From the submitter description:
The process is forked transitioning a map entry to COW
Thread A writes to a page on the map entry, faults, updates the pmap to
writable at a new phys addr, and starts TLB invalidations...
Thread B acquires a lock, writes to a location on the new phys addr, and
releases the lock
Thread C acquires the lock, reads from the location on the old phys addr...
Thread A ...continues the TLB invalidations which are completed
Thread C ...reads from the location on the new phys addr, and releases
the lock
In this example Thread B and C [lock, use and unlock] properly and
neither own the lock at the same time. Thread A was writing somewhere
else on the page and so never had/needed the lock. Thread C sees a
location that is only ever read|modified under a lock change beneath
it while it is the lock owner.
To fix this, perform the two-stage update of the copied PTE. First,
the PTE is updated with the address of the new physical page with
copied content, but in read-only mode. The pmap locking and the page
busy state during PTE update and TLB invalidation IPIs ensure that any
writer to the page cannot upgrade the PTE to the writable state until
all CPUs updated their TLB to not cache old mapping. Then, after the
busy state of the page is lifted, the faults for write can proceed and
do not violate the consistency of the reads.
The change is done in vm_fault because most architectures do need IPIs
to invalidate remote TLBs. More, I think that hardware guarantees of
atomicity of the remote TLB invalidation are not enough to prevent the
inconsistent reads of non-atomic reads, like multi-word accesses
protected by a lock. So instead of modifying each pmap invalidation
code, I did it there.
Discovered and analyzed by: Elliott.Rabe@dell.com
Reviewed by: markj
PR: 225584 (appeared to have the same cause)
Tested by: Elliott.Rabe@dell.com, emaste, Mike Tancsa <mike@sentex.net>, truckman
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D14347
If the map entry elookup was performed due to the mapping changes, we
need to ensure that there is still some access permission bit
requested which is compatible with the current vm_map_entry mode. If
not, restart the handler from scratch instead of trying to save the
current progress.
Also adjust fault_type to not include cleared permission bits.
Reviewed by: markj
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D14347
Suppose that we have an object with a mapped superpage, and that all
pages in the superpages are held (by some driver). Additionally,
suppose that the object is terminated, e.g. because the only process
mapping it is exiting. Then the reservation is broken, but the pages
cannot be freed until later, when they are unheld. In this situation,
the reservation code cannot clean psind, since no pages are freed, and
the page is freed and then reused with invalid psind.
Clean psind on vm_reserv_break() to avoid the situation.
Reported and tested by: Slava Shwartsman
Reviewed by: markj
Sponsored by: Mellanox Technologies
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D14335
significant source of cache line contention from vm_page_alloc(). Use
accessors and vm_page_unwire_noq() so that the mechanism can be easily
changed in the future.
Reviewed by: markj
Discussed with: kib, glebius
Tested by: pho (earlier version)
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D14273
size of UMA zone allocation is greater than page size. In this case zone
of zones can not use UMA_MD_SMALL_ALLOC, and we need to postpone switch
off of this zone from startup_alloc() until full launch of VM.
o Always supply number of VM zones to uma_startup_count(). On machines
with UMA_MD_SMALL_ALLOC ignore it completely, unless zsize goes over
a page. In the latter case account VM zones for number of allocations
from the zone of zones.
o Rewrite startup_alloc() so that it will immediately switch off from
itself any zone that is already capable of running real alloc.
In worst case scenario we may leak a single page here. See comment
in uma_startup_count().
o Hardcode call to uma_startup2() into vm_mem_init(). Otherwise some
extra SYSINITs, e.g. vm_page_init() may sneak in before.
o While here, remove uma_boot_pages_mtx. With recent changes to boot
pages calculation, we are guaranteed to use all of the boot_pages
in the early single threaded stage.
Reported & tested by: mav
o Most of startup zones have struct uma_slab embedded into the slab,
so provide macro UMA_SLAB_SPACE and use it instead of UMA_SLAB_SIZE,
when calculating how many pages would certain kind of allocations
require. Some zones are offpage, so we might have a positive inaccuracy.
o The keg for the zone of zones is allocated "dynamically", so we
need +1 when calculating amount of pages for kegs. [1]
o The zones of zones and zones of kegs have arbitrary alignment of 32,
and this also needs to be accounted for. [2]
While here, spread more comments and improve diagnostic messages.
Reported by: pho [1], jtl [2]
Previously, wiring a page would cause it to be removed from its page
queue. In the common case, unwiring causes it to be enqueued at the tail
of that page queue. This change modifies vm_page_wire() to not dequeue
the page, thus avoiding the highly contended page queue locks. Instead,
vm_page_unwire() takes care of requeuing the page as a single operation,
and the page daemon dequeues wired pages as they are encountered during
a queue scan to avoid needlessly revisiting them later. For pages in
PQ_ACTIVE we do even better, since a requeue is unnecessary.
The change improves scalability for some common workloads. For instance,
threads wiring pages into the buffer cache no longer need to modify
global page queues, and unwiring is usually done by the bufspace thread,
so concurrency is not as much of an issue. As another example, many
sysctl handlers wire the output buffer to avoid faults on copyout, and
since the buffer is likely to be in PQ_ACTIVE, we now entirely avoid
modifying the page queue in this case.
The change also adds a block comment describing some properties of
struct vm_page's reference counters, and the busy lock.
Reviewed by: jeff
Discussed with: alc, kib
MFC after: 1 month
Differential Revision: https://reviews.freebsd.org/D11943
global to per-domain state. Protect reservations with the free lock
from the domain that they belong to. Refactor to make vm domains more
of a first class object.
Reviewed by: markj, kib, gallatin
Tested by: pho
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D14000
o Call uma_startup1() after initializing kmem, vmem and domains.
o Include 8 eight VM startup pages into uma_startup_count() calculation.
o Account for vmem_startup() and vm_map_startup() preallocating pages.
o Account for extra two allocations done by kmem_init() and vmem_create().
o Hardcode the place of execution of vm_radix_reserve_kva(). Using SYSINIT
allowed several other SYSINITs to sneak in before it, thus bumping
requirement for amount of boot pages.
for UMA startup.
o Introduce another stage of UMA startup, which is entered after
vm_page_startup() finishes. After this stage we don't yet enable buckets,
but we can ask VM for pages. Rename stages to meaningful names while here.
New list of stages: BOOT_COLD, BOOT_STRAPPED, BOOT_PAGEALLOC, BOOT_BUCKETS,
BOOT_RUNNING.
Enabling page alloc earlier allows us to dramatically reduce number of
boot pages required. What is more important number of zones becomes
consistent across different machines, as no MD allocations are done before
the BOOT_PAGEALLOC stage. Now only UMA internal zones actually need to use
startup_alloc(), however that may change, so vm_page_startup() provides
its need for early zones as argument.
o Introduce uma_startup_count() function, to avoid code duplication. The
functions calculates sizes of zones zone and kegs zone, and calculates how
many pages UMA will need to bootstrap.
It counts not only of zone structures, but also of kegs, slabs and hashes.
o Hide uma_startup_foo() declarations from public file.
o Provide several DIAGNOSTIC printfs on boot_pages usage.
o Bugfix: when calculating zone of zones size use (mp_maxid + 1) instead of
mp_ncpus. Use resulting number not only in the size argument to zone_ctor()
but also as args.size.
Reviewed by: imp, gallatin (earlier version)
Differential Revision: https://reviews.freebsd.org/D14054
It is possible, for complex fork()/collapse situations, to have
sibling address spaces to partially share shadow chains. If one
sibling performs wiring, it can happen that a transient page, invalid
and busy, is installed into a shadow object which is visible to other
sibling for the duration of vm_fault_hold(). When the backing object
contains the valid page, and the wiring is performed on read-only
entry, the transient page is eventually removed.
But the sibling which observed the transient page might perform the
unwire, executing vm_object_unwire(). There, the first page found in
the shadow chain is considered as the page that was wired for the
mapping. It is really the page below it which is wired. So we unwire
the wrong page, either triggering the asserts of breaking the page'
wire counter.
As the fix, wait for the busy state to finish if we find such page
during unwire, and restart the shadow chain walk after the sleep.
Reported and tested by: pho
Reviewed by: markj
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D14184
clear dirty bits for completely invalid blocks.
Otherwise we might not write out the last chunk that is shorter than
512 bytes, if the file end is not aligned on disk block boundary.
This become important after the r324794.
PR: 225586
Reported by: tris_vern@hotmail.com
Tested by: pho
Sponsored by: The FreeBSD Foundation
MFC after: 3 days
In several places, entry start and end field are checked, after
excluding the possibility that the entry is map->header. By assigning
max and min values to the start and end fields of map->header in
vm_map_init, the explicit map->header checks become unnecessary.
Submitted by: Doug Moore <dougm@rice.edu>
Reviewed by: alc, kib, markj (previous version)
Tested by: pho (previous version)
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D13735
kernel by PHYS_TO_DMAP() as previously present on amd64, arm64, riscv, and
powerpc64. This introduces a new MI macro (PMAP_HAS_DMAP) that can be
evaluated at runtime to determine if the architecture has a direct map;
if it does not (or does) unconditionally and PMAP_HAS_DMAP is either 0 or
1, the compiler can remove the conditional logic.
As part of this, implement PHYS_TO_DMAP() on sparc64 and mips64, which had
similar things but spelled differently. 32-bit MIPS has a partial direct-map
that maps poorly to this concept and is unchanged.
Reviewed by: kib
Suggestions from: marius, alc, kib
Runtime tested on: amd64, powerpc64, powerpc, mips64
allocated with a tag to come from the specified domain if it meets the
other constraints provided by the tag. Automatically create a tag at
the root of each bus specifying the domain local to that bus if
available.
Reviewed by: jhb, kib
Tested by: pho
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D13545
domains can be done by the _domain() API variants. UMA also supports a
first-touch policy via the NUMA zone flag.
The slab layer is now segregated by VM domains and is precise. It handles
iteration for round-robin directly. The per-cpu cache layer remains
a mix of domains according to where memory is allocated and freed. Well
behaved clients can achieve perfect locality with no performance penalty.
The direct domain allocation functions have to visit the slab layer and
so require per-zone locks which come at some expense.
Reviewed by: Attilio (a slightly older version)
Tested by: pho
Sponsored by: Netflix, Dell/EMC Isilon
reservations by giving each memory domain its own KVA space in vmem that
is naturally aligned on superpage boundaries.
Reviewed by: alc, markj, kib (some objections)
Sponsored by: Netflix, Dell/EMC Isilon
Tested by; pho
Differential Revision: https://reviews.freebsd.org/D13289
Implement 'domainset', a cpuset based NUMA policy mechanism. This allows
userspace to control NUMA policy administratively and programmatically.
Implement domainset based iterators in the page layer.
Remove the now legacy numa_* syscalls.
Cleanup some header polution created by having seq.h in proc.h.
Reviewed by: markj, kib
Discussed with: alc
Tested by: pho
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D13403
userspace to control NUMA policy administratively and programmatically.
Implement domainset based iterators in the page layer.
Remove the now legacy numa_* syscalls.
Cleanup some header polution created by having seq.h in proc.h.
Reviewed by: markj, kib
Discussed with: alc
Tested by: pho
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D13403
Consolidate the regions covered by the process lock.
Combine similar conditions tests into one, e.g. all process flags can
be test with one logical operation.
Add check for in-exec state, since p_vmspace is dererenced.
Remove labels and goto by explicitly tracking state.
Update comments.
Reviewed by: alc, markj (previous version)
Tested by: pho (previous version)
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
Differential revision: https://reviews.freebsd.org/D13693
its per-thread kernel stack pages by making them pass through the inactive
queue first. Instead, immediately place them in the laundry so that they
might be cleaned and made available for reclamation sooner.
Reviewed by: kib, markj
MFC after: 1 week
rather than kmem arena size to determine available memory.
Initialize the UMA limit to LONG_MAX to avoid spurious wakeups on boot before
the real limit is set.
PR: 224330 (partial), 224080
Reviewed by: markj, avg
Sponsored by: Netflix / Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D13494
On a load where single anonymous object consumes almost all memory on
the large system, swapout code executes the iteration over the
corresponding object page queue for long time, owning the map and
object locks. This blocks pagedaemon which tries to lock the object,
and blocks other threads in the process in vm_fault() waiting for the
map lock.
Handle the issue by terminating the deactivation loop if we executed
too long and by yielding at the top level in vm_daemon.
Reported by: peterj, pho
Reviewed by: alc
Tested by: pho (as part of the larger patch)
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D13671
introduction in r83366. (At that time, this code appeared in vm/vm_glue.c,
because vm/vm_swapout.c did not exist.) When the FOREACH_THREAD loop
completes, we know that the sleep time for every thread is above whichever
threshold is being applied.
Reviewed by: kib
X-MFC with: r327354
swp_pager_meta_ctl(), with no opportunity to recognize freeing of
consecutive blocks and free fewer block ranges. To open that opportunity,
this change removes the SWM_FREE option from swp_pager_meta_ctl(), and
compels the caller to do the freeing when a valid block address is returned.
In swap_pager_copy(), these frees are aggregated, so that a sequence of them
can be done at one time.
The only other caller to swp_pager_meta_ctl() that passed SWM_FREE,
swp_pager_unswapped(), is also modified to handle its single free
explicitly.
Submitted by: Doug Moore <dougm@rice.edu>
Reviewed by: kib (an earlier version)
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D13290
Neither swapout_procs() nor swapout() access the map. Since the
process' vmspace is referenced only to obtain the pointer to the
vm_map, the reference is not needed as well.
Reviewed by: alc
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D13681
Reviewed by: alc, markj (as part of the larger patch)
Tested by: pho (again, as part of the larger patch)
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D13671
for finding aligned free space in the given map. With this change, we
always return KERN_NO_SPACE when we fail to find free space. Whereas,
previously, we might return KERN_INVALID_ADDRESS. Also, with this change,
we explicitly check for address wrap, rather than relying upon the map's
min and max addresses to establish sentinel-like regions.
This refactoring was inspired by the problem that we addressed in r326098.
Reviewed by: kib
Tested by: pho
Discussed with: markj
MFC after: 3 weeks
Differential Revision: https://reviews.freebsd.org/D13346
Otherwise the page daemon will not reclaim pages and thus will not
wake threads sleeping in VM_WAIT.
Reported and tested by: pho
Reviewed by: alc, kib
X-MFC with: r327168
Differential Revision: https://reviews.freebsd.org/D13640
Both issues caused the page daemon to erroneously go to sleep when
applications are consuming free pages at a high rate, leaving the
application threads blocked in VM_WAIT.
1) After completing an inactive queue scan, concurrent allocations may
have prevented the page daemon from meeting the v_free_min threshold.
In this case, the page daemon was going to sleep even when the
inactive queue contained plenty of clean pages.
2) pagedaemon_wakeup() may be called without the free queues lock held.
This can lead to a lost wakeup if a call occurs after the page daemon
clears vm_pageout_wanted but before going to sleep.
Fix 1) by ensuring that we start a new inactive queue scan immediately
if v_free_count < v_free_min after a prior scan.
Fix 2) by adding a new subroutine, pagedaemon_wait(), called from
vm_wait() and vm_waitpfault(). It wakes up the page daemon if either
vm_pages_needed or vm_pageout_wanted is false, and atomically sleeps
on v_free_count.
Reported by: jeff
Reviewed by: alc
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D13424
The laundry thread keeps track of the number of inactive queue scans
performed by the page daemon, and was previously using the v_pdwakeups
counter to count them. However, in some cases the inactive queue may
be scanned multiple times after a single wakeup, so it's more accurate
to use a dedicated counter.
Reviewed by: alc, kib (previous version)
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D13422
r292392 modified the active queue scan to weigh clean pages differently
from dirty pages when attempting to meet the inactive queue target. When
r306706 was merged into the PQ_LAUNDRY branch, this mechanism was
broken. Fix it by scalaing the correct page shortage variable.
Reviewed by: alc, kib
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D13423
atomic_set_*() sets a bit in the target memory location, so
atomic_set_int(&uma_reclaim_needed, 0) does not do what it looks like
it does.
PR: 224080
Reviewed by: jeff, kib
Differential Revision: https://reviews.freebsd.org/D13412
Commit r326346 moved domain iterators from physical layer to vm_page one,
but it also removed translation of freelist to flind for
vm_page_alloc_freelist() call. Before it expects VM_FREELIST_ parameter,
but after it expect freelist index.
On small WiFi boxes with few megabytes of RAM, there is only one freelist
VM_FREELIST_LOWMEM (1) and there is no VM_FREELIST_DEFAULT(0) (see file
sys/mips/include/vmparam.h). It results in freelist 1 with flind 0.
At first, this commit renames flind to freelist in vm_page_alloc_freelist
to avoid misunderstanding about input parameters. Then on physical layer it
restores translation for correct handling of freelist parameter.
Reported by: landonf
Reviewed by: jeff
Differential Revision: https://reviews.freebsd.org/D13351
It's theoretically possible for the vnode and object to be disassociated
while locks are dropped around the vget() call, in which case we
shouldn't proceed with laundering.
Noted and reviewed by: kib
MFC after: 1 week
