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_int vm_page_max_user_wired;
198 SYSCTL_UINT(_vm, OID_AUTO, max_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->aflags & 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->aflags & 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->aflags & 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->aflags & 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);
308 /* Return the next page to be scanned, or NULL if the scan is complete. */
309 static __always_inline vm_page_t
310 vm_pageout_next(struct scan_state *ss, const bool dequeue)
313 if (ss->bq.bq_cnt == 0)
314 vm_pageout_collect_batch(ss, dequeue);
315 return (vm_batchqueue_pop(&ss->bq));
319 * Scan for pages at adjacent offsets within the given page's object that are
320 * eligible for laundering, form a cluster of these pages and the given page,
321 * and launder that cluster.
324 vm_pageout_cluster(vm_page_t m)
327 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
329 int ib, is, page_base, pageout_count;
331 vm_page_assert_locked(m);
333 VM_OBJECT_ASSERT_WLOCKED(object);
336 vm_page_assert_unbusied(m);
337 KASSERT(!vm_page_held(m), ("page %p is held", m));
339 pmap_remove_write(m);
342 mc[vm_pageout_page_count] = pb = ps = m;
344 page_base = vm_pageout_page_count;
349 * We can cluster only if the page is not clean, busy, or held, and
350 * the page is in the laundry queue.
352 * During heavy mmap/modification loads the pageout
353 * daemon can really fragment the underlying file
354 * due to flushing pages out of order and not trying to
355 * align the clusters (which leaves sporadic out-of-order
356 * holes). To solve this problem we do the reverse scan
357 * first and attempt to align our cluster, then do a
358 * forward scan if room remains.
361 while (ib != 0 && pageout_count < vm_pageout_page_count) {
366 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
370 vm_page_test_dirty(p);
376 if (vm_page_held(p) || !vm_page_in_laundry(p)) {
381 pmap_remove_write(p);
383 mc[--page_base] = pb = p;
388 * We are at an alignment boundary. Stop here, and switch
389 * directions. Do not clear ib.
391 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
394 while (pageout_count < vm_pageout_page_count &&
395 pindex + is < object->size) {
396 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
398 vm_page_test_dirty(p);
402 if (vm_page_held(p) || !vm_page_in_laundry(p)) {
406 pmap_remove_write(p);
408 mc[page_base + pageout_count] = ps = p;
414 * If we exhausted our forward scan, continue with the reverse scan
415 * when possible, even past an alignment boundary. This catches
416 * boundary conditions.
418 if (ib != 0 && pageout_count < vm_pageout_page_count)
421 return (vm_pageout_flush(&mc[page_base], pageout_count,
422 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
426 * vm_pageout_flush() - launder the given pages
428 * The given pages are laundered. Note that we setup for the start of
429 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
430 * reference count all in here rather then in the parent. If we want
431 * the parent to do more sophisticated things we may have to change
434 * Returned runlen is the count of pages between mreq and first
435 * page after mreq with status VM_PAGER_AGAIN.
436 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
437 * for any page in runlen set.
440 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
443 vm_object_t object = mc[0]->object;
444 int pageout_status[count];
448 VM_OBJECT_ASSERT_WLOCKED(object);
451 * Initiate I/O. Mark the pages busy and verify that they're valid
454 * We do not have to fixup the clean/dirty bits here... we can
455 * allow the pager to do it after the I/O completes.
457 * NOTE! mc[i]->dirty may be partial or fragmented due to an
458 * edge case with file fragments.
460 for (i = 0; i < count; i++) {
461 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
462 ("vm_pageout_flush: partially invalid page %p index %d/%d",
464 KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
465 ("vm_pageout_flush: writeable page %p", mc[i]));
466 vm_page_sbusy(mc[i]);
468 vm_object_pip_add(object, count);
470 vm_pager_put_pages(object, mc, count, flags, pageout_status);
472 runlen = count - mreq;
475 for (i = 0; i < count; i++) {
476 vm_page_t mt = mc[i];
478 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
479 !pmap_page_is_write_mapped(mt),
480 ("vm_pageout_flush: page %p is not write protected", mt));
481 switch (pageout_status[i]) {
484 if (vm_page_in_laundry(mt))
485 vm_page_deactivate_noreuse(mt);
493 * The page is outside the object's range. We pretend
494 * that the page out worked and clean the page, so the
495 * changes will be lost if the page is reclaimed by
500 if (vm_page_in_laundry(mt))
501 vm_page_deactivate_noreuse(mt);
507 * If the page couldn't be paged out to swap because the
508 * pager wasn't able to find space, place the page in
509 * the PQ_UNSWAPPABLE holding queue. This is an
510 * optimization that prevents the page daemon from
511 * wasting CPU cycles on pages that cannot be reclaimed
512 * becase no swap device is configured.
514 * Otherwise, reactivate the page so that it doesn't
515 * clog the laundry and inactive queues. (We will try
516 * paging it out again later.)
519 if (object->type == OBJT_SWAP &&
520 pageout_status[i] == VM_PAGER_FAIL) {
521 vm_page_unswappable(mt);
524 vm_page_activate(mt);
526 if (eio != NULL && i >= mreq && i - mreq < runlen)
530 if (i >= mreq && i - mreq < runlen)
536 * If the operation is still going, leave the page busy to
537 * block all other accesses. Also, leave the paging in
538 * progress indicator set so that we don't attempt an object
541 if (pageout_status[i] != VM_PAGER_PEND) {
542 vm_object_pip_wakeup(object);
548 return (numpagedout);
552 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
555 atomic_store_rel_int(&swapdev_enabled, 1);
559 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
562 if (swap_pager_nswapdev() == 1)
563 atomic_store_rel_int(&swapdev_enabled, 0);
567 * Attempt to acquire all of the necessary locks to launder a page and
568 * then call through the clustering layer to PUTPAGES. Wait a short
569 * time for a vnode lock.
571 * Requires the page and object lock on entry, releases both before return.
572 * Returns 0 on success and an errno otherwise.
575 vm_pageout_clean(vm_page_t m, int *numpagedout)
583 vm_page_assert_locked(m);
585 VM_OBJECT_ASSERT_WLOCKED(object);
591 * The object is already known NOT to be dead. It
592 * is possible for the vget() to block the whole
593 * pageout daemon, but the new low-memory handling
594 * code should prevent it.
596 * We can't wait forever for the vnode lock, we might
597 * deadlock due to a vn_read() getting stuck in
598 * vm_wait while holding this vnode. We skip the
599 * vnode if we can't get it in a reasonable amount
602 if (object->type == OBJT_VNODE) {
605 if (vp->v_type == VREG &&
606 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
612 ("vp %p with NULL v_mount", vp));
613 vm_object_reference_locked(object);
615 VM_OBJECT_WUNLOCK(object);
616 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
617 LK_SHARED : LK_EXCLUSIVE;
618 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
623 VM_OBJECT_WLOCK(object);
626 * Ensure that the object and vnode were not disassociated
627 * while locks were dropped.
629 if (vp->v_object != object) {
636 * While the object and page were unlocked, the page
638 * (1) moved to a different queue,
639 * (2) reallocated to a different object,
640 * (3) reallocated to a different offset, or
643 if (!vm_page_in_laundry(m) || m->object != object ||
644 m->pindex != pindex || m->dirty == 0) {
651 * The page may have been busied or referenced while the object
652 * and page locks were released.
654 if (vm_page_busied(m) || vm_page_held(m)) {
662 * If a page is dirty, then it is either being washed
663 * (but not yet cleaned) or it is still in the
664 * laundry. If it is still in the laundry, then we
665 * start the cleaning operation.
667 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
671 VM_OBJECT_WUNLOCK(object);
674 vm_page_lock_assert(m, MA_NOTOWNED);
678 vm_object_deallocate(object);
679 vn_finished_write(mp);
686 * Attempt to launder the specified number of pages.
688 * Returns the number of pages successfully laundered.
691 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
693 struct scan_state ss;
694 struct vm_pagequeue *pq;
698 int act_delta, error, numpagedout, queue, starting_target;
704 starting_target = launder;
708 * Scan the laundry queues for pages eligible to be laundered. We stop
709 * once the target number of dirty pages have been laundered, or once
710 * we've reached the end of the queue. A single iteration of this loop
711 * may cause more than one page to be laundered because of clustering.
713 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
714 * swap devices are configured.
716 if (atomic_load_acq_int(&swapdev_enabled))
717 queue = PQ_UNSWAPPABLE;
722 marker = &vmd->vmd_markers[queue];
723 pq = &vmd->vmd_pagequeues[queue];
724 vm_pagequeue_lock(pq);
725 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
726 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
727 if (__predict_false((m->flags & PG_MARKER) != 0))
730 vm_page_change_lock(m, &mtx);
734 * The page may have been disassociated from the queue
735 * while locks were dropped.
737 if (vm_page_queue(m) != queue)
741 * A requeue was requested, so this page gets a second
744 if ((m->aflags & PGA_REQUEUE) != 0) {
750 * Held pages are essentially stuck in the queue.
752 * Wired pages may not be freed. Complete their removal
753 * from the queue now to avoid needless revisits during
756 if (m->hold_count != 0)
758 if (vm_page_wired(m)) {
759 vm_page_dequeue_deferred(m);
763 if (object != m->object) {
765 VM_OBJECT_WUNLOCK(object);
767 if (!VM_OBJECT_TRYWLOCK(object)) {
769 /* Depends on type-stability. */
770 VM_OBJECT_WLOCK(object);
776 if (vm_page_busied(m))
780 * Invalid pages can be easily freed. They cannot be
781 * mapped; vm_page_free() asserts this.
787 * If the page has been referenced and the object is not dead,
788 * reactivate or requeue the page depending on whether the
791 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
792 * that a reference from a concurrently destroyed mapping is
793 * observed here and now.
795 if (object->ref_count != 0)
796 act_delta = pmap_ts_referenced(m);
798 KASSERT(!pmap_page_is_mapped(m),
799 ("page %p is mapped", m));
802 if ((m->aflags & PGA_REFERENCED) != 0) {
803 vm_page_aflag_clear(m, PGA_REFERENCED);
806 if (act_delta != 0) {
807 if (object->ref_count != 0) {
808 VM_CNT_INC(v_reactivated);
812 * Increase the activation count if the page
813 * was referenced while in the laundry queue.
814 * This makes it less likely that the page will
815 * be returned prematurely to the inactive
818 m->act_count += act_delta + ACT_ADVANCE;
821 * If this was a background laundering, count
822 * activated pages towards our target. The
823 * purpose of background laundering is to ensure
824 * that pages are eventually cycled through the
825 * laundry queue, and an activation is a valid
831 } else if ((object->flags & OBJ_DEAD) == 0) {
838 * If the page appears to be clean at the machine-independent
839 * layer, then remove all of its mappings from the pmap in
840 * anticipation of freeing it. If, however, any of the page's
841 * mappings allow write access, then the page may still be
842 * modified until the last of those mappings are removed.
844 if (object->ref_count != 0) {
845 vm_page_test_dirty(m);
851 * Clean pages are freed, and dirty pages are paged out unless
852 * they belong to a dead object. Requeueing dirty pages from
853 * dead objects is pointless, as they are being paged out and
854 * freed by the thread that destroyed the object.
860 } else if ((object->flags & OBJ_DEAD) == 0) {
861 if (object->type != OBJT_SWAP &&
862 object->type != OBJT_DEFAULT)
864 else if (disable_swap_pageouts)
874 * Form a cluster with adjacent, dirty pages from the
875 * same object, and page out that entire cluster.
877 * The adjacent, dirty pages must also be in the
878 * laundry. However, their mappings are not checked
879 * for new references. Consequently, a recently
880 * referenced page may be paged out. However, that
881 * page will not be prematurely reclaimed. After page
882 * out, the page will be placed in the inactive queue,
883 * where any new references will be detected and the
886 error = vm_pageout_clean(m, &numpagedout);
888 launder -= numpagedout;
889 ss.scanned += numpagedout;
890 } else if (error == EDEADLK) {
902 if (object != NULL) {
903 VM_OBJECT_WUNLOCK(object);
906 vm_pagequeue_lock(pq);
907 vm_pageout_end_scan(&ss);
908 vm_pagequeue_unlock(pq);
910 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
916 * Wakeup the sync daemon if we skipped a vnode in a writeable object
917 * and we didn't launder enough pages.
919 if (vnodes_skipped > 0 && launder > 0)
920 (void)speedup_syncer();
922 return (starting_target - launder);
926 * Compute the integer square root.
931 u_int bit, root, tmp;
933 bit = 1u << ((NBBY * sizeof(u_int)) - 2);
950 * Perform the work of the laundry thread: periodically wake up and determine
951 * whether any pages need to be laundered. If so, determine the number of pages
952 * that need to be laundered, and launder them.
955 vm_pageout_laundry_worker(void *arg)
957 struct vm_domain *vmd;
958 struct vm_pagequeue *pq;
959 uint64_t nclean, ndirty, nfreed;
960 int domain, last_target, launder, shortfall, shortfall_cycle, target;
963 domain = (uintptr_t)arg;
964 vmd = VM_DOMAIN(domain);
965 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
966 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
969 in_shortfall = false;
971 last_target = target = 0;
975 * Calls to these handlers are serialized by the swap syscall lock.
977 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
978 EVENTHANDLER_PRI_ANY);
979 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
980 EVENTHANDLER_PRI_ANY);
983 * The pageout laundry worker is never done, so loop forever.
986 KASSERT(target >= 0, ("negative target %d", target));
987 KASSERT(shortfall_cycle >= 0,
988 ("negative cycle %d", shortfall_cycle));
992 * First determine whether we need to launder pages to meet a
993 * shortage of free pages.
997 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
999 } else if (!in_shortfall)
1001 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1003 * We recently entered shortfall and began laundering
1004 * pages. If we have completed that laundering run
1005 * (and we are no longer in shortfall) or we have met
1006 * our laundry target through other activity, then we
1007 * can stop laundering pages.
1009 in_shortfall = false;
1013 launder = target / shortfall_cycle--;
1017 * There's no immediate need to launder any pages; see if we
1018 * meet the conditions to perform background laundering:
1020 * 1. The ratio of dirty to clean inactive pages exceeds the
1021 * background laundering threshold, or
1022 * 2. we haven't yet reached the target of the current
1023 * background laundering run.
1025 * The background laundering threshold is not a constant.
1026 * Instead, it is a slowly growing function of the number of
1027 * clean pages freed by the page daemon since the last
1028 * background laundering. Thus, as the ratio of dirty to
1029 * clean inactive pages grows, the amount of memory pressure
1030 * required to trigger laundering decreases. We ensure
1031 * that the threshold is non-zero after an inactive queue
1032 * scan, even if that scan failed to free a single clean page.
1035 nclean = vmd->vmd_free_count +
1036 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1037 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1038 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1039 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1040 target = vmd->vmd_background_launder_target;
1044 * We have a non-zero background laundering target. If we've
1045 * laundered up to our maximum without observing a page daemon
1046 * request, just stop. This is a safety belt that ensures we
1047 * don't launder an excessive amount if memory pressure is low
1048 * and the ratio of dirty to clean pages is large. Otherwise,
1049 * proceed at the background laundering rate.
1054 last_target = target;
1055 } else if (last_target - target >=
1056 vm_background_launder_max * PAGE_SIZE / 1024) {
1059 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1060 launder /= VM_LAUNDER_RATE;
1061 if (launder > target)
1068 * Because of I/O clustering, the number of laundered
1069 * pages could exceed "target" by the maximum size of
1070 * a cluster minus one.
1072 target -= min(vm_pageout_launder(vmd, launder,
1073 in_shortfall), target);
1074 pause("laundp", hz / VM_LAUNDER_RATE);
1078 * If we're not currently laundering pages and the page daemon
1079 * hasn't posted a new request, sleep until the page daemon
1082 vm_pagequeue_lock(pq);
1083 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1084 (void)mtx_sleep(&vmd->vmd_laundry_request,
1085 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1088 * If the pagedaemon has indicated that it's in shortfall, start
1089 * a shortfall laundering unless we're already in the middle of
1090 * one. This may preempt a background laundering.
1092 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1093 (!in_shortfall || shortfall_cycle == 0)) {
1094 shortfall = vm_laundry_target(vmd) +
1095 vmd->vmd_pageout_deficit;
1101 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1102 nfreed += vmd->vmd_clean_pages_freed;
1103 vmd->vmd_clean_pages_freed = 0;
1104 vm_pagequeue_unlock(pq);
1109 * Compute the number of pages we want to try to move from the
1110 * active queue to either the inactive or laundry queue.
1112 * When scanning active pages during a shortage, we make clean pages
1113 * count more heavily towards the page shortage than dirty pages.
1114 * This is because dirty pages must be laundered before they can be
1115 * reused and thus have less utility when attempting to quickly
1116 * alleviate a free page shortage. However, this weighting also
1117 * causes the scan to deactivate dirty pages more aggressively,
1118 * improving the effectiveness of clustering.
1121 vm_pageout_active_target(struct vm_domain *vmd)
1125 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1126 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1127 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1128 shortage *= act_scan_laundry_weight;
1133 * Scan the active queue. If there is no shortage of inactive pages, scan a
1134 * small portion of the queue in order to maintain quasi-LRU.
1137 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1139 struct scan_state ss;
1141 vm_page_t m, marker;
1142 struct vm_pagequeue *pq;
1144 int act_delta, max_scan, scan_tick;
1146 marker = &vmd->vmd_markers[PQ_ACTIVE];
1147 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1148 vm_pagequeue_lock(pq);
1151 * If we're just idle polling attempt to visit every
1152 * active page within 'update_period' seconds.
1155 if (vm_pageout_update_period != 0) {
1156 min_scan = pq->pq_cnt;
1157 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1158 min_scan /= hz * vm_pageout_update_period;
1161 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1162 vmd->vmd_last_active_scan = scan_tick;
1165 * Scan the active queue for pages that can be deactivated. Update
1166 * the per-page activity counter and use it to identify deactivation
1167 * candidates. Held pages may be deactivated.
1169 * To avoid requeuing each page that remains in the active queue, we
1170 * implement the CLOCK algorithm. To keep the implementation of the
1171 * enqueue operation consistent for all page queues, we use two hands,
1172 * represented by marker pages. Scans begin at the first hand, which
1173 * precedes the second hand in the queue. When the two hands meet,
1174 * they are moved back to the head and tail of the queue, respectively,
1175 * and scanning resumes.
1177 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1180 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1181 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1182 if (__predict_false(m == &vmd->vmd_clock[1])) {
1183 vm_pagequeue_lock(pq);
1184 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1185 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1186 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1188 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1190 max_scan -= ss.scanned;
1191 vm_pageout_end_scan(&ss);
1194 if (__predict_false((m->flags & PG_MARKER) != 0))
1197 vm_page_change_lock(m, &mtx);
1200 * The page may have been disassociated from the queue
1201 * while locks were dropped.
1203 if (vm_page_queue(m) != PQ_ACTIVE)
1207 * Wired pages are dequeued lazily.
1209 if (vm_page_wired(m)) {
1210 vm_page_dequeue_deferred(m);
1215 * Check to see "how much" the page has been used.
1217 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1218 * that a reference from a concurrently destroyed mapping is
1219 * observed here and now.
1221 * Perform an unsynchronized object ref count check. While
1222 * the page lock ensures that the page is not reallocated to
1223 * another object, in particular, one with unmanaged mappings
1224 * that cannot support pmap_ts_referenced(), two races are,
1225 * nonetheless, possible:
1226 * 1) The count was transitioning to zero, but we saw a non-
1227 * zero value. pmap_ts_referenced() will return zero
1228 * because the page is not mapped.
1229 * 2) The count was transitioning to one, but we saw zero.
1230 * This race delays the detection of a new reference. At
1231 * worst, we will deactivate and reactivate the page.
1233 if (m->object->ref_count != 0)
1234 act_delta = pmap_ts_referenced(m);
1237 if ((m->aflags & PGA_REFERENCED) != 0) {
1238 vm_page_aflag_clear(m, PGA_REFERENCED);
1243 * Advance or decay the act_count based on recent usage.
1245 if (act_delta != 0) {
1246 m->act_count += ACT_ADVANCE + act_delta;
1247 if (m->act_count > ACT_MAX)
1248 m->act_count = ACT_MAX;
1250 m->act_count -= min(m->act_count, ACT_DECLINE);
1252 if (m->act_count == 0) {
1254 * When not short for inactive pages, let dirty pages go
1255 * through the inactive queue before moving to the
1256 * laundry queues. This gives them some extra time to
1257 * be reactivated, potentially avoiding an expensive
1258 * pageout. However, during a page shortage, the
1259 * inactive queue is necessarily small, and so dirty
1260 * pages would only spend a trivial amount of time in
1261 * the inactive queue. Therefore, we might as well
1262 * place them directly in the laundry queue to reduce
1265 if (page_shortage <= 0)
1266 vm_page_deactivate(m);
1269 * Calling vm_page_test_dirty() here would
1270 * require acquisition of the object's write
1271 * lock. However, during a page shortage,
1272 * directing dirty pages into the laundry
1273 * queue is only an optimization and not a
1274 * requirement. Therefore, we simply rely on
1275 * the opportunistic updates to the page's
1276 * dirty field by the pmap.
1278 if (m->dirty == 0) {
1279 vm_page_deactivate(m);
1281 act_scan_laundry_weight;
1293 vm_pagequeue_lock(pq);
1294 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1295 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1296 vm_pageout_end_scan(&ss);
1297 vm_pagequeue_unlock(pq);
1301 vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m)
1303 struct vm_domain *vmd;
1305 if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0)
1307 vm_page_aflag_set(m, PGA_ENQUEUED);
1308 if ((m->aflags & PGA_REQUEUE_HEAD) != 0) {
1309 vmd = vm_pagequeue_domain(m);
1310 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
1311 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1312 } else if ((m->aflags & PGA_REQUEUE) != 0) {
1313 TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q);
1314 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1316 TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q);
1321 * Re-add stuck pages to the inactive queue. We will examine them again
1322 * during the next scan. If the queue state of a page has changed since
1323 * it was physically removed from the page queue in
1324 * vm_pageout_collect_batch(), don't do anything with that page.
1327 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1330 struct vm_pagequeue *pq;
1337 if (vm_batchqueue_insert(bq, m))
1339 vm_pagequeue_lock(pq);
1340 delta += vm_pageout_reinsert_inactive_page(ss, m);
1342 vm_pagequeue_lock(pq);
1343 while ((m = vm_batchqueue_pop(bq)) != NULL)
1344 delta += vm_pageout_reinsert_inactive_page(ss, m);
1345 vm_pagequeue_cnt_add(pq, delta);
1346 vm_pagequeue_unlock(pq);
1347 vm_batchqueue_init(bq);
1351 * Attempt to reclaim the requested number of pages from the inactive queue.
1352 * Returns true if the shortage was addressed.
1355 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1358 struct scan_state ss;
1359 struct vm_batchqueue rq;
1361 vm_page_t m, marker;
1362 struct vm_pagequeue *pq;
1364 int act_delta, addl_page_shortage, deficit, page_shortage;
1365 int starting_page_shortage;
1368 * The addl_page_shortage is an estimate of the number of temporarily
1369 * stuck pages in the inactive queue. In other words, the
1370 * number of pages from the inactive count that should be
1371 * discounted in setting the target for the active queue scan.
1373 addl_page_shortage = 0;
1376 * vmd_pageout_deficit counts the number of pages requested in
1377 * allocations that failed because of a free page shortage. We assume
1378 * that the allocations will be reattempted and thus include the deficit
1379 * in our scan target.
1381 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1382 starting_page_shortage = page_shortage = shortage + deficit;
1386 vm_batchqueue_init(&rq);
1389 * Start scanning the inactive queue for pages that we can free. The
1390 * scan will stop when we reach the target or we have scanned the
1391 * entire queue. (Note that m->act_count is not used to make
1392 * decisions for the inactive queue, only for the active queue.)
1394 marker = &vmd->vmd_markers[PQ_INACTIVE];
1395 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1396 vm_pagequeue_lock(pq);
1397 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1398 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1399 KASSERT((m->flags & PG_MARKER) == 0,
1400 ("marker page %p was dequeued", m));
1402 vm_page_change_lock(m, &mtx);
1406 * The page may have been disassociated from the queue
1407 * while locks were dropped.
1409 if (vm_page_queue(m) != PQ_INACTIVE) {
1410 addl_page_shortage++;
1415 * The page was re-enqueued after the page queue lock was
1416 * dropped, or a requeue was requested. This page gets a second
1419 if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE |
1420 PGA_REQUEUE_HEAD)) != 0)
1424 * Held pages are essentially stuck in the queue. So,
1425 * they ought to be discounted from the inactive count.
1426 * See the description of addl_page_shortage above.
1428 * Wired pages may not be freed. Complete their removal
1429 * from the queue now to avoid needless revisits during
1432 if (m->hold_count != 0) {
1433 addl_page_shortage++;
1436 if (vm_page_wired(m)) {
1437 vm_page_dequeue_deferred(m);
1441 if (object != m->object) {
1443 VM_OBJECT_WUNLOCK(object);
1445 if (!VM_OBJECT_TRYWLOCK(object)) {
1447 /* Depends on type-stability. */
1448 VM_OBJECT_WLOCK(object);
1454 if (vm_page_busied(m)) {
1456 * Don't mess with busy pages. Leave them at
1457 * the front of the queue. Most likely, they
1458 * are being paged out and will leave the
1459 * queue shortly after the scan finishes. So,
1460 * they ought to be discounted from the
1463 addl_page_shortage++;
1468 * Invalid pages can be easily freed. They cannot be
1469 * mapped, vm_page_free() asserts this.
1475 * If the page has been referenced and the object is not dead,
1476 * reactivate or requeue the page depending on whether the
1479 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1480 * that a reference from a concurrently destroyed mapping is
1481 * observed here and now.
1483 if (object->ref_count != 0)
1484 act_delta = pmap_ts_referenced(m);
1486 KASSERT(!pmap_page_is_mapped(m),
1487 ("page %p is mapped", m));
1490 if ((m->aflags & PGA_REFERENCED) != 0) {
1491 vm_page_aflag_clear(m, PGA_REFERENCED);
1494 if (act_delta != 0) {
1495 if (object->ref_count != 0) {
1496 VM_CNT_INC(v_reactivated);
1497 vm_page_activate(m);
1500 * Increase the activation count if the page
1501 * was referenced while in the inactive queue.
1502 * This makes it less likely that the page will
1503 * be returned prematurely to the inactive
1506 m->act_count += act_delta + ACT_ADVANCE;
1508 } else if ((object->flags & OBJ_DEAD) == 0) {
1509 vm_page_aflag_set(m, PGA_REQUEUE);
1515 * If the page appears to be clean at the machine-independent
1516 * layer, then remove all of its mappings from the pmap in
1517 * anticipation of freeing it. If, however, any of the page's
1518 * mappings allow write access, then the page may still be
1519 * modified until the last of those mappings are removed.
1521 if (object->ref_count != 0) {
1522 vm_page_test_dirty(m);
1528 * Clean pages can be freed, but dirty pages must be sent back
1529 * to the laundry, unless they belong to a dead object.
1530 * Requeueing dirty pages from dead objects is pointless, as
1531 * they are being paged out and freed by the thread that
1532 * destroyed the object.
1534 if (m->dirty == 0) {
1537 * Because we dequeued the page and have already
1538 * checked for concurrent dequeue and enqueue
1539 * requests, we can safely disassociate the page
1540 * from the inactive queue.
1542 KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
1543 ("page %p has queue state", m));
1547 } else if ((object->flags & OBJ_DEAD) == 0)
1551 vm_pageout_reinsert_inactive(&ss, &rq, m);
1556 VM_OBJECT_WUNLOCK(object);
1557 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1558 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1559 vm_pagequeue_lock(pq);
1560 vm_pageout_end_scan(&ss);
1561 vm_pagequeue_unlock(pq);
1563 VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1566 * Wake up the laundry thread so that it can perform any needed
1567 * laundering. If we didn't meet our target, we're in shortfall and
1568 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1569 * swap devices are configured, the laundry thread has no work to do, so
1570 * don't bother waking it up.
1572 * The laundry thread uses the number of inactive queue scans elapsed
1573 * since the last laundering to determine whether to launder again, so
1576 if (starting_page_shortage > 0) {
1577 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1578 vm_pagequeue_lock(pq);
1579 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1580 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1581 if (page_shortage > 0) {
1582 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1583 VM_CNT_INC(v_pdshortfalls);
1584 } else if (vmd->vmd_laundry_request !=
1585 VM_LAUNDRY_SHORTFALL)
1586 vmd->vmd_laundry_request =
1587 VM_LAUNDRY_BACKGROUND;
1588 wakeup(&vmd->vmd_laundry_request);
1590 vmd->vmd_clean_pages_freed +=
1591 starting_page_shortage - page_shortage;
1592 vm_pagequeue_unlock(pq);
1596 * Wakeup the swapout daemon if we didn't free the targeted number of
1599 if (page_shortage > 0)
1603 * If the inactive queue scan fails repeatedly to meet its
1604 * target, kill the largest process.
1606 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1609 * Reclaim pages by swapping out idle processes, if configured to do so.
1611 vm_swapout_run_idle();
1614 * See the description of addl_page_shortage above.
1616 *addl_shortage = addl_page_shortage + deficit;
1618 return (page_shortage <= 0);
1621 static int vm_pageout_oom_vote;
1624 * The pagedaemon threads randlomly select one to perform the
1625 * OOM. Trying to kill processes before all pagedaemons
1626 * failed to reach free target is premature.
1629 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1630 int starting_page_shortage)
1634 if (starting_page_shortage <= 0 || starting_page_shortage !=
1636 vmd->vmd_oom_seq = 0;
1639 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1641 vmd->vmd_oom = FALSE;
1642 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1648 * Do not follow the call sequence until OOM condition is
1651 vmd->vmd_oom_seq = 0;
1656 vmd->vmd_oom = TRUE;
1657 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1658 if (old_vote != vm_ndomains - 1)
1662 * The current pagedaemon thread is the last in the quorum to
1663 * start OOM. Initiate the selection and signaling of the
1666 vm_pageout_oom(VM_OOM_MEM);
1669 * After one round of OOM terror, recall our vote. On the
1670 * next pass, current pagedaemon would vote again if the low
1671 * memory condition is still there, due to vmd_oom being
1674 vmd->vmd_oom = FALSE;
1675 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1679 * The OOM killer is the page daemon's action of last resort when
1680 * memory allocation requests have been stalled for a prolonged period
1681 * of time because it cannot reclaim memory. This function computes
1682 * the approximate number of physical pages that could be reclaimed if
1683 * the specified address space is destroyed.
1685 * Private, anonymous memory owned by the address space is the
1686 * principal resource that we expect to recover after an OOM kill.
1687 * Since the physical pages mapped by the address space's COW entries
1688 * are typically shared pages, they are unlikely to be released and so
1689 * they are not counted.
1691 * To get to the point where the page daemon runs the OOM killer, its
1692 * efforts to write-back vnode-backed pages may have stalled. This
1693 * could be caused by a memory allocation deadlock in the write path
1694 * that might be resolved by an OOM kill. Therefore, physical pages
1695 * belonging to vnode-backed objects are counted, because they might
1696 * be freed without being written out first if the address space holds
1697 * the last reference to an unlinked vnode.
1699 * Similarly, physical pages belonging to OBJT_PHYS objects are
1700 * counted because the address space might hold the last reference to
1704 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1707 vm_map_entry_t entry;
1711 map = &vmspace->vm_map;
1712 KASSERT(!map->system_map, ("system map"));
1713 sx_assert(&map->lock, SA_LOCKED);
1715 for (entry = map->header.next; entry != &map->header;
1716 entry = entry->next) {
1717 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1719 obj = entry->object.vm_object;
1722 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1723 obj->ref_count != 1)
1725 switch (obj->type) {
1730 res += obj->resident_page_count;
1737 static int vm_oom_ratelim_last;
1738 static int vm_oom_pf_secs = 10;
1739 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1741 static struct mtx vm_oom_ratelim_mtx;
1744 vm_pageout_oom(int shortage)
1746 struct proc *p, *bigproc;
1747 vm_offset_t size, bigsize;
1754 * For OOM requests originating from vm_fault(), there is a high
1755 * chance that a single large process faults simultaneously in
1756 * several threads. Also, on an active system running many
1757 * processes of middle-size, like buildworld, all of them
1758 * could fault almost simultaneously as well.
1760 * To avoid killing too many processes, rate-limit OOMs
1761 * initiated by vm_fault() time-outs on the waits for free
1764 mtx_lock(&vm_oom_ratelim_mtx);
1766 if (shortage == VM_OOM_MEM_PF &&
1767 (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1768 mtx_unlock(&vm_oom_ratelim_mtx);
1771 vm_oom_ratelim_last = now;
1772 mtx_unlock(&vm_oom_ratelim_mtx);
1775 * We keep the process bigproc locked once we find it to keep anyone
1776 * from messing with it; however, there is a possibility of
1777 * deadlock if process B is bigproc and one of its child processes
1778 * attempts to propagate a signal to B while we are waiting for A's
1779 * lock while walking this list. To avoid this, we don't block on
1780 * the process lock but just skip a process if it is already locked.
1784 sx_slock(&allproc_lock);
1785 FOREACH_PROC_IN_SYSTEM(p) {
1789 * If this is a system, protected or killed process, skip it.
1791 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1792 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1793 p->p_pid == 1 || P_KILLED(p) ||
1794 (p->p_pid < 48 && swap_pager_avail != 0)) {
1799 * If the process is in a non-running type state,
1800 * don't touch it. Check all the threads individually.
1803 FOREACH_THREAD_IN_PROC(p, td) {
1805 if (!TD_ON_RUNQ(td) &&
1806 !TD_IS_RUNNING(td) &&
1807 !TD_IS_SLEEPING(td) &&
1808 !TD_IS_SUSPENDED(td) &&
1809 !TD_IS_SWAPPED(td)) {
1821 * get the process size
1823 vm = vmspace_acquire_ref(p);
1830 sx_sunlock(&allproc_lock);
1831 if (!vm_map_trylock_read(&vm->vm_map)) {
1833 sx_slock(&allproc_lock);
1837 size = vmspace_swap_count(vm);
1838 if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
1839 size += vm_pageout_oom_pagecount(vm);
1840 vm_map_unlock_read(&vm->vm_map);
1842 sx_slock(&allproc_lock);
1845 * If this process is bigger than the biggest one,
1848 if (size > bigsize) {
1849 if (bigproc != NULL)
1857 sx_sunlock(&allproc_lock);
1858 if (bigproc != NULL) {
1859 if (vm_panic_on_oom != 0)
1860 panic("out of swap space");
1862 killproc(bigproc, "out of swap space");
1863 sched_nice(bigproc, PRIO_MIN);
1865 PROC_UNLOCK(bigproc);
1870 vm_pageout_lowmem(void)
1872 static int lowmem_ticks = 0;
1875 last = atomic_load_int(&lowmem_ticks);
1876 while ((u_int)(ticks - last) / hz >= lowmem_period) {
1877 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
1881 * Decrease registered cache sizes.
1883 SDT_PROBE0(vm, , , vm__lowmem_scan);
1884 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1887 * We do this explicitly after the caches have been
1897 vm_pageout_worker(void *arg)
1899 struct vm_domain *vmd;
1901 int addl_shortage, domain, shortage;
1904 domain = (uintptr_t)arg;
1905 vmd = VM_DOMAIN(domain);
1910 * XXXKIB It could be useful to bind pageout daemon threads to
1911 * the cores belonging to the domain, from which vm_page_array
1915 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1916 vmd->vmd_last_active_scan = ticks;
1919 * The pageout daemon worker is never done, so loop forever.
1922 vm_domain_pageout_lock(vmd);
1925 * We need to clear wanted before we check the limits. This
1926 * prevents races with wakers who will check wanted after they
1929 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
1932 * Might the page daemon need to run again?
1934 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
1936 * Yes. If the scan failed to produce enough free
1937 * pages, sleep uninterruptibly for some time in the
1938 * hope that the laundry thread will clean some pages.
1940 vm_domain_pageout_unlock(vmd);
1942 pause("pwait", hz / VM_INACT_SCAN_RATE);
1945 * No, sleep until the next wakeup or until pages
1946 * need to have their reference stats updated.
1948 if (mtx_sleep(&vmd->vmd_pageout_wanted,
1949 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
1950 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
1951 VM_CNT_INC(v_pdwakeups);
1954 /* Prevent spurious wakeups by ensuring that wanted is set. */
1955 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1958 * Use the controller to calculate how many pages to free in
1959 * this interval, and scan the inactive queue. If the lowmem
1960 * handlers appear to have freed up some pages, subtract the
1961 * difference from the inactive queue scan target.
1963 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
1965 ofree = vmd->vmd_free_count;
1966 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
1967 shortage -= min(vmd->vmd_free_count - ofree,
1969 target_met = vm_pageout_scan_inactive(vmd, shortage,
1975 * Scan the active queue. A positive value for shortage
1976 * indicates that we must aggressively deactivate pages to avoid
1979 shortage = vm_pageout_active_target(vmd) + addl_shortage;
1980 vm_pageout_scan_active(vmd, shortage);
1985 * vm_pageout_init initialises basic pageout daemon settings.
1988 vm_pageout_init_domain(int domain)
1990 struct vm_domain *vmd;
1991 struct sysctl_oid *oid;
1993 vmd = VM_DOMAIN(domain);
1994 vmd->vmd_interrupt_free_min = 2;
1997 * v_free_reserved needs to include enough for the largest
1998 * swap pager structures plus enough for any pv_entry structs
2001 if (vmd->vmd_page_count > 1024)
2002 vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200;
2004 vmd->vmd_free_min = 4;
2005 vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
2006 vmd->vmd_interrupt_free_min;
2007 vmd->vmd_free_reserved = vm_pageout_page_count +
2008 vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768);
2009 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2010 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2011 vmd->vmd_free_min += vmd->vmd_free_reserved;
2012 vmd->vmd_free_severe += vmd->vmd_free_reserved;
2013 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2014 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2015 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2018 * Set the default wakeup threshold to be 10% below the paging
2019 * target. This keeps the steady state out of shortfall.
2021 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2024 * Target amount of memory to move out of the laundry queue during a
2025 * background laundering. This is proportional to the amount of system
2028 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2029 vmd->vmd_free_min) / 10;
2031 /* Initialize the pageout daemon pid controller. */
2032 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2033 vmd->vmd_free_target, PIDCTRL_BOUND,
2034 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2035 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2036 "pidctrl", CTLFLAG_RD, NULL, "");
2037 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2041 vm_pageout_init(void)
2047 * Initialize some paging parameters.
2049 if (vm_cnt.v_page_count < 2000)
2050 vm_pageout_page_count = 8;
2053 for (i = 0; i < vm_ndomains; i++) {
2054 struct vm_domain *vmd;
2056 vm_pageout_init_domain(i);
2058 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2059 vm_cnt.v_free_target += vmd->vmd_free_target;
2060 vm_cnt.v_free_min += vmd->vmd_free_min;
2061 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2062 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2063 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2064 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2065 freecount += vmd->vmd_free_count;
2069 * Set interval in seconds for active scan. We want to visit each
2070 * page at least once every ten minutes. This is to prevent worst
2071 * case paging behaviors with stale active LRU.
2073 if (vm_pageout_update_period == 0)
2074 vm_pageout_update_period = 600;
2077 * Set the maximum number of user-wired virtual pages. Historically the
2078 * main source of such pages was mlock(2) and mlockall(2). Hypervisors
2079 * may also request user-wired memory.
2081 if (vm_page_max_user_wired == 0)
2082 vm_page_max_user_wired = 4 * freecount / 5;
2086 * vm_pageout is the high level pageout daemon.
2093 int error, first, i;
2098 mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2099 swap_pager_swap_init();
2100 for (first = -1, i = 0; i < vm_ndomains; i++) {
2101 if (VM_DOMAIN_EMPTY(i)) {
2103 printf("domain %d empty; skipping pageout\n",
2110 error = kthread_add(vm_pageout_worker,
2111 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2113 panic("starting pageout for domain %d: %d\n",
2116 error = kthread_add(vm_pageout_laundry_worker,
2117 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2119 panic("starting laundry for domain %d: %d", i, error);
2121 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2123 panic("starting uma_reclaim helper, error %d\n", error);
2125 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2126 vm_pageout_worker((void *)(uintptr_t)first);
2130 * Perform an advisory wakeup of the page daemon.
2133 pagedaemon_wakeup(int domain)
2135 struct vm_domain *vmd;
2137 vmd = VM_DOMAIN(domain);
2138 vm_domain_pageout_assert_unlocked(vmd);
2139 if (curproc == pageproc)
2142 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2143 vm_domain_pageout_lock(vmd);
2144 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2145 wakeup(&vmd->vmd_pageout_wanted);
2146 vm_domain_pageout_unlock(vmd);