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 bool vm_pageout_scan(struct vm_domain *vmd, int pass, int shortage);
128 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
129 int starting_page_shortage);
131 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
134 struct proc *pageproc;
136 static struct kproc_desc page_kp = {
141 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
144 SDT_PROVIDER_DEFINE(vm);
145 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
147 /* Pagedaemon activity rates, in subdivisions of one second. */
148 #define VM_LAUNDER_RATE 10
149 #define VM_INACT_SCAN_RATE 10
151 static int vm_pageout_oom_seq = 12;
153 static int vm_pageout_update_period;
154 static int disable_swap_pageouts;
155 static int lowmem_period = 10;
156 static time_t lowmem_uptime;
157 static int swapdev_enabled;
159 static int vm_panic_on_oom = 0;
161 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
162 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
163 "panic on out of memory instead of killing the largest process");
165 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
166 CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
167 "Maximum active LRU update period");
169 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
170 "Low memory callback period");
172 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
173 CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
175 static int pageout_lock_miss;
176 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
177 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
179 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
180 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
181 "back-to-back calls to oom detector to start OOM");
183 static int act_scan_laundry_weight = 3;
184 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
185 &act_scan_laundry_weight, 0,
186 "weight given to clean vs. dirty pages in active queue scans");
188 static u_int vm_background_launder_rate = 4096;
189 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
190 &vm_background_launder_rate, 0,
191 "background laundering rate, in kilobytes per second");
193 static u_int vm_background_launder_max = 20 * 1024;
194 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
195 &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
197 int vm_pageout_page_count = 32;
199 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
200 SYSCTL_INT(_vm, OID_AUTO, max_wired,
201 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
203 static u_int isqrt(u_int num);
204 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
206 static void vm_pageout_laundry_worker(void *arg);
209 struct vm_batchqueue bq;
210 struct vm_pagequeue *pq;
217 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
218 vm_page_t marker, vm_page_t after, int maxscan)
221 vm_pagequeue_assert_locked(pq);
222 KASSERT((marker->aflags & PGA_ENQUEUED) == 0,
223 ("marker %p already enqueued", marker));
226 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
228 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
229 vm_page_aflag_set(marker, PGA_ENQUEUED);
231 vm_batchqueue_init(&ss->bq);
234 ss->maxscan = maxscan;
236 vm_pagequeue_unlock(pq);
240 vm_pageout_end_scan(struct scan_state *ss)
242 struct vm_pagequeue *pq;
245 vm_pagequeue_assert_locked(pq);
246 KASSERT((ss->marker->aflags & PGA_ENQUEUED) != 0,
247 ("marker %p not enqueued", ss->marker));
249 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
250 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
251 VM_CNT_ADD(v_pdpages, ss->scanned);
255 * Add a small number of queued pages to a batch queue for later processing
256 * without the corresponding queue lock held. The caller must have enqueued a
257 * marker page at the desired start point for the scan. Pages will be
258 * physically dequeued if the caller so requests. Otherwise, the returned
259 * batch may contain marker pages, and it is up to the caller to handle them.
261 * When processing the batch queue, vm_page_queue() must be used to
262 * determine whether the page has been logically dequeued by another thread.
263 * Once this check is performed, the page lock guarantees that the page will
264 * not be disassociated from the queue.
266 static __always_inline void
267 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
269 struct vm_pagequeue *pq;
275 KASSERT((marker->aflags & PGA_ENQUEUED) != 0,
276 ("marker %p not enqueued", ss->marker));
278 vm_pagequeue_lock(pq);
279 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
280 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
281 m = TAILQ_NEXT(m, plinks.q), ss->scanned++) {
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;
700 bool obj_locked, pageout_ok;
705 starting_target = launder;
709 * Scan the laundry queues for pages eligible to be laundered. We stop
710 * once the target number of dirty pages have been laundered, or once
711 * we've reached the end of the queue. A single iteration of this loop
712 * may cause more than one page to be laundered because of clustering.
714 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
715 * swap devices are configured.
717 if (atomic_load_acq_int(&swapdev_enabled))
718 queue = PQ_UNSWAPPABLE;
723 marker = &vmd->vmd_markers[queue];
724 pq = &vmd->vmd_pagequeues[queue];
725 vm_pagequeue_lock(pq);
726 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
727 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
728 if (__predict_false((m->flags & PG_MARKER) != 0))
731 vm_page_change_lock(m, &mtx);
735 * The page may have been disassociated from the queue
736 * while locks were dropped.
738 if (vm_page_queue(m) != queue)
742 * A requeue was requested, so this page gets a second
745 if ((m->aflags & PGA_REQUEUE) != 0) {
751 * Held pages are essentially stuck in the queue.
753 * Wired pages may not be freed. Complete their removal
754 * from the queue now to avoid needless revisits during
757 if (m->hold_count != 0)
759 if (m->wire_count != 0) {
760 vm_page_dequeue_deferred(m);
764 if (object != m->object) {
766 VM_OBJECT_WUNLOCK(object);
772 if (!VM_OBJECT_TRYWLOCK(object)) {
774 /* Depends on type-stability. */
775 VM_OBJECT_WLOCK(object);
783 if (vm_page_busied(m))
787 * Invalid pages can be easily freed. They cannot be
788 * mapped; vm_page_free() asserts this.
794 * If the page has been referenced and the object is not dead,
795 * reactivate or requeue the page depending on whether the
798 if ((m->aflags & PGA_REFERENCED) != 0) {
799 vm_page_aflag_clear(m, PGA_REFERENCED);
803 if (object->ref_count != 0)
804 act_delta += pmap_ts_referenced(m);
806 KASSERT(!pmap_page_is_mapped(m),
807 ("page %p is mapped", m));
809 if (act_delta != 0) {
810 if (object->ref_count != 0) {
811 VM_CNT_INC(v_reactivated);
815 * Increase the activation count if the page
816 * was referenced while in the laundry queue.
817 * This makes it less likely that the page will
818 * be returned prematurely to the inactive
821 m->act_count += act_delta + ACT_ADVANCE;
824 * If this was a background laundering, count
825 * activated pages towards our target. The
826 * purpose of background laundering is to ensure
827 * that pages are eventually cycled through the
828 * laundry queue, and an activation is a valid
834 } else if ((object->flags & OBJ_DEAD) == 0) {
841 * If the page appears to be clean at the machine-independent
842 * layer, then remove all of its mappings from the pmap in
843 * anticipation of freeing it. If, however, any of the page's
844 * mappings allow write access, then the page may still be
845 * modified until the last of those mappings are removed.
847 if (object->ref_count != 0) {
848 vm_page_test_dirty(m);
854 * Clean pages are freed, and dirty pages are paged out unless
855 * they belong to a dead object. Requeueing dirty pages from
856 * dead objects is pointless, as they are being paged out and
857 * freed by the thread that destroyed the object.
863 } else if ((object->flags & OBJ_DEAD) == 0) {
864 if (object->type != OBJT_SWAP &&
865 object->type != OBJT_DEFAULT)
867 else if (disable_swap_pageouts)
877 * Form a cluster with adjacent, dirty pages from the
878 * same object, and page out that entire cluster.
880 * The adjacent, dirty pages must also be in the
881 * laundry. However, their mappings are not checked
882 * for new references. Consequently, a recently
883 * referenced page may be paged out. However, that
884 * page will not be prematurely reclaimed. After page
885 * out, the page will be placed in the inactive queue,
886 * where any new references will be detected and the
889 error = vm_pageout_clean(m, &numpagedout);
891 launder -= numpagedout;
892 ss.scanned += numpagedout;
893 } else if (error == EDEADLK) {
906 VM_OBJECT_WUNLOCK(object);
909 vm_pagequeue_lock(pq);
910 vm_pageout_end_scan(&ss);
911 vm_pagequeue_unlock(pq);
913 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
919 * Wakeup the sync daemon if we skipped a vnode in a writeable object
920 * and we didn't launder enough pages.
922 if (vnodes_skipped > 0 && launder > 0)
923 (void)speedup_syncer();
925 return (starting_target - launder);
929 * Compute the integer square root.
934 u_int bit, root, tmp;
936 bit = 1u << ((NBBY * sizeof(u_int)) - 2);
953 * Perform the work of the laundry thread: periodically wake up and determine
954 * whether any pages need to be laundered. If so, determine the number of pages
955 * that need to be laundered, and launder them.
958 vm_pageout_laundry_worker(void *arg)
960 struct vm_domain *vmd;
961 struct vm_pagequeue *pq;
962 uint64_t nclean, ndirty, nfreed;
963 int domain, last_target, launder, shortfall, shortfall_cycle, target;
966 domain = (uintptr_t)arg;
967 vmd = VM_DOMAIN(domain);
968 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
969 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
972 in_shortfall = false;
978 * Calls to these handlers are serialized by the swap syscall lock.
980 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
981 EVENTHANDLER_PRI_ANY);
982 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
983 EVENTHANDLER_PRI_ANY);
986 * The pageout laundry worker is never done, so loop forever.
989 KASSERT(target >= 0, ("negative target %d", target));
990 KASSERT(shortfall_cycle >= 0,
991 ("negative cycle %d", shortfall_cycle));
995 * First determine whether we need to launder pages to meet a
996 * shortage of free pages.
1000 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1002 } else if (!in_shortfall)
1004 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1006 * We recently entered shortfall and began laundering
1007 * pages. If we have completed that laundering run
1008 * (and we are no longer in shortfall) or we have met
1009 * our laundry target through other activity, then we
1010 * can stop laundering pages.
1012 in_shortfall = false;
1016 launder = target / shortfall_cycle--;
1020 * There's no immediate need to launder any pages; see if we
1021 * meet the conditions to perform background laundering:
1023 * 1. The ratio of dirty to clean inactive pages exceeds the
1024 * background laundering threshold, or
1025 * 2. we haven't yet reached the target of the current
1026 * background laundering run.
1028 * The background laundering threshold is not a constant.
1029 * Instead, it is a slowly growing function of the number of
1030 * clean pages freed by the page daemon since the last
1031 * background laundering. Thus, as the ratio of dirty to
1032 * clean inactive pages grows, the amount of memory pressure
1033 * required to trigger laundering decreases. We ensure
1034 * that the threshold is non-zero after an inactive queue
1035 * scan, even if that scan failed to free a single clean page.
1038 nclean = vmd->vmd_free_count +
1039 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1040 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1041 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1042 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1043 target = vmd->vmd_background_launder_target;
1047 * We have a non-zero background laundering target. If we've
1048 * laundered up to our maximum without observing a page daemon
1049 * request, just stop. This is a safety belt that ensures we
1050 * don't launder an excessive amount if memory pressure is low
1051 * and the ratio of dirty to clean pages is large. Otherwise,
1052 * proceed at the background laundering rate.
1057 last_target = target;
1058 } else if (last_target - target >=
1059 vm_background_launder_max * PAGE_SIZE / 1024) {
1062 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1063 launder /= VM_LAUNDER_RATE;
1064 if (launder > target)
1071 * Because of I/O clustering, the number of laundered
1072 * pages could exceed "target" by the maximum size of
1073 * a cluster minus one.
1075 target -= min(vm_pageout_launder(vmd, launder,
1076 in_shortfall), target);
1077 pause("laundp", hz / VM_LAUNDER_RATE);
1081 * If we're not currently laundering pages and the page daemon
1082 * hasn't posted a new request, sleep until the page daemon
1085 vm_pagequeue_lock(pq);
1086 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1087 (void)mtx_sleep(&vmd->vmd_laundry_request,
1088 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1091 * If the pagedaemon has indicated that it's in shortfall, start
1092 * a shortfall laundering unless we're already in the middle of
1093 * one. This may preempt a background laundering.
1095 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1096 (!in_shortfall || shortfall_cycle == 0)) {
1097 shortfall = vm_laundry_target(vmd) +
1098 vmd->vmd_pageout_deficit;
1104 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1105 nfreed += vmd->vmd_clean_pages_freed;
1106 vmd->vmd_clean_pages_freed = 0;
1107 vm_pagequeue_unlock(pq);
1112 vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m)
1114 struct vm_domain *vmd;
1116 if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0)
1118 vm_page_aflag_set(m, PGA_ENQUEUED);
1119 if ((m->aflags & PGA_REQUEUE_HEAD) != 0) {
1120 vmd = vm_pagequeue_domain(m);
1121 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
1122 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1123 } else if ((m->aflags & PGA_REQUEUE) != 0) {
1124 TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q);
1125 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1127 TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q);
1132 * Re-add stuck pages to the inactive queue. We will examine them again
1133 * during the next scan. If the queue state of a page has changed since
1134 * it was physically removed from the page queue in
1135 * vm_pageout_collect_batch(), don't do anything with that page.
1138 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1141 struct vm_pagequeue *pq;
1148 if (vm_batchqueue_insert(bq, m))
1150 vm_pagequeue_lock(pq);
1151 delta += vm_pageout_reinsert_inactive_page(ss, m);
1153 vm_pagequeue_lock(pq);
1154 while ((m = vm_batchqueue_pop(bq)) != NULL)
1155 delta += vm_pageout_reinsert_inactive_page(ss, m);
1156 vm_pagequeue_cnt_add(pq, delta);
1157 vm_pagequeue_unlock(pq);
1158 vm_batchqueue_init(bq);
1162 * vm_pageout_scan does the dirty work for the pageout daemon.
1164 * pass == 0: Update active LRU/deactivate pages
1165 * pass >= 1: Free inactive pages
1167 * Returns true if pass was zero or enough pages were freed by the inactive
1168 * queue scan to meet the target.
1171 vm_pageout_scan(struct vm_domain *vmd, int pass, int shortage)
1173 struct scan_state ss;
1174 struct vm_batchqueue rq;
1176 vm_page_t m, marker;
1177 struct vm_pagequeue *pq;
1180 int act_delta, addl_page_shortage, deficit, inactq_shortage, max_scan;
1181 int page_shortage, scan_tick, starting_page_shortage;
1185 * If we need to reclaim memory ask kernel caches to return
1186 * some. We rate limit to avoid thrashing.
1188 if (vmd == VM_DOMAIN(0) && pass > 0 &&
1189 (time_uptime - lowmem_uptime) >= lowmem_period) {
1191 * Decrease registered cache sizes.
1193 SDT_PROBE0(vm, , , vm__lowmem_scan);
1194 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1196 * We do this explicitly after the caches have been
1200 lowmem_uptime = time_uptime;
1204 * The addl_page_shortage is an estimate of the number of temporarily
1205 * stuck pages in the inactive queue. In other words, the
1206 * number of pages from the inactive count that should be
1207 * discounted in setting the target for the active queue scan.
1209 addl_page_shortage = 0;
1212 * Calculate the number of pages that we want to free. This number
1213 * can be negative if many pages are freed between the wakeup call to
1214 * the page daemon and this calculation.
1217 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1218 page_shortage = shortage + deficit;
1220 page_shortage = deficit = 0;
1221 starting_page_shortage = page_shortage;
1226 vm_batchqueue_init(&rq);
1229 * Start scanning the inactive queue for pages that we can free. The
1230 * scan will stop when we reach the target or we have scanned the
1231 * entire queue. (Note that m->act_count is not used to make
1232 * decisions for the inactive queue, only for the active queue.)
1234 marker = &vmd->vmd_markers[PQ_INACTIVE];
1235 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1236 vm_pagequeue_lock(pq);
1237 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1238 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1239 KASSERT((m->flags & PG_MARKER) == 0,
1240 ("marker page %p was dequeued", m));
1242 vm_page_change_lock(m, &mtx);
1246 * The page may have been disassociated from the queue
1247 * while locks were dropped.
1249 if (vm_page_queue(m) != PQ_INACTIVE) {
1250 addl_page_shortage++;
1255 * The page was re-enqueued after the page queue lock was
1256 * dropped, or a requeue was requested. This page gets a second
1259 if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE |
1260 PGA_REQUEUE_HEAD)) != 0)
1264 * Held pages are essentially stuck in the queue. So,
1265 * they ought to be discounted from the inactive count.
1266 * See the calculation of inactq_shortage before the
1267 * loop over the active queue below.
1269 * Wired pages may not be freed. Complete their removal
1270 * from the queue now to avoid needless revisits during
1273 if (m->hold_count != 0) {
1274 addl_page_shortage++;
1277 if (m->wire_count != 0) {
1278 vm_page_dequeue_deferred(m);
1282 if (object != m->object) {
1284 VM_OBJECT_WUNLOCK(object);
1290 if (!VM_OBJECT_TRYWLOCK(object)) {
1292 /* Depends on type-stability. */
1293 VM_OBJECT_WLOCK(object);
1301 if (vm_page_busied(m)) {
1303 * Don't mess with busy pages. Leave them at
1304 * the front of the queue. Most likely, they
1305 * are being paged out and will leave the
1306 * queue shortly after the scan finishes. So,
1307 * they ought to be discounted from the
1310 addl_page_shortage++;
1315 * Invalid pages can be easily freed. They cannot be
1316 * mapped, vm_page_free() asserts this.
1322 * If the page has been referenced and the object is not dead,
1323 * reactivate or requeue the page depending on whether the
1326 if ((m->aflags & PGA_REFERENCED) != 0) {
1327 vm_page_aflag_clear(m, PGA_REFERENCED);
1331 if (object->ref_count != 0) {
1332 act_delta += pmap_ts_referenced(m);
1334 KASSERT(!pmap_page_is_mapped(m),
1335 ("vm_pageout_scan: page %p is mapped", m));
1337 if (act_delta != 0) {
1338 if (object->ref_count != 0) {
1339 VM_CNT_INC(v_reactivated);
1340 vm_page_activate(m);
1343 * Increase the activation count if the page
1344 * was referenced while in the inactive queue.
1345 * This makes it less likely that the page will
1346 * be returned prematurely to the inactive
1349 m->act_count += act_delta + ACT_ADVANCE;
1351 } else if ((object->flags & OBJ_DEAD) == 0) {
1352 vm_page_aflag_set(m, PGA_REQUEUE);
1358 * If the page appears to be clean at the machine-independent
1359 * layer, then remove all of its mappings from the pmap in
1360 * anticipation of freeing it. If, however, any of the page's
1361 * mappings allow write access, then the page may still be
1362 * modified until the last of those mappings are removed.
1364 if (object->ref_count != 0) {
1365 vm_page_test_dirty(m);
1371 * Clean pages can be freed, but dirty pages must be sent back
1372 * to the laundry, unless they belong to a dead object.
1373 * Requeueing dirty pages from dead objects is pointless, as
1374 * they are being paged out and freed by the thread that
1375 * destroyed the object.
1377 if (m->dirty == 0) {
1380 * Because we dequeued the page and have already
1381 * checked for concurrent dequeue and enqueue
1382 * requests, we can safely disassociate the page
1383 * from the inactive queue.
1385 KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
1386 ("page %p has queue state", m));
1390 } else if ((object->flags & OBJ_DEAD) == 0)
1394 vm_pageout_reinsert_inactive(&ss, &rq, m);
1401 VM_OBJECT_WUNLOCK(object);
1404 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1405 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1406 vm_pagequeue_lock(pq);
1407 vm_pageout_end_scan(&ss);
1408 vm_pagequeue_unlock(pq);
1410 VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1413 * Wake up the laundry thread so that it can perform any needed
1414 * laundering. If we didn't meet our target, we're in shortfall and
1415 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1416 * swap devices are configured, the laundry thread has no work to do, so
1417 * don't bother waking it up.
1419 * The laundry thread uses the number of inactive queue scans elapsed
1420 * since the last laundering to determine whether to launder again, so
1423 if (starting_page_shortage > 0) {
1424 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1425 vm_pagequeue_lock(pq);
1426 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1427 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1428 if (page_shortage > 0) {
1429 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1430 VM_CNT_INC(v_pdshortfalls);
1431 } else if (vmd->vmd_laundry_request !=
1432 VM_LAUNDRY_SHORTFALL)
1433 vmd->vmd_laundry_request =
1434 VM_LAUNDRY_BACKGROUND;
1435 wakeup(&vmd->vmd_laundry_request);
1437 vmd->vmd_clean_pages_freed +=
1438 starting_page_shortage - page_shortage;
1439 vm_pagequeue_unlock(pq);
1443 * Wakeup the swapout daemon if we didn't free the targeted number of
1446 if (page_shortage > 0)
1450 * If the inactive queue scan fails repeatedly to meet its
1451 * target, kill the largest process.
1453 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1456 * Compute the number of pages we want to try to move from the
1457 * active queue to either the inactive or laundry queue.
1459 * When scanning active pages, we make clean pages count more heavily
1460 * towards the page shortage than dirty pages. This is because dirty
1461 * pages must be laundered before they can be reused and thus have less
1462 * utility when attempting to quickly alleviate a shortage. However,
1463 * this weighting also causes the scan to deactivate dirty pages more
1464 * more aggressively, improving the effectiveness of clustering and
1465 * ensuring that they can eventually be reused.
1467 inactq_shortage = vmd->vmd_inactive_target - (pq->pq_cnt +
1468 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight) +
1469 vm_paging_target(vmd) + deficit + addl_page_shortage;
1470 inactq_shortage *= act_scan_laundry_weight;
1472 marker = &vmd->vmd_markers[PQ_ACTIVE];
1473 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1474 vm_pagequeue_lock(pq);
1477 * If we're just idle polling attempt to visit every
1478 * active page within 'update_period' seconds.
1481 if (vm_pageout_update_period != 0) {
1482 min_scan = pq->pq_cnt;
1483 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1484 min_scan /= hz * vm_pageout_update_period;
1487 if (min_scan > 0 || (inactq_shortage > 0 && pq->pq_cnt > 0))
1488 vmd->vmd_last_active_scan = scan_tick;
1491 * Scan the active queue for pages that can be deactivated. Update
1492 * the per-page activity counter and use it to identify deactivation
1493 * candidates. Held pages may be deactivated.
1495 * To avoid requeuing each page that remains in the active queue, we
1496 * implement the CLOCK algorithm. To maintain consistency in the
1497 * generic page queue code, pages are inserted at the tail of the
1498 * active queue. We thus use two hands, represented by marker pages:
1499 * scans begin at the first hand, which precedes the second hand in
1500 * the queue. When the two hands meet, they are moved back to the
1501 * head and tail of the queue, respectively, and scanning resumes.
1503 max_scan = inactq_shortage > 0 ? pq->pq_cnt : min_scan;
1505 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1506 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1507 if (__predict_false(m == &vmd->vmd_clock[1])) {
1508 vm_pagequeue_lock(pq);
1509 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1510 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1511 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1513 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1515 max_scan -= ss.scanned;
1516 vm_pageout_end_scan(&ss);
1519 if (__predict_false((m->flags & PG_MARKER) != 0))
1522 vm_page_change_lock(m, &mtx);
1525 * The page may have been disassociated from the queue
1526 * while locks were dropped.
1528 if (vm_page_queue(m) != PQ_ACTIVE)
1532 * Wired pages are dequeued lazily.
1534 if (m->wire_count != 0) {
1535 vm_page_dequeue_deferred(m);
1540 * Check to see "how much" the page has been used.
1542 if ((m->aflags & PGA_REFERENCED) != 0) {
1543 vm_page_aflag_clear(m, PGA_REFERENCED);
1549 * Perform an unsynchronized object ref count check. While
1550 * the page lock ensures that the page is not reallocated to
1551 * another object, in particular, one with unmanaged mappings
1552 * that cannot support pmap_ts_referenced(), two races are,
1553 * nonetheless, possible:
1554 * 1) The count was transitioning to zero, but we saw a non-
1555 * zero value. pmap_ts_referenced() will return zero
1556 * because the page is not mapped.
1557 * 2) The count was transitioning to one, but we saw zero.
1558 * This race delays the detection of a new reference. At
1559 * worst, we will deactivate and reactivate the page.
1561 if (m->object->ref_count != 0)
1562 act_delta += pmap_ts_referenced(m);
1565 * Advance or decay the act_count based on recent usage.
1567 if (act_delta != 0) {
1568 m->act_count += ACT_ADVANCE + act_delta;
1569 if (m->act_count > ACT_MAX)
1570 m->act_count = ACT_MAX;
1572 m->act_count -= min(m->act_count, ACT_DECLINE);
1574 if (m->act_count == 0) {
1576 * When not short for inactive pages, let dirty pages go
1577 * through the inactive queue before moving to the
1578 * laundry queues. This gives them some extra time to
1579 * be reactivated, potentially avoiding an expensive
1580 * pageout. During a page shortage, the inactive queue
1581 * is necessarily small, so we may move dirty pages
1582 * directly to the laundry queue.
1584 if (inactq_shortage <= 0)
1585 vm_page_deactivate(m);
1588 * Calling vm_page_test_dirty() here would
1589 * require acquisition of the object's write
1590 * lock. However, during a page shortage,
1591 * directing dirty pages into the laundry
1592 * queue is only an optimization and not a
1593 * requirement. Therefore, we simply rely on
1594 * the opportunistic updates to the page's
1595 * dirty field by the pmap.
1597 if (m->dirty == 0) {
1598 vm_page_deactivate(m);
1600 act_scan_laundry_weight;
1612 vm_pagequeue_lock(pq);
1613 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1614 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1615 vm_pageout_end_scan(&ss);
1616 vm_pagequeue_unlock(pq);
1619 vm_swapout_run_idle();
1620 return (page_shortage <= 0);
1623 static int vm_pageout_oom_vote;
1626 * The pagedaemon threads randlomly select one to perform the
1627 * OOM. Trying to kill processes before all pagedaemons
1628 * failed to reach free target is premature.
1631 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1632 int starting_page_shortage)
1636 if (starting_page_shortage <= 0 || starting_page_shortage !=
1638 vmd->vmd_oom_seq = 0;
1641 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1643 vmd->vmd_oom = FALSE;
1644 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1650 * Do not follow the call sequence until OOM condition is
1653 vmd->vmd_oom_seq = 0;
1658 vmd->vmd_oom = TRUE;
1659 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1660 if (old_vote != vm_ndomains - 1)
1664 * The current pagedaemon thread is the last in the quorum to
1665 * start OOM. Initiate the selection and signaling of the
1668 vm_pageout_oom(VM_OOM_MEM);
1671 * After one round of OOM terror, recall our vote. On the
1672 * next pass, current pagedaemon would vote again if the low
1673 * memory condition is still there, due to vmd_oom being
1676 vmd->vmd_oom = FALSE;
1677 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1681 * The OOM killer is the page daemon's action of last resort when
1682 * memory allocation requests have been stalled for a prolonged period
1683 * of time because it cannot reclaim memory. This function computes
1684 * the approximate number of physical pages that could be reclaimed if
1685 * the specified address space is destroyed.
1687 * Private, anonymous memory owned by the address space is the
1688 * principal resource that we expect to recover after an OOM kill.
1689 * Since the physical pages mapped by the address space's COW entries
1690 * are typically shared pages, they are unlikely to be released and so
1691 * they are not counted.
1693 * To get to the point where the page daemon runs the OOM killer, its
1694 * efforts to write-back vnode-backed pages may have stalled. This
1695 * could be caused by a memory allocation deadlock in the write path
1696 * that might be resolved by an OOM kill. Therefore, physical pages
1697 * belonging to vnode-backed objects are counted, because they might
1698 * be freed without being written out first if the address space holds
1699 * the last reference to an unlinked vnode.
1701 * Similarly, physical pages belonging to OBJT_PHYS objects are
1702 * counted because the address space might hold the last reference to
1706 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1709 vm_map_entry_t entry;
1713 map = &vmspace->vm_map;
1714 KASSERT(!map->system_map, ("system map"));
1715 sx_assert(&map->lock, SA_LOCKED);
1717 for (entry = map->header.next; entry != &map->header;
1718 entry = entry->next) {
1719 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1721 obj = entry->object.vm_object;
1724 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1725 obj->ref_count != 1)
1727 switch (obj->type) {
1732 res += obj->resident_page_count;
1740 vm_pageout_oom(int shortage)
1742 struct proc *p, *bigproc;
1743 vm_offset_t size, bigsize;
1749 * We keep the process bigproc locked once we find it to keep anyone
1750 * from messing with it; however, there is a possibility of
1751 * deadlock if process B is bigproc and one of its child processes
1752 * attempts to propagate a signal to B while we are waiting for A's
1753 * lock while walking this list. To avoid this, we don't block on
1754 * the process lock but just skip a process if it is already locked.
1758 sx_slock(&allproc_lock);
1759 FOREACH_PROC_IN_SYSTEM(p) {
1763 * If this is a system, protected or killed process, skip it.
1765 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1766 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1767 p->p_pid == 1 || P_KILLED(p) ||
1768 (p->p_pid < 48 && swap_pager_avail != 0)) {
1773 * If the process is in a non-running type state,
1774 * don't touch it. Check all the threads individually.
1777 FOREACH_THREAD_IN_PROC(p, td) {
1779 if (!TD_ON_RUNQ(td) &&
1780 !TD_IS_RUNNING(td) &&
1781 !TD_IS_SLEEPING(td) &&
1782 !TD_IS_SUSPENDED(td) &&
1783 !TD_IS_SWAPPED(td)) {
1795 * get the process size
1797 vm = vmspace_acquire_ref(p);
1804 sx_sunlock(&allproc_lock);
1805 if (!vm_map_trylock_read(&vm->vm_map)) {
1807 sx_slock(&allproc_lock);
1811 size = vmspace_swap_count(vm);
1812 if (shortage == VM_OOM_MEM)
1813 size += vm_pageout_oom_pagecount(vm);
1814 vm_map_unlock_read(&vm->vm_map);
1816 sx_slock(&allproc_lock);
1819 * If this process is bigger than the biggest one,
1822 if (size > bigsize) {
1823 if (bigproc != NULL)
1831 sx_sunlock(&allproc_lock);
1832 if (bigproc != NULL) {
1833 if (vm_panic_on_oom != 0)
1834 panic("out of swap space");
1836 killproc(bigproc, "out of swap space");
1837 sched_nice(bigproc, PRIO_MIN);
1839 PROC_UNLOCK(bigproc);
1844 vm_pageout_worker(void *arg)
1846 struct vm_domain *vmd;
1847 int domain, pass, shortage;
1850 domain = (uintptr_t)arg;
1851 vmd = VM_DOMAIN(domain);
1857 * XXXKIB It could be useful to bind pageout daemon threads to
1858 * the cores belonging to the domain, from which vm_page_array
1862 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1863 vmd->vmd_last_active_scan = ticks;
1866 * The pageout daemon worker is never done, so loop forever.
1869 vm_domain_pageout_lock(vmd);
1871 * We need to clear wanted before we check the limits. This
1872 * prevents races with wakers who will check wanted after they
1875 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
1878 * Might the page daemon need to run again?
1880 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
1882 * Yes, the scan failed to free enough pages. If
1883 * we have performed a level >= 1 (page reclamation)
1884 * scan, then sleep a bit and try again.
1886 vm_domain_pageout_unlock(vmd);
1888 pause("pwait", hz / VM_INACT_SCAN_RATE);
1891 * No, sleep until the next wakeup or until pages
1892 * need to have their reference stats updated.
1894 if (mtx_sleep(&vmd->vmd_pageout_wanted,
1895 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
1896 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
1897 VM_CNT_INC(v_pdwakeups);
1899 /* Prevent spurious wakeups by ensuring that wanted is set. */
1900 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1903 * Use the controller to calculate how many pages to free in
1906 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
1907 if (shortage && pass == 0)
1910 target_met = vm_pageout_scan(vmd, pass, shortage);
1912 * If the target was not met we must increase the pass to
1913 * more aggressively reclaim.
1921 * vm_pageout_init initialises basic pageout daemon settings.
1924 vm_pageout_init_domain(int domain)
1926 struct vm_domain *vmd;
1927 struct sysctl_oid *oid;
1929 vmd = VM_DOMAIN(domain);
1930 vmd->vmd_interrupt_free_min = 2;
1933 * v_free_reserved needs to include enough for the largest
1934 * swap pager structures plus enough for any pv_entry structs
1937 if (vmd->vmd_page_count > 1024)
1938 vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200;
1940 vmd->vmd_free_min = 4;
1941 vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1942 vmd->vmd_interrupt_free_min;
1943 vmd->vmd_free_reserved = vm_pageout_page_count +
1944 vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768);
1945 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
1946 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
1947 vmd->vmd_free_min += vmd->vmd_free_reserved;
1948 vmd->vmd_free_severe += vmd->vmd_free_reserved;
1949 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
1950 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
1951 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
1954 * Set the default wakeup threshold to be 10% below the paging
1955 * target. This keeps the steady state out of shortfall.
1957 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
1960 * Target amount of memory to move out of the laundry queue during a
1961 * background laundering. This is proportional to the amount of system
1964 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
1965 vmd->vmd_free_min) / 10;
1967 /* Initialize the pageout daemon pid controller. */
1968 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
1969 vmd->vmd_free_target, PIDCTRL_BOUND,
1970 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
1971 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
1972 "pidctrl", CTLFLAG_RD, NULL, "");
1973 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
1977 vm_pageout_init(void)
1983 * Initialize some paging parameters.
1985 if (vm_cnt.v_page_count < 2000)
1986 vm_pageout_page_count = 8;
1989 for (i = 0; i < vm_ndomains; i++) {
1990 struct vm_domain *vmd;
1992 vm_pageout_init_domain(i);
1994 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
1995 vm_cnt.v_free_target += vmd->vmd_free_target;
1996 vm_cnt.v_free_min += vmd->vmd_free_min;
1997 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
1998 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
1999 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2000 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2001 freecount += vmd->vmd_free_count;
2005 * Set interval in seconds for active scan. We want to visit each
2006 * page at least once every ten minutes. This is to prevent worst
2007 * case paging behaviors with stale active LRU.
2009 if (vm_pageout_update_period == 0)
2010 vm_pageout_update_period = 600;
2012 if (vm_page_max_wired == 0)
2013 vm_page_max_wired = freecount / 3;
2017 * vm_pageout is the high level pageout daemon.
2025 swap_pager_swap_init();
2026 snprintf(curthread->td_name, sizeof(curthread->td_name), "dom0");
2027 error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
2028 0, 0, "laundry: dom0");
2030 panic("starting laundry for domain 0, error %d", error);
2031 for (i = 1; i < vm_ndomains; i++) {
2032 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
2033 curproc, NULL, 0, 0, "dom%d", i);
2035 panic("starting pageout for domain %d, error %d\n",
2038 error = kthread_add(vm_pageout_laundry_worker,
2039 (void *)(uintptr_t)i, curproc, NULL, 0, 0,
2040 "laundry: dom%d", i);
2042 panic("starting laundry for domain %d, error %d",
2045 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
2048 panic("starting uma_reclaim helper, error %d\n", error);
2049 vm_pageout_worker((void *)(uintptr_t)0);
2053 * Perform an advisory wakeup of the page daemon.
2056 pagedaemon_wakeup(int domain)
2058 struct vm_domain *vmd;
2060 vmd = VM_DOMAIN(domain);
2061 vm_domain_pageout_assert_unlocked(vmd);
2062 if (curproc == pageproc)
2065 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2066 vm_domain_pageout_lock(vmd);
2067 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2068 wakeup(&vmd->vmd_pageout_wanted);
2069 vm_domain_pageout_unlock(vmd);