
In the other function, sc_bell() it seems to get the period from
the KDMKTONE ioctl in terms if 1/1193182th second, so we hardcode
the 1193182 and leave it at that. It's probably not important.
Change a few other sysbeep() uses which obviously knew that the
argument was in terms of i8254 frequency, and leave alone those
that look like people thought sysbeep() took frequency in hertz.
This eliminates the knowledge of i8254_freq from all but the actual
clock.c code and the prof_machdep.c on amd64 and i386, where I think
it would be smart to ask for help from the timecounters anyway [TBD].
- Always include the ie_disable and ie_eoi methods in 'struct intr_event'
and collapse down to one intr_event_create() routine. The disable and
eoi hooks simply aren't used currently in the !INTR_FILTER case.
- Expand 'disab' to 'disable' in a few places.
- Use function casts for arm and i386:intr_eoi_src() instead of wrapper
routines since to trim one extra indirection.
Compiled on: {arm,amd64,i386,ia64,ppc,sparc64} x {FILTER, !FILTER}
Tested on: {amd64,i386} x {FILTER, !FILTER}
after each SYSINIT() macro invocation. This makes a number of
lightweight C parsers much happier with the FreeBSD kernel
source, including cflow's prcc and lxr.
MFC after: 1 month
Discussed with: imp, rink
all. The reference in ia64 code is due to cutNpaste in its history
and can safely be removed.
Revired by: cognet, raj, marcel, jhb and maybe one other whom I'm forgetting
- Add a new intr_event method ie_assign_cpu() that is invoked when the MI
code wishes to bind an interrupt source to an individual CPU. The MD
code may reject the binding with an error. If an assign_cpu function
is not provided, then the kernel assumes the platform does not support
binding interrupts to CPUs and fails all requests to do so.
- Bind ithreads to CPUs on their next execution loop once an interrupt
event is bound to a CPU. Only shared ithreads are bound. We currently
leave private ithreads for drivers using filters + ithreads in the
INTR_FILTER case unbound.
- A new intr_event_bind() routine is used to bind an interrupt event to
a CPU.
- Implement binding on amd64 and i386 by way of the existing pic_assign_cpu
PIC method.
- For x86, provide a 'intr_bind(IRQ, cpu)' wrapper routine that looks up
an interrupt source and binds its interrupt event to the specified CPU.
MI code can currently (ab)use this by doing:
intr_bind(rman_get_start(irq_res), cpu);
however, I plan to add a truly MI interface (probably a bus_bind_intr(9))
where the implementation in the x86 nexus(4) driver would end up calling
intr_bind() internally.
Requested by: kmacy, gallatin, jeff
Tested on: {amd64, i386} x {regular, INTR_FILTER}
different "platforms" on x86 machines. The existing code already handles
having two platforms: ACPI and legacy. However, the existing approach was
rather hardcoded and difficult to extend. These changes take the approach
that each x86 hardware platform should provide its own nexus(4) driver (it
can inherit most of its behavior from the default legacy nexus(4) driver)
which is responsible for probing for the platform and performing
appropriate platform-specific setup during attach (such as adding a
platform-specific bus device). This does mean changing the x86 platform
busses to no longer use an identify routine for probing, but to move that
logic into their matching nexus(4) driver instead.
- Make the default nexus(4) driver in nexus.c on i386 and amd64 handle the
legacy platform. It's probe routine now returns BUS_PROBE_GENERIC so it
can be overriden.
- Expose a nexus_init_resources() routine which initializes the various
resource managers so that subclassed nexus(4) drivers can invoke it from
their attach routine.
- The legacy nexus(4) driver explicitly adds a legacy0 device in its
attach routine.
- The ACPI driver no longer contains an new-bus identify method. Instead
it exposes a public function (acpi_identify()) which is a probe routine
that the MD nexus(4) drivers can use to probe for ACPI. All of the
probe logic in acpi_probe() is now moved into acpi_identify() and
acpi_probe() is just a stub.
- On i386 and amd64, an ACPI-specific nexus(4) driver checks for ACPI via
acpi_identify() and claims the nexus0 device if the probe succeeds. It
then explicitly adds an acpi0 device in its attach routine.
- The legacy(4) driver no longer knows anything about the acpi0 device.
- On ia64 if acpi_identify() fails you basically end up with no devices.
This matches the previous behavior where the old acpi_identify() would
fail to add an acpi0 device again leaving you with no devices.
Discussed with: imp
Silence on: arch@
While the KSE project was quite successful in bringing threading to
FreeBSD, the M:N approach taken by the kse library was never developed
to its full potential. Backwards compatibility will be provided via
libmap.conf for dynamically linked binaries and static binaries will
be broken.
tree structure that encodes the level of cache sharing and other
properties.
- Provide several convenience functions for creating one and two level
cpu trees as well as a default flat topology. The system now always
has some topology.
- On i386 and amd64 create a seperate level in the hierarchy for HTT
and multi-core cpus. This will allow the scheduler to intelligently
load balance non-uniform cores. Presently we don't detect what level
of the cache hierarchy is shared at each level in the topology.
- Add a mechanism for testing common topologies that have more information
than the MD code is able to provide via the kern.smp.topology tunable.
This should be considered a debugging tool only and not a stable api.
Sponsored by: Nokia
coherent with the data caches. Implement a quick fix to allow
us to boot on Montecito, while I'm working on a better fix in
the mean time.
Commit made on Montecito-based Itanium...
for that argument. This will allow DDB to detect the broad category of
reason why the debugger has been entered, which it can use for the
purposes of deciding which DDB script to run.
Assign approximate why values to all current consumers of the
kdb_enter() interface.
o Disklabels can have between 8 and 20 partitions (inclusive).
o No device special file is created for the raw partition.
o Switch ia64 to use this backend.
o No support for boot code yet.
- Introduce per-architecture stack_machdep.c to hold stack_save(9).
- Introduce per-architecture machine/stack.h to capture any common
definitions required between db_trace.c and stack_machdep.c.
- Add new kernel option "options STACK"; we will build in stack(9) if it is
defined, or also if "options DDB" is defined to provide compatibility
with existing users of stack(9).
Add new stack_save_td(9) function, which allows the capture of a stacktrace
of another thread rather than the current thread, which the existing
stack_save(9) was limited to. It requires that the thread be neither
swapped out nor running, which is the responsibility of the consumer to
enforce.
Update stack(9) man page.
Build tested: amd64, arm, i386, ia64, powerpc, sparc64, sun4v
Runtime tested: amd64 (rwatson), arm (cognet), i386 (rwatson)
routine. It is not needed as the existing tests for segment coalescing
already handle bounced addresses and it prevents legal segment coalescing
in certain edge cases.
MFC after: 1 week
Reviewed by: scottl
include the ithread scheduling step. Without this, a preemption might
occur in between the interrupt getting masked and the ithread getting
scheduled. Since the interrupt handler runs in the context of curthread,
the scheudler might see it as having a such a low priority on a busy system
that it doesn't get to run for a _long_ time, leaving the interrupt stranded
in a disabled state. The only way that the preemption can happen is by
a fast/filter handler triggering a schduling event earlier in the handler,
so this problem can only happen for cases where an interrupt is being
shared by both a fast/filter handler and an ithread handler. Unfortunately,
it seems to be common for this sharing to happen with network and USB
devices, for example. This fixes many of the mysterious TCP session
timeouts and NIC watchdogs that were being reported. Many thanks to Sam
Lefler for getting to the bottom of this problem.
Reviewed by: jhb, jeff, silby
First, a file is mmap(2)ed and then mlock(2)ed. Later, it is truncated.
Under "normal" circumstances, i.e., when the file is not mlock(2)ed, the
pages beyond the EOF are unmapped and freed. However, when the file is
mlock(2)ed, the pages beyond the EOF are unmapped but not freed because
they have a non-zero wire count. This can be a mistake. Specifically,
it is a mistake if the sole reason why the pages are wired is because of
wired, managed mappings. Previously, unmapping the pages destroys these
wired, managed mappings, but does not reduce the pages' wire count.
Consequently, when the file is unmapped, the pages are not unwired
because the wired mapping has been destroyed. Moreover, when the vm
object is finally destroyed, the pages are leaked because they are still
wired. The fix is to reduce the pages' wired count by the number of
wired, managed mappings destroyed. To do this, I introduce a new pmap
function pmap_page_wired_mappings() that returns the number of managed
mappings to the given physical page that are wired, and I use this
function in vm_object_page_remove().
Reviewed by: tegge
MFC after: 6 weeks
communicate that it relates to (is called by) thread_alloc()
o Add cpu_thread_free() which is called from thread_free()
to counter-act cpu_thread_alloc().
i386: Have cpu_thread_free() call cpu_thread_clean() to
preserve behaviour.
ia64: Have cpu_thread_free() call mtx_destroy() for the
mutex initialized in cpu_thread_alloc().
PR: ia64/118024
opposed to what process. Since threads by default have teh name of the
process unless over-written with more useful information, just print the
thread name instead.
silent NULL pointer dereference in the i386 and sparc64 pmap_pinit()
when the kmem_alloc_nofault() failed to allocate address space. Both
functions now return error instead of panicing or dereferencing NULL.
As consequence, vmspace_exec() and vmspace_unshare() returns the errno
int. struct vmspace arg was added to vm_forkproc() to avoid dealing
with failed allocation when most of the fork1() job is already done.
The kernel stack for the thread is now set up in the thread_alloc(),
that itself may return NULL. Also, allocation of the first process
thread is performed in the fork1() to properly deal with stack
allocation failure. proc_linkup() is separated into proc_linkup()
called from fork1(), and proc_linkup0(), that is used to set up the
kernel process (was known as swapper).
In collaboration with: Peter Holm
Reviewed by: jhb
This avoids back-to-back faults for all TLB misses. This can be
improved further in the future by also setting PTE_DIRTY for TLB
misses for write accesses.
MFC after: 1 week
ways:
(1) Cached pages are no longer kept in the object's resident page
splay tree and memq. Instead, they are kept in a separate per-object
splay tree of cached pages. However, access to this new per-object
splay tree is synchronized by the _free_ page queues lock, not to be
confused with the heavily contended page queues lock. Consequently, a
cached page can be reclaimed by vm_page_alloc(9) without acquiring the
object's lock or the page queues lock.
This solves a problem independently reported by tegge@ and Isilon.
Specifically, they observed the page daemon consuming a great deal of
CPU time because of pages bouncing back and forth between the cache
queue (PQ_CACHE) and the inactive queue (PQ_INACTIVE). The source of
this problem turned out to be a deadlock avoidance strategy employed
when selecting a cached page to reclaim in vm_page_select_cache().
However, the root cause was really that reclaiming a cached page
required the acquisition of an object lock while the page queues lock
was already held. Thus, this change addresses the problem at its
root, by eliminating the need to acquire the object's lock.
Moreover, keeping cached pages in the object's primary splay tree and
memq was, in effect, optimizing for the uncommon case. Cached pages
are reclaimed far, far more often than they are reactivated. Instead,
this change makes reclamation cheaper, especially in terms of
synchronization overhead, and reactivation more expensive, because
reactivated pages will have to be reentered into the object's primary
splay tree and memq.
(2) Cached pages are now stored alongside free pages in the physical
memory allocator's buddy queues, increasing the likelihood that large
allocations of contiguous physical memory (i.e., superpages) will
succeed.
Finally, as a result of this change long-standing restrictions on when
and where a cached page can be reclaimed and returned by
vm_page_alloc(9) are eliminated. Specifically, calls to
vm_page_alloc(9) specifying VM_ALLOC_INTERRUPT can now reclaim and
return a formerly cached page. Consequently, a call to malloc(9)
specifying M_NOWAIT is less likely to fail.
Discussed with: many over the course of the summer, including jeff@,
Justin Husted @ Isilon, peter@, tegge@
Tested by: an earlier version by kris@
Approved by: re (kensmith)
of pages don't sum to anywhere near the total number of pages on amd64.
This is for the most part because uma_small_alloc() pages have never been
counted as wired pages, like their kmem_malloc() brethren. They should
be. This changes fixes that.
It is no longer necessary for the page queues lock to be held to free
pages allocated by uma_small_alloc(). I removed the acquisition and
release of the page queues lock from uma_small_free() on amd64 and ia64
weeks ago. This patch updates the other architectures that have
uma_small_alloc() and uma_small_free().
Approved by: re (kensmith)
Recently the AP in my Merced box seems to have grown a habit
of getting unexpected interrupts, such as redundant wake-ups
and legacy interrupts that require an INTA cycle.
While here, replace DELAY(0) with cpu_spinwait() so that it's
clear what we're doing as well as enable the code to take
advantage of cpu_spinwait() when it gets implemented.
Approved by: re (blanket)
