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 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
198 SYSCTL_INT(_vm, OID_AUTO, max_wired,
199 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
201 static u_int isqrt(u_int num);
202 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
204 static void vm_pageout_laundry_worker(void *arg);
207 struct vm_batchqueue bq;
208 struct vm_pagequeue *pq;
215 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
216 vm_page_t marker, vm_page_t after, int maxscan)
219 vm_pagequeue_assert_locked(pq);
220 KASSERT((marker->aflags & PGA_ENQUEUED) == 0,
221 ("marker %p already enqueued", marker));
224 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
226 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
227 vm_page_aflag_set(marker, PGA_ENQUEUED);
229 vm_batchqueue_init(&ss->bq);
232 ss->maxscan = maxscan;
234 vm_pagequeue_unlock(pq);
238 vm_pageout_end_scan(struct scan_state *ss)
240 struct vm_pagequeue *pq;
243 vm_pagequeue_assert_locked(pq);
244 KASSERT((ss->marker->aflags & PGA_ENQUEUED) != 0,
245 ("marker %p not enqueued", ss->marker));
247 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
248 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
249 pq->pq_pdpages += ss->scanned;
253 * Add a small number of queued pages to a batch queue for later processing
254 * without the corresponding queue lock held. The caller must have enqueued a
255 * marker page at the desired start point for the scan. Pages will be
256 * physically dequeued if the caller so requests. Otherwise, the returned
257 * batch may contain marker pages, and it is up to the caller to handle them.
259 * When processing the batch queue, vm_page_queue() must be used to
260 * determine whether the page has been logically dequeued by another thread.
261 * Once this check is performed, the page lock guarantees that the page will
262 * not be disassociated from the queue.
264 static __always_inline void
265 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
267 struct vm_pagequeue *pq;
273 KASSERT((marker->aflags & PGA_ENQUEUED) != 0,
274 ("marker %p not enqueued", ss->marker));
276 vm_pagequeue_lock(pq);
277 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
278 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
279 m = TAILQ_NEXT(m, plinks.q), ss->scanned++) {
280 if ((m->flags & PG_MARKER) == 0) {
281 KASSERT((m->aflags & PGA_ENQUEUED) != 0,
282 ("page %p not enqueued", m));
283 KASSERT((m->flags & PG_FICTITIOUS) == 0,
284 ("Fictitious page %p cannot be in page queue", m));
285 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
286 ("Unmanaged page %p cannot be in page queue", m));
290 (void)vm_batchqueue_insert(&ss->bq, m);
292 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
293 vm_page_aflag_clear(m, PGA_ENQUEUED);
296 TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
297 if (__predict_true(m != NULL))
298 TAILQ_INSERT_BEFORE(m, marker, plinks.q);
300 TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
302 vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
303 vm_pagequeue_unlock(pq);
306 /* Return the next page to be scanned, or NULL if the scan is complete. */
307 static __always_inline vm_page_t
308 vm_pageout_next(struct scan_state *ss, const bool dequeue)
311 if (ss->bq.bq_cnt == 0)
312 vm_pageout_collect_batch(ss, dequeue);
313 return (vm_batchqueue_pop(&ss->bq));
317 * Scan for pages at adjacent offsets within the given page's object that are
318 * eligible for laundering, form a cluster of these pages and the given page,
319 * and launder that cluster.
322 vm_pageout_cluster(vm_page_t m)
325 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
327 int ib, is, page_base, pageout_count;
329 vm_page_assert_locked(m);
331 VM_OBJECT_ASSERT_WLOCKED(object);
334 vm_page_assert_unbusied(m);
335 KASSERT(!vm_page_held(m), ("page %p is held", m));
337 pmap_remove_write(m);
340 mc[vm_pageout_page_count] = pb = ps = m;
342 page_base = vm_pageout_page_count;
347 * We can cluster only if the page is not clean, busy, or held, and
348 * the page is in the laundry queue.
350 * During heavy mmap/modification loads the pageout
351 * daemon can really fragment the underlying file
352 * due to flushing pages out of order and not trying to
353 * align the clusters (which leaves sporadic out-of-order
354 * holes). To solve this problem we do the reverse scan
355 * first and attempt to align our cluster, then do a
356 * forward scan if room remains.
359 while (ib != 0 && pageout_count < vm_pageout_page_count) {
364 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
368 vm_page_test_dirty(p);
374 if (vm_page_held(p) || !vm_page_in_laundry(p)) {
379 pmap_remove_write(p);
381 mc[--page_base] = pb = p;
386 * We are at an alignment boundary. Stop here, and switch
387 * directions. Do not clear ib.
389 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
392 while (pageout_count < vm_pageout_page_count &&
393 pindex + is < object->size) {
394 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
396 vm_page_test_dirty(p);
400 if (vm_page_held(p) || !vm_page_in_laundry(p)) {
404 pmap_remove_write(p);
406 mc[page_base + pageout_count] = ps = p;
412 * If we exhausted our forward scan, continue with the reverse scan
413 * when possible, even past an alignment boundary. This catches
414 * boundary conditions.
416 if (ib != 0 && pageout_count < vm_pageout_page_count)
419 return (vm_pageout_flush(&mc[page_base], pageout_count,
420 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
424 * vm_pageout_flush() - launder the given pages
426 * The given pages are laundered. Note that we setup for the start of
427 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
428 * reference count all in here rather then in the parent. If we want
429 * the parent to do more sophisticated things we may have to change
432 * Returned runlen is the count of pages between mreq and first
433 * page after mreq with status VM_PAGER_AGAIN.
434 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
435 * for any page in runlen set.
438 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
441 vm_object_t object = mc[0]->object;
442 int pageout_status[count];
446 VM_OBJECT_ASSERT_WLOCKED(object);
449 * Initiate I/O. Mark the pages busy and verify that they're valid
452 * We do not have to fixup the clean/dirty bits here... we can
453 * allow the pager to do it after the I/O completes.
455 * NOTE! mc[i]->dirty may be partial or fragmented due to an
456 * edge case with file fragments.
458 for (i = 0; i < count; i++) {
459 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
460 ("vm_pageout_flush: partially invalid page %p index %d/%d",
462 KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
463 ("vm_pageout_flush: writeable page %p", mc[i]));
464 vm_page_sbusy(mc[i]);
466 vm_object_pip_add(object, count);
468 vm_pager_put_pages(object, mc, count, flags, pageout_status);
470 runlen = count - mreq;
473 for (i = 0; i < count; i++) {
474 vm_page_t mt = mc[i];
476 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
477 !pmap_page_is_write_mapped(mt),
478 ("vm_pageout_flush: page %p is not write protected", mt));
479 switch (pageout_status[i]) {
482 if (vm_page_in_laundry(mt))
483 vm_page_deactivate_noreuse(mt);
491 * The page is outside the object's range. We pretend
492 * that the page out worked and clean the page, so the
493 * changes will be lost if the page is reclaimed by
498 if (vm_page_in_laundry(mt))
499 vm_page_deactivate_noreuse(mt);
505 * If the page couldn't be paged out to swap because the
506 * pager wasn't able to find space, place the page in
507 * the PQ_UNSWAPPABLE holding queue. This is an
508 * optimization that prevents the page daemon from
509 * wasting CPU cycles on pages that cannot be reclaimed
510 * becase no swap device is configured.
512 * Otherwise, reactivate the page so that it doesn't
513 * clog the laundry and inactive queues. (We will try
514 * paging it out again later.)
517 if (object->type == OBJT_SWAP &&
518 pageout_status[i] == VM_PAGER_FAIL) {
519 vm_page_unswappable(mt);
522 vm_page_activate(mt);
524 if (eio != NULL && i >= mreq && i - mreq < runlen)
528 if (i >= mreq && i - mreq < runlen)
534 * If the operation is still going, leave the page busy to
535 * block all other accesses. Also, leave the paging in
536 * progress indicator set so that we don't attempt an object
539 if (pageout_status[i] != VM_PAGER_PEND) {
540 vm_object_pip_wakeup(object);
546 return (numpagedout);
550 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
553 atomic_store_rel_int(&swapdev_enabled, 1);
557 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
560 if (swap_pager_nswapdev() == 1)
561 atomic_store_rel_int(&swapdev_enabled, 0);
565 * Attempt to acquire all of the necessary locks to launder a page and
566 * then call through the clustering layer to PUTPAGES. Wait a short
567 * time for a vnode lock.
569 * Requires the page and object lock on entry, releases both before return.
570 * Returns 0 on success and an errno otherwise.
573 vm_pageout_clean(vm_page_t m, int *numpagedout)
581 vm_page_assert_locked(m);
583 VM_OBJECT_ASSERT_WLOCKED(object);
589 * The object is already known NOT to be dead. It
590 * is possible for the vget() to block the whole
591 * pageout daemon, but the new low-memory handling
592 * code should prevent it.
594 * We can't wait forever for the vnode lock, we might
595 * deadlock due to a vn_read() getting stuck in
596 * vm_wait while holding this vnode. We skip the
597 * vnode if we can't get it in a reasonable amount
600 if (object->type == OBJT_VNODE) {
603 if (vp->v_type == VREG &&
604 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
610 ("vp %p with NULL v_mount", vp));
611 vm_object_reference_locked(object);
613 VM_OBJECT_WUNLOCK(object);
614 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
615 LK_SHARED : LK_EXCLUSIVE;
616 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
621 VM_OBJECT_WLOCK(object);
624 * Ensure that the object and vnode were not disassociated
625 * while locks were dropped.
627 if (vp->v_object != object) {
634 * While the object and page were unlocked, the page
636 * (1) moved to a different queue,
637 * (2) reallocated to a different object,
638 * (3) reallocated to a different offset, or
641 if (!vm_page_in_laundry(m) || m->object != object ||
642 m->pindex != pindex || m->dirty == 0) {
649 * The page may have been busied or referenced while the object
650 * and page locks were released.
652 if (vm_page_busied(m) || vm_page_held(m)) {
660 * If a page is dirty, then it is either being washed
661 * (but not yet cleaned) or it is still in the
662 * laundry. If it is still in the laundry, then we
663 * start the cleaning operation.
665 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
669 VM_OBJECT_WUNLOCK(object);
672 vm_page_lock_assert(m, MA_NOTOWNED);
676 vm_object_deallocate(object);
677 vn_finished_write(mp);
684 * Attempt to launder the specified number of pages.
686 * Returns the number of pages successfully laundered.
689 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
691 struct scan_state ss;
692 struct vm_pagequeue *pq;
696 int act_delta, error, numpagedout, queue, starting_target;
702 starting_target = launder;
706 * Scan the laundry queues for pages eligible to be laundered. We stop
707 * once the target number of dirty pages have been laundered, or once
708 * we've reached the end of the queue. A single iteration of this loop
709 * may cause more than one page to be laundered because of clustering.
711 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
712 * swap devices are configured.
714 if (atomic_load_acq_int(&swapdev_enabled))
715 queue = PQ_UNSWAPPABLE;
720 marker = &vmd->vmd_markers[queue];
721 pq = &vmd->vmd_pagequeues[queue];
722 vm_pagequeue_lock(pq);
723 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
724 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
725 if (__predict_false((m->flags & PG_MARKER) != 0))
728 vm_page_change_lock(m, &mtx);
732 * The page may have been disassociated from the queue
733 * while locks were dropped.
735 if (vm_page_queue(m) != queue)
739 * A requeue was requested, so this page gets a second
742 if ((m->aflags & PGA_REQUEUE) != 0) {
748 * Held pages are essentially stuck in the queue.
750 * Wired pages may not be freed. Complete their removal
751 * from the queue now to avoid needless revisits during
754 if (m->hold_count != 0)
756 if (m->wire_count != 0) {
757 vm_page_dequeue_deferred(m);
761 if (object != m->object) {
763 VM_OBJECT_WUNLOCK(object);
765 if (!VM_OBJECT_TRYWLOCK(object)) {
767 /* Depends on type-stability. */
768 VM_OBJECT_WLOCK(object);
774 if (vm_page_busied(m))
778 * Invalid pages can be easily freed. They cannot be
779 * mapped; vm_page_free() asserts this.
785 * If the page has been referenced and the object is not dead,
786 * reactivate or requeue the page depending on whether the
789 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
790 * that a reference from a concurrently destroyed mapping is
791 * observed here and now.
793 if (object->ref_count != 0)
794 act_delta = pmap_ts_referenced(m);
796 KASSERT(!pmap_page_is_mapped(m),
797 ("page %p is mapped", m));
800 if ((m->aflags & PGA_REFERENCED) != 0) {
801 vm_page_aflag_clear(m, PGA_REFERENCED);
804 if (act_delta != 0) {
805 if (object->ref_count != 0) {
806 VM_CNT_INC(v_reactivated);
810 * Increase the activation count if the page
811 * was referenced while in the laundry queue.
812 * This makes it less likely that the page will
813 * be returned prematurely to the inactive
816 m->act_count += act_delta + ACT_ADVANCE;
819 * If this was a background laundering, count
820 * activated pages towards our target. The
821 * purpose of background laundering is to ensure
822 * that pages are eventually cycled through the
823 * laundry queue, and an activation is a valid
829 } else if ((object->flags & OBJ_DEAD) == 0) {
836 * If the page appears to be clean at the machine-independent
837 * layer, then remove all of its mappings from the pmap in
838 * anticipation of freeing it. If, however, any of the page's
839 * mappings allow write access, then the page may still be
840 * modified until the last of those mappings are removed.
842 if (object->ref_count != 0) {
843 vm_page_test_dirty(m);
849 * Clean pages are freed, and dirty pages are paged out unless
850 * they belong to a dead object. Requeueing dirty pages from
851 * dead objects is pointless, as they are being paged out and
852 * freed by the thread that destroyed the object.
858 } else if ((object->flags & OBJ_DEAD) == 0) {
859 if (object->type != OBJT_SWAP &&
860 object->type != OBJT_DEFAULT)
862 else if (disable_swap_pageouts)
872 * Form a cluster with adjacent, dirty pages from the
873 * same object, and page out that entire cluster.
875 * The adjacent, dirty pages must also be in the
876 * laundry. However, their mappings are not checked
877 * for new references. Consequently, a recently
878 * referenced page may be paged out. However, that
879 * page will not be prematurely reclaimed. After page
880 * out, the page will be placed in the inactive queue,
881 * where any new references will be detected and the
884 error = vm_pageout_clean(m, &numpagedout);
886 launder -= numpagedout;
887 ss.scanned += numpagedout;
888 } else if (error == EDEADLK) {
900 if (object != NULL) {
901 VM_OBJECT_WUNLOCK(object);
904 vm_pagequeue_lock(pq);
905 vm_pageout_end_scan(&ss);
906 vm_pagequeue_unlock(pq);
908 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
914 * Wakeup the sync daemon if we skipped a vnode in a writeable object
915 * and we didn't launder enough pages.
917 if (vnodes_skipped > 0 && launder > 0)
918 (void)speedup_syncer();
920 return (starting_target - launder);
924 * Compute the integer square root.
929 u_int bit, root, tmp;
931 bit = 1u << ((NBBY * sizeof(u_int)) - 2);
948 * Perform the work of the laundry thread: periodically wake up and determine
949 * whether any pages need to be laundered. If so, determine the number of pages
950 * that need to be laundered, and launder them.
953 vm_pageout_laundry_worker(void *arg)
955 struct vm_domain *vmd;
956 struct vm_pagequeue *pq;
957 uint64_t nclean, ndirty, nfreed;
958 int domain, last_target, launder, shortfall, shortfall_cycle, target;
961 domain = (uintptr_t)arg;
962 vmd = VM_DOMAIN(domain);
963 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
964 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
967 in_shortfall = false;
969 last_target = target = 0;
973 * Calls to these handlers are serialized by the swap syscall lock.
975 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
976 EVENTHANDLER_PRI_ANY);
977 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
978 EVENTHANDLER_PRI_ANY);
981 * The pageout laundry worker is never done, so loop forever.
984 KASSERT(target >= 0, ("negative target %d", target));
985 KASSERT(shortfall_cycle >= 0,
986 ("negative cycle %d", shortfall_cycle));
990 * First determine whether we need to launder pages to meet a
991 * shortage of free pages.
995 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
997 } else if (!in_shortfall)
999 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1001 * We recently entered shortfall and began laundering
1002 * pages. If we have completed that laundering run
1003 * (and we are no longer in shortfall) or we have met
1004 * our laundry target through other activity, then we
1005 * can stop laundering pages.
1007 in_shortfall = false;
1011 launder = target / shortfall_cycle--;
1015 * There's no immediate need to launder any pages; see if we
1016 * meet the conditions to perform background laundering:
1018 * 1. The ratio of dirty to clean inactive pages exceeds the
1019 * background laundering threshold, or
1020 * 2. we haven't yet reached the target of the current
1021 * background laundering run.
1023 * The background laundering threshold is not a constant.
1024 * Instead, it is a slowly growing function of the number of
1025 * clean pages freed by the page daemon since the last
1026 * background laundering. Thus, as the ratio of dirty to
1027 * clean inactive pages grows, the amount of memory pressure
1028 * required to trigger laundering decreases. We ensure
1029 * that the threshold is non-zero after an inactive queue
1030 * scan, even if that scan failed to free a single clean page.
1033 nclean = vmd->vmd_free_count +
1034 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1035 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1036 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1037 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1038 target = vmd->vmd_background_launder_target;
1042 * We have a non-zero background laundering target. If we've
1043 * laundered up to our maximum without observing a page daemon
1044 * request, just stop. This is a safety belt that ensures we
1045 * don't launder an excessive amount if memory pressure is low
1046 * and the ratio of dirty to clean pages is large. Otherwise,
1047 * proceed at the background laundering rate.
1052 last_target = target;
1053 } else if (last_target - target >=
1054 vm_background_launder_max * PAGE_SIZE / 1024) {
1057 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1058 launder /= VM_LAUNDER_RATE;
1059 if (launder > target)
1066 * Because of I/O clustering, the number of laundered
1067 * pages could exceed "target" by the maximum size of
1068 * a cluster minus one.
1070 target -= min(vm_pageout_launder(vmd, launder,
1071 in_shortfall), target);
1072 pause("laundp", hz / VM_LAUNDER_RATE);
1076 * If we're not currently laundering pages and the page daemon
1077 * hasn't posted a new request, sleep until the page daemon
1080 vm_pagequeue_lock(pq);
1081 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1082 (void)mtx_sleep(&vmd->vmd_laundry_request,
1083 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1086 * If the pagedaemon has indicated that it's in shortfall, start
1087 * a shortfall laundering unless we're already in the middle of
1088 * one. This may preempt a background laundering.
1090 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1091 (!in_shortfall || shortfall_cycle == 0)) {
1092 shortfall = vm_laundry_target(vmd) +
1093 vmd->vmd_pageout_deficit;
1099 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1100 nfreed += vmd->vmd_clean_pages_freed;
1101 vmd->vmd_clean_pages_freed = 0;
1102 vm_pagequeue_unlock(pq);
1107 * Compute the number of pages we want to try to move from the
1108 * active queue to either the inactive or laundry queue.
1110 * When scanning active pages during a shortage, we make clean pages
1111 * count more heavily towards the page shortage than dirty pages.
1112 * This is because dirty pages must be laundered before they can be
1113 * reused and thus have less utility when attempting to quickly
1114 * alleviate a free page shortage. However, this weighting also
1115 * causes the scan to deactivate dirty pages more aggressively,
1116 * improving the effectiveness of clustering.
1119 vm_pageout_active_target(struct vm_domain *vmd)
1123 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1124 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1125 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1126 shortage *= act_scan_laundry_weight;
1131 * Scan the active queue. If there is no shortage of inactive pages, scan a
1132 * small portion of the queue in order to maintain quasi-LRU.
1135 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1137 struct scan_state ss;
1139 vm_page_t m, marker;
1140 struct vm_pagequeue *pq;
1142 int act_delta, max_scan, scan_tick;
1144 marker = &vmd->vmd_markers[PQ_ACTIVE];
1145 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1146 vm_pagequeue_lock(pq);
1149 * If we're just idle polling attempt to visit every
1150 * active page within 'update_period' seconds.
1153 if (vm_pageout_update_period != 0) {
1154 min_scan = pq->pq_cnt;
1155 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1156 min_scan /= hz * vm_pageout_update_period;
1159 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1160 vmd->vmd_last_active_scan = scan_tick;
1163 * Scan the active queue for pages that can be deactivated. Update
1164 * the per-page activity counter and use it to identify deactivation
1165 * candidates. Held pages may be deactivated.
1167 * To avoid requeuing each page that remains in the active queue, we
1168 * implement the CLOCK algorithm. To keep the implementation of the
1169 * enqueue operation consistent for all page queues, we use two hands,
1170 * represented by marker pages. Scans begin at the first hand, which
1171 * precedes the second hand in the queue. When the two hands meet,
1172 * they are moved back to the head and tail of the queue, respectively,
1173 * and scanning resumes.
1175 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1178 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1179 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1180 if (__predict_false(m == &vmd->vmd_clock[1])) {
1181 vm_pagequeue_lock(pq);
1182 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1183 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1184 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1186 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1188 max_scan -= ss.scanned;
1189 vm_pageout_end_scan(&ss);
1192 if (__predict_false((m->flags & PG_MARKER) != 0))
1195 vm_page_change_lock(m, &mtx);
1198 * The page may have been disassociated from the queue
1199 * while locks were dropped.
1201 if (vm_page_queue(m) != PQ_ACTIVE)
1205 * Wired pages are dequeued lazily.
1207 if (m->wire_count != 0) {
1208 vm_page_dequeue_deferred(m);
1213 * Check to see "how much" the page has been used.
1215 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1216 * that a reference from a concurrently destroyed mapping is
1217 * observed here and now.
1219 * Perform an unsynchronized object ref count check. While
1220 * the page lock ensures that the page is not reallocated to
1221 * another object, in particular, one with unmanaged mappings
1222 * that cannot support pmap_ts_referenced(), two races are,
1223 * nonetheless, possible:
1224 * 1) The count was transitioning to zero, but we saw a non-
1225 * zero value. pmap_ts_referenced() will return zero
1226 * because the page is not mapped.
1227 * 2) The count was transitioning to one, but we saw zero.
1228 * This race delays the detection of a new reference. At
1229 * worst, we will deactivate and reactivate the page.
1231 if (m->object->ref_count != 0)
1232 act_delta = pmap_ts_referenced(m);
1235 if ((m->aflags & PGA_REFERENCED) != 0) {
1236 vm_page_aflag_clear(m, PGA_REFERENCED);
1241 * Advance or decay the act_count based on recent usage.
1243 if (act_delta != 0) {
1244 m->act_count += ACT_ADVANCE + act_delta;
1245 if (m->act_count > ACT_MAX)
1246 m->act_count = ACT_MAX;
1248 m->act_count -= min(m->act_count, ACT_DECLINE);
1250 if (m->act_count == 0) {
1252 * When not short for inactive pages, let dirty pages go
1253 * through the inactive queue before moving to the
1254 * laundry queues. This gives them some extra time to
1255 * be reactivated, potentially avoiding an expensive
1256 * pageout. However, during a page shortage, the
1257 * inactive queue is necessarily small, and so dirty
1258 * pages would only spend a trivial amount of time in
1259 * the inactive queue. Therefore, we might as well
1260 * place them directly in the laundry queue to reduce
1263 if (page_shortage <= 0)
1264 vm_page_deactivate(m);
1267 * Calling vm_page_test_dirty() here would
1268 * require acquisition of the object's write
1269 * lock. However, during a page shortage,
1270 * directing dirty pages into the laundry
1271 * queue is only an optimization and not a
1272 * requirement. Therefore, we simply rely on
1273 * the opportunistic updates to the page's
1274 * dirty field by the pmap.
1276 if (m->dirty == 0) {
1277 vm_page_deactivate(m);
1279 act_scan_laundry_weight;
1291 vm_pagequeue_lock(pq);
1292 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1293 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1294 vm_pageout_end_scan(&ss);
1295 vm_pagequeue_unlock(pq);
1299 vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m)
1301 struct vm_domain *vmd;
1303 if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0)
1305 vm_page_aflag_set(m, PGA_ENQUEUED);
1306 if ((m->aflags & PGA_REQUEUE_HEAD) != 0) {
1307 vmd = vm_pagequeue_domain(m);
1308 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
1309 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1310 } else if ((m->aflags & PGA_REQUEUE) != 0) {
1311 TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q);
1312 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1314 TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q);
1319 * Re-add stuck pages to the inactive queue. We will examine them again
1320 * during the next scan. If the queue state of a page has changed since
1321 * it was physically removed from the page queue in
1322 * vm_pageout_collect_batch(), don't do anything with that page.
1325 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1328 struct vm_pagequeue *pq;
1335 if (vm_batchqueue_insert(bq, m))
1337 vm_pagequeue_lock(pq);
1338 delta += vm_pageout_reinsert_inactive_page(ss, m);
1340 vm_pagequeue_lock(pq);
1341 while ((m = vm_batchqueue_pop(bq)) != NULL)
1342 delta += vm_pageout_reinsert_inactive_page(ss, m);
1343 vm_pagequeue_cnt_add(pq, delta);
1344 vm_pagequeue_unlock(pq);
1345 vm_batchqueue_init(bq);
1349 * Attempt to reclaim the requested number of pages from the inactive queue.
1350 * Returns true if the shortage was addressed.
1353 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1356 struct scan_state ss;
1357 struct vm_batchqueue rq;
1359 vm_page_t m, marker;
1360 struct vm_pagequeue *pq;
1362 int act_delta, addl_page_shortage, deficit, page_shortage;
1363 int starting_page_shortage;
1366 * The addl_page_shortage is an estimate of the number of temporarily
1367 * stuck pages in the inactive queue. In other words, the
1368 * number of pages from the inactive count that should be
1369 * discounted in setting the target for the active queue scan.
1371 addl_page_shortage = 0;
1374 * vmd_pageout_deficit counts the number of pages requested in
1375 * allocations that failed because of a free page shortage. We assume
1376 * that the allocations will be reattempted and thus include the deficit
1377 * in our scan target.
1379 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1380 starting_page_shortage = page_shortage = shortage + deficit;
1384 vm_batchqueue_init(&rq);
1387 * Start scanning the inactive queue for pages that we can free. The
1388 * scan will stop when we reach the target or we have scanned the
1389 * entire queue. (Note that m->act_count is not used to make
1390 * decisions for the inactive queue, only for the active queue.)
1392 marker = &vmd->vmd_markers[PQ_INACTIVE];
1393 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1394 vm_pagequeue_lock(pq);
1395 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1396 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1397 KASSERT((m->flags & PG_MARKER) == 0,
1398 ("marker page %p was dequeued", m));
1400 vm_page_change_lock(m, &mtx);
1404 * The page may have been disassociated from the queue
1405 * while locks were dropped.
1407 if (vm_page_queue(m) != PQ_INACTIVE) {
1408 addl_page_shortage++;
1413 * The page was re-enqueued after the page queue lock was
1414 * dropped, or a requeue was requested. This page gets a second
1417 if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE |
1418 PGA_REQUEUE_HEAD)) != 0)
1422 * Held pages are essentially stuck in the queue. So,
1423 * they ought to be discounted from the inactive count.
1424 * See the description of addl_page_shortage above.
1426 * Wired pages may not be freed. Complete their removal
1427 * from the queue now to avoid needless revisits during
1430 if (m->hold_count != 0) {
1431 addl_page_shortage++;
1434 if (m->wire_count != 0) {
1435 vm_page_dequeue_deferred(m);
1439 if (object != m->object) {
1441 VM_OBJECT_WUNLOCK(object);
1443 if (!VM_OBJECT_TRYWLOCK(object)) {
1445 /* Depends on type-stability. */
1446 VM_OBJECT_WLOCK(object);
1452 if (vm_page_busied(m)) {
1454 * Don't mess with busy pages. Leave them at
1455 * the front of the queue. Most likely, they
1456 * are being paged out and will leave the
1457 * queue shortly after the scan finishes. So,
1458 * they ought to be discounted from the
1461 addl_page_shortage++;
1466 * Invalid pages can be easily freed. They cannot be
1467 * mapped, vm_page_free() asserts this.
1473 * If the page has been referenced and the object is not dead,
1474 * reactivate or requeue the page depending on whether the
1477 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1478 * that a reference from a concurrently destroyed mapping is
1479 * observed here and now.
1481 if (object->ref_count != 0)
1482 act_delta = pmap_ts_referenced(m);
1484 KASSERT(!pmap_page_is_mapped(m),
1485 ("page %p is mapped", m));
1488 if ((m->aflags & PGA_REFERENCED) != 0) {
1489 vm_page_aflag_clear(m, PGA_REFERENCED);
1492 if (act_delta != 0) {
1493 if (object->ref_count != 0) {
1494 VM_CNT_INC(v_reactivated);
1495 vm_page_activate(m);
1498 * Increase the activation count if the page
1499 * was referenced while in the inactive queue.
1500 * This makes it less likely that the page will
1501 * be returned prematurely to the inactive
1504 m->act_count += act_delta + ACT_ADVANCE;
1506 } else if ((object->flags & OBJ_DEAD) == 0) {
1507 vm_page_aflag_set(m, PGA_REQUEUE);
1513 * If the page appears to be clean at the machine-independent
1514 * layer, then remove all of its mappings from the pmap in
1515 * anticipation of freeing it. If, however, any of the page's
1516 * mappings allow write access, then the page may still be
1517 * modified until the last of those mappings are removed.
1519 if (object->ref_count != 0) {
1520 vm_page_test_dirty(m);
1526 * Clean pages can be freed, but dirty pages must be sent back
1527 * to the laundry, unless they belong to a dead object.
1528 * Requeueing dirty pages from dead objects is pointless, as
1529 * they are being paged out and freed by the thread that
1530 * destroyed the object.
1532 if (m->dirty == 0) {
1535 * Because we dequeued the page and have already
1536 * checked for concurrent dequeue and enqueue
1537 * requests, we can safely disassociate the page
1538 * from the inactive queue.
1540 KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
1541 ("page %p has queue state", m));
1545 } else if ((object->flags & OBJ_DEAD) == 0)
1549 vm_pageout_reinsert_inactive(&ss, &rq, m);
1554 VM_OBJECT_WUNLOCK(object);
1555 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1556 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1557 vm_pagequeue_lock(pq);
1558 vm_pageout_end_scan(&ss);
1559 vm_pagequeue_unlock(pq);
1561 VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1564 * Wake up the laundry thread so that it can perform any needed
1565 * laundering. If we didn't meet our target, we're in shortfall and
1566 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1567 * swap devices are configured, the laundry thread has no work to do, so
1568 * don't bother waking it up.
1570 * The laundry thread uses the number of inactive queue scans elapsed
1571 * since the last laundering to determine whether to launder again, so
1574 if (starting_page_shortage > 0) {
1575 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1576 vm_pagequeue_lock(pq);
1577 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1578 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1579 if (page_shortage > 0) {
1580 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1581 VM_CNT_INC(v_pdshortfalls);
1582 } else if (vmd->vmd_laundry_request !=
1583 VM_LAUNDRY_SHORTFALL)
1584 vmd->vmd_laundry_request =
1585 VM_LAUNDRY_BACKGROUND;
1586 wakeup(&vmd->vmd_laundry_request);
1588 vmd->vmd_clean_pages_freed +=
1589 starting_page_shortage - page_shortage;
1590 vm_pagequeue_unlock(pq);
1594 * Wakeup the swapout daemon if we didn't free the targeted number of
1597 if (page_shortage > 0)
1601 * If the inactive queue scan fails repeatedly to meet its
1602 * target, kill the largest process.
1604 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1607 * Reclaim pages by swapping out idle processes, if configured to do so.
1609 vm_swapout_run_idle();
1612 * See the description of addl_page_shortage above.
1614 *addl_shortage = addl_page_shortage + deficit;
1616 return (page_shortage <= 0);
1619 static int vm_pageout_oom_vote;
1622 * The pagedaemon threads randlomly select one to perform the
1623 * OOM. Trying to kill processes before all pagedaemons
1624 * failed to reach free target is premature.
1627 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1628 int starting_page_shortage)
1632 if (starting_page_shortage <= 0 || starting_page_shortage !=
1634 vmd->vmd_oom_seq = 0;
1637 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1639 vmd->vmd_oom = FALSE;
1640 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1646 * Do not follow the call sequence until OOM condition is
1649 vmd->vmd_oom_seq = 0;
1654 vmd->vmd_oom = TRUE;
1655 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1656 if (old_vote != vm_ndomains - 1)
1660 * The current pagedaemon thread is the last in the quorum to
1661 * start OOM. Initiate the selection and signaling of the
1664 vm_pageout_oom(VM_OOM_MEM);
1667 * After one round of OOM terror, recall our vote. On the
1668 * next pass, current pagedaemon would vote again if the low
1669 * memory condition is still there, due to vmd_oom being
1672 vmd->vmd_oom = FALSE;
1673 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1677 * The OOM killer is the page daemon's action of last resort when
1678 * memory allocation requests have been stalled for a prolonged period
1679 * of time because it cannot reclaim memory. This function computes
1680 * the approximate number of physical pages that could be reclaimed if
1681 * the specified address space is destroyed.
1683 * Private, anonymous memory owned by the address space is the
1684 * principal resource that we expect to recover after an OOM kill.
1685 * Since the physical pages mapped by the address space's COW entries
1686 * are typically shared pages, they are unlikely to be released and so
1687 * they are not counted.
1689 * To get to the point where the page daemon runs the OOM killer, its
1690 * efforts to write-back vnode-backed pages may have stalled. This
1691 * could be caused by a memory allocation deadlock in the write path
1692 * that might be resolved by an OOM kill. Therefore, physical pages
1693 * belonging to vnode-backed objects are counted, because they might
1694 * be freed without being written out first if the address space holds
1695 * the last reference to an unlinked vnode.
1697 * Similarly, physical pages belonging to OBJT_PHYS objects are
1698 * counted because the address space might hold the last reference to
1702 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1705 vm_map_entry_t entry;
1709 map = &vmspace->vm_map;
1710 KASSERT(!map->system_map, ("system map"));
1711 sx_assert(&map->lock, SA_LOCKED);
1713 for (entry = map->header.next; entry != &map->header;
1714 entry = entry->next) {
1715 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1717 obj = entry->object.vm_object;
1720 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1721 obj->ref_count != 1)
1723 switch (obj->type) {
1728 res += obj->resident_page_count;
1736 vm_pageout_oom(int shortage)
1738 struct proc *p, *bigproc;
1739 vm_offset_t size, bigsize;
1745 * We keep the process bigproc locked once we find it to keep anyone
1746 * from messing with it; however, there is a possibility of
1747 * deadlock if process B is bigproc and one of its child processes
1748 * attempts to propagate a signal to B while we are waiting for A's
1749 * lock while walking this list. To avoid this, we don't block on
1750 * the process lock but just skip a process if it is already locked.
1754 sx_slock(&allproc_lock);
1755 FOREACH_PROC_IN_SYSTEM(p) {
1759 * If this is a system, protected or killed process, skip it.
1761 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1762 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1763 p->p_pid == 1 || P_KILLED(p) ||
1764 (p->p_pid < 48 && swap_pager_avail != 0)) {
1769 * If the process is in a non-running type state,
1770 * don't touch it. Check all the threads individually.
1773 FOREACH_THREAD_IN_PROC(p, td) {
1775 if (!TD_ON_RUNQ(td) &&
1776 !TD_IS_RUNNING(td) &&
1777 !TD_IS_SLEEPING(td) &&
1778 !TD_IS_SUSPENDED(td) &&
1779 !TD_IS_SWAPPED(td)) {
1791 * get the process size
1793 vm = vmspace_acquire_ref(p);
1800 sx_sunlock(&allproc_lock);
1801 if (!vm_map_trylock_read(&vm->vm_map)) {
1803 sx_slock(&allproc_lock);
1807 size = vmspace_swap_count(vm);
1808 if (shortage == VM_OOM_MEM)
1809 size += vm_pageout_oom_pagecount(vm);
1810 vm_map_unlock_read(&vm->vm_map);
1812 sx_slock(&allproc_lock);
1815 * If this process is bigger than the biggest one,
1818 if (size > bigsize) {
1819 if (bigproc != NULL)
1827 sx_sunlock(&allproc_lock);
1828 if (bigproc != NULL) {
1829 if (vm_panic_on_oom != 0)
1830 panic("out of swap space");
1832 killproc(bigproc, "out of swap space");
1833 sched_nice(bigproc, PRIO_MIN);
1835 PROC_UNLOCK(bigproc);
1840 vm_pageout_lowmem(void)
1842 static int lowmem_ticks = 0;
1845 last = atomic_load_int(&lowmem_ticks);
1846 while ((u_int)(ticks - last) / hz >= lowmem_period) {
1847 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
1851 * Decrease registered cache sizes.
1853 SDT_PROBE0(vm, , , vm__lowmem_scan);
1854 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1857 * We do this explicitly after the caches have been
1867 vm_pageout_worker(void *arg)
1869 struct vm_domain *vmd;
1871 int addl_shortage, domain, shortage;
1874 domain = (uintptr_t)arg;
1875 vmd = VM_DOMAIN(domain);
1880 * XXXKIB It could be useful to bind pageout daemon threads to
1881 * the cores belonging to the domain, from which vm_page_array
1885 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1886 vmd->vmd_last_active_scan = ticks;
1889 * The pageout daemon worker is never done, so loop forever.
1892 vm_domain_pageout_lock(vmd);
1895 * We need to clear wanted before we check the limits. This
1896 * prevents races with wakers who will check wanted after they
1899 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
1902 * Might the page daemon need to run again?
1904 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
1906 * Yes. If the scan failed to produce enough free
1907 * pages, sleep uninterruptibly for some time in the
1908 * hope that the laundry thread will clean some pages.
1910 vm_domain_pageout_unlock(vmd);
1912 pause("pwait", hz / VM_INACT_SCAN_RATE);
1915 * No, sleep until the next wakeup or until pages
1916 * need to have their reference stats updated.
1918 if (mtx_sleep(&vmd->vmd_pageout_wanted,
1919 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
1920 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
1921 VM_CNT_INC(v_pdwakeups);
1924 /* Prevent spurious wakeups by ensuring that wanted is set. */
1925 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1928 * Use the controller to calculate how many pages to free in
1929 * this interval, and scan the inactive queue. If the lowmem
1930 * handlers appear to have freed up some pages, subtract the
1931 * difference from the inactive queue scan target.
1933 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
1935 ofree = vmd->vmd_free_count;
1936 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
1937 shortage -= min(vmd->vmd_free_count - ofree,
1939 target_met = vm_pageout_scan_inactive(vmd, shortage,
1945 * Scan the active queue. A positive value for shortage
1946 * indicates that we must aggressively deactivate pages to avoid
1949 shortage = vm_pageout_active_target(vmd) + addl_shortage;
1950 vm_pageout_scan_active(vmd, shortage);
1955 * vm_pageout_init initialises basic pageout daemon settings.
1958 vm_pageout_init_domain(int domain)
1960 struct vm_domain *vmd;
1961 struct sysctl_oid *oid;
1963 vmd = VM_DOMAIN(domain);
1964 vmd->vmd_interrupt_free_min = 2;
1967 * v_free_reserved needs to include enough for the largest
1968 * swap pager structures plus enough for any pv_entry structs
1971 if (vmd->vmd_page_count > 1024)
1972 vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200;
1974 vmd->vmd_free_min = 4;
1975 vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1976 vmd->vmd_interrupt_free_min;
1977 vmd->vmd_free_reserved = vm_pageout_page_count +
1978 vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768);
1979 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
1980 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
1981 vmd->vmd_free_min += vmd->vmd_free_reserved;
1982 vmd->vmd_free_severe += vmd->vmd_free_reserved;
1983 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
1984 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
1985 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
1988 * Set the default wakeup threshold to be 10% below the paging
1989 * target. This keeps the steady state out of shortfall.
1991 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
1994 * Target amount of memory to move out of the laundry queue during a
1995 * background laundering. This is proportional to the amount of system
1998 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
1999 vmd->vmd_free_min) / 10;
2001 /* Initialize the pageout daemon pid controller. */
2002 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2003 vmd->vmd_free_target, PIDCTRL_BOUND,
2004 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2005 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2006 "pidctrl", CTLFLAG_RD, NULL, "");
2007 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2011 vm_pageout_init(void)
2017 * Initialize some paging parameters.
2019 if (vm_cnt.v_page_count < 2000)
2020 vm_pageout_page_count = 8;
2023 for (i = 0; i < vm_ndomains; i++) {
2024 struct vm_domain *vmd;
2026 vm_pageout_init_domain(i);
2028 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2029 vm_cnt.v_free_target += vmd->vmd_free_target;
2030 vm_cnt.v_free_min += vmd->vmd_free_min;
2031 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2032 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2033 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2034 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2035 freecount += vmd->vmd_free_count;
2039 * Set interval in seconds for active scan. We want to visit each
2040 * page at least once every ten minutes. This is to prevent worst
2041 * case paging behaviors with stale active LRU.
2043 if (vm_pageout_update_period == 0)
2044 vm_pageout_update_period = 600;
2046 if (vm_page_max_wired == 0)
2047 vm_page_max_wired = freecount / 3;
2051 * vm_pageout is the high level pageout daemon.
2058 int error, first, i;
2063 swap_pager_swap_init();
2064 for (first = -1, i = 0; i < vm_ndomains; i++) {
2065 if (VM_DOMAIN_EMPTY(i)) {
2067 printf("domain %d empty; skipping pageout\n",
2074 error = kthread_add(vm_pageout_worker,
2075 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2077 panic("starting pageout for domain %d: %d\n",
2080 error = kthread_add(vm_pageout_laundry_worker,
2081 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2083 panic("starting laundry for domain %d: %d", i, error);
2085 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2087 panic("starting uma_reclaim helper, error %d\n", error);
2089 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2090 vm_pageout_worker((void *)(uintptr_t)first);
2094 * Perform an advisory wakeup of the page daemon.
2097 pagedaemon_wakeup(int domain)
2099 struct vm_domain *vmd;
2101 vmd = VM_DOMAIN(domain);
2102 vm_domain_pageout_assert_unlocked(vmd);
2103 if (curproc == pageproc)
2106 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2107 vm_domain_pageout_lock(vmd);
2108 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2109 wakeup(&vmd->vmd_pageout_wanted);
2110 vm_domain_pageout_unlock(vmd);