checks for the swap space consumption plus checks that the amount of
the free pages exceeds some limit, in case pagedeamon did not coped
with the page shortage in one of the late passes. This is wrong
because it does not account for the presence of the reclamaible pages
in the queues which are not selectable for reclaim immediately. E.g.,
on the swap-less systems, large active queue easily triggered OOM.
Instead, only raise OOM when pagedaemon is unable to produce a free
page in several back-to-back passes. Track the failed passes per
pagedaemon thread.
The number of passes to trigger OOM was selected empirically and
tested both on small (32M-64M i386 VM) and large (32G amd64)
configurations. If the specifics of the load require tuning, sysctl
vm.pageout_oom_seq sets the number of back-to-back passes which must
fail before OOM is raised. Each pass takes 1/2 of seconds. Less the
value, more sensible the pagedaemon is to the page shortage.
In future, some heuristic to calculate the value of the tunable might
be designed based on the system configuration and load. But before it
can be done, the i/o system must be fixed to reliably time-out
pagedaemon writes, even if waiting for the memory to proceed. Then,
code can account for the in-flight page-outs and postpone OOM until
all of them finished, which should reduce the need in tuning. Right
now, ignoring the in-flight writes and the counter allows to break
deadlocks due to write path doing sleepable memory allocations.
Reported by: Dmitry Sivachenko, bde, many others
Tested by: pho, bde, tuexen (arm)
Reviewed by: alc
Discussed with: bde, imp
Sponsored by: The FreeBSD Foundation
MFC after: 3 weeks
Residency count track the number of pte entries installed into the
current pmap, which does not reflect the consumption of the physical
memory by the address map. Due to several mechanisms like pv entries
reclamation, copy on write etc. the resident pte entries count may be
much less than the amount of physical memory kept by the process.
Provide the OOM-specific vm_pageout_oom_pagecount() function which
estimates the amount of reclamaible memory which could be stolen if
the process is killed.
Reported and tested by: pho
Reviewed by: alc
Comments text by: alc
Sponsored by: The FreeBSD Foundation
MFC after: 3 weeks
other actions, swaps out kernel stacks of the processes. On the other
hand, currentl OOM logic which selects a process to kill in the
critical condition, skips process with swapped-out thread. Under some
loads, this results in the big(gest) process being ignored by OOM.
Do not skip a process which has inhibited thread due to the swap-out,
in the OOM selection loop. Note that killing such process requires
the thread stack page-in, but sometimes this is the only way to
recover.
Reported and tested by: pho
Reviewed by: alc
Sponsored by: The FreeBSD Foundation
MFC after: 3 weeks
certain kernel structures for use by debuggers. This mostly aids
in examining cores from a kernel without debug symbols as a debugger
can infer these values if debug symbols are available.
One set of variables describes the layout of 'struct linker_file' to
walk the list of loaded kernel modules.
A second set of variables describes the layout of 'struct proc' and
'struct thread' to walk the list of processes in the kernel and the
threads in each process.
The 'pcb_size' variable is used to index into the stoppcbs[] array.
The 'vm_maxuser_address' is used to distinguish kernel virtual addresses
from user addresses. This doesn't have to be perfect, and
'vm_maxuser_address' is a cheap and simple way to differentiate kernel
pointers from simple values like TIDs and PIDs.
While here, annotate the fields in struct pcb used by kgdb on amd64
and i386 to note that their ABI should be preserved. Annotations for
other platforms will be added in the future.
Reviewed by: kib
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D3773
reclaimed in FIFO order by the pagedaemon. Previously we would enqueue
such pages at the head of the inactive queue, yielding a LIFO reclaim order.
Reviewed by: alc
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Division
pager. It is enough to execute VOP_BMAP() once to obtain both the
disk block address for the requested page, and the before/after limits
for the contiguous run. The clipping of the vm_page_t array passed to
the vnode_pager_generic_getpages() and the disk address for the first
page in the clipped array can be deduced from the call results.
While there, remove some noise (like if (1) {...}) and adjust nearby
code.
Reviewed by: alc
Discussed with: glebius
Tested by: pho
Sponsored by: The FreeBSD Foundation
MFC after: 3 weeks
in vm_pageout_fallback_object_lock() and vm_pageout_page_lock(). The
check for the m->queue == queue assumes that the page does belong to a
queue.
Modify the 'unchanged' calculation bu dereferencing the marker tailq
pointers, which is known to belong to the queue. Since for a page m
linked to the queue, m->queue must be equal to the queue index, assert
this instead of checking.
In collaboration with: alc
Sponsored by: The FreeBSD Foundation (kib)
MFC after: 2 weeks
8x performance improvement in a micro benchmark on a 4 socket machine.
- Get buffer headers from a per-cpu uma cache that sits in from of the
free queue.
- Use a per-cpu quantum cache in vmem to eliminate contention for kva.
- Use multiple clean queues according to buffer cache size to eliminate
clean queue lock contention.
- Introduce a bufspace daemon that attempts to prevent getnewbuf() callers
from blocking or doing direct recycling.
- Close some bufspace allocation races that could lead to endless
recycling.
- Further the transition to a more modern style of small functions grouped
by prefix in order to improve growing complexity.
Sponsored by: EMC / Isilon
Reviewed by: kib
Tested by: pho
the kernel or kmem object can't be paged out. Since they can't be paged
out, they are never enqueued in a paging queue. Nonetheless, passing
PQ_INACTIVE to vm_page_unwire() in kmem_unback() creates the appearance
that these pages are being enqueued in the inactive queue. As of r288122,
we can avoid giving this false impression by passing PQ_NONE.
Submitted by: kmacy
Differential Revision: https://reviews.freebsd.org/D1674
of POSIX_FADV_DONTNEED so that it causes the backing pages to be moved to
the head of the inactive queue instead of being cached.
This affects the implementation of POSIX_FADV_NOREUSE as well, since it
works by applying POSIX_FADV_DONTNEED to file ranges after they have been
read or written. At that point the corresponding buffers may still be
dirty, so the previous implementation would coalesce successive ranges and
apply POSIX_FADV_DONTNEED to the result, ensuring that pages backing the
dirty buffers would eventually be cached. To preserve this behaviour in an
efficient manner, this change adds a new buf flag, B_NOREUSE, which causes
the pages backing a VMIO buf to be placed at the head of the inactive queue
when the buf is released. POSIX_FADV_NOREUSE then works by setting this
flag in bufs that underlie the specified range.
Reviewed by: alc, kib
Sponsored by: EMC / Isilon Storage Division
Differential Revision: https://reviews.freebsd.org/D3726
arena in r254025 introduced a bug in the case when an allocation is only
partially successful. Specifically, the vm object lock was not being
acquired before freeing the allocated pages. To address this bug, replace
the existing code by a call to kmem_unback().
Change the type of a variable in kmem_alloc_attr() so that an allocation
of two or more gigabytes won't fail.
Replace the error handling code in kmem_back() by a call to kmem_unback().
Reviewed by: kib (an earlier version)
MFC after: 1 week
Sponsored by: EMC / Isilon Storage Division
by noobj_alloc() don't belong to a vm object, they can't be paged out.
Since they can't be paged out, they are never enqueued in a paging queue.
Nonetheless, passing PQ_INACTIVE to vm_page_unwire() creates the appearance
that these pages are being enqueued in the inactive queue. As of r288122,
we can avoid giving this false impression by passing PQ_NONE.
Submitted by: kmacy
Differential Revision: https://reviews.freebsd.org/D1674
queue and (2) returns a Boolean indicating whether the page's wire count
transitioned to zero.
Exploit this change in vfs_vmio_release() to avoid pointlessly enqueueing
a page that is about to be freed.
(An earlier version of this change was developed by attilio@ and kmacy@.
Any errors in this version are my own.)
Reviewed by: kib
Sponsored by: EMC / Isilon Storage Division
should not assume that vm_pages_needed will remain set while it sleeps.
Other threads can clear vm_pages_needed by performing a sufficient
number of vm_page_free() calls, e.g., process termination. The effect
of this error was that vm_pageout_worker() would free and/or launder
pages when, in fact, there was no shortage of free pages.
Rewrite a nearby comment to describe all of the possible cases and not
just the most common case. The problem being that the comment made
the most common case seem like the only case.
Reviewed by: kib
MFC after: 1 week
Sponsored by: EMC / Isilon Storage Division
with a reference count of zero can't possibly be mapped, so there is never a
need for vm_page_set_invalid() to call pmap_remove_all() on them.
Reviewed by: kib
MFC after: 1 week
Sponsored by: EMC / Isilon Storage Division
mapped address without valid pte installed, when parallel wiring of
the entry happen. The entry must be copy on write. If entry is COW
but was already copied, and parallel wiring set
MAP_ENTRY_IN_TRANSITION, vm_fault() would sleep waiting for the
MAP_ENTRY_IN_TRANSITION flag to clear. After that, the fault handler
is restarted and vm_map_lookup() or vm_map_lookup_locked() trip over
the check. Note that this is race, if the address is accessed after
the wiring is done, the entry does not fault at all.
There is no reason in the current kernel to disallow write access to
the COW wired entry if the entry permissions allow it. Initially this
was done in r24666, since that kernel did not supported proper
copy-on-write for wired text, which was fixed in r199869. The r251901
revision re-introduced the r24666 fix for the current VM.
Note that write access must clear MAP_ENTRY_NEEDS_COPY entry flag by
performing COW. In reverse, when MAP_ENTRY_NEEDS_COPY is set in
vmspace_fork(), the MAP_ENTRY_USER_WIRED flag is cleared. Put the
assert stating the invariant, instead of returning the error.
Reported and debugging help by: peter
Reviewed by: alc
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
locking and doesn't sleep. Flag the consumer we create as such. In
addition, decrement the in flight index when we have an out of memory
error after having incremented it previously. This would have
prevented swapoff from working if the swap pager ever hit a resource
shortage trying to swap out something (the swap in path always waits
for a bio, so won't have this issue). Simplify the close logic by
abandoning the use of private and initializing the index to 1 and
dropping that reference when we previously set private.
Also, set sw_id only while sw_dev_mtx is held. This should only affect
swapping to a vnode, as opposed to a geom whose close always sets it to
NULL with sw_dev_mtx held.
Differential Review: https://reviews.freebsd.org/D3547
so that there is only one place where pages are freed and only one place
where pages are moved to the tail of the queue.
Reviewed by: kib
Sponsored by: EMC / Isilon Storage Division
pages will have left the inactive queue before the page daemon performs
its next scan. Also, ignore references to pages from terminated objects.
This allows the clean pages to be freed a little sooner.
Move some comments to their proper place, i.e., next to the code that
they describe, and update other nearby comments.
Reviewed by: kib
Sponsored by: EMC / Isilon Storage Division
Objects obtained from such zones are supposed to retain type stability,
which was violated by aforementioned trashing.
This is a follow-up to r284861.
Discussed with: kib
This was added in r51337 as part of the implementation of
madvise(MADV_DONTNEED). Its objective was to ensure that the page daemon
would eventually reclaim other unreferenced pages (i.e., unreferenced pages
not touched by madvise()) from the active queue.
Now that the pagedaemon performs steady scanning of the active page queue,
this weighted handling is unnecessary. Instead, always "cache" clean pages
by moving them to the head of the inactive page queue. This simplifies the
implementation of vm_page_advise() and eliminates the fragmentation that
resulted from the distribution of pages among multiple queues.
Suggested by: alc
Reviewed by: alc
Sponsored by: EMC / Isilon Storage Division
Differential Revision: https://reviews.freebsd.org/D3401
it may involve a pmap operation that iterates over the page's PV list, so
unnecessarily holding the page lock is undesirable.
MFC after: 1 week
Sponsored by: EMC / Isilon Storage Division
Provide and document the RANDOM_ENABLE_UMA option.
Change RANDOM_FAST to RANDOM_UMA to clarify the harvesting.
Remove RANDOM_DEBUG option, replace with SDT probes. These will be of
use to folks measuring the harvesting effect when deciding whether to
use RANDOM_ENABLE_UMA.
Requested by: scottl and others.
Approved by: so (/dev/random blanket)
Differential Revision: https://reviews.freebsd.org/D3197
Currently vm_pageout_scan() uses a ticks-based scheme to rate-limit
the number of times that the vm_lowmem event will happen. However
if no events happen for long enough for ticks to roll over, this
leaves us in a long window in which vm_lowmem events will not
happen.
Replace the use of ticks with time_t to prevent rollover from ever
being an issue.
Reviewed by: ian
MFC after: 3 weeks
Sponsored by: EMC / Isilon Storage Division
Differential Revision: https://reviews.freebsd.org/D3439
to vm_page_try_to_cache() from vm_pageout_flush(). Other changes, most
recently r286814, have made this call unnecessary.
Reviewed by: kib
Discussed with: jeff
Tested by: pho
Sponsored by: EMC / Isilon Storage Division
initial thread stack is not adjusted by the tunable, the stack is
allocated too early to get access to the kernel environment. See
TD0_KSTACK_PAGES for the thread0 stack sizing on i386.
The tunable was tested on x86 only. From the visual inspection, it
seems that it might work on arm and powerpc. The arm
USPACE_SVC_STACK_TOP and powerpc USPACE macros seems to be already
incorrect for the threads with non-default kstack size. I only
changed the macros to use variable instead of constant, since I cannot
test.
On arm64, mips and sparc64, some static data structures are sized by
KSTACK_PAGES, so the tunable is disabled.
Sponsored by: The FreeBSD Foundation
MFC after: 2 week
Fixes "panic: vm_radix_reserve_kva: unable to reserve KVA" caused by sign
extention of "pages * UMA_SLAB_SIZE" value passed to kva_alloc() which
takes unsigned long argument.
In the erroneus case that triggered this bug, the number of pages
to allocate in uma_zone_reserve_kva() was 0x8ebe6, that gave the
total number of bytes to allocate equal to 0x8ebe6000 (int).
This was then sign extended in kva_alloc() to 0xffffffff8ebe6000
(unsigned long).
Reviewed by: alc, kib
Submitted by: Zbigniew Bodek <zbb@semihalf.com>
Obtained from: Semihalf
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D3346
vm_offset_t pmap_quick_enter_page(vm_page_t m)
void pmap_quick_remove_page(vm_offset_t kva)
These will create and destroy a temporary, CPU-local KVA mapping of a specified page.
Guarantees:
--Will not sleep and will not fail.
--Safe to call under a non-sleepable lock or from an ithread
Restrictions:
--Not guaranteed to be safe to call from an interrupt filter or under a spin mutex on all platforms
--Current implementation does not guarantee more than one page of mapping space across all platforms. MI code should not make nested calls to pmap_quick_enter_page.
--MI code should not perform locking while holding onto a mapping created by pmap_quick_enter_page
The idea is to use this in busdma, for bounce buffer copies as well as virtually-indexed cache maintenance on mips and arm.
NOTE: the non-i386, non-amd64 implementations of these functions still need review and testing.
Reviewed by: kib
Approved by: kib (mentor)
Differential Revision: http://reviews.freebsd.org/D3013
which constitute the majority of the pages that are processed by
vm_fault_dontneed(), are already near the tail of the inactive queue. Only
the pages at faulting virtual addresses are actually moved by
vm_page_advise(..., MADV_DONTNEED). However, vm_page_advise(...,
MADV_DONTNEED) is simultaneously too aggressive and passive for the moved
pages. It makes most of these pages too easily reclaimable, and at the same
time it leaves enough pages in the active queue to trigger pageouts by the
page daemon. Instead, with this change, the pages at faulting virtual
addresses are moved to the tail of the inactive queue, where they are
relatively close to the pages prefetched by the same page fault.
Discussed with: jeff
Sponsored by: EMC / Isilon Storage Division
the VM_FAULT_CHANGE_WIRING flag to VM_FAULT_WIRE. Assert that the
flag is only passed when faulting on the wired map entry. Remove the
vm_page_unwire() call, which should be never reachable.
Since VM_FAULT_WIRE flag implies wired map entry, the TRYPAGER() macro
is reduced to the testing of the fs.object having a default pager.
Inline the check.
Suggested and reviewed by: alc
Tested by: pho (previous version)
MFC after: 1 week
'buf' is inconvenient and has lead me to some irritating to discover
bugs over the years. It also makes it more challenging to refactor
the buf allocation system.
- Move swbuf and declare it as an extern in vfs_bio.c. This is still
not perfect but better than it was before.
- Eliminate the unused ffs function that relied on knowledge of the buf
array.
- Move the shutdown code that iterates over the buf array into vfs_bio.c.
Reviewed by: kib
Sponsored by: EMC / Isilon Storage Division
Assume that a vnode is mapped shared and mlocked(), and then the vnode
is truncated, or truncated and then again extended past the mapping
point EOF. Truncation removes the pages past the truncation point,
and if pages are later created at this range, they are not properly
mapped into the mlocked region, and their wiring count is wrong.
The revert leaves the invalidated but wired pages on the object queue,
which means that the pages are found by vm_object_unwire() when the
mapped range is munlock()ed, and reused by the buffer cache when the
vnode is extended again.
The changes in r173708 were required since then vm_map_unwire() looked
at the page tables to find the page to unwire. This is no longer
needed with the vm_object_unwire() introduction, which follows the
objects shadow chain.
Also eliminate OBJPR_NOTWIRED flag for vm_object_page_remove(), which
is now redundand, we do not remove wired pages.
Reported by: trasz, Dmitry Sivachenko <trtrmitya@gmail.com>
Suggested and reviewed by: alc
Tested by: pho
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
- Use pointer assignment rather than a combination of pointers and
flags to switch buffers between unmapped and mapped. This eliminates
multiple flags and generally simplifies the logic.
- Eliminate b_saveaddr since it is only used with pager bufs which have
their b_data re-initialized on each allocation.
- Gather up some convenience routines in the buffer cache for
manipulating buf space and buf malloc space.
- Add an inline, buf_mapped(), to standardize checks around unmapped
buffers.
In collaboration with: mlaier
Reviewed by: kib
Tested by: pho (many small revisions ago)
Sponsored by: EMC / Isilon Storage Division
This is based on work done by jeff@ and jhb@, as well as the numa.diff
patch that has been circulating when someone asks for first-touch NUMA
on -10 or -11.
* Introduce a simple set of VM policy and iterator types.
* tie the policy types into the vm_phys path for now, mirroring how
the initial first-touch allocation work was enabled.
* add syscalls to control changing thread and process defaults.
* add a global NUMA VM domain policy.
* implement a simple cascade policy order - if a thread policy exists, use it;
if a process policy exists, use it; use the default policy.
* processes inherit policies from their parent processes, threads inherit
policies from their parent threads.
* add a simple tool (numactl) to query and modify default thread/process
policities.
* add documentation for the new syscalls, for numa and for numactl.
* re-enable first touch NUMA again by default, as now policies can be
set in a variety of methods.
This is only relevant for very specific workloads.
This doesn't pretend to be a final NUMA solution.
The previous defaults in -HEAD (with MAXMEMDOM set) can be achieved by
'sysctl vm.default_policy=rr'.
This is only relevant if MAXMEMDOM is set to something other than 1.
Ie, if you're using GENERIC or a modified kernel with non-NUMA, then
this is a glorified no-op for you.
Thank you to Norse Corp for giving me access to rather large
(for FreeBSD!) NUMA machines in order to develop and verify this.
Thank you to Dell for providing me with dual socket sandybridge
and westmere v3 hardware to do NUMA development with.
Thank you to Scott Long at Netflix for providing me with access
to the two-socket, four-domain haswell v3 hardware.
Thank you to Peter Holm for running the stress testing suite
against the NUMA branch during various stages of development!
Tested:
* MIPS (regression testing; non-NUMA)
* i386 (regression testing; non-NUMA GENERIC)
* amd64 (regression testing; non-NUMA GENERIC)
* westmere, 2 socket (thankyou norse!)
* sandy bridge, 2 socket (thankyou dell!)
* ivy bridge, 2 socket (thankyou norse!)
* westmere-EX, 4 socket / 1TB RAM (thankyou norse!)
* haswell, 2 socket (thankyou norse!)
* haswell v3, 2 socket (thankyou dell)
* haswell v3, 2x18 core (thankyou scott long / netflix!)
* Peter Holm ran a stress test suite on this work and found one
issue, but has not been able to verify it (it doesn't look NUMA
related, and he only saw it once over many testing runs.)
* I've tested bhyve instances running in fixed NUMA domains and cpusets;
all seems to work correctly.
Verified:
* intel-pcm - pcm-numa.x and pcm-memory.x, whilst selecting different
NUMA policies for processes under test.
Review:
This was reviewed through phabricator (https://reviews.freebsd.org/D2559)
as well as privately and via emails to freebsd-arch@. The git history
with specific attributes is available at https://github.com/erikarn/freebsd/
in the NUMA branch (https://github.com/erikarn/freebsd/compare/local/adrian_numa_policy).
This has been reviewed by a number of people (stas, rpaulo, kib, ngie,
wblock) but not achieved a clear consensus. My hope is that with further
exposure and testing more functionality can be implemented and evaluated.
Notes:
* The VM doesn't handle unbalanced domains very well, and if you have an overly
unbalanced memory setup whilst under high memory pressure, VM page allocation
may fail leading to a kernel panic. This was a problem in the past, but it's
much more easily triggered now with these tools.
* This work only controls the path through vm_phys; it doesn't yet strongly/predictably
affect contigmalloc, KVA placement, UMA, etc. So, driver placement of memory
isn't really guaranteed in any way. That's next on my plate.
Sponsored by: Norse Corp, Inc.; Dell
However, I've observed the active queue scan stopping when there are
frequent free page shortages and the inactive queue is steadily refilled
by other mechanisms, such as the sequential access heuristic in vm_fault()
or madvise(2). To remedy this problem, record the time of the last active
queue scan, and always scan a number of pages proportional to the time
since the last scan, regardless of whether that last scan was a
timeout-triggered ("pass == 0") or free-page-shortage-triggered ("pass >
0") scan.
Also, on a timeout-triggered scan, allow a full scan of the active queue
when the system is short of inactive pages.
Reviewed by: kib
MFC after: 6 weeks
Sponsored by: EMC / Isilon Storage Division
