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 * Scan for pages at adjacent offsets within the given page's object that are
322 * eligible for laundering, form a cluster of these pages and the given page,
323 * and launder that cluster.
326 vm_pageout_cluster(vm_page_t m)
329 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
331 int ib, is, page_base, pageout_count;
334 VM_OBJECT_ASSERT_WLOCKED(object);
337 vm_page_assert_xbusied(m);
339 mc[vm_pageout_page_count] = pb = ps = m;
341 page_base = vm_pageout_page_count;
346 * We can cluster only if the page is not clean, busy, or held, and
347 * the page is in the laundry queue.
349 * During heavy mmap/modification loads the pageout
350 * daemon can really fragment the underlying file
351 * due to flushing pages out of order and not trying to
352 * align the clusters (which leaves sporadic out-of-order
353 * holes). To solve this problem we do the reverse scan
354 * first and attempt to align our cluster, then do a
355 * forward scan if room remains.
358 while (ib != 0 && pageout_count < vm_pageout_page_count) {
363 if ((p = vm_page_prev(pb)) == NULL ||
364 vm_page_tryxbusy(p) == 0) {
368 if (vm_page_wired(p)) {
373 vm_page_test_dirty(p);
380 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
387 mc[--page_base] = pb = p;
392 * We are at an alignment boundary. Stop here, and switch
393 * directions. Do not clear ib.
395 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
398 while (pageout_count < vm_pageout_page_count &&
399 pindex + is < object->size) {
400 if ((p = vm_page_next(ps)) == NULL ||
401 vm_page_tryxbusy(p) == 0)
403 if (vm_page_wired(p)) {
407 vm_page_test_dirty(p);
413 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
419 mc[page_base + pageout_count] = ps = p;
425 * If we exhausted our forward scan, continue with the reverse scan
426 * when possible, even past an alignment boundary. This catches
427 * boundary conditions.
429 if (ib != 0 && pageout_count < vm_pageout_page_count)
432 return (vm_pageout_flush(&mc[page_base], pageout_count,
433 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
437 * vm_pageout_flush() - launder the given pages
439 * The given pages are laundered. Note that we setup for the start of
440 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
441 * reference count all in here rather then in the parent. If we want
442 * the parent to do more sophisticated things we may have to change
445 * Returned runlen is the count of pages between mreq and first
446 * page after mreq with status VM_PAGER_AGAIN.
447 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
448 * for any page in runlen set.
451 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
454 vm_object_t object = mc[0]->object;
455 int pageout_status[count];
459 VM_OBJECT_ASSERT_WLOCKED(object);
462 * Initiate I/O. Mark the pages shared busy and verify that they're
463 * valid and read-only.
465 * We do not have to fixup the clean/dirty bits here... we can
466 * allow the pager to do it after the I/O completes.
468 * NOTE! mc[i]->dirty may be partial or fragmented due to an
469 * edge case with file fragments.
471 for (i = 0; i < count; i++) {
472 KASSERT(vm_page_all_valid(mc[i]),
473 ("vm_pageout_flush: partially invalid page %p index %d/%d",
475 KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
476 ("vm_pageout_flush: writeable page %p", mc[i]));
477 vm_page_busy_downgrade(mc[i]);
479 vm_object_pip_add(object, count);
481 vm_pager_put_pages(object, mc, count, flags, pageout_status);
483 runlen = count - mreq;
486 for (i = 0; i < count; i++) {
487 vm_page_t mt = mc[i];
489 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
490 !pmap_page_is_write_mapped(mt),
491 ("vm_pageout_flush: page %p is not write protected", mt));
492 switch (pageout_status[i]) {
495 if (vm_page_in_laundry(mt))
496 vm_page_deactivate_noreuse(mt);
504 * The page is outside the object's range. We pretend
505 * that the page out worked and clean the page, so the
506 * changes will be lost if the page is reclaimed by
511 if (vm_page_in_laundry(mt))
512 vm_page_deactivate_noreuse(mt);
518 * If the page couldn't be paged out to swap because the
519 * pager wasn't able to find space, place the page in
520 * the PQ_UNSWAPPABLE holding queue. This is an
521 * optimization that prevents the page daemon from
522 * wasting CPU cycles on pages that cannot be reclaimed
523 * becase no swap device is configured.
525 * Otherwise, reactivate the page so that it doesn't
526 * clog the laundry and inactive queues. (We will try
527 * paging it out again later.)
530 if (object->type == OBJT_SWAP &&
531 pageout_status[i] == VM_PAGER_FAIL) {
532 vm_page_unswappable(mt);
535 vm_page_activate(mt);
537 if (eio != NULL && i >= mreq && i - mreq < runlen)
541 if (i >= mreq && i - mreq < runlen)
547 * If the operation is still going, leave the page busy to
548 * block all other accesses. Also, leave the paging in
549 * progress indicator set so that we don't attempt an object
552 if (pageout_status[i] != VM_PAGER_PEND) {
553 vm_object_pip_wakeup(object);
559 return (numpagedout);
563 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
566 atomic_store_rel_int(&swapdev_enabled, 1);
570 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
573 if (swap_pager_nswapdev() == 1)
574 atomic_store_rel_int(&swapdev_enabled, 0);
578 * Attempt to acquire all of the necessary locks to launder a page and
579 * then call through the clustering layer to PUTPAGES. Wait a short
580 * time for a vnode lock.
582 * Requires the page and object lock on entry, releases both before return.
583 * Returns 0 on success and an errno otherwise.
586 vm_pageout_clean(vm_page_t m, int *numpagedout)
594 vm_page_assert_locked(m);
596 VM_OBJECT_ASSERT_WLOCKED(object);
602 * The object is already known NOT to be dead. It
603 * is possible for the vget() to block the whole
604 * pageout daemon, but the new low-memory handling
605 * code should prevent it.
607 * We can't wait forever for the vnode lock, we might
608 * deadlock due to a vn_read() getting stuck in
609 * vm_wait while holding this vnode. We skip the
610 * vnode if we can't get it in a reasonable amount
613 if (object->type == OBJT_VNODE) {
617 if (vp->v_type == VREG &&
618 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
624 ("vp %p with NULL v_mount", vp));
625 vm_object_reference_locked(object);
627 VM_OBJECT_WUNLOCK(object);
628 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
629 LK_SHARED : LK_EXCLUSIVE;
630 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
635 VM_OBJECT_WLOCK(object);
638 * Ensure that the object and vnode were not disassociated
639 * while locks were dropped.
641 if (vp->v_object != object) {
648 * While the object and page were unlocked, the page
650 * (1) moved to a different queue,
651 * (2) reallocated to a different object,
652 * (3) reallocated to a different offset, or
655 if (!vm_page_in_laundry(m) || m->object != object ||
656 m->pindex != pindex || m->dirty == 0) {
663 * The page may have been busied while the object and page
664 * locks were released.
666 if (vm_page_tryxbusy(m) == 0) {
674 * Remove all writeable mappings, failing if the page is wired.
676 if (!vm_page_try_remove_write(m)) {
685 * If a page is dirty, then it is either being washed
686 * (but not yet cleaned) or it is still in the
687 * laundry. If it is still in the laundry, then we
688 * start the cleaning operation.
690 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
694 VM_OBJECT_WUNLOCK(object);
697 vm_page_lock_assert(m, MA_NOTOWNED);
701 vm_object_deallocate(object);
702 vn_finished_write(mp);
709 * Attempt to launder the specified number of pages.
711 * Returns the number of pages successfully laundered.
714 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
716 struct scan_state ss;
717 struct vm_pagequeue *pq;
721 vm_page_astate_t new, old;
722 int act_delta, error, numpagedout, queue, refs, starting_target;
728 starting_target = launder;
732 * Scan the laundry queues for pages eligible to be laundered. We stop
733 * once the target number of dirty pages have been laundered, or once
734 * we've reached the end of the queue. A single iteration of this loop
735 * may cause more than one page to be laundered because of clustering.
737 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
738 * swap devices are configured.
740 if (atomic_load_acq_int(&swapdev_enabled))
741 queue = PQ_UNSWAPPABLE;
746 marker = &vmd->vmd_markers[queue];
747 pq = &vmd->vmd_pagequeues[queue];
748 vm_pagequeue_lock(pq);
749 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
750 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
751 if (__predict_false((m->flags & PG_MARKER) != 0))
754 vm_page_change_lock(m, &mtx);
758 * The page may have been disassociated from the queue
759 * or even freed while locks were dropped. We thus must be
760 * careful whenever modifying page state. Once the object lock
761 * has been acquired, we have a stable reference to the page.
763 if (vm_page_queue(m) != queue)
767 * A requeue was requested, so this page gets a second
770 if ((m->a.flags & PGA_REQUEUE) != 0) {
771 vm_page_pqbatch_submit(m, queue);
776 * Wired pages may not be freed. Complete their removal
777 * from the queue now to avoid needless revisits during
778 * future scans. This check is racy and must be reverified once
779 * we hold the object lock and have verified that the page
782 if (vm_page_wired(m)) {
783 vm_page_dequeue_deferred(m);
787 if (object != m->object) {
789 VM_OBJECT_WUNLOCK(object);
792 * A page's object pointer may be set to NULL before
793 * the object lock is acquired.
795 object = (vm_object_t)atomic_load_ptr(&m->object);
796 if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) {
798 /* Depends on type-stability. */
799 VM_OBJECT_WLOCK(object);
804 if (__predict_false(m->object == NULL))
806 * The page has been removed from its object.
809 KASSERT(m->object == object, ("page %p does not belong to %p",
812 if (vm_page_tryxbusy(m) == 0)
816 * Re-check for wirings now that we hold the object lock and
817 * have verified that the page is unbusied. If the page is
818 * mapped, it may still be wired by pmap lookups. The call to
819 * vm_page_try_remove_all() below atomically checks for such
820 * wirings and removes mappings. If the page is unmapped, the
821 * wire count is guaranteed not to increase.
823 if (__predict_false(vm_page_wired(m))) {
824 vm_page_dequeue_deferred(m);
829 * Invalid pages can be easily freed. They cannot be
830 * mapped; vm_page_free() asserts this.
832 if (vm_page_none_valid(m))
835 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
837 for (old = vm_page_astate_load(m);;) {
839 * Check to see if the page has been removed from the
840 * queue since the first such check. Leave it alone if
841 * so, discarding any references collected by
842 * pmap_ts_referenced().
844 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
849 if ((old.flags & PGA_REFERENCED) != 0) {
850 new.flags &= ~PGA_REFERENCED;
853 if (act_delta == 0) {
855 } else if (object->ref_count != 0) {
857 * Increase the activation count if the page was
858 * referenced while in the laundry queue. This
859 * makes it less likely that the page will be
860 * returned prematurely to the laundry queue.
862 new.act_count += ACT_ADVANCE +
864 if (new.act_count > ACT_MAX)
865 new.act_count = ACT_MAX;
867 new.flags |= PGA_REQUEUE;
868 new.queue = PQ_ACTIVE;
869 if (!vm_page_pqstate_commit(m, &old, new))
873 * If this was a background laundering, count
874 * activated pages towards our target. The
875 * purpose of background laundering is to ensure
876 * that pages are eventually cycled through the
877 * laundry queue, and an activation is a valid
882 VM_CNT_INC(v_reactivated);
884 } else if ((object->flags & OBJ_DEAD) == 0) {
885 new.flags |= PGA_REQUEUE;
886 if (!vm_page_pqstate_commit(m, &old, new))
894 * If the page appears to be clean at the machine-independent
895 * layer, then remove all of its mappings from the pmap in
896 * anticipation of freeing it. If, however, any of the page's
897 * mappings allow write access, then the page may still be
898 * modified until the last of those mappings are removed.
900 if (object->ref_count != 0) {
901 vm_page_test_dirty(m);
902 if (m->dirty == 0 && !vm_page_try_remove_all(m)) {
903 vm_page_dequeue_deferred(m);
909 * Clean pages are freed, and dirty pages are paged out unless
910 * they belong to a dead object. Requeueing dirty pages from
911 * dead objects is pointless, as they are being paged out and
912 * freed by the thread that destroyed the object.
918 } else if ((object->flags & OBJ_DEAD) == 0) {
919 if (object->type != OBJT_SWAP &&
920 object->type != OBJT_DEFAULT)
922 else if (disable_swap_pageouts)
932 * Form a cluster with adjacent, dirty pages from the
933 * same object, and page out that entire cluster.
935 * The adjacent, dirty pages must also be in the
936 * laundry. However, their mappings are not checked
937 * for new references. Consequently, a recently
938 * referenced page may be paged out. However, that
939 * page will not be prematurely reclaimed. After page
940 * out, the page will be placed in the inactive queue,
941 * where any new references will be detected and the
944 error = vm_pageout_clean(m, &numpagedout);
946 launder -= numpagedout;
947 ss.scanned += numpagedout;
948 } else if (error == EDEADLK) {
963 if (object != NULL) {
964 VM_OBJECT_WUNLOCK(object);
967 vm_pagequeue_lock(pq);
968 vm_pageout_end_scan(&ss);
969 vm_pagequeue_unlock(pq);
971 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
977 * Wakeup the sync daemon if we skipped a vnode in a writeable object
978 * and we didn't launder enough pages.
980 if (vnodes_skipped > 0 && launder > 0)
981 (void)speedup_syncer();
983 return (starting_target - launder);
987 * Compute the integer square root.
992 u_int bit, root, tmp;
994 bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
1009 * Perform the work of the laundry thread: periodically wake up and determine
1010 * whether any pages need to be laundered. If so, determine the number of pages
1011 * that need to be laundered, and launder them.
1014 vm_pageout_laundry_worker(void *arg)
1016 struct vm_domain *vmd;
1017 struct vm_pagequeue *pq;
1018 uint64_t nclean, ndirty, nfreed;
1019 int domain, last_target, launder, shortfall, shortfall_cycle, target;
1022 domain = (uintptr_t)arg;
1023 vmd = VM_DOMAIN(domain);
1024 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1025 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1028 in_shortfall = false;
1029 shortfall_cycle = 0;
1030 last_target = target = 0;
1034 * Calls to these handlers are serialized by the swap syscall lock.
1036 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
1037 EVENTHANDLER_PRI_ANY);
1038 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
1039 EVENTHANDLER_PRI_ANY);
1042 * The pageout laundry worker is never done, so loop forever.
1045 KASSERT(target >= 0, ("negative target %d", target));
1046 KASSERT(shortfall_cycle >= 0,
1047 ("negative cycle %d", shortfall_cycle));
1051 * First determine whether we need to launder pages to meet a
1052 * shortage of free pages.
1054 if (shortfall > 0) {
1055 in_shortfall = true;
1056 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1058 } else if (!in_shortfall)
1060 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1062 * We recently entered shortfall and began laundering
1063 * pages. If we have completed that laundering run
1064 * (and we are no longer in shortfall) or we have met
1065 * our laundry target through other activity, then we
1066 * can stop laundering pages.
1068 in_shortfall = false;
1072 launder = target / shortfall_cycle--;
1076 * There's no immediate need to launder any pages; see if we
1077 * meet the conditions to perform background laundering:
1079 * 1. The ratio of dirty to clean inactive pages exceeds the
1080 * background laundering threshold, or
1081 * 2. we haven't yet reached the target of the current
1082 * background laundering run.
1084 * The background laundering threshold is not a constant.
1085 * Instead, it is a slowly growing function of the number of
1086 * clean pages freed by the page daemon since the last
1087 * background laundering. Thus, as the ratio of dirty to
1088 * clean inactive pages grows, the amount of memory pressure
1089 * required to trigger laundering decreases. We ensure
1090 * that the threshold is non-zero after an inactive queue
1091 * scan, even if that scan failed to free a single clean page.
1094 nclean = vmd->vmd_free_count +
1095 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1096 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1097 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1098 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1099 target = vmd->vmd_background_launder_target;
1103 * We have a non-zero background laundering target. If we've
1104 * laundered up to our maximum without observing a page daemon
1105 * request, just stop. This is a safety belt that ensures we
1106 * don't launder an excessive amount if memory pressure is low
1107 * and the ratio of dirty to clean pages is large. Otherwise,
1108 * proceed at the background laundering rate.
1113 last_target = target;
1114 } else if (last_target - target >=
1115 vm_background_launder_max * PAGE_SIZE / 1024) {
1118 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1119 launder /= VM_LAUNDER_RATE;
1120 if (launder > target)
1127 * Because of I/O clustering, the number of laundered
1128 * pages could exceed "target" by the maximum size of
1129 * a cluster minus one.
1131 target -= min(vm_pageout_launder(vmd, launder,
1132 in_shortfall), target);
1133 pause("laundp", hz / VM_LAUNDER_RATE);
1137 * If we're not currently laundering pages and the page daemon
1138 * hasn't posted a new request, sleep until the page daemon
1141 vm_pagequeue_lock(pq);
1142 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1143 (void)mtx_sleep(&vmd->vmd_laundry_request,
1144 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1147 * If the pagedaemon has indicated that it's in shortfall, start
1148 * a shortfall laundering unless we're already in the middle of
1149 * one. This may preempt a background laundering.
1151 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1152 (!in_shortfall || shortfall_cycle == 0)) {
1153 shortfall = vm_laundry_target(vmd) +
1154 vmd->vmd_pageout_deficit;
1160 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1161 nfreed += vmd->vmd_clean_pages_freed;
1162 vmd->vmd_clean_pages_freed = 0;
1163 vm_pagequeue_unlock(pq);
1168 * Compute the number of pages we want to try to move from the
1169 * active queue to either the inactive or laundry queue.
1171 * When scanning active pages during a shortage, we make clean pages
1172 * count more heavily towards the page shortage than dirty pages.
1173 * This is because dirty pages must be laundered before they can be
1174 * reused and thus have less utility when attempting to quickly
1175 * alleviate a free page shortage. However, this weighting also
1176 * causes the scan to deactivate dirty pages more aggressively,
1177 * improving the effectiveness of clustering.
1180 vm_pageout_active_target(struct vm_domain *vmd)
1184 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1185 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1186 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1187 shortage *= act_scan_laundry_weight;
1192 * Scan the active queue. If there is no shortage of inactive pages, scan a
1193 * small portion of the queue in order to maintain quasi-LRU.
1196 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1198 struct scan_state ss;
1201 vm_page_t m, marker;
1202 struct vm_pagequeue *pq;
1203 vm_page_astate_t old, new;
1205 int act_delta, max_scan, ps_delta, refs, scan_tick;
1208 marker = &vmd->vmd_markers[PQ_ACTIVE];
1209 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1210 vm_pagequeue_lock(pq);
1213 * If we're just idle polling attempt to visit every
1214 * active page within 'update_period' seconds.
1217 if (vm_pageout_update_period != 0) {
1218 min_scan = pq->pq_cnt;
1219 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1220 min_scan /= hz * vm_pageout_update_period;
1223 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1224 vmd->vmd_last_active_scan = scan_tick;
1227 * Scan the active queue for pages that can be deactivated. Update
1228 * the per-page activity counter and use it to identify deactivation
1229 * candidates. Held pages may be deactivated.
1231 * To avoid requeuing each page that remains in the active queue, we
1232 * implement the CLOCK algorithm. To keep the implementation of the
1233 * enqueue operation consistent for all page queues, we use two hands,
1234 * represented by marker pages. Scans begin at the first hand, which
1235 * precedes the second hand in the queue. When the two hands meet,
1236 * they are moved back to the head and tail of the queue, respectively,
1237 * and scanning resumes.
1239 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1242 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1243 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1244 if (__predict_false(m == &vmd->vmd_clock[1])) {
1245 vm_pagequeue_lock(pq);
1246 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1247 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1248 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1250 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1252 max_scan -= ss.scanned;
1253 vm_pageout_end_scan(&ss);
1256 if (__predict_false((m->flags & PG_MARKER) != 0))
1259 vm_page_change_lock(m, &mtx);
1262 * The page may have been disassociated from the queue
1263 * or even freed while locks were dropped. We thus must be
1264 * careful whenever modifying page state. Once the object lock
1265 * has been acquired, we have a stable reference to the page.
1267 if (vm_page_queue(m) != PQ_ACTIVE)
1271 * Wired pages are dequeued lazily.
1273 if (vm_page_wired(m)) {
1274 vm_page_dequeue_deferred(m);
1279 * A page's object pointer may be set to NULL before
1280 * the object lock is acquired.
1282 object = (vm_object_t)atomic_load_ptr(&m->object);
1283 if (__predict_false(object == NULL))
1285 * The page has been removed from its object.
1289 /* Deferred free of swap space. */
1290 if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
1291 VM_OBJECT_TRYWLOCK(object)) {
1292 if (m->object == object)
1293 vm_pager_page_unswapped(m);
1294 VM_OBJECT_WUNLOCK(object);
1298 * Check to see "how much" the page has been used.
1300 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1301 * that a reference from a concurrently destroyed mapping is
1302 * observed here and now.
1304 * Perform an unsynchronized object ref count check. While
1305 * the page lock ensures that the page is not reallocated to
1306 * another object, in particular, one with unmanaged mappings
1307 * that cannot support pmap_ts_referenced(), two races are,
1308 * nonetheless, possible:
1309 * 1) The count was transitioning to zero, but we saw a non-
1310 * zero value. pmap_ts_referenced() will return zero
1311 * because the page is not mapped.
1312 * 2) The count was transitioning to one, but we saw zero.
1313 * This race delays the detection of a new reference. At
1314 * worst, we will deactivate and reactivate the page.
1316 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1318 old = vm_page_astate_load(m);
1321 * Check to see if the page has been removed from the
1322 * queue since the first such check. Leave it alone if
1323 * so, discarding any references collected by
1324 * pmap_ts_referenced().
1326 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1330 * Advance or decay the act_count based on recent usage.
1334 if ((old.flags & PGA_REFERENCED) != 0) {
1335 new.flags &= ~PGA_REFERENCED;
1338 if (act_delta != 0) {
1339 new.act_count += ACT_ADVANCE + act_delta;
1340 if (new.act_count > ACT_MAX)
1341 new.act_count = ACT_MAX;
1343 new.act_count -= min(new.act_count,
1347 if (new.act_count > 0) {
1349 * Adjust the activation count and keep the page
1350 * in the active queue. The count might be left
1351 * unchanged if it is saturated. The page may
1352 * have been moved to a different queue since we
1353 * started the scan, in which case we move it
1357 if (old.queue != PQ_ACTIVE) {
1358 old.queue = PQ_ACTIVE;
1359 old.flags |= PGA_REQUEUE;
1363 * When not short for inactive pages, let dirty
1364 * pages go through the inactive queue before
1365 * moving to the laundry queue. This gives them
1366 * some extra time to be reactivated,
1367 * potentially avoiding an expensive pageout.
1368 * However, during a page shortage, the inactive
1369 * queue is necessarily small, and so dirty
1370 * pages would only spend a trivial amount of
1371 * time in the inactive queue. Therefore, we
1372 * might as well place them directly in the
1373 * laundry queue to reduce queuing overhead.
1375 * Calling vm_page_test_dirty() here would
1376 * require acquisition of the object's write
1377 * lock. However, during a page shortage,
1378 * directing dirty pages into the laundry queue
1379 * is only an optimization and not a
1380 * requirement. Therefore, we simply rely on
1381 * the opportunistic updates to the page's dirty
1382 * field by the pmap.
1384 if (page_shortage <= 0) {
1385 nqueue = PQ_INACTIVE;
1387 } else if (m->dirty == 0) {
1388 nqueue = PQ_INACTIVE;
1389 ps_delta = act_scan_laundry_weight;
1391 nqueue = PQ_LAUNDRY;
1395 new.flags |= PGA_REQUEUE;
1398 } while (!vm_page_pqstate_commit(m, &old, new));
1400 page_shortage -= ps_delta;
1406 vm_pagequeue_lock(pq);
1407 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1408 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1409 vm_pageout_end_scan(&ss);
1410 vm_pagequeue_unlock(pq);
1414 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker,
1417 vm_page_astate_t as;
1419 vm_pagequeue_assert_locked(pq);
1421 as = vm_page_astate_load(m);
1422 if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
1424 vm_page_aflag_set(m, PGA_ENQUEUED);
1425 TAILQ_INSERT_BEFORE(marker, m, plinks.q);
1430 * Re-add stuck pages to the inactive queue. We will examine them again
1431 * during the next scan. If the queue state of a page has changed since
1432 * it was physically removed from the page queue in
1433 * vm_pageout_collect_batch(), don't do anything with that page.
1436 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1439 struct vm_pagequeue *pq;
1444 marker = ss->marker;
1448 if (vm_batchqueue_insert(bq, m))
1450 vm_pagequeue_lock(pq);
1451 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1453 vm_pagequeue_lock(pq);
1454 while ((m = vm_batchqueue_pop(bq)) != NULL)
1455 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1456 vm_pagequeue_cnt_add(pq, delta);
1457 vm_pagequeue_unlock(pq);
1458 vm_batchqueue_init(bq);
1462 * Attempt to reclaim the requested number of pages from the inactive queue.
1463 * Returns true if the shortage was addressed.
1466 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1469 struct scan_state ss;
1470 struct vm_batchqueue rq;
1472 vm_page_t m, marker;
1473 struct vm_pagequeue *pq;
1475 vm_page_astate_t old, new;
1476 int act_delta, addl_page_shortage, deficit, page_shortage, refs;
1477 int starting_page_shortage;
1480 * The addl_page_shortage is an estimate of the number of temporarily
1481 * stuck pages in the inactive queue. In other words, the
1482 * number of pages from the inactive count that should be
1483 * discounted in setting the target for the active queue scan.
1485 addl_page_shortage = 0;
1488 * vmd_pageout_deficit counts the number of pages requested in
1489 * allocations that failed because of a free page shortage. We assume
1490 * that the allocations will be reattempted and thus include the deficit
1491 * in our scan target.
1493 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1494 starting_page_shortage = page_shortage = shortage + deficit;
1498 vm_batchqueue_init(&rq);
1501 * Start scanning the inactive queue for pages that we can free. The
1502 * scan will stop when we reach the target or we have scanned the
1503 * entire queue. (Note that m->a.act_count is not used to make
1504 * decisions for the inactive queue, only for the active queue.)
1506 marker = &vmd->vmd_markers[PQ_INACTIVE];
1507 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1508 vm_pagequeue_lock(pq);
1509 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1510 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1511 KASSERT((m->flags & PG_MARKER) == 0,
1512 ("marker page %p was dequeued", m));
1514 vm_page_change_lock(m, &mtx);
1518 * The page may have been disassociated from the queue
1519 * or even freed while locks were dropped. We thus must be
1520 * careful whenever modifying page state. Once the object lock
1521 * has been acquired, we have a stable reference to the page.
1523 old = vm_page_astate_load(m);
1524 if (old.queue != PQ_INACTIVE ||
1525 (old.flags & PGA_QUEUE_STATE_MASK) != 0)
1529 * Wired pages may not be freed. Complete their removal
1530 * from the queue now to avoid needless revisits during
1531 * future scans. This check is racy and must be reverified once
1532 * we hold the object lock and have verified that the page
1535 if (vm_page_wired(m)) {
1536 vm_page_dequeue_deferred(m);
1540 if (object != m->object) {
1542 VM_OBJECT_WUNLOCK(object);
1545 * A page's object pointer may be set to NULL before
1546 * the object lock is acquired.
1548 object = (vm_object_t)atomic_load_ptr(&m->object);
1549 if (object != NULL && !VM_OBJECT_TRYWLOCK(object)) {
1551 /* Depends on type-stability. */
1552 VM_OBJECT_WLOCK(object);
1557 if (__predict_false(m->object == NULL))
1559 * The page has been removed from its object.
1562 KASSERT(m->object == object, ("page %p does not belong to %p",
1565 if (vm_page_tryxbusy(m) == 0) {
1567 * Don't mess with busy pages. Leave them at
1568 * the front of the queue. Most likely, they
1569 * are being paged out and will leave the
1570 * queue shortly after the scan finishes. So,
1571 * they ought to be discounted from the
1574 addl_page_shortage++;
1578 /* Deferred free of swap space. */
1579 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1580 vm_pager_page_unswapped(m);
1583 * Re-check for wirings now that we hold the object lock and
1584 * have verified that the page is unbusied. If the page is
1585 * mapped, it may still be wired by pmap lookups. The call to
1586 * vm_page_try_remove_all() below atomically checks for such
1587 * wirings and removes mappings. If the page is unmapped, the
1588 * wire count is guaranteed not to increase.
1590 if (__predict_false(vm_page_wired(m))) {
1591 vm_page_dequeue_deferred(m);
1596 * Invalid pages can be easily freed. They cannot be
1597 * mapped, vm_page_free() asserts this.
1599 if (vm_page_none_valid(m))
1602 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1604 for (old = vm_page_astate_load(m);;) {
1606 * Check to see if the page has been removed from the
1607 * queue since the first such check. Leave it alone if
1608 * so, discarding any references collected by
1609 * pmap_ts_referenced().
1611 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1616 if ((old.flags & PGA_REFERENCED) != 0) {
1617 new.flags &= ~PGA_REFERENCED;
1620 if (act_delta == 0) {
1622 } else if (object->ref_count != 0) {
1624 * Increase the activation count if the
1625 * page was referenced while in the
1626 * inactive queue. This makes it less
1627 * likely that the page will be returned
1628 * prematurely to the inactive queue.
1630 new.act_count += ACT_ADVANCE +
1632 if (new.act_count > ACT_MAX)
1633 new.act_count = ACT_MAX;
1635 new.flags |= PGA_REQUEUE;
1636 new.queue = PQ_ACTIVE;
1637 if (!vm_page_pqstate_commit(m, &old, new))
1640 VM_CNT_INC(v_reactivated);
1642 } else if ((object->flags & OBJ_DEAD) == 0) {
1643 new.queue = PQ_INACTIVE;
1644 new.flags |= PGA_REQUEUE;
1645 if (!vm_page_pqstate_commit(m, &old, new))
1653 * If the page appears to be clean at the machine-independent
1654 * layer, then remove all of its mappings from the pmap in
1655 * anticipation of freeing it. If, however, any of the page's
1656 * mappings allow write access, then the page may still be
1657 * modified until the last of those mappings are removed.
1659 if (object->ref_count != 0) {
1660 vm_page_test_dirty(m);
1661 if (m->dirty == 0 && !vm_page_try_remove_all(m)) {
1662 vm_page_dequeue_deferred(m);
1668 * Clean pages can be freed, but dirty pages must be sent back
1669 * to the laundry, unless they belong to a dead object.
1670 * Requeueing dirty pages from dead objects is pointless, as
1671 * they are being paged out and freed by the thread that
1672 * destroyed the object.
1674 if (m->dirty == 0) {
1677 * Because we dequeued the page and have already
1678 * checked for concurrent dequeue and enqueue
1679 * requests, we can safely disassociate the page
1680 * from the inactive queue.
1682 KASSERT((m->a.flags & PGA_QUEUE_STATE_MASK) == 0,
1683 ("page %p has queue state", m));
1684 m->a.queue = PQ_NONE;
1689 if ((object->flags & OBJ_DEAD) == 0)
1695 vm_pageout_reinsert_inactive(&ss, &rq, m);
1700 VM_OBJECT_WUNLOCK(object);
1701 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1702 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1703 vm_pagequeue_lock(pq);
1704 vm_pageout_end_scan(&ss);
1705 vm_pagequeue_unlock(pq);
1707 VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1710 * Wake up the laundry thread so that it can perform any needed
1711 * laundering. If we didn't meet our target, we're in shortfall and
1712 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1713 * swap devices are configured, the laundry thread has no work to do, so
1714 * don't bother waking it up.
1716 * The laundry thread uses the number of inactive queue scans elapsed
1717 * since the last laundering to determine whether to launder again, so
1720 if (starting_page_shortage > 0) {
1721 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1722 vm_pagequeue_lock(pq);
1723 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1724 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1725 if (page_shortage > 0) {
1726 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1727 VM_CNT_INC(v_pdshortfalls);
1728 } else if (vmd->vmd_laundry_request !=
1729 VM_LAUNDRY_SHORTFALL)
1730 vmd->vmd_laundry_request =
1731 VM_LAUNDRY_BACKGROUND;
1732 wakeup(&vmd->vmd_laundry_request);
1734 vmd->vmd_clean_pages_freed +=
1735 starting_page_shortage - page_shortage;
1736 vm_pagequeue_unlock(pq);
1740 * Wakeup the swapout daemon if we didn't free the targeted number of
1743 if (page_shortage > 0)
1747 * If the inactive queue scan fails repeatedly to meet its
1748 * target, kill the largest process.
1750 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1753 * Reclaim pages by swapping out idle processes, if configured to do so.
1755 vm_swapout_run_idle();
1758 * See the description of addl_page_shortage above.
1760 *addl_shortage = addl_page_shortage + deficit;
1762 return (page_shortage <= 0);
1765 static int vm_pageout_oom_vote;
1768 * The pagedaemon threads randlomly select one to perform the
1769 * OOM. Trying to kill processes before all pagedaemons
1770 * failed to reach free target is premature.
1773 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1774 int starting_page_shortage)
1778 if (starting_page_shortage <= 0 || starting_page_shortage !=
1780 vmd->vmd_oom_seq = 0;
1783 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1785 vmd->vmd_oom = FALSE;
1786 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1792 * Do not follow the call sequence until OOM condition is
1795 vmd->vmd_oom_seq = 0;
1800 vmd->vmd_oom = TRUE;
1801 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1802 if (old_vote != vm_ndomains - 1)
1806 * The current pagedaemon thread is the last in the quorum to
1807 * start OOM. Initiate the selection and signaling of the
1810 vm_pageout_oom(VM_OOM_MEM);
1813 * After one round of OOM terror, recall our vote. On the
1814 * next pass, current pagedaemon would vote again if the low
1815 * memory condition is still there, due to vmd_oom being
1818 vmd->vmd_oom = FALSE;
1819 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1823 * The OOM killer is the page daemon's action of last resort when
1824 * memory allocation requests have been stalled for a prolonged period
1825 * of time because it cannot reclaim memory. This function computes
1826 * the approximate number of physical pages that could be reclaimed if
1827 * the specified address space is destroyed.
1829 * Private, anonymous memory owned by the address space is the
1830 * principal resource that we expect to recover after an OOM kill.
1831 * Since the physical pages mapped by the address space's COW entries
1832 * are typically shared pages, they are unlikely to be released and so
1833 * they are not counted.
1835 * To get to the point where the page daemon runs the OOM killer, its
1836 * efforts to write-back vnode-backed pages may have stalled. This
1837 * could be caused by a memory allocation deadlock in the write path
1838 * that might be resolved by an OOM kill. Therefore, physical pages
1839 * belonging to vnode-backed objects are counted, because they might
1840 * be freed without being written out first if the address space holds
1841 * the last reference to an unlinked vnode.
1843 * Similarly, physical pages belonging to OBJT_PHYS objects are
1844 * counted because the address space might hold the last reference to
1848 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1851 vm_map_entry_t entry;
1855 map = &vmspace->vm_map;
1856 KASSERT(!map->system_map, ("system map"));
1857 sx_assert(&map->lock, SA_LOCKED);
1859 VM_MAP_ENTRY_FOREACH(entry, map) {
1860 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1862 obj = entry->object.vm_object;
1865 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1866 obj->ref_count != 1)
1868 switch (obj->type) {
1873 res += obj->resident_page_count;
1880 static int vm_oom_ratelim_last;
1881 static int vm_oom_pf_secs = 10;
1882 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1884 static struct mtx vm_oom_ratelim_mtx;
1887 vm_pageout_oom(int shortage)
1889 struct proc *p, *bigproc;
1890 vm_offset_t size, bigsize;
1897 * For OOM requests originating from vm_fault(), there is a high
1898 * chance that a single large process faults simultaneously in
1899 * several threads. Also, on an active system running many
1900 * processes of middle-size, like buildworld, all of them
1901 * could fault almost simultaneously as well.
1903 * To avoid killing too many processes, rate-limit OOMs
1904 * initiated by vm_fault() time-outs on the waits for free
1907 mtx_lock(&vm_oom_ratelim_mtx);
1909 if (shortage == VM_OOM_MEM_PF &&
1910 (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1911 mtx_unlock(&vm_oom_ratelim_mtx);
1914 vm_oom_ratelim_last = now;
1915 mtx_unlock(&vm_oom_ratelim_mtx);
1918 * We keep the process bigproc locked once we find it to keep anyone
1919 * from messing with it; however, there is a possibility of
1920 * deadlock if process B is bigproc and one of its child processes
1921 * attempts to propagate a signal to B while we are waiting for A's
1922 * lock while walking this list. To avoid this, we don't block on
1923 * the process lock but just skip a process if it is already locked.
1927 sx_slock(&allproc_lock);
1928 FOREACH_PROC_IN_SYSTEM(p) {
1932 * If this is a system, protected or killed process, skip it.
1934 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1935 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1936 p->p_pid == 1 || P_KILLED(p) ||
1937 (p->p_pid < 48 && swap_pager_avail != 0)) {
1942 * If the process is in a non-running type state,
1943 * don't touch it. Check all the threads individually.
1946 FOREACH_THREAD_IN_PROC(p, td) {
1948 if (!TD_ON_RUNQ(td) &&
1949 !TD_IS_RUNNING(td) &&
1950 !TD_IS_SLEEPING(td) &&
1951 !TD_IS_SUSPENDED(td) &&
1952 !TD_IS_SWAPPED(td)) {
1964 * get the process size
1966 vm = vmspace_acquire_ref(p);
1973 sx_sunlock(&allproc_lock);
1974 if (!vm_map_trylock_read(&vm->vm_map)) {
1976 sx_slock(&allproc_lock);
1980 size = vmspace_swap_count(vm);
1981 if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
1982 size += vm_pageout_oom_pagecount(vm);
1983 vm_map_unlock_read(&vm->vm_map);
1985 sx_slock(&allproc_lock);
1988 * If this process is bigger than the biggest one,
1991 if (size > bigsize) {
1992 if (bigproc != NULL)
2000 sx_sunlock(&allproc_lock);
2001 if (bigproc != NULL) {
2002 if (vm_panic_on_oom != 0)
2003 panic("out of swap space");
2005 killproc(bigproc, "out of swap space");
2006 sched_nice(bigproc, PRIO_MIN);
2008 PROC_UNLOCK(bigproc);
2013 * Signal a free page shortage to subsystems that have registered an event
2014 * handler. Reclaim memory from UMA in the event of a severe shortage.
2015 * Return true if the free page count should be re-evaluated.
2018 vm_pageout_lowmem(void)
2020 static int lowmem_ticks = 0;
2026 last = atomic_load_int(&lowmem_ticks);
2027 while ((u_int)(ticks - last) / hz >= lowmem_period) {
2028 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
2032 * Decrease registered cache sizes.
2034 SDT_PROBE0(vm, , , vm__lowmem_scan);
2035 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
2038 * We do this explicitly after the caches have been
2041 uma_reclaim(UMA_RECLAIM_TRIM);
2046 * Kick off an asynchronous reclaim of cached memory if one of the
2047 * page daemons is failing to keep up with demand. Use the "severe"
2048 * threshold instead of "min" to ensure that we do not blow away the
2049 * caches if a subset of the NUMA domains are depleted by kernel memory
2050 * allocations; the domainset iterators automatically skip domains
2051 * below the "min" threshold on the first pass.
2053 * UMA reclaim worker has its own rate-limiting mechanism, so don't
2054 * worry about kicking it too often.
2056 if (vm_page_count_severe())
2057 uma_reclaim_wakeup();
2063 vm_pageout_worker(void *arg)
2065 struct vm_domain *vmd;
2067 int addl_shortage, domain, shortage;
2070 domain = (uintptr_t)arg;
2071 vmd = VM_DOMAIN(domain);
2076 * XXXKIB It could be useful to bind pageout daemon threads to
2077 * the cores belonging to the domain, from which vm_page_array
2081 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
2082 vmd->vmd_last_active_scan = ticks;
2085 * The pageout daemon worker is never done, so loop forever.
2088 vm_domain_pageout_lock(vmd);
2091 * We need to clear wanted before we check the limits. This
2092 * prevents races with wakers who will check wanted after they
2095 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
2098 * Might the page daemon need to run again?
2100 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
2102 * Yes. If the scan failed to produce enough free
2103 * pages, sleep uninterruptibly for some time in the
2104 * hope that the laundry thread will clean some pages.
2106 vm_domain_pageout_unlock(vmd);
2108 pause("pwait", hz / VM_INACT_SCAN_RATE);
2111 * No, sleep until the next wakeup or until pages
2112 * need to have their reference stats updated.
2114 if (mtx_sleep(&vmd->vmd_pageout_wanted,
2115 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
2116 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
2117 VM_CNT_INC(v_pdwakeups);
2120 /* Prevent spurious wakeups by ensuring that wanted is set. */
2121 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2124 * Use the controller to calculate how many pages to free in
2125 * this interval, and scan the inactive queue. If the lowmem
2126 * handlers appear to have freed up some pages, subtract the
2127 * difference from the inactive queue scan target.
2129 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
2131 ofree = vmd->vmd_free_count;
2132 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
2133 shortage -= min(vmd->vmd_free_count - ofree,
2135 target_met = vm_pageout_scan_inactive(vmd, shortage,
2141 * Scan the active queue. A positive value for shortage
2142 * indicates that we must aggressively deactivate pages to avoid
2145 shortage = vm_pageout_active_target(vmd) + addl_shortage;
2146 vm_pageout_scan_active(vmd, shortage);
2151 * Initialize basic pageout daemon settings. See the comment above the
2152 * definition of vm_domain for some explanation of how these thresholds are
2156 vm_pageout_init_domain(int domain)
2158 struct vm_domain *vmd;
2159 struct sysctl_oid *oid;
2161 vmd = VM_DOMAIN(domain);
2162 vmd->vmd_interrupt_free_min = 2;
2165 * v_free_reserved needs to include enough for the largest
2166 * swap pager structures plus enough for any pv_entry structs
2169 vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE +
2170 vmd->vmd_interrupt_free_min;
2171 vmd->vmd_free_reserved = vm_pageout_page_count +
2172 vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768;
2173 vmd->vmd_free_min = vmd->vmd_page_count / 200;
2174 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2175 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2176 vmd->vmd_free_min += vmd->vmd_free_reserved;
2177 vmd->vmd_free_severe += vmd->vmd_free_reserved;
2178 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2179 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2180 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2183 * Set the default wakeup threshold to be 10% below the paging
2184 * target. This keeps the steady state out of shortfall.
2186 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2189 * Target amount of memory to move out of the laundry queue during a
2190 * background laundering. This is proportional to the amount of system
2193 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2194 vmd->vmd_free_min) / 10;
2196 /* Initialize the pageout daemon pid controller. */
2197 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2198 vmd->vmd_free_target, PIDCTRL_BOUND,
2199 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2200 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2201 "pidctrl", CTLFLAG_RD, NULL, "");
2202 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2206 vm_pageout_init(void)
2212 * Initialize some paging parameters.
2214 if (vm_cnt.v_page_count < 2000)
2215 vm_pageout_page_count = 8;
2218 for (i = 0; i < vm_ndomains; i++) {
2219 struct vm_domain *vmd;
2221 vm_pageout_init_domain(i);
2223 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2224 vm_cnt.v_free_target += vmd->vmd_free_target;
2225 vm_cnt.v_free_min += vmd->vmd_free_min;
2226 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2227 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2228 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2229 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2230 freecount += vmd->vmd_free_count;
2234 * Set interval in seconds for active scan. We want to visit each
2235 * page at least once every ten minutes. This is to prevent worst
2236 * case paging behaviors with stale active LRU.
2238 if (vm_pageout_update_period == 0)
2239 vm_pageout_update_period = 600;
2241 if (vm_page_max_user_wired == 0)
2242 vm_page_max_user_wired = freecount / 3;
2246 * vm_pageout is the high level pageout daemon.
2253 int error, first, i;
2258 mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2259 swap_pager_swap_init();
2260 for (first = -1, i = 0; i < vm_ndomains; i++) {
2261 if (VM_DOMAIN_EMPTY(i)) {
2263 printf("domain %d empty; skipping pageout\n",
2270 error = kthread_add(vm_pageout_worker,
2271 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2273 panic("starting pageout for domain %d: %d\n",
2276 error = kthread_add(vm_pageout_laundry_worker,
2277 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2279 panic("starting laundry for domain %d: %d", i, error);
2281 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2283 panic("starting uma_reclaim helper, error %d\n", error);
2285 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2286 vm_pageout_worker((void *)(uintptr_t)first);
2290 * Perform an advisory wakeup of the page daemon.
2293 pagedaemon_wakeup(int domain)
2295 struct vm_domain *vmd;
2297 vmd = VM_DOMAIN(domain);
2298 vm_domain_pageout_assert_unlocked(vmd);
2299 if (curproc == pageproc)
2302 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2303 vm_domain_pageout_lock(vmd);
2304 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2305 wakeup(&vmd->vmd_pageout_wanted);
2306 vm_domain_pageout_unlock(vmd);