2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
10 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * The Mach Operating System project at Carnegie-Mellon University.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
44 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
47 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
48 * All rights reserved.
50 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
52 * Permission to use, copy, modify and distribute this software and
53 * its documentation is hereby granted, provided that both the copyright
54 * notice and this permission notice appear in all copies of the
55 * software, derivative works or modified versions, and any portions
56 * thereof, and that both notices appear in supporting documentation.
58 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
59 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
60 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
62 * Carnegie Mellon requests users of this software to return to
64 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
65 * School of Computer Science
66 * Carnegie Mellon University
67 * Pittsburgh PA 15213-3890
69 * any improvements or extensions that they make and grant Carnegie the
70 * rights to redistribute these changes.
74 * The proverbial page-out daemon.
77 #include <sys/cdefs.h>
78 __FBSDID("$FreeBSD$");
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
85 #include <sys/eventhandler.h>
87 #include <sys/mutex.h>
89 #include <sys/kthread.h>
91 #include <sys/mount.h>
92 #include <sys/racct.h>
93 #include <sys/resourcevar.h>
94 #include <sys/sched.h>
96 #include <sys/signalvar.h>
99 #include <sys/vnode.h>
100 #include <sys/vmmeter.h>
101 #include <sys/rwlock.h>
103 #include <sys/sysctl.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/vm_pagequeue.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
119 * System initialization
122 /* the kernel process "vm_pageout"*/
123 static void vm_pageout(void);
124 static void vm_pageout_init(void);
125 static int vm_pageout_clean(vm_page_t m, int *numpagedout);
126 static int vm_pageout_cluster(vm_page_t m);
127 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
128 int starting_page_shortage);
130 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
133 struct proc *pageproc;
135 static struct kproc_desc page_kp = {
140 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
143 SDT_PROVIDER_DEFINE(vm);
144 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
146 /* Pagedaemon activity rates, in subdivisions of one second. */
147 #define VM_LAUNDER_RATE 10
148 #define VM_INACT_SCAN_RATE 10
150 static int vm_pageout_oom_seq = 12;
152 static int vm_pageout_update_period;
153 static int disable_swap_pageouts;
154 static int lowmem_period = 10;
155 static int swapdev_enabled;
157 static int vm_panic_on_oom = 0;
159 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
160 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
161 "panic on out of memory instead of killing the largest process");
163 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
164 CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
165 "Maximum active LRU update period");
167 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
168 "Low memory callback period");
170 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
171 CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
173 static int pageout_lock_miss;
174 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
175 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
177 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
178 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
179 "back-to-back calls to oom detector to start OOM");
181 static int act_scan_laundry_weight = 3;
182 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
183 &act_scan_laundry_weight, 0,
184 "weight given to clean vs. dirty pages in active queue scans");
186 static u_int vm_background_launder_rate = 4096;
187 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
188 &vm_background_launder_rate, 0,
189 "background laundering rate, in kilobytes per second");
191 static u_int vm_background_launder_max = 20 * 1024;
192 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
193 &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
195 int vm_pageout_page_count = 32;
197 u_long vm_page_max_user_wired;
198 SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW,
199 &vm_page_max_user_wired, 0,
200 "system-wide limit to user-wired page count");
202 static u_int isqrt(u_int num);
203 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
205 static void vm_pageout_laundry_worker(void *arg);
208 struct vm_batchqueue bq;
209 struct vm_pagequeue *pq;
216 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
217 vm_page_t marker, vm_page_t after, int maxscan)
220 vm_pagequeue_assert_locked(pq);
221 KASSERT((marker->a.flags & PGA_ENQUEUED) == 0,
222 ("marker %p already enqueued", marker));
225 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
227 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
228 vm_page_aflag_set(marker, PGA_ENQUEUED);
230 vm_batchqueue_init(&ss->bq);
233 ss->maxscan = maxscan;
235 vm_pagequeue_unlock(pq);
239 vm_pageout_end_scan(struct scan_state *ss)
241 struct vm_pagequeue *pq;
244 vm_pagequeue_assert_locked(pq);
245 KASSERT((ss->marker->a.flags & PGA_ENQUEUED) != 0,
246 ("marker %p not enqueued", ss->marker));
248 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
249 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
250 pq->pq_pdpages += ss->scanned;
254 * Add a small number of queued pages to a batch queue for later processing
255 * without the corresponding queue lock held. The caller must have enqueued a
256 * marker page at the desired start point for the scan. Pages will be
257 * physically dequeued if the caller so requests. Otherwise, the returned
258 * batch may contain marker pages, and it is up to the caller to handle them.
260 * When processing the batch queue, vm_page_queue() must be used to
261 * determine whether the page has been logically dequeued by another thread.
262 * Once this check is performed, the page lock guarantees that the page will
263 * not be disassociated from the queue.
265 static __always_inline void
266 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
268 struct vm_pagequeue *pq;
269 vm_page_t m, marker, n;
274 KASSERT((marker->a.flags & PGA_ENQUEUED) != 0,
275 ("marker %p not enqueued", ss->marker));
277 vm_pagequeue_lock(pq);
278 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
279 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
280 m = n, ss->scanned++) {
281 n = TAILQ_NEXT(m, plinks.q);
282 if ((m->flags & PG_MARKER) == 0) {
283 KASSERT((m->a.flags & PGA_ENQUEUED) != 0,
284 ("page %p not enqueued", m));
285 KASSERT((m->flags & PG_FICTITIOUS) == 0,
286 ("Fictitious page %p cannot be in page queue", m));
287 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
288 ("Unmanaged page %p cannot be in page queue", m));
292 (void)vm_batchqueue_insert(&ss->bq, m);
294 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
295 vm_page_aflag_clear(m, PGA_ENQUEUED);
298 TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
299 if (__predict_true(m != NULL))
300 TAILQ_INSERT_BEFORE(m, marker, plinks.q);
302 TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
304 vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
305 vm_pagequeue_unlock(pq);
309 * Return the next page to be scanned, or NULL if the scan is complete.
311 static __always_inline vm_page_t
312 vm_pageout_next(struct scan_state *ss, const bool dequeue)
315 if (ss->bq.bq_cnt == 0)
316 vm_pageout_collect_batch(ss, dequeue);
317 return (vm_batchqueue_pop(&ss->bq));
321 * Determine whether processing of a page should be deferred and ensure that any
322 * outstanding queue operations are processed.
324 static __always_inline bool
325 vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued)
329 as = vm_page_astate_load(m);
330 if (__predict_false(as.queue != queue ||
331 ((as.flags & PGA_ENQUEUED) != 0) != enqueued))
333 if ((as.flags & PGA_QUEUE_OP_MASK) != 0) {
334 vm_page_pqbatch_submit(m, queue);
341 * Scan for pages at adjacent offsets within the given page's object that are
342 * eligible for laundering, form a cluster of these pages and the given page,
343 * and launder that cluster.
346 vm_pageout_cluster(vm_page_t m)
349 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
351 int ib, is, page_base, pageout_count;
354 VM_OBJECT_ASSERT_WLOCKED(object);
357 vm_page_assert_xbusied(m);
359 mc[vm_pageout_page_count] = pb = ps = m;
361 page_base = vm_pageout_page_count;
366 * We can cluster only if the page is not clean, busy, or held, and
367 * the page is in the laundry queue.
369 * During heavy mmap/modification loads the pageout
370 * daemon can really fragment the underlying file
371 * due to flushing pages out of order and not trying to
372 * align the clusters (which leaves sporadic out-of-order
373 * holes). To solve this problem we do the reverse scan
374 * first and attempt to align our cluster, then do a
375 * forward scan if room remains.
378 while (ib != 0 && pageout_count < vm_pageout_page_count) {
383 if ((p = vm_page_prev(pb)) == NULL ||
384 vm_page_tryxbusy(p) == 0) {
388 if (vm_page_wired(p)) {
393 vm_page_test_dirty(p);
400 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
407 mc[--page_base] = pb = p;
412 * We are at an alignment boundary. Stop here, and switch
413 * directions. Do not clear ib.
415 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
418 while (pageout_count < vm_pageout_page_count &&
419 pindex + is < object->size) {
420 if ((p = vm_page_next(ps)) == NULL ||
421 vm_page_tryxbusy(p) == 0)
423 if (vm_page_wired(p)) {
427 vm_page_test_dirty(p);
433 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
439 mc[page_base + pageout_count] = ps = p;
445 * If we exhausted our forward scan, continue with the reverse scan
446 * when possible, even past an alignment boundary. This catches
447 * boundary conditions.
449 if (ib != 0 && pageout_count < vm_pageout_page_count)
452 return (vm_pageout_flush(&mc[page_base], pageout_count,
453 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
457 * vm_pageout_flush() - launder the given pages
459 * The given pages are laundered. Note that we setup for the start of
460 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
461 * reference count all in here rather then in the parent. If we want
462 * the parent to do more sophisticated things we may have to change
465 * Returned runlen is the count of pages between mreq and first
466 * page after mreq with status VM_PAGER_AGAIN.
467 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
468 * for any page in runlen set.
471 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
474 vm_object_t object = mc[0]->object;
475 int pageout_status[count];
479 VM_OBJECT_ASSERT_WLOCKED(object);
482 * Initiate I/O. Mark the pages shared busy and verify that they're
483 * valid and read-only.
485 * We do not have to fixup the clean/dirty bits here... we can
486 * allow the pager to do it after the I/O completes.
488 * NOTE! mc[i]->dirty may be partial or fragmented due to an
489 * edge case with file fragments.
491 for (i = 0; i < count; i++) {
492 KASSERT(vm_page_all_valid(mc[i]),
493 ("vm_pageout_flush: partially invalid page %p index %d/%d",
495 KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
496 ("vm_pageout_flush: writeable page %p", mc[i]));
497 vm_page_busy_downgrade(mc[i]);
499 vm_object_pip_add(object, count);
501 vm_pager_put_pages(object, mc, count, flags, pageout_status);
503 runlen = count - mreq;
506 for (i = 0; i < count; i++) {
507 vm_page_t mt = mc[i];
509 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
510 !pmap_page_is_write_mapped(mt),
511 ("vm_pageout_flush: page %p is not write protected", mt));
512 switch (pageout_status[i]) {
515 if (vm_page_in_laundry(mt))
516 vm_page_deactivate_noreuse(mt);
524 * The page is outside the object's range. We pretend
525 * that the page out worked and clean the page, so the
526 * changes will be lost if the page is reclaimed by
531 if (vm_page_in_laundry(mt))
532 vm_page_deactivate_noreuse(mt);
538 * If the page couldn't be paged out to swap because the
539 * pager wasn't able to find space, place the page in
540 * the PQ_UNSWAPPABLE holding queue. This is an
541 * optimization that prevents the page daemon from
542 * wasting CPU cycles on pages that cannot be reclaimed
543 * becase no swap device is configured.
545 * Otherwise, reactivate the page so that it doesn't
546 * clog the laundry and inactive queues. (We will try
547 * paging it out again later.)
550 if (object->type == OBJT_SWAP &&
551 pageout_status[i] == VM_PAGER_FAIL) {
552 vm_page_unswappable(mt);
555 vm_page_activate(mt);
557 if (eio != NULL && i >= mreq && i - mreq < runlen)
561 if (i >= mreq && i - mreq < runlen)
567 * If the operation is still going, leave the page busy to
568 * block all other accesses. Also, leave the paging in
569 * progress indicator set so that we don't attempt an object
572 if (pageout_status[i] != VM_PAGER_PEND) {
573 vm_object_pip_wakeup(object);
579 return (numpagedout);
583 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
586 atomic_store_rel_int(&swapdev_enabled, 1);
590 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
593 if (swap_pager_nswapdev() == 1)
594 atomic_store_rel_int(&swapdev_enabled, 0);
598 * Attempt to acquire all of the necessary locks to launder a page and
599 * then call through the clustering layer to PUTPAGES. Wait a short
600 * time for a vnode lock.
602 * Requires the page and object lock on entry, releases both before return.
603 * Returns 0 on success and an errno otherwise.
606 vm_pageout_clean(vm_page_t m, int *numpagedout)
614 vm_page_assert_locked(m);
616 VM_OBJECT_ASSERT_WLOCKED(object);
622 * The object is already known NOT to be dead. It
623 * is possible for the vget() to block the whole
624 * pageout daemon, but the new low-memory handling
625 * code should prevent it.
627 * We can't wait forever for the vnode lock, we might
628 * deadlock due to a vn_read() getting stuck in
629 * vm_wait while holding this vnode. We skip the
630 * vnode if we can't get it in a reasonable amount
633 if (object->type == OBJT_VNODE) {
637 if (vp->v_type == VREG &&
638 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
644 ("vp %p with NULL v_mount", vp));
645 vm_object_reference_locked(object);
647 VM_OBJECT_WUNLOCK(object);
648 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
649 LK_SHARED : LK_EXCLUSIVE;
650 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
655 VM_OBJECT_WLOCK(object);
658 * Ensure that the object and vnode were not disassociated
659 * while locks were dropped.
661 if (vp->v_object != object) {
668 * While the object and page were unlocked, the page
670 * (1) moved to a different queue,
671 * (2) reallocated to a different object,
672 * (3) reallocated to a different offset, or
675 if (!vm_page_in_laundry(m) || m->object != object ||
676 m->pindex != pindex || m->dirty == 0) {
683 * The page may have been busied while the object and page
684 * locks were released.
686 if (vm_page_tryxbusy(m) == 0) {
694 * Remove all writeable mappings, failing if the page is wired.
696 if (!vm_page_try_remove_write(m)) {
705 * If a page is dirty, then it is either being washed
706 * (but not yet cleaned) or it is still in the
707 * laundry. If it is still in the laundry, then we
708 * start the cleaning operation.
710 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
714 VM_OBJECT_WUNLOCK(object);
717 vm_page_lock_assert(m, MA_NOTOWNED);
721 vm_object_deallocate(object);
722 vn_finished_write(mp);
729 * Attempt to launder the specified number of pages.
731 * Returns the number of pages successfully laundered.
734 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
736 struct scan_state ss;
737 struct vm_pagequeue *pq;
741 vm_page_astate_t new, old;
742 int act_delta, error, numpagedout, queue, refs, starting_target;
748 starting_target = launder;
752 * Scan the laundry queues for pages eligible to be laundered. We stop
753 * once the target number of dirty pages have been laundered, or once
754 * we've reached the end of the queue. A single iteration of this loop
755 * may cause more than one page to be laundered because of clustering.
757 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
758 * swap devices are configured.
760 if (atomic_load_acq_int(&swapdev_enabled))
761 queue = PQ_UNSWAPPABLE;
766 marker = &vmd->vmd_markers[queue];
767 pq = &vmd->vmd_pagequeues[queue];
768 vm_pagequeue_lock(pq);
769 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
770 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
771 if (__predict_false((m->flags & PG_MARKER) != 0))
774 vm_page_change_lock(m, &mtx);
778 * Don't touch a page that was removed from the queue after the
779 * page queue lock was released. Otherwise, ensure that any
780 * pending queue operations, such as dequeues for wired pages,
783 if (vm_pageout_defer(m, queue, true))
786 if (object != m->object) {
788 VM_OBJECT_WUNLOCK(object);
791 * A page's object pointer may be set to NULL before
792 * the object lock is acquired.
794 object = (vm_object_t)atomic_load_ptr(&m->object);
795 if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) {
797 /* Depends on type-stability. */
798 VM_OBJECT_WLOCK(object);
803 if (__predict_false(m->object == NULL))
805 * The page has been removed from its object.
808 KASSERT(m->object == object, ("page %p does not belong to %p",
811 if (vm_page_tryxbusy(m) == 0)
815 * Check for wirings now that we hold the object lock and have
816 * verified that the page is unbusied. If the page is mapped,
817 * it may still be wired by pmap lookups. The call to
818 * vm_page_try_remove_all() below atomically checks for such
819 * wirings and removes mappings. If the page is unmapped, the
820 * wire count is guaranteed not to increase.
822 if (__predict_false(vm_page_wired(m))) {
823 vm_page_dequeue_deferred(m);
828 * Invalid pages can be easily freed. They cannot be
829 * mapped; vm_page_free() asserts this.
831 if (vm_page_none_valid(m))
834 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
836 for (old = vm_page_astate_load(m);;) {
838 * Check to see if the page has been removed from the
839 * queue since the first such check. Leave it alone if
840 * so, discarding any references collected by
841 * pmap_ts_referenced().
843 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
848 if ((old.flags & PGA_REFERENCED) != 0) {
849 new.flags &= ~PGA_REFERENCED;
852 if (act_delta == 0) {
854 } else if (object->ref_count != 0) {
856 * Increase the activation count if the page was
857 * referenced while in the laundry queue. This
858 * makes it less likely that the page will be
859 * returned prematurely to the laundry queue.
861 new.act_count += ACT_ADVANCE +
863 if (new.act_count > ACT_MAX)
864 new.act_count = ACT_MAX;
866 new.flags |= PGA_REQUEUE;
867 new.queue = PQ_ACTIVE;
868 if (!vm_page_pqstate_commit(m, &old, new))
872 * If this was a background laundering, count
873 * activated pages towards our target. The
874 * purpose of background laundering is to ensure
875 * that pages are eventually cycled through the
876 * laundry queue, and an activation is a valid
881 VM_CNT_INC(v_reactivated);
883 } else if ((object->flags & OBJ_DEAD) == 0) {
884 new.flags |= PGA_REQUEUE;
885 if (!vm_page_pqstate_commit(m, &old, new))
893 * If the page appears to be clean at the machine-independent
894 * layer, then remove all of its mappings from the pmap in
895 * anticipation of freeing it. If, however, any of the page's
896 * mappings allow write access, then the page may still be
897 * modified until the last of those mappings are removed.
899 if (object->ref_count != 0) {
900 vm_page_test_dirty(m);
901 if (m->dirty == 0 && !vm_page_try_remove_all(m)) {
902 vm_page_dequeue_deferred(m);
908 * Clean pages are freed, and dirty pages are paged out unless
909 * they belong to a dead object. Requeueing dirty pages from
910 * dead objects is pointless, as they are being paged out and
911 * freed by the thread that destroyed the object.
917 } else if ((object->flags & OBJ_DEAD) == 0) {
918 if (object->type != OBJT_SWAP &&
919 object->type != OBJT_DEFAULT)
921 else if (disable_swap_pageouts)
931 * Form a cluster with adjacent, dirty pages from the
932 * same object, and page out that entire cluster.
934 * The adjacent, dirty pages must also be in the
935 * laundry. However, their mappings are not checked
936 * for new references. Consequently, a recently
937 * referenced page may be paged out. However, that
938 * page will not be prematurely reclaimed. After page
939 * out, the page will be placed in the inactive queue,
940 * where any new references will be detected and the
943 error = vm_pageout_clean(m, &numpagedout);
945 launder -= numpagedout;
946 ss.scanned += numpagedout;
947 } else if (error == EDEADLK) {
962 if (object != NULL) {
963 VM_OBJECT_WUNLOCK(object);
966 vm_pagequeue_lock(pq);
967 vm_pageout_end_scan(&ss);
968 vm_pagequeue_unlock(pq);
970 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
976 * Wakeup the sync daemon if we skipped a vnode in a writeable object
977 * and we didn't launder enough pages.
979 if (vnodes_skipped > 0 && launder > 0)
980 (void)speedup_syncer();
982 return (starting_target - launder);
986 * Compute the integer square root.
991 u_int bit, root, tmp;
993 bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
1008 * Perform the work of the laundry thread: periodically wake up and determine
1009 * whether any pages need to be laundered. If so, determine the number of pages
1010 * that need to be laundered, and launder them.
1013 vm_pageout_laundry_worker(void *arg)
1015 struct vm_domain *vmd;
1016 struct vm_pagequeue *pq;
1017 uint64_t nclean, ndirty, nfreed;
1018 int domain, last_target, launder, shortfall, shortfall_cycle, target;
1021 domain = (uintptr_t)arg;
1022 vmd = VM_DOMAIN(domain);
1023 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1024 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1027 in_shortfall = false;
1028 shortfall_cycle = 0;
1029 last_target = target = 0;
1033 * Calls to these handlers are serialized by the swap syscall lock.
1035 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
1036 EVENTHANDLER_PRI_ANY);
1037 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
1038 EVENTHANDLER_PRI_ANY);
1041 * The pageout laundry worker is never done, so loop forever.
1044 KASSERT(target >= 0, ("negative target %d", target));
1045 KASSERT(shortfall_cycle >= 0,
1046 ("negative cycle %d", shortfall_cycle));
1050 * First determine whether we need to launder pages to meet a
1051 * shortage of free pages.
1053 if (shortfall > 0) {
1054 in_shortfall = true;
1055 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1057 } else if (!in_shortfall)
1059 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1061 * We recently entered shortfall and began laundering
1062 * pages. If we have completed that laundering run
1063 * (and we are no longer in shortfall) or we have met
1064 * our laundry target through other activity, then we
1065 * can stop laundering pages.
1067 in_shortfall = false;
1071 launder = target / shortfall_cycle--;
1075 * There's no immediate need to launder any pages; see if we
1076 * meet the conditions to perform background laundering:
1078 * 1. The ratio of dirty to clean inactive pages exceeds the
1079 * background laundering threshold, or
1080 * 2. we haven't yet reached the target of the current
1081 * background laundering run.
1083 * The background laundering threshold is not a constant.
1084 * Instead, it is a slowly growing function of the number of
1085 * clean pages freed by the page daemon since the last
1086 * background laundering. Thus, as the ratio of dirty to
1087 * clean inactive pages grows, the amount of memory pressure
1088 * required to trigger laundering decreases. We ensure
1089 * that the threshold is non-zero after an inactive queue
1090 * scan, even if that scan failed to free a single clean page.
1093 nclean = vmd->vmd_free_count +
1094 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1095 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1096 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1097 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1098 target = vmd->vmd_background_launder_target;
1102 * We have a non-zero background laundering target. If we've
1103 * laundered up to our maximum without observing a page daemon
1104 * request, just stop. This is a safety belt that ensures we
1105 * don't launder an excessive amount if memory pressure is low
1106 * and the ratio of dirty to clean pages is large. Otherwise,
1107 * proceed at the background laundering rate.
1112 last_target = target;
1113 } else if (last_target - target >=
1114 vm_background_launder_max * PAGE_SIZE / 1024) {
1117 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1118 launder /= VM_LAUNDER_RATE;
1119 if (launder > target)
1126 * Because of I/O clustering, the number of laundered
1127 * pages could exceed "target" by the maximum size of
1128 * a cluster minus one.
1130 target -= min(vm_pageout_launder(vmd, launder,
1131 in_shortfall), target);
1132 pause("laundp", hz / VM_LAUNDER_RATE);
1136 * If we're not currently laundering pages and the page daemon
1137 * hasn't posted a new request, sleep until the page daemon
1140 vm_pagequeue_lock(pq);
1141 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1142 (void)mtx_sleep(&vmd->vmd_laundry_request,
1143 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1146 * If the pagedaemon has indicated that it's in shortfall, start
1147 * a shortfall laundering unless we're already in the middle of
1148 * one. This may preempt a background laundering.
1150 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1151 (!in_shortfall || shortfall_cycle == 0)) {
1152 shortfall = vm_laundry_target(vmd) +
1153 vmd->vmd_pageout_deficit;
1159 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1160 nfreed += vmd->vmd_clean_pages_freed;
1161 vmd->vmd_clean_pages_freed = 0;
1162 vm_pagequeue_unlock(pq);
1167 * Compute the number of pages we want to try to move from the
1168 * active queue to either the inactive or laundry queue.
1170 * When scanning active pages during a shortage, we make clean pages
1171 * count more heavily towards the page shortage than dirty pages.
1172 * This is because dirty pages must be laundered before they can be
1173 * reused and thus have less utility when attempting to quickly
1174 * alleviate a free page shortage. However, this weighting also
1175 * causes the scan to deactivate dirty pages more aggressively,
1176 * improving the effectiveness of clustering.
1179 vm_pageout_active_target(struct vm_domain *vmd)
1183 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1184 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1185 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1186 shortage *= act_scan_laundry_weight;
1191 * Scan the active queue. If there is no shortage of inactive pages, scan a
1192 * small portion of the queue in order to maintain quasi-LRU.
1195 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1197 struct scan_state ss;
1200 vm_page_t m, marker;
1201 struct vm_pagequeue *pq;
1202 vm_page_astate_t old, new;
1204 int act_delta, max_scan, ps_delta, refs, scan_tick;
1207 marker = &vmd->vmd_markers[PQ_ACTIVE];
1208 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1209 vm_pagequeue_lock(pq);
1212 * If we're just idle polling attempt to visit every
1213 * active page within 'update_period' seconds.
1216 if (vm_pageout_update_period != 0) {
1217 min_scan = pq->pq_cnt;
1218 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1219 min_scan /= hz * vm_pageout_update_period;
1222 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1223 vmd->vmd_last_active_scan = scan_tick;
1226 * Scan the active queue for pages that can be deactivated. Update
1227 * the per-page activity counter and use it to identify deactivation
1228 * candidates. Held pages may be deactivated.
1230 * To avoid requeuing each page that remains in the active queue, we
1231 * implement the CLOCK algorithm. To keep the implementation of the
1232 * enqueue operation consistent for all page queues, we use two hands,
1233 * represented by marker pages. Scans begin at the first hand, which
1234 * precedes the second hand in the queue. When the two hands meet,
1235 * they are moved back to the head and tail of the queue, respectively,
1236 * and scanning resumes.
1238 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1241 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1242 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1243 if (__predict_false(m == &vmd->vmd_clock[1])) {
1244 vm_pagequeue_lock(pq);
1245 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1246 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1247 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1249 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1251 max_scan -= ss.scanned;
1252 vm_pageout_end_scan(&ss);
1255 if (__predict_false((m->flags & PG_MARKER) != 0))
1258 vm_page_change_lock(m, &mtx);
1261 * Don't touch a page that was removed from the queue after the
1262 * page queue lock was released. Otherwise, ensure that any
1263 * pending queue operations, such as dequeues for wired pages,
1266 if (vm_pageout_defer(m, PQ_ACTIVE, true))
1270 * A page's object pointer may be set to NULL before
1271 * the object lock is acquired.
1273 object = (vm_object_t)atomic_load_ptr(&m->object);
1274 if (__predict_false(object == NULL))
1276 * The page has been removed from its object.
1280 /* Deferred free of swap space. */
1281 if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
1282 VM_OBJECT_TRYWLOCK(object)) {
1283 if (m->object == object)
1284 vm_pager_page_unswapped(m);
1285 VM_OBJECT_WUNLOCK(object);
1289 * Check to see "how much" the page has been used.
1291 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1292 * that a reference from a concurrently destroyed mapping is
1293 * observed here and now.
1295 * Perform an unsynchronized object ref count check. While
1296 * the page lock ensures that the page is not reallocated to
1297 * another object, in particular, one with unmanaged mappings
1298 * that cannot support pmap_ts_referenced(), two races are,
1299 * nonetheless, possible:
1300 * 1) The count was transitioning to zero, but we saw a non-
1301 * zero value. pmap_ts_referenced() will return zero
1302 * because the page is not mapped.
1303 * 2) The count was transitioning to one, but we saw zero.
1304 * This race delays the detection of a new reference. At
1305 * worst, we will deactivate and reactivate the page.
1307 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1309 old = vm_page_astate_load(m);
1312 * Check to see if the page has been removed from the
1313 * queue since the first such check. Leave it alone if
1314 * so, discarding any references collected by
1315 * pmap_ts_referenced().
1317 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1321 * Advance or decay the act_count based on recent usage.
1325 if ((old.flags & PGA_REFERENCED) != 0) {
1326 new.flags &= ~PGA_REFERENCED;
1329 if (act_delta != 0) {
1330 new.act_count += ACT_ADVANCE + act_delta;
1331 if (new.act_count > ACT_MAX)
1332 new.act_count = ACT_MAX;
1334 new.act_count -= min(new.act_count,
1338 if (new.act_count > 0) {
1340 * Adjust the activation count and keep the page
1341 * in the active queue. The count might be left
1342 * unchanged if it is saturated. The page may
1343 * have been moved to a different queue since we
1344 * started the scan, in which case we move it
1348 if (old.queue != PQ_ACTIVE) {
1349 old.queue = PQ_ACTIVE;
1350 old.flags |= PGA_REQUEUE;
1354 * When not short for inactive pages, let dirty
1355 * pages go through the inactive queue before
1356 * moving to the laundry queue. This gives them
1357 * some extra time to be reactivated,
1358 * potentially avoiding an expensive pageout.
1359 * However, during a page shortage, the inactive
1360 * queue is necessarily small, and so dirty
1361 * pages would only spend a trivial amount of
1362 * time in the inactive queue. Therefore, we
1363 * might as well place them directly in the
1364 * laundry queue to reduce queuing overhead.
1366 * Calling vm_page_test_dirty() here would
1367 * require acquisition of the object's write
1368 * lock. However, during a page shortage,
1369 * directing dirty pages into the laundry queue
1370 * is only an optimization and not a
1371 * requirement. Therefore, we simply rely on
1372 * the opportunistic updates to the page's dirty
1373 * field by the pmap.
1375 if (page_shortage <= 0) {
1376 nqueue = PQ_INACTIVE;
1378 } else if (m->dirty == 0) {
1379 nqueue = PQ_INACTIVE;
1380 ps_delta = act_scan_laundry_weight;
1382 nqueue = PQ_LAUNDRY;
1386 new.flags |= PGA_REQUEUE;
1389 } while (!vm_page_pqstate_commit(m, &old, new));
1391 page_shortage -= ps_delta;
1397 vm_pagequeue_lock(pq);
1398 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1399 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1400 vm_pageout_end_scan(&ss);
1401 vm_pagequeue_unlock(pq);
1405 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker,
1408 vm_page_astate_t as;
1410 vm_pagequeue_assert_locked(pq);
1412 as = vm_page_astate_load(m);
1413 if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
1415 vm_page_aflag_set(m, PGA_ENQUEUED);
1416 TAILQ_INSERT_BEFORE(marker, m, plinks.q);
1421 * Re-add stuck pages to the inactive queue. We will examine them again
1422 * during the next scan. If the queue state of a page has changed since
1423 * it was physically removed from the page queue in
1424 * vm_pageout_collect_batch(), don't do anything with that page.
1427 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1430 struct vm_pagequeue *pq;
1435 marker = ss->marker;
1439 if (vm_batchqueue_insert(bq, m))
1441 vm_pagequeue_lock(pq);
1442 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1444 vm_pagequeue_lock(pq);
1445 while ((m = vm_batchqueue_pop(bq)) != NULL)
1446 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1447 vm_pagequeue_cnt_add(pq, delta);
1448 vm_pagequeue_unlock(pq);
1449 vm_batchqueue_init(bq);
1453 * Attempt to reclaim the requested number of pages from the inactive queue.
1454 * Returns true if the shortage was addressed.
1457 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1460 struct scan_state ss;
1461 struct vm_batchqueue rq;
1463 vm_page_t m, marker;
1464 struct vm_pagequeue *pq;
1466 vm_page_astate_t old, new;
1467 int act_delta, addl_page_shortage, deficit, page_shortage, refs;
1468 int starting_page_shortage;
1471 * The addl_page_shortage is an estimate of the number of temporarily
1472 * stuck pages in the inactive queue. In other words, the
1473 * number of pages from the inactive count that should be
1474 * discounted in setting the target for the active queue scan.
1476 addl_page_shortage = 0;
1479 * vmd_pageout_deficit counts the number of pages requested in
1480 * allocations that failed because of a free page shortage. We assume
1481 * that the allocations will be reattempted and thus include the deficit
1482 * in our scan target.
1484 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1485 starting_page_shortage = page_shortage = shortage + deficit;
1489 vm_batchqueue_init(&rq);
1492 * Start scanning the inactive queue for pages that we can free. The
1493 * scan will stop when we reach the target or we have scanned the
1494 * entire queue. (Note that m->a.act_count is not used to make
1495 * decisions for the inactive queue, only for the active queue.)
1497 marker = &vmd->vmd_markers[PQ_INACTIVE];
1498 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1499 vm_pagequeue_lock(pq);
1500 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1501 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1502 KASSERT((m->flags & PG_MARKER) == 0,
1503 ("marker page %p was dequeued", m));
1505 vm_page_change_lock(m, &mtx);
1509 * Don't touch a page that was removed from the queue after the
1510 * page queue lock was released. Otherwise, ensure that any
1511 * pending queue operations, such as dequeues for wired pages,
1514 if (vm_pageout_defer(m, PQ_INACTIVE, false))
1517 if (object != m->object) {
1519 VM_OBJECT_WUNLOCK(object);
1522 * A page's object pointer may be set to NULL before
1523 * the object lock is acquired.
1525 object = (vm_object_t)atomic_load_ptr(&m->object);
1526 if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) {
1528 /* Depends on type-stability. */
1529 VM_OBJECT_WLOCK(object);
1534 if (__predict_false(m->object == NULL))
1536 * The page has been removed from its object.
1539 KASSERT(m->object == object, ("page %p does not belong to %p",
1542 if (vm_page_tryxbusy(m) == 0) {
1544 * Don't mess with busy pages. Leave them at
1545 * the front of the queue. Most likely, they
1546 * are being paged out and will leave the
1547 * queue shortly after the scan finishes. So,
1548 * they ought to be discounted from the
1551 addl_page_shortage++;
1555 /* Deferred free of swap space. */
1556 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1557 vm_pager_page_unswapped(m);
1560 * Re-check for wirings now that we hold the object lock and
1561 * have verified that the page is unbusied. If the page is
1562 * mapped, it may still be wired by pmap lookups. The call to
1563 * vm_page_try_remove_all() below atomically checks for such
1564 * wirings and removes mappings. If the page is unmapped, the
1565 * wire count is guaranteed not to increase.
1567 if (__predict_false(vm_page_wired(m))) {
1568 vm_page_dequeue_deferred(m);
1573 * Invalid pages can be easily freed. They cannot be
1574 * mapped, vm_page_free() asserts this.
1576 if (vm_page_none_valid(m))
1579 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1581 for (old = vm_page_astate_load(m);;) {
1583 * Check to see if the page has been removed from the
1584 * queue since the first such check. Leave it alone if
1585 * so, discarding any references collected by
1586 * pmap_ts_referenced().
1588 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1593 if ((old.flags & PGA_REFERENCED) != 0) {
1594 new.flags &= ~PGA_REFERENCED;
1597 if (act_delta == 0) {
1599 } else if (object->ref_count != 0) {
1601 * Increase the activation count if the
1602 * page was referenced while in the
1603 * inactive queue. This makes it less
1604 * likely that the page will be returned
1605 * prematurely to the inactive queue.
1607 new.act_count += ACT_ADVANCE +
1609 if (new.act_count > ACT_MAX)
1610 new.act_count = ACT_MAX;
1612 new.flags |= PGA_REQUEUE;
1613 new.queue = PQ_ACTIVE;
1614 if (!vm_page_pqstate_commit(m, &old, new))
1617 VM_CNT_INC(v_reactivated);
1619 } else if ((object->flags & OBJ_DEAD) == 0) {
1620 new.queue = PQ_INACTIVE;
1621 new.flags |= PGA_REQUEUE;
1622 if (!vm_page_pqstate_commit(m, &old, new))
1630 * If the page appears to be clean at the machine-independent
1631 * layer, then remove all of its mappings from the pmap in
1632 * anticipation of freeing it. If, however, any of the page's
1633 * mappings allow write access, then the page may still be
1634 * modified until the last of those mappings are removed.
1636 if (object->ref_count != 0) {
1637 vm_page_test_dirty(m);
1638 if (m->dirty == 0 && !vm_page_try_remove_all(m)) {
1639 vm_page_dequeue_deferred(m);
1645 * Clean pages can be freed, but dirty pages must be sent back
1646 * to the laundry, unless they belong to a dead object.
1647 * Requeueing dirty pages from dead objects is pointless, as
1648 * they are being paged out and freed by the thread that
1649 * destroyed the object.
1651 if (m->dirty == 0) {
1654 * Because we dequeued the page and have already
1655 * checked for concurrent dequeue and enqueue
1656 * requests, we can safely disassociate the page
1657 * from the inactive queue.
1659 KASSERT((m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
1660 ("page %p has queue state", m));
1661 m->a.queue = PQ_NONE;
1666 if ((object->flags & OBJ_DEAD) == 0)
1672 vm_pageout_reinsert_inactive(&ss, &rq, m);
1677 VM_OBJECT_WUNLOCK(object);
1678 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1679 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1680 vm_pagequeue_lock(pq);
1681 vm_pageout_end_scan(&ss);
1682 vm_pagequeue_unlock(pq);
1684 VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1687 * Wake up the laundry thread so that it can perform any needed
1688 * laundering. If we didn't meet our target, we're in shortfall and
1689 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1690 * swap devices are configured, the laundry thread has no work to do, so
1691 * don't bother waking it up.
1693 * The laundry thread uses the number of inactive queue scans elapsed
1694 * since the last laundering to determine whether to launder again, so
1697 if (starting_page_shortage > 0) {
1698 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1699 vm_pagequeue_lock(pq);
1700 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1701 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1702 if (page_shortage > 0) {
1703 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1704 VM_CNT_INC(v_pdshortfalls);
1705 } else if (vmd->vmd_laundry_request !=
1706 VM_LAUNDRY_SHORTFALL)
1707 vmd->vmd_laundry_request =
1708 VM_LAUNDRY_BACKGROUND;
1709 wakeup(&vmd->vmd_laundry_request);
1711 vmd->vmd_clean_pages_freed +=
1712 starting_page_shortage - page_shortage;
1713 vm_pagequeue_unlock(pq);
1717 * Wakeup the swapout daemon if we didn't free the targeted number of
1720 if (page_shortage > 0)
1724 * If the inactive queue scan fails repeatedly to meet its
1725 * target, kill the largest process.
1727 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1730 * Reclaim pages by swapping out idle processes, if configured to do so.
1732 vm_swapout_run_idle();
1735 * See the description of addl_page_shortage above.
1737 *addl_shortage = addl_page_shortage + deficit;
1739 return (page_shortage <= 0);
1742 static int vm_pageout_oom_vote;
1745 * The pagedaemon threads randlomly select one to perform the
1746 * OOM. Trying to kill processes before all pagedaemons
1747 * failed to reach free target is premature.
1750 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1751 int starting_page_shortage)
1755 if (starting_page_shortage <= 0 || starting_page_shortage !=
1757 vmd->vmd_oom_seq = 0;
1760 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1762 vmd->vmd_oom = FALSE;
1763 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1769 * Do not follow the call sequence until OOM condition is
1772 vmd->vmd_oom_seq = 0;
1777 vmd->vmd_oom = TRUE;
1778 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1779 if (old_vote != vm_ndomains - 1)
1783 * The current pagedaemon thread is the last in the quorum to
1784 * start OOM. Initiate the selection and signaling of the
1787 vm_pageout_oom(VM_OOM_MEM);
1790 * After one round of OOM terror, recall our vote. On the
1791 * next pass, current pagedaemon would vote again if the low
1792 * memory condition is still there, due to vmd_oom being
1795 vmd->vmd_oom = FALSE;
1796 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1800 * The OOM killer is the page daemon's action of last resort when
1801 * memory allocation requests have been stalled for a prolonged period
1802 * of time because it cannot reclaim memory. This function computes
1803 * the approximate number of physical pages that could be reclaimed if
1804 * the specified address space is destroyed.
1806 * Private, anonymous memory owned by the address space is the
1807 * principal resource that we expect to recover after an OOM kill.
1808 * Since the physical pages mapped by the address space's COW entries
1809 * are typically shared pages, they are unlikely to be released and so
1810 * they are not counted.
1812 * To get to the point where the page daemon runs the OOM killer, its
1813 * efforts to write-back vnode-backed pages may have stalled. This
1814 * could be caused by a memory allocation deadlock in the write path
1815 * that might be resolved by an OOM kill. Therefore, physical pages
1816 * belonging to vnode-backed objects are counted, because they might
1817 * be freed without being written out first if the address space holds
1818 * the last reference to an unlinked vnode.
1820 * Similarly, physical pages belonging to OBJT_PHYS objects are
1821 * counted because the address space might hold the last reference to
1825 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1828 vm_map_entry_t entry;
1832 map = &vmspace->vm_map;
1833 KASSERT(!map->system_map, ("system map"));
1834 sx_assert(&map->lock, SA_LOCKED);
1836 VM_MAP_ENTRY_FOREACH(entry, map) {
1837 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1839 obj = entry->object.vm_object;
1842 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1843 obj->ref_count != 1)
1845 switch (obj->type) {
1850 res += obj->resident_page_count;
1857 static int vm_oom_ratelim_last;
1858 static int vm_oom_pf_secs = 10;
1859 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1861 static struct mtx vm_oom_ratelim_mtx;
1864 vm_pageout_oom(int shortage)
1866 struct proc *p, *bigproc;
1867 vm_offset_t size, bigsize;
1874 * For OOM requests originating from vm_fault(), there is a high
1875 * chance that a single large process faults simultaneously in
1876 * several threads. Also, on an active system running many
1877 * processes of middle-size, like buildworld, all of them
1878 * could fault almost simultaneously as well.
1880 * To avoid killing too many processes, rate-limit OOMs
1881 * initiated by vm_fault() time-outs on the waits for free
1884 mtx_lock(&vm_oom_ratelim_mtx);
1886 if (shortage == VM_OOM_MEM_PF &&
1887 (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1888 mtx_unlock(&vm_oom_ratelim_mtx);
1891 vm_oom_ratelim_last = now;
1892 mtx_unlock(&vm_oom_ratelim_mtx);
1895 * We keep the process bigproc locked once we find it to keep anyone
1896 * from messing with it; however, there is a possibility of
1897 * deadlock if process B is bigproc and one of its child processes
1898 * attempts to propagate a signal to B while we are waiting for A's
1899 * lock while walking this list. To avoid this, we don't block on
1900 * the process lock but just skip a process if it is already locked.
1904 sx_slock(&allproc_lock);
1905 FOREACH_PROC_IN_SYSTEM(p) {
1909 * If this is a system, protected or killed process, skip it.
1911 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1912 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1913 p->p_pid == 1 || P_KILLED(p) ||
1914 (p->p_pid < 48 && swap_pager_avail != 0)) {
1919 * If the process is in a non-running type state,
1920 * don't touch it. Check all the threads individually.
1923 FOREACH_THREAD_IN_PROC(p, td) {
1925 if (!TD_ON_RUNQ(td) &&
1926 !TD_IS_RUNNING(td) &&
1927 !TD_IS_SLEEPING(td) &&
1928 !TD_IS_SUSPENDED(td) &&
1929 !TD_IS_SWAPPED(td)) {
1941 * get the process size
1943 vm = vmspace_acquire_ref(p);
1950 sx_sunlock(&allproc_lock);
1951 if (!vm_map_trylock_read(&vm->vm_map)) {
1953 sx_slock(&allproc_lock);
1957 size = vmspace_swap_count(vm);
1958 if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
1959 size += vm_pageout_oom_pagecount(vm);
1960 vm_map_unlock_read(&vm->vm_map);
1962 sx_slock(&allproc_lock);
1965 * If this process is bigger than the biggest one,
1968 if (size > bigsize) {
1969 if (bigproc != NULL)
1977 sx_sunlock(&allproc_lock);
1978 if (bigproc != NULL) {
1979 if (vm_panic_on_oom != 0)
1980 panic("out of swap space");
1982 killproc(bigproc, "out of swap space");
1983 sched_nice(bigproc, PRIO_MIN);
1985 PROC_UNLOCK(bigproc);
1990 * Signal a free page shortage to subsystems that have registered an event
1991 * handler. Reclaim memory from UMA in the event of a severe shortage.
1992 * Return true if the free page count should be re-evaluated.
1995 vm_pageout_lowmem(void)
1997 static int lowmem_ticks = 0;
2003 last = atomic_load_int(&lowmem_ticks);
2004 while ((u_int)(ticks - last) / hz >= lowmem_period) {
2005 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
2009 * Decrease registered cache sizes.
2011 SDT_PROBE0(vm, , , vm__lowmem_scan);
2012 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
2015 * We do this explicitly after the caches have been
2018 uma_reclaim(UMA_RECLAIM_TRIM);
2023 * Kick off an asynchronous reclaim of cached memory if one of the
2024 * page daemons is failing to keep up with demand. Use the "severe"
2025 * threshold instead of "min" to ensure that we do not blow away the
2026 * caches if a subset of the NUMA domains are depleted by kernel memory
2027 * allocations; the domainset iterators automatically skip domains
2028 * below the "min" threshold on the first pass.
2030 * UMA reclaim worker has its own rate-limiting mechanism, so don't
2031 * worry about kicking it too often.
2033 if (vm_page_count_severe())
2034 uma_reclaim_wakeup();
2040 vm_pageout_worker(void *arg)
2042 struct vm_domain *vmd;
2044 int addl_shortage, domain, shortage;
2047 domain = (uintptr_t)arg;
2048 vmd = VM_DOMAIN(domain);
2053 * XXXKIB It could be useful to bind pageout daemon threads to
2054 * the cores belonging to the domain, from which vm_page_array
2058 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
2059 vmd->vmd_last_active_scan = ticks;
2062 * The pageout daemon worker is never done, so loop forever.
2065 vm_domain_pageout_lock(vmd);
2068 * We need to clear wanted before we check the limits. This
2069 * prevents races with wakers who will check wanted after they
2072 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
2075 * Might the page daemon need to run again?
2077 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
2079 * Yes. If the scan failed to produce enough free
2080 * pages, sleep uninterruptibly for some time in the
2081 * hope that the laundry thread will clean some pages.
2083 vm_domain_pageout_unlock(vmd);
2085 pause("pwait", hz / VM_INACT_SCAN_RATE);
2088 * No, sleep until the next wakeup or until pages
2089 * need to have their reference stats updated.
2091 if (mtx_sleep(&vmd->vmd_pageout_wanted,
2092 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
2093 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
2094 VM_CNT_INC(v_pdwakeups);
2097 /* Prevent spurious wakeups by ensuring that wanted is set. */
2098 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2101 * Use the controller to calculate how many pages to free in
2102 * this interval, and scan the inactive queue. If the lowmem
2103 * handlers appear to have freed up some pages, subtract the
2104 * difference from the inactive queue scan target.
2106 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
2108 ofree = vmd->vmd_free_count;
2109 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
2110 shortage -= min(vmd->vmd_free_count - ofree,
2112 target_met = vm_pageout_scan_inactive(vmd, shortage,
2118 * Scan the active queue. A positive value for shortage
2119 * indicates that we must aggressively deactivate pages to avoid
2122 shortage = vm_pageout_active_target(vmd) + addl_shortage;
2123 vm_pageout_scan_active(vmd, shortage);
2128 * Initialize basic pageout daemon settings. See the comment above the
2129 * definition of vm_domain for some explanation of how these thresholds are
2133 vm_pageout_init_domain(int domain)
2135 struct vm_domain *vmd;
2136 struct sysctl_oid *oid;
2138 vmd = VM_DOMAIN(domain);
2139 vmd->vmd_interrupt_free_min = 2;
2142 * v_free_reserved needs to include enough for the largest
2143 * swap pager structures plus enough for any pv_entry structs
2146 vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE +
2147 vmd->vmd_interrupt_free_min;
2148 vmd->vmd_free_reserved = vm_pageout_page_count +
2149 vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768;
2150 vmd->vmd_free_min = vmd->vmd_page_count / 200;
2151 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2152 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2153 vmd->vmd_free_min += vmd->vmd_free_reserved;
2154 vmd->vmd_free_severe += vmd->vmd_free_reserved;
2155 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2156 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2157 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2160 * Set the default wakeup threshold to be 10% below the paging
2161 * target. This keeps the steady state out of shortfall.
2163 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2166 * Target amount of memory to move out of the laundry queue during a
2167 * background laundering. This is proportional to the amount of system
2170 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2171 vmd->vmd_free_min) / 10;
2173 /* Initialize the pageout daemon pid controller. */
2174 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2175 vmd->vmd_free_target, PIDCTRL_BOUND,
2176 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2177 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2178 "pidctrl", CTLFLAG_RD, NULL, "");
2179 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2183 vm_pageout_init(void)
2189 * Initialize some paging parameters.
2191 if (vm_cnt.v_page_count < 2000)
2192 vm_pageout_page_count = 8;
2195 for (i = 0; i < vm_ndomains; i++) {
2196 struct vm_domain *vmd;
2198 vm_pageout_init_domain(i);
2200 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2201 vm_cnt.v_free_target += vmd->vmd_free_target;
2202 vm_cnt.v_free_min += vmd->vmd_free_min;
2203 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2204 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2205 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2206 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2207 freecount += vmd->vmd_free_count;
2211 * Set interval in seconds for active scan. We want to visit each
2212 * page at least once every ten minutes. This is to prevent worst
2213 * case paging behaviors with stale active LRU.
2215 if (vm_pageout_update_period == 0)
2216 vm_pageout_update_period = 600;
2218 if (vm_page_max_user_wired == 0)
2219 vm_page_max_user_wired = freecount / 3;
2223 * vm_pageout is the high level pageout daemon.
2230 int error, first, i;
2235 mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2236 swap_pager_swap_init();
2237 for (first = -1, i = 0; i < vm_ndomains; i++) {
2238 if (VM_DOMAIN_EMPTY(i)) {
2240 printf("domain %d empty; skipping pageout\n",
2247 error = kthread_add(vm_pageout_worker,
2248 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2250 panic("starting pageout for domain %d: %d\n",
2253 error = kthread_add(vm_pageout_laundry_worker,
2254 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2256 panic("starting laundry for domain %d: %d", i, error);
2258 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2260 panic("starting uma_reclaim helper, error %d\n", error);
2262 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2263 vm_pageout_worker((void *)(uintptr_t)first);
2267 * Perform an advisory wakeup of the page daemon.
2270 pagedaemon_wakeup(int domain)
2272 struct vm_domain *vmd;
2274 vmd = VM_DOMAIN(domain);
2275 vm_domain_pageout_assert_unlocked(vmd);
2276 if (curproc == pageproc)
2279 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2280 vm_domain_pageout_lock(vmd);
2281 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2282 wakeup(&vmd->vmd_pageout_wanted);
2283 vm_domain_pageout_unlock(vmd);