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/blockcount.h>
86 #include <sys/eventhandler.h>
88 #include <sys/mutex.h>
90 #include <sys/kthread.h>
92 #include <sys/mount.h>
93 #include <sys/racct.h>
94 #include <sys/resourcevar.h>
95 #include <sys/sched.h>
97 #include <sys/signalvar.h>
100 #include <sys/vnode.h>
101 #include <sys/vmmeter.h>
102 #include <sys/rwlock.h>
104 #include <sys/sysctl.h>
107 #include <vm/vm_param.h>
108 #include <vm/vm_object.h>
109 #include <vm/vm_page.h>
110 #include <vm/vm_map.h>
111 #include <vm/vm_pageout.h>
112 #include <vm/vm_pager.h>
113 #include <vm/vm_phys.h>
114 #include <vm/vm_pagequeue.h>
115 #include <vm/swap_pager.h>
116 #include <vm/vm_extern.h>
120 * System initialization
123 /* the kernel process "vm_pageout"*/
124 static void vm_pageout(void);
125 static void vm_pageout_init(void);
126 static int vm_pageout_clean(vm_page_t m, int *numpagedout);
127 static int vm_pageout_cluster(vm_page_t m);
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 int swapdev_enabled;
158 static int vm_panic_on_oom = 0;
160 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
161 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
162 "Panic on the given number of out-of-memory errors instead of killing the largest process");
164 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
165 CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
166 "Maximum active LRU update period");
168 static int pageout_cpus_per_thread = 16;
169 SYSCTL_INT(_vm, OID_AUTO, pageout_cpus_per_thread, CTLFLAG_RDTUN,
170 &pageout_cpus_per_thread, 0,
171 "Number of CPUs per pagedaemon worker thread");
173 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
174 "Low memory callback period");
176 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
177 CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
179 static int pageout_lock_miss;
180 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
181 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
183 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
184 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
185 "back-to-back calls to oom detector to start OOM");
187 static int act_scan_laundry_weight = 3;
188 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
189 &act_scan_laundry_weight, 0,
190 "weight given to clean vs. dirty pages in active queue scans");
192 static u_int vm_background_launder_rate = 4096;
193 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
194 &vm_background_launder_rate, 0,
195 "background laundering rate, in kilobytes per second");
197 static u_int vm_background_launder_max = 20 * 1024;
198 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
199 &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
201 int vm_pageout_page_count = 32;
203 u_long vm_page_max_user_wired;
204 SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW,
205 &vm_page_max_user_wired, 0,
206 "system-wide limit to user-wired page count");
208 static u_int isqrt(u_int num);
209 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
211 static void vm_pageout_laundry_worker(void *arg);
214 struct vm_batchqueue bq;
215 struct vm_pagequeue *pq;
222 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
223 vm_page_t marker, vm_page_t after, int maxscan)
226 vm_pagequeue_assert_locked(pq);
227 KASSERT((marker->a.flags & PGA_ENQUEUED) == 0,
228 ("marker %p already enqueued", marker));
231 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
233 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
234 vm_page_aflag_set(marker, PGA_ENQUEUED);
236 vm_batchqueue_init(&ss->bq);
239 ss->maxscan = maxscan;
241 vm_pagequeue_unlock(pq);
245 vm_pageout_end_scan(struct scan_state *ss)
247 struct vm_pagequeue *pq;
250 vm_pagequeue_assert_locked(pq);
251 KASSERT((ss->marker->a.flags & PGA_ENQUEUED) != 0,
252 ("marker %p not enqueued", ss->marker));
254 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
255 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
256 pq->pq_pdpages += ss->scanned;
260 * Add a small number of queued pages to a batch queue for later processing
261 * without the corresponding queue lock held. The caller must have enqueued a
262 * marker page at the desired start point for the scan. Pages will be
263 * physically dequeued if the caller so requests. Otherwise, the returned
264 * batch may contain marker pages, and it is up to the caller to handle them.
266 * When processing the batch queue, vm_pageout_defer() must be used to
267 * determine whether the page has been logically dequeued since the batch was
270 static __always_inline void
271 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
273 struct vm_pagequeue *pq;
274 vm_page_t m, marker, n;
279 KASSERT((marker->a.flags & PGA_ENQUEUED) != 0,
280 ("marker %p not enqueued", ss->marker));
282 vm_pagequeue_lock(pq);
283 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
284 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
285 m = n, ss->scanned++) {
286 n = TAILQ_NEXT(m, plinks.q);
287 if ((m->flags & PG_MARKER) == 0) {
288 KASSERT((m->a.flags & PGA_ENQUEUED) != 0,
289 ("page %p not enqueued", m));
290 KASSERT((m->flags & PG_FICTITIOUS) == 0,
291 ("Fictitious page %p cannot be in page queue", m));
292 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
293 ("Unmanaged page %p cannot be in page queue", m));
297 (void)vm_batchqueue_insert(&ss->bq, m);
299 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
300 vm_page_aflag_clear(m, PGA_ENQUEUED);
303 TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
304 if (__predict_true(m != NULL))
305 TAILQ_INSERT_BEFORE(m, marker, plinks.q);
307 TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
309 vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
310 vm_pagequeue_unlock(pq);
314 * Return the next page to be scanned, or NULL if the scan is complete.
316 static __always_inline vm_page_t
317 vm_pageout_next(struct scan_state *ss, const bool dequeue)
320 if (ss->bq.bq_cnt == 0)
321 vm_pageout_collect_batch(ss, dequeue);
322 return (vm_batchqueue_pop(&ss->bq));
326 * Determine whether processing of a page should be deferred and ensure that any
327 * outstanding queue operations are processed.
329 static __always_inline bool
330 vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued)
334 as = vm_page_astate_load(m);
335 if (__predict_false(as.queue != queue ||
336 ((as.flags & PGA_ENQUEUED) != 0) != enqueued))
338 if ((as.flags & PGA_QUEUE_OP_MASK) != 0) {
339 vm_page_pqbatch_submit(m, queue);
346 * Scan for pages at adjacent offsets within the given page's object that are
347 * eligible for laundering, form a cluster of these pages and the given page,
348 * and launder that cluster.
351 vm_pageout_cluster(vm_page_t m)
354 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
356 int ib, is, page_base, pageout_count;
359 VM_OBJECT_ASSERT_WLOCKED(object);
362 vm_page_assert_xbusied(m);
364 mc[vm_pageout_page_count] = pb = ps = m;
366 page_base = vm_pageout_page_count;
371 * We can cluster only if the page is not clean, busy, or held, and
372 * the page is in the laundry queue.
374 * During heavy mmap/modification loads the pageout
375 * daemon can really fragment the underlying file
376 * due to flushing pages out of order and not trying to
377 * align the clusters (which leaves sporadic out-of-order
378 * holes). To solve this problem we do the reverse scan
379 * first and attempt to align our cluster, then do a
380 * forward scan if room remains.
383 while (ib != 0 && pageout_count < vm_pageout_page_count) {
388 if ((p = vm_page_prev(pb)) == NULL ||
389 vm_page_tryxbusy(p) == 0) {
393 if (vm_page_wired(p)) {
398 vm_page_test_dirty(p);
404 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
409 mc[--page_base] = pb = p;
414 * We are at an alignment boundary. Stop here, and switch
415 * directions. Do not clear ib.
417 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
420 while (pageout_count < vm_pageout_page_count &&
421 pindex + is < object->size) {
422 if ((p = vm_page_next(ps)) == NULL ||
423 vm_page_tryxbusy(p) == 0)
425 if (vm_page_wired(p)) {
429 vm_page_test_dirty(p);
434 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
438 mc[page_base + pageout_count] = ps = p;
444 * If we exhausted our forward scan, continue with the reverse scan
445 * when possible, even past an alignment boundary. This catches
446 * boundary conditions.
448 if (ib != 0 && pageout_count < vm_pageout_page_count)
451 return (vm_pageout_flush(&mc[page_base], pageout_count,
452 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
456 * vm_pageout_flush() - launder the given pages
458 * The given pages are laundered. Note that we setup for the start of
459 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
460 * reference count all in here rather then in the parent. If we want
461 * the parent to do more sophisticated things we may have to change
464 * Returned runlen is the count of pages between mreq and first
465 * page after mreq with status VM_PAGER_AGAIN.
466 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
467 * for any page in runlen set.
470 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
473 vm_object_t object = mc[0]->object;
474 int pageout_status[count];
478 VM_OBJECT_ASSERT_WLOCKED(object);
481 * Initiate I/O. Mark the pages shared busy and verify that they're
482 * valid and read-only.
484 * We do not have to fixup the clean/dirty bits here... we can
485 * allow the pager to do it after the I/O completes.
487 * NOTE! mc[i]->dirty may be partial or fragmented due to an
488 * edge case with file fragments.
490 for (i = 0; i < count; i++) {
491 KASSERT(vm_page_all_valid(mc[i]),
492 ("vm_pageout_flush: partially invalid page %p index %d/%d",
494 KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
495 ("vm_pageout_flush: writeable page %p", mc[i]));
496 vm_page_busy_downgrade(mc[i]);
498 vm_object_pip_add(object, count);
500 vm_pager_put_pages(object, mc, count, flags, pageout_status);
502 runlen = count - mreq;
505 for (i = 0; i < count; i++) {
506 vm_page_t mt = mc[i];
508 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
509 !pmap_page_is_write_mapped(mt),
510 ("vm_pageout_flush: page %p is not write protected", mt));
511 switch (pageout_status[i]) {
514 * The page may have moved since laundering started, in
515 * which case it should be left alone.
517 if (vm_page_in_laundry(mt))
518 vm_page_deactivate_noreuse(mt);
525 * The page is outside the object's range. We pretend
526 * that the page out worked and clean the page, so the
527 * changes will be lost if the page is reclaimed by
531 if (vm_page_in_laundry(mt))
532 vm_page_deactivate_noreuse(mt);
537 * If the page couldn't be paged out to swap because the
538 * pager wasn't able to find space, place the page in
539 * the PQ_UNSWAPPABLE holding queue. This is an
540 * optimization that prevents the page daemon from
541 * wasting CPU cycles on pages that cannot be reclaimed
542 * because no swap device is configured.
544 * Otherwise, reactivate the page so that it doesn't
545 * clog the laundry and inactive queues. (We will try
546 * paging it out again later.)
548 if ((object->flags & OBJ_SWAP) != 0 &&
549 pageout_status[i] == VM_PAGER_FAIL) {
550 vm_page_unswappable(mt);
553 vm_page_activate(mt);
554 if (eio != NULL && i >= mreq && i - mreq < runlen)
558 if (i >= mreq && i - mreq < runlen)
564 * If the operation is still going, leave the page busy to
565 * block all other accesses. Also, leave the paging in
566 * progress indicator set so that we don't attempt an object
569 if (pageout_status[i] != VM_PAGER_PEND) {
570 vm_object_pip_wakeup(object);
576 return (numpagedout);
580 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
583 atomic_store_rel_int(&swapdev_enabled, 1);
587 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
590 if (swap_pager_nswapdev() == 1)
591 atomic_store_rel_int(&swapdev_enabled, 0);
595 * Attempt to acquire all of the necessary locks to launder a page and
596 * then call through the clustering layer to PUTPAGES. Wait a short
597 * time for a vnode lock.
599 * Requires the page and object lock on entry, releases both before return.
600 * Returns 0 on success and an errno otherwise.
603 vm_pageout_clean(vm_page_t m, int *numpagedout)
612 VM_OBJECT_ASSERT_WLOCKED(object);
618 * The object is already known NOT to be dead. It
619 * is possible for the vget() to block the whole
620 * pageout daemon, but the new low-memory handling
621 * code should prevent it.
623 * We can't wait forever for the vnode lock, we might
624 * deadlock due to a vn_read() getting stuck in
625 * vm_wait while holding this vnode. We skip the
626 * vnode if we can't get it in a reasonable amount
629 if (object->type == OBJT_VNODE) {
632 if (vp->v_type == VREG &&
633 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
639 ("vp %p with NULL v_mount", vp));
640 vm_object_reference_locked(object);
642 VM_OBJECT_WUNLOCK(object);
643 if (vget(vp, vn_lktype_write(NULL, vp) | LK_TIMELOCK) != 0) {
648 VM_OBJECT_WLOCK(object);
651 * Ensure that the object and vnode were not disassociated
652 * while locks were dropped.
654 if (vp->v_object != object) {
660 * While the object was unlocked, the page may have been:
661 * (1) moved to a different queue,
662 * (2) reallocated to a different object,
663 * (3) reallocated to a different offset, or
666 if (!vm_page_in_laundry(m) || m->object != object ||
667 m->pindex != pindex || m->dirty == 0) {
673 * The page may have been busied while the object lock was
676 if (vm_page_tryxbusy(m) == 0) {
683 * Remove all writeable mappings, failing if the page is wired.
685 if (!vm_page_try_remove_write(m)) {
692 * If a page is dirty, then it is either being washed
693 * (but not yet cleaned) or it is still in the
694 * laundry. If it is still in the laundry, then we
695 * start the cleaning operation.
697 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
701 VM_OBJECT_WUNLOCK(object);
707 vm_object_deallocate(object);
708 vn_finished_write(mp);
715 * Attempt to launder the specified number of pages.
717 * Returns the number of pages successfully laundered.
720 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
722 struct scan_state ss;
723 struct vm_pagequeue *pq;
726 vm_page_astate_t new, old;
727 int act_delta, error, numpagedout, queue, refs, starting_target;
732 starting_target = launder;
736 * Scan the laundry queues for pages eligible to be laundered. We stop
737 * once the target number of dirty pages have been laundered, or once
738 * we've reached the end of the queue. A single iteration of this loop
739 * may cause more than one page to be laundered because of clustering.
741 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
742 * swap devices are configured.
744 if (atomic_load_acq_int(&swapdev_enabled))
745 queue = PQ_UNSWAPPABLE;
750 marker = &vmd->vmd_markers[queue];
751 pq = &vmd->vmd_pagequeues[queue];
752 vm_pagequeue_lock(pq);
753 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
754 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
755 if (__predict_false((m->flags & PG_MARKER) != 0))
759 * Don't touch a page that was removed from the queue after the
760 * page queue lock was released. Otherwise, ensure that any
761 * pending queue operations, such as dequeues for wired pages,
764 if (vm_pageout_defer(m, queue, true))
768 * Lock the page's object.
770 if (object == NULL || object != m->object) {
772 VM_OBJECT_WUNLOCK(object);
773 object = atomic_load_ptr(&m->object);
774 if (__predict_false(object == NULL))
775 /* The page is being freed by another thread. */
778 /* Depends on type-stability. */
779 VM_OBJECT_WLOCK(object);
780 if (__predict_false(m->object != object)) {
781 VM_OBJECT_WUNLOCK(object);
787 if (vm_page_tryxbusy(m) == 0)
791 * Check for wirings now that we hold the object lock and have
792 * exclusively busied the page. If the page is mapped, it may
793 * still be wired by pmap lookups. The call to
794 * vm_page_try_remove_all() below atomically checks for such
795 * wirings and removes mappings. If the page is unmapped, the
796 * wire count is guaranteed not to increase after this check.
798 if (__predict_false(vm_page_wired(m)))
802 * Invalid pages can be easily freed. They cannot be
803 * mapped; vm_page_free() asserts this.
805 if (vm_page_none_valid(m))
808 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
810 for (old = vm_page_astate_load(m);;) {
812 * Check to see if the page has been removed from the
813 * queue since the first such check. Leave it alone if
814 * so, discarding any references collected by
815 * pmap_ts_referenced().
817 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
822 if ((old.flags & PGA_REFERENCED) != 0) {
823 new.flags &= ~PGA_REFERENCED;
826 if (act_delta == 0) {
828 } else if (object->ref_count != 0) {
830 * Increase the activation count if the page was
831 * referenced while in the laundry queue. This
832 * makes it less likely that the page will be
833 * returned prematurely to the laundry queue.
835 new.act_count += ACT_ADVANCE +
837 if (new.act_count > ACT_MAX)
838 new.act_count = ACT_MAX;
840 new.flags &= ~PGA_QUEUE_OP_MASK;
841 new.flags |= PGA_REQUEUE;
842 new.queue = PQ_ACTIVE;
843 if (!vm_page_pqstate_commit(m, &old, new))
847 * If this was a background laundering, count
848 * activated pages towards our target. The
849 * purpose of background laundering is to ensure
850 * that pages are eventually cycled through the
851 * laundry queue, and an activation is a valid
856 VM_CNT_INC(v_reactivated);
858 } else if ((object->flags & OBJ_DEAD) == 0) {
859 new.flags |= PGA_REQUEUE;
860 if (!vm_page_pqstate_commit(m, &old, new))
868 * If the page appears to be clean at the machine-independent
869 * layer, then remove all of its mappings from the pmap in
870 * anticipation of freeing it. If, however, any of the page's
871 * mappings allow write access, then the page may still be
872 * modified until the last of those mappings are removed.
874 if (object->ref_count != 0) {
875 vm_page_test_dirty(m);
876 if (m->dirty == 0 && !vm_page_try_remove_all(m))
881 * Clean pages are freed, and dirty pages are paged out unless
882 * they belong to a dead object. Requeueing dirty pages from
883 * dead objects is pointless, as they are being paged out and
884 * freed by the thread that destroyed the object.
889 * Now we are guaranteed that no other threads are
890 * manipulating the page, check for a last-second
893 if (vm_pageout_defer(m, queue, true))
897 } else if ((object->flags & OBJ_DEAD) == 0) {
898 if ((object->flags & OBJ_SWAP) == 0 &&
899 object->type != OBJT_DEFAULT)
901 else if (disable_swap_pageouts)
911 * Form a cluster with adjacent, dirty pages from the
912 * same object, and page out that entire cluster.
914 * The adjacent, dirty pages must also be in the
915 * laundry. However, their mappings are not checked
916 * for new references. Consequently, a recently
917 * referenced page may be paged out. However, that
918 * page will not be prematurely reclaimed. After page
919 * out, the page will be placed in the inactive queue,
920 * where any new references will be detected and the
923 error = vm_pageout_clean(m, &numpagedout);
925 launder -= numpagedout;
926 ss.scanned += numpagedout;
927 } else if (error == EDEADLK) {
937 if (object != NULL) {
938 VM_OBJECT_WUNLOCK(object);
941 vm_pagequeue_lock(pq);
942 vm_pageout_end_scan(&ss);
943 vm_pagequeue_unlock(pq);
945 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
951 * Wakeup the sync daemon if we skipped a vnode in a writeable object
952 * and we didn't launder enough pages.
954 if (vnodes_skipped > 0 && launder > 0)
955 (void)speedup_syncer();
957 return (starting_target - launder);
961 * Compute the integer square root.
966 u_int bit, root, tmp;
968 bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
983 * Perform the work of the laundry thread: periodically wake up and determine
984 * whether any pages need to be laundered. If so, determine the number of pages
985 * that need to be laundered, and launder them.
988 vm_pageout_laundry_worker(void *arg)
990 struct vm_domain *vmd;
991 struct vm_pagequeue *pq;
992 uint64_t nclean, ndirty, nfreed;
993 int domain, last_target, launder, shortfall, shortfall_cycle, target;
996 domain = (uintptr_t)arg;
997 vmd = VM_DOMAIN(domain);
998 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
999 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1002 in_shortfall = false;
1003 shortfall_cycle = 0;
1004 last_target = target = 0;
1008 * Calls to these handlers are serialized by the swap syscall lock.
1010 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
1011 EVENTHANDLER_PRI_ANY);
1012 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
1013 EVENTHANDLER_PRI_ANY);
1016 * The pageout laundry worker is never done, so loop forever.
1019 KASSERT(target >= 0, ("negative target %d", target));
1020 KASSERT(shortfall_cycle >= 0,
1021 ("negative cycle %d", shortfall_cycle));
1025 * First determine whether we need to launder pages to meet a
1026 * shortage of free pages.
1028 if (shortfall > 0) {
1029 in_shortfall = true;
1030 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1032 } else if (!in_shortfall)
1034 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1036 * We recently entered shortfall and began laundering
1037 * pages. If we have completed that laundering run
1038 * (and we are no longer in shortfall) or we have met
1039 * our laundry target through other activity, then we
1040 * can stop laundering pages.
1042 in_shortfall = false;
1046 launder = target / shortfall_cycle--;
1050 * There's no immediate need to launder any pages; see if we
1051 * meet the conditions to perform background laundering:
1053 * 1. The ratio of dirty to clean inactive pages exceeds the
1054 * background laundering threshold, or
1055 * 2. we haven't yet reached the target of the current
1056 * background laundering run.
1058 * The background laundering threshold is not a constant.
1059 * Instead, it is a slowly growing function of the number of
1060 * clean pages freed by the page daemon since the last
1061 * background laundering. Thus, as the ratio of dirty to
1062 * clean inactive pages grows, the amount of memory pressure
1063 * required to trigger laundering decreases. We ensure
1064 * that the threshold is non-zero after an inactive queue
1065 * scan, even if that scan failed to free a single clean page.
1068 nclean = vmd->vmd_free_count +
1069 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1070 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1071 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1072 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1073 target = vmd->vmd_background_launder_target;
1077 * We have a non-zero background laundering target. If we've
1078 * laundered up to our maximum without observing a page daemon
1079 * request, just stop. This is a safety belt that ensures we
1080 * don't launder an excessive amount if memory pressure is low
1081 * and the ratio of dirty to clean pages is large. Otherwise,
1082 * proceed at the background laundering rate.
1087 last_target = target;
1088 } else if (last_target - target >=
1089 vm_background_launder_max * PAGE_SIZE / 1024) {
1092 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1093 launder /= VM_LAUNDER_RATE;
1094 if (launder > target)
1101 * Because of I/O clustering, the number of laundered
1102 * pages could exceed "target" by the maximum size of
1103 * a cluster minus one.
1105 target -= min(vm_pageout_launder(vmd, launder,
1106 in_shortfall), target);
1107 pause("laundp", hz / VM_LAUNDER_RATE);
1111 * If we're not currently laundering pages and the page daemon
1112 * hasn't posted a new request, sleep until the page daemon
1115 vm_pagequeue_lock(pq);
1116 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1117 (void)mtx_sleep(&vmd->vmd_laundry_request,
1118 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1121 * If the pagedaemon has indicated that it's in shortfall, start
1122 * a shortfall laundering unless we're already in the middle of
1123 * one. This may preempt a background laundering.
1125 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1126 (!in_shortfall || shortfall_cycle == 0)) {
1127 shortfall = vm_laundry_target(vmd) +
1128 vmd->vmd_pageout_deficit;
1134 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1135 nfreed += vmd->vmd_clean_pages_freed;
1136 vmd->vmd_clean_pages_freed = 0;
1137 vm_pagequeue_unlock(pq);
1142 * Compute the number of pages we want to try to move from the
1143 * active queue to either the inactive or laundry queue.
1145 * When scanning active pages during a shortage, we make clean pages
1146 * count more heavily towards the page shortage than dirty pages.
1147 * This is because dirty pages must be laundered before they can be
1148 * reused and thus have less utility when attempting to quickly
1149 * alleviate a free page shortage. However, this weighting also
1150 * causes the scan to deactivate dirty pages more aggressively,
1151 * improving the effectiveness of clustering.
1154 vm_pageout_active_target(struct vm_domain *vmd)
1158 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1159 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1160 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1161 shortage *= act_scan_laundry_weight;
1166 * Scan the active queue. If there is no shortage of inactive pages, scan a
1167 * small portion of the queue in order to maintain quasi-LRU.
1170 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1172 struct scan_state ss;
1174 vm_page_t m, marker;
1175 struct vm_pagequeue *pq;
1176 vm_page_astate_t old, new;
1178 int act_delta, max_scan, ps_delta, refs, scan_tick;
1181 marker = &vmd->vmd_markers[PQ_ACTIVE];
1182 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1183 vm_pagequeue_lock(pq);
1186 * If we're just idle polling attempt to visit every
1187 * active page within 'update_period' seconds.
1190 if (vm_pageout_update_period != 0) {
1191 min_scan = pq->pq_cnt;
1192 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1193 min_scan /= hz * vm_pageout_update_period;
1196 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1197 vmd->vmd_last_active_scan = scan_tick;
1200 * Scan the active queue for pages that can be deactivated. Update
1201 * the per-page activity counter and use it to identify deactivation
1202 * candidates. Held pages may be deactivated.
1204 * To avoid requeuing each page that remains in the active queue, we
1205 * implement the CLOCK algorithm. To keep the implementation of the
1206 * enqueue operation consistent for all page queues, we use two hands,
1207 * represented by marker pages. Scans begin at the first hand, which
1208 * precedes the second hand in the queue. When the two hands meet,
1209 * they are moved back to the head and tail of the queue, respectively,
1210 * and scanning resumes.
1212 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1214 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1215 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1216 if (__predict_false(m == &vmd->vmd_clock[1])) {
1217 vm_pagequeue_lock(pq);
1218 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1219 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1220 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1222 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1224 max_scan -= ss.scanned;
1225 vm_pageout_end_scan(&ss);
1228 if (__predict_false((m->flags & PG_MARKER) != 0))
1232 * Don't touch a page that was removed from the queue after the
1233 * page queue lock was released. Otherwise, ensure that any
1234 * pending queue operations, such as dequeues for wired pages,
1237 if (vm_pageout_defer(m, PQ_ACTIVE, true))
1241 * A page's object pointer may be set to NULL before
1242 * the object lock is acquired.
1244 object = atomic_load_ptr(&m->object);
1245 if (__predict_false(object == NULL))
1247 * The page has been removed from its object.
1251 /* Deferred free of swap space. */
1252 if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
1253 VM_OBJECT_TRYWLOCK(object)) {
1254 if (m->object == object)
1255 vm_pager_page_unswapped(m);
1256 VM_OBJECT_WUNLOCK(object);
1260 * Check to see "how much" the page has been used.
1262 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1263 * that a reference from a concurrently destroyed mapping is
1264 * observed here and now.
1266 * Perform an unsynchronized object ref count check. While
1267 * the page lock ensures that the page is not reallocated to
1268 * another object, in particular, one with unmanaged mappings
1269 * that cannot support pmap_ts_referenced(), two races are,
1270 * nonetheless, possible:
1271 * 1) The count was transitioning to zero, but we saw a non-
1272 * zero value. pmap_ts_referenced() will return zero
1273 * because the page is not mapped.
1274 * 2) The count was transitioning to one, but we saw zero.
1275 * This race delays the detection of a new reference. At
1276 * worst, we will deactivate and reactivate the page.
1278 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1280 old = vm_page_astate_load(m);
1283 * Check to see if the page has been removed from the
1284 * queue since the first such check. Leave it alone if
1285 * so, discarding any references collected by
1286 * pmap_ts_referenced().
1288 if (__predict_false(_vm_page_queue(old) == PQ_NONE)) {
1294 * Advance or decay the act_count based on recent usage.
1298 if ((old.flags & PGA_REFERENCED) != 0) {
1299 new.flags &= ~PGA_REFERENCED;
1302 if (act_delta != 0) {
1303 new.act_count += ACT_ADVANCE + act_delta;
1304 if (new.act_count > ACT_MAX)
1305 new.act_count = ACT_MAX;
1307 new.act_count -= min(new.act_count,
1311 if (new.act_count > 0) {
1313 * Adjust the activation count and keep the page
1314 * in the active queue. The count might be left
1315 * unchanged if it is saturated. The page may
1316 * have been moved to a different queue since we
1317 * started the scan, in which case we move it
1321 if (old.queue != PQ_ACTIVE) {
1322 new.flags &= ~PGA_QUEUE_OP_MASK;
1323 new.flags |= PGA_REQUEUE;
1324 new.queue = PQ_ACTIVE;
1328 * When not short for inactive pages, let dirty
1329 * pages go through the inactive queue before
1330 * moving to the laundry queue. This gives them
1331 * some extra time to be reactivated,
1332 * potentially avoiding an expensive pageout.
1333 * However, during a page shortage, the inactive
1334 * queue is necessarily small, and so dirty
1335 * pages would only spend a trivial amount of
1336 * time in the inactive queue. Therefore, we
1337 * might as well place them directly in the
1338 * laundry queue to reduce queuing overhead.
1340 * Calling vm_page_test_dirty() here would
1341 * require acquisition of the object's write
1342 * lock. However, during a page shortage,
1343 * directing dirty pages into the laundry queue
1344 * is only an optimization and not a
1345 * requirement. Therefore, we simply rely on
1346 * the opportunistic updates to the page's dirty
1347 * field by the pmap.
1349 if (page_shortage <= 0) {
1350 nqueue = PQ_INACTIVE;
1352 } else if (m->dirty == 0) {
1353 nqueue = PQ_INACTIVE;
1354 ps_delta = act_scan_laundry_weight;
1356 nqueue = PQ_LAUNDRY;
1360 new.flags &= ~PGA_QUEUE_OP_MASK;
1361 new.flags |= PGA_REQUEUE;
1364 } while (!vm_page_pqstate_commit(m, &old, new));
1366 page_shortage -= ps_delta;
1368 vm_pagequeue_lock(pq);
1369 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1370 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1371 vm_pageout_end_scan(&ss);
1372 vm_pagequeue_unlock(pq);
1376 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker,
1379 vm_page_astate_t as;
1381 vm_pagequeue_assert_locked(pq);
1383 as = vm_page_astate_load(m);
1384 if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
1386 vm_page_aflag_set(m, PGA_ENQUEUED);
1387 TAILQ_INSERT_BEFORE(marker, m, plinks.q);
1392 * Re-add stuck pages to the inactive queue. We will examine them again
1393 * during the next scan. If the queue state of a page has changed since
1394 * it was physically removed from the page queue in
1395 * vm_pageout_collect_batch(), don't do anything with that page.
1398 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1401 struct vm_pagequeue *pq;
1406 marker = ss->marker;
1410 if (vm_batchqueue_insert(bq, m))
1412 vm_pagequeue_lock(pq);
1413 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1415 vm_pagequeue_lock(pq);
1416 while ((m = vm_batchqueue_pop(bq)) != NULL)
1417 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1418 vm_pagequeue_cnt_add(pq, delta);
1419 vm_pagequeue_unlock(pq);
1420 vm_batchqueue_init(bq);
1424 vm_pageout_scan_inactive(struct vm_domain *vmd, int page_shortage)
1426 struct timeval start, end;
1427 struct scan_state ss;
1428 struct vm_batchqueue rq;
1429 struct vm_page marker_page;
1430 vm_page_t m, marker;
1431 struct vm_pagequeue *pq;
1433 vm_page_astate_t old, new;
1434 int act_delta, addl_page_shortage, starting_page_shortage, refs;
1437 vm_batchqueue_init(&rq);
1438 getmicrouptime(&start);
1441 * The addl_page_shortage is an estimate of the number of temporarily
1442 * stuck pages in the inactive queue. In other words, the
1443 * number of pages from the inactive count that should be
1444 * discounted in setting the target for the active queue scan.
1446 addl_page_shortage = 0;
1449 * Start scanning the inactive queue for pages that we can free. The
1450 * scan will stop when we reach the target or we have scanned the
1451 * entire queue. (Note that m->a.act_count is not used to make
1452 * decisions for the inactive queue, only for the active queue.)
1454 starting_page_shortage = page_shortage;
1455 marker = &marker_page;
1456 vm_page_init_marker(marker, PQ_INACTIVE, 0);
1457 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1458 vm_pagequeue_lock(pq);
1459 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1460 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1461 KASSERT((m->flags & PG_MARKER) == 0,
1462 ("marker page %p was dequeued", m));
1465 * Don't touch a page that was removed from the queue after the
1466 * page queue lock was released. Otherwise, ensure that any
1467 * pending queue operations, such as dequeues for wired pages,
1470 if (vm_pageout_defer(m, PQ_INACTIVE, false))
1474 * Lock the page's object.
1476 if (object == NULL || object != m->object) {
1478 VM_OBJECT_WUNLOCK(object);
1479 object = atomic_load_ptr(&m->object);
1480 if (__predict_false(object == NULL))
1481 /* The page is being freed by another thread. */
1484 /* Depends on type-stability. */
1485 VM_OBJECT_WLOCK(object);
1486 if (__predict_false(m->object != object)) {
1487 VM_OBJECT_WUNLOCK(object);
1493 if (vm_page_tryxbusy(m) == 0) {
1495 * Don't mess with busy pages. Leave them at
1496 * the front of the queue. Most likely, they
1497 * are being paged out and will leave the
1498 * queue shortly after the scan finishes. So,
1499 * they ought to be discounted from the
1502 addl_page_shortage++;
1506 /* Deferred free of swap space. */
1507 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1508 vm_pager_page_unswapped(m);
1511 * Check for wirings now that we hold the object lock and have
1512 * exclusively busied the page. If the page is mapped, it may
1513 * still be wired by pmap lookups. The call to
1514 * vm_page_try_remove_all() below atomically checks for such
1515 * wirings and removes mappings. If the page is unmapped, the
1516 * wire count is guaranteed not to increase after this check.
1518 if (__predict_false(vm_page_wired(m)))
1522 * Invalid pages can be easily freed. They cannot be
1523 * mapped, vm_page_free() asserts this.
1525 if (vm_page_none_valid(m))
1528 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1530 for (old = vm_page_astate_load(m);;) {
1532 * Check to see if the page has been removed from the
1533 * queue since the first such check. Leave it alone if
1534 * so, discarding any references collected by
1535 * pmap_ts_referenced().
1537 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1542 if ((old.flags & PGA_REFERENCED) != 0) {
1543 new.flags &= ~PGA_REFERENCED;
1546 if (act_delta == 0) {
1548 } else if (object->ref_count != 0) {
1550 * Increase the activation count if the
1551 * page was referenced while in the
1552 * inactive queue. This makes it less
1553 * likely that the page will be returned
1554 * prematurely to the inactive queue.
1556 new.act_count += ACT_ADVANCE +
1558 if (new.act_count > ACT_MAX)
1559 new.act_count = ACT_MAX;
1561 new.flags &= ~PGA_QUEUE_OP_MASK;
1562 new.flags |= PGA_REQUEUE;
1563 new.queue = PQ_ACTIVE;
1564 if (!vm_page_pqstate_commit(m, &old, new))
1567 VM_CNT_INC(v_reactivated);
1569 } else if ((object->flags & OBJ_DEAD) == 0) {
1570 new.queue = PQ_INACTIVE;
1571 new.flags |= PGA_REQUEUE;
1572 if (!vm_page_pqstate_commit(m, &old, new))
1580 * If the page appears to be clean at the machine-independent
1581 * layer, then remove all of its mappings from the pmap in
1582 * anticipation of freeing it. If, however, any of the page's
1583 * mappings allow write access, then the page may still be
1584 * modified until the last of those mappings are removed.
1586 if (object->ref_count != 0) {
1587 vm_page_test_dirty(m);
1588 if (m->dirty == 0 && !vm_page_try_remove_all(m))
1593 * Clean pages can be freed, but dirty pages must be sent back
1594 * to the laundry, unless they belong to a dead object.
1595 * Requeueing dirty pages from dead objects is pointless, as
1596 * they are being paged out and freed by the thread that
1597 * destroyed the object.
1599 if (m->dirty == 0) {
1602 * Now we are guaranteed that no other threads are
1603 * manipulating the page, check for a last-second
1604 * reference that would save it from doom.
1606 if (vm_pageout_defer(m, PQ_INACTIVE, false))
1610 * Because we dequeued the page and have already checked
1611 * for pending dequeue and enqueue requests, we can
1612 * safely disassociate the page from the inactive queue
1613 * without holding the queue lock.
1615 m->a.queue = PQ_NONE;
1620 if ((object->flags & OBJ_DEAD) == 0)
1626 vm_pageout_reinsert_inactive(&ss, &rq, m);
1629 VM_OBJECT_WUNLOCK(object);
1630 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1631 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1632 vm_pagequeue_lock(pq);
1633 vm_pageout_end_scan(&ss);
1634 vm_pagequeue_unlock(pq);
1637 * Record the remaining shortage and the progress and rate it was made.
1639 atomic_add_int(&vmd->vmd_addl_shortage, addl_page_shortage);
1640 getmicrouptime(&end);
1641 timevalsub(&end, &start);
1642 atomic_add_int(&vmd->vmd_inactive_us,
1643 end.tv_sec * 1000000 + end.tv_usec);
1644 atomic_add_int(&vmd->vmd_inactive_freed,
1645 starting_page_shortage - page_shortage);
1649 * Dispatch a number of inactive threads according to load and collect the
1650 * results to present a coherent view of paging activity on this domain.
1653 vm_pageout_inactive_dispatch(struct vm_domain *vmd, int shortage)
1655 u_int freed, pps, slop, threads, us;
1657 vmd->vmd_inactive_shortage = shortage;
1661 * If we have more work than we can do in a quarter of our interval, we
1662 * fire off multiple threads to process it.
1664 threads = vmd->vmd_inactive_threads;
1665 if (threads > 1 && vmd->vmd_inactive_pps != 0 &&
1666 shortage > vmd->vmd_inactive_pps / VM_INACT_SCAN_RATE / 4) {
1667 vmd->vmd_inactive_shortage /= threads;
1668 slop = shortage % threads;
1669 vm_domain_pageout_lock(vmd);
1670 blockcount_acquire(&vmd->vmd_inactive_starting, threads - 1);
1671 blockcount_acquire(&vmd->vmd_inactive_running, threads - 1);
1672 wakeup(&vmd->vmd_inactive_shortage);
1673 vm_domain_pageout_unlock(vmd);
1676 /* Run the local thread scan. */
1677 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage + slop);
1680 * Block until helper threads report results and then accumulate
1683 blockcount_wait(&vmd->vmd_inactive_running, NULL, "vmpoid", PVM);
1684 freed = atomic_readandclear_int(&vmd->vmd_inactive_freed);
1685 VM_CNT_ADD(v_dfree, freed);
1688 * Calculate the per-thread paging rate with an exponential decay of
1689 * prior results. Careful to avoid integer rounding errors with large
1692 us = max(atomic_readandclear_int(&vmd->vmd_inactive_us), 1);
1694 /* Keep rounding to tenths */
1695 pps = (freed * 10) / ((us * 10) / 1000000);
1697 pps = (1000000 / us) * freed;
1698 vmd->vmd_inactive_pps = (vmd->vmd_inactive_pps / 2) + (pps / 2);
1700 return (shortage - freed);
1704 * Attempt to reclaim the requested number of pages from the inactive queue.
1705 * Returns true if the shortage was addressed.
1708 vm_pageout_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage)
1710 struct vm_pagequeue *pq;
1711 u_int addl_page_shortage, deficit, page_shortage;
1712 u_int starting_page_shortage;
1715 * vmd_pageout_deficit counts the number of pages requested in
1716 * allocations that failed because of a free page shortage. We assume
1717 * that the allocations will be reattempted and thus include the deficit
1718 * in our scan target.
1720 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1721 starting_page_shortage = shortage + deficit;
1724 * Run the inactive scan on as many threads as is necessary.
1726 page_shortage = vm_pageout_inactive_dispatch(vmd, starting_page_shortage);
1727 addl_page_shortage = atomic_readandclear_int(&vmd->vmd_addl_shortage);
1730 * Wake up the laundry thread so that it can perform any needed
1731 * laundering. If we didn't meet our target, we're in shortfall and
1732 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1733 * swap devices are configured, the laundry thread has no work to do, so
1734 * don't bother waking it up.
1736 * The laundry thread uses the number of inactive queue scans elapsed
1737 * since the last laundering to determine whether to launder again, so
1740 if (starting_page_shortage > 0) {
1741 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1742 vm_pagequeue_lock(pq);
1743 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1744 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1745 if (page_shortage > 0) {
1746 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1747 VM_CNT_INC(v_pdshortfalls);
1748 } else if (vmd->vmd_laundry_request !=
1749 VM_LAUNDRY_SHORTFALL)
1750 vmd->vmd_laundry_request =
1751 VM_LAUNDRY_BACKGROUND;
1752 wakeup(&vmd->vmd_laundry_request);
1754 vmd->vmd_clean_pages_freed +=
1755 starting_page_shortage - page_shortage;
1756 vm_pagequeue_unlock(pq);
1760 * Wakeup the swapout daemon if we didn't free the targeted number of
1763 if (page_shortage > 0)
1767 * If the inactive queue scan fails repeatedly to meet its
1768 * target, kill the largest process.
1770 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1773 * Reclaim pages by swapping out idle processes, if configured to do so.
1775 vm_swapout_run_idle();
1778 * See the description of addl_page_shortage above.
1780 *addl_shortage = addl_page_shortage + deficit;
1782 return (page_shortage <= 0);
1785 static int vm_pageout_oom_vote;
1788 * The pagedaemon threads randlomly select one to perform the
1789 * OOM. Trying to kill processes before all pagedaemons
1790 * failed to reach free target is premature.
1793 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1794 int starting_page_shortage)
1798 if (starting_page_shortage <= 0 || starting_page_shortage !=
1800 vmd->vmd_oom_seq = 0;
1803 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1805 vmd->vmd_oom = FALSE;
1806 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1812 * Do not follow the call sequence until OOM condition is
1815 vmd->vmd_oom_seq = 0;
1820 vmd->vmd_oom = TRUE;
1821 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1822 if (old_vote != vm_ndomains - 1)
1826 * The current pagedaemon thread is the last in the quorum to
1827 * start OOM. Initiate the selection and signaling of the
1830 vm_pageout_oom(VM_OOM_MEM);
1833 * After one round of OOM terror, recall our vote. On the
1834 * next pass, current pagedaemon would vote again if the low
1835 * memory condition is still there, due to vmd_oom being
1838 vmd->vmd_oom = FALSE;
1839 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1843 * The OOM killer is the page daemon's action of last resort when
1844 * memory allocation requests have been stalled for a prolonged period
1845 * of time because it cannot reclaim memory. This function computes
1846 * the approximate number of physical pages that could be reclaimed if
1847 * the specified address space is destroyed.
1849 * Private, anonymous memory owned by the address space is the
1850 * principal resource that we expect to recover after an OOM kill.
1851 * Since the physical pages mapped by the address space's COW entries
1852 * are typically shared pages, they are unlikely to be released and so
1853 * they are not counted.
1855 * To get to the point where the page daemon runs the OOM killer, its
1856 * efforts to write-back vnode-backed pages may have stalled. This
1857 * could be caused by a memory allocation deadlock in the write path
1858 * that might be resolved by an OOM kill. Therefore, physical pages
1859 * belonging to vnode-backed objects are counted, because they might
1860 * be freed without being written out first if the address space holds
1861 * the last reference to an unlinked vnode.
1863 * Similarly, physical pages belonging to OBJT_PHYS objects are
1864 * counted because the address space might hold the last reference to
1868 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1871 vm_map_entry_t entry;
1875 map = &vmspace->vm_map;
1876 KASSERT(!map->system_map, ("system map"));
1877 sx_assert(&map->lock, SA_LOCKED);
1879 VM_MAP_ENTRY_FOREACH(entry, map) {
1880 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1882 obj = entry->object.vm_object;
1885 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1886 obj->ref_count != 1)
1888 if (obj->type == OBJT_DEFAULT || obj->type == OBJT_PHYS ||
1889 obj->type == OBJT_VNODE || (obj->flags & OBJ_SWAP) != 0)
1890 res += obj->resident_page_count;
1895 static int vm_oom_ratelim_last;
1896 static int vm_oom_pf_secs = 10;
1897 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1899 static struct mtx vm_oom_ratelim_mtx;
1902 vm_pageout_oom(int shortage)
1904 struct proc *p, *bigproc;
1905 vm_offset_t size, bigsize;
1912 * For OOM requests originating from vm_fault(), there is a high
1913 * chance that a single large process faults simultaneously in
1914 * several threads. Also, on an active system running many
1915 * processes of middle-size, like buildworld, all of them
1916 * could fault almost simultaneously as well.
1918 * To avoid killing too many processes, rate-limit OOMs
1919 * initiated by vm_fault() time-outs on the waits for free
1922 mtx_lock(&vm_oom_ratelim_mtx);
1924 if (shortage == VM_OOM_MEM_PF &&
1925 (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1926 mtx_unlock(&vm_oom_ratelim_mtx);
1929 vm_oom_ratelim_last = now;
1930 mtx_unlock(&vm_oom_ratelim_mtx);
1933 * We keep the process bigproc locked once we find it to keep anyone
1934 * from messing with it; however, there is a possibility of
1935 * deadlock if process B is bigproc and one of its child processes
1936 * attempts to propagate a signal to B while we are waiting for A's
1937 * lock while walking this list. To avoid this, we don't block on
1938 * the process lock but just skip a process if it is already locked.
1942 sx_slock(&allproc_lock);
1943 FOREACH_PROC_IN_SYSTEM(p) {
1947 * If this is a system, protected or killed process, skip it.
1949 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1950 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1951 p->p_pid == 1 || P_KILLED(p) ||
1952 (p->p_pid < 48 && swap_pager_avail != 0)) {
1957 * If the process is in a non-running type state,
1958 * don't touch it. Check all the threads individually.
1961 FOREACH_THREAD_IN_PROC(p, td) {
1963 if (!TD_ON_RUNQ(td) &&
1964 !TD_IS_RUNNING(td) &&
1965 !TD_IS_SLEEPING(td) &&
1966 !TD_IS_SUSPENDED(td) &&
1967 !TD_IS_SWAPPED(td)) {
1979 * get the process size
1981 vm = vmspace_acquire_ref(p);
1988 sx_sunlock(&allproc_lock);
1989 if (!vm_map_trylock_read(&vm->vm_map)) {
1991 sx_slock(&allproc_lock);
1995 size = vmspace_swap_count(vm);
1996 if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
1997 size += vm_pageout_oom_pagecount(vm);
1998 vm_map_unlock_read(&vm->vm_map);
2000 sx_slock(&allproc_lock);
2003 * If this process is bigger than the biggest one,
2006 if (size > bigsize) {
2007 if (bigproc != NULL)
2015 sx_sunlock(&allproc_lock);
2016 if (bigproc != NULL) {
2017 if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0)
2018 panic("out of swap space");
2020 killproc(bigproc, "out of swap space");
2021 sched_nice(bigproc, PRIO_MIN);
2023 PROC_UNLOCK(bigproc);
2028 * Signal a free page shortage to subsystems that have registered an event
2029 * handler. Reclaim memory from UMA in the event of a severe shortage.
2030 * Return true if the free page count should be re-evaluated.
2033 vm_pageout_lowmem(void)
2035 static int lowmem_ticks = 0;
2041 last = atomic_load_int(&lowmem_ticks);
2042 while ((u_int)(ticks - last) / hz >= lowmem_period) {
2043 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
2047 * Decrease registered cache sizes.
2049 SDT_PROBE0(vm, , , vm__lowmem_scan);
2050 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
2053 * We do this explicitly after the caches have been
2056 uma_reclaim(UMA_RECLAIM_TRIM);
2062 * Kick off an asynchronous reclaim of cached memory if one of the
2063 * page daemons is failing to keep up with demand. Use the "severe"
2064 * threshold instead of "min" to ensure that we do not blow away the
2065 * caches if a subset of the NUMA domains are depleted by kernel memory
2066 * allocations; the domainset iterators automatically skip domains
2067 * below the "min" threshold on the first pass.
2069 * UMA reclaim worker has its own rate-limiting mechanism, so don't
2070 * worry about kicking it too often.
2072 if (vm_page_count_severe())
2073 uma_reclaim_wakeup();
2079 vm_pageout_worker(void *arg)
2081 struct vm_domain *vmd;
2083 int addl_shortage, domain, shortage;
2086 domain = (uintptr_t)arg;
2087 vmd = VM_DOMAIN(domain);
2092 * XXXKIB It could be useful to bind pageout daemon threads to
2093 * the cores belonging to the domain, from which vm_page_array
2097 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
2098 vmd->vmd_last_active_scan = ticks;
2101 * The pageout daemon worker is never done, so loop forever.
2104 vm_domain_pageout_lock(vmd);
2107 * We need to clear wanted before we check the limits. This
2108 * prevents races with wakers who will check wanted after they
2111 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
2114 * Might the page daemon need to run again?
2116 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
2118 * Yes. If the scan failed to produce enough free
2119 * pages, sleep uninterruptibly for some time in the
2120 * hope that the laundry thread will clean some pages.
2122 vm_domain_pageout_unlock(vmd);
2124 pause("pwait", hz / VM_INACT_SCAN_RATE);
2127 * No, sleep until the next wakeup or until pages
2128 * need to have their reference stats updated.
2130 if (mtx_sleep(&vmd->vmd_pageout_wanted,
2131 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
2132 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
2133 VM_CNT_INC(v_pdwakeups);
2136 /* Prevent spurious wakeups by ensuring that wanted is set. */
2137 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2140 * Use the controller to calculate how many pages to free in
2141 * this interval, and scan the inactive queue. If the lowmem
2142 * handlers appear to have freed up some pages, subtract the
2143 * difference from the inactive queue scan target.
2145 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
2147 ofree = vmd->vmd_free_count;
2148 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
2149 shortage -= min(vmd->vmd_free_count - ofree,
2151 target_met = vm_pageout_inactive(vmd, shortage,
2157 * Scan the active queue. A positive value for shortage
2158 * indicates that we must aggressively deactivate pages to avoid
2161 shortage = vm_pageout_active_target(vmd) + addl_shortage;
2162 vm_pageout_scan_active(vmd, shortage);
2167 * vm_pageout_helper runs additional pageout daemons in times of high paging
2171 vm_pageout_helper(void *arg)
2173 struct vm_domain *vmd;
2176 domain = (uintptr_t)arg;
2177 vmd = VM_DOMAIN(domain);
2179 vm_domain_pageout_lock(vmd);
2181 msleep(&vmd->vmd_inactive_shortage,
2182 vm_domain_pageout_lockptr(vmd), PVM, "psleep", 0);
2183 blockcount_release(&vmd->vmd_inactive_starting, 1);
2185 vm_domain_pageout_unlock(vmd);
2186 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage);
2187 vm_domain_pageout_lock(vmd);
2190 * Release the running count while the pageout lock is held to
2191 * prevent wakeup races.
2193 blockcount_release(&vmd->vmd_inactive_running, 1);
2198 get_pageout_threads_per_domain(const struct vm_domain *vmd)
2200 unsigned total_pageout_threads, eligible_cpus, domain_cpus;
2202 if (VM_DOMAIN_EMPTY(vmd->vmd_domain))
2206 * Semi-arbitrarily constrain pagedaemon threads to less than half the
2207 * total number of CPUs in the system as an upper limit.
2209 if (pageout_cpus_per_thread < 2)
2210 pageout_cpus_per_thread = 2;
2211 else if (pageout_cpus_per_thread > mp_ncpus)
2212 pageout_cpus_per_thread = mp_ncpus;
2214 total_pageout_threads = howmany(mp_ncpus, pageout_cpus_per_thread);
2215 domain_cpus = CPU_COUNT(&cpuset_domain[vmd->vmd_domain]);
2217 /* Pagedaemons are not run in empty domains. */
2218 eligible_cpus = mp_ncpus;
2219 for (unsigned i = 0; i < vm_ndomains; i++)
2220 if (VM_DOMAIN_EMPTY(i))
2221 eligible_cpus -= CPU_COUNT(&cpuset_domain[i]);
2224 * Assign a portion of the total pageout threads to this domain
2225 * corresponding to the fraction of pagedaemon-eligible CPUs in the
2226 * domain. In asymmetric NUMA systems, domains with more CPUs may be
2227 * allocated more threads than domains with fewer CPUs.
2229 return (howmany(total_pageout_threads * domain_cpus, eligible_cpus));
2233 * Initialize basic pageout daemon settings. See the comment above the
2234 * definition of vm_domain for some explanation of how these thresholds are
2238 vm_pageout_init_domain(int domain)
2240 struct vm_domain *vmd;
2241 struct sysctl_oid *oid;
2243 vmd = VM_DOMAIN(domain);
2244 vmd->vmd_interrupt_free_min = 2;
2247 * v_free_reserved needs to include enough for the largest
2248 * swap pager structures plus enough for any pv_entry structs
2251 vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE +
2252 vmd->vmd_interrupt_free_min;
2253 vmd->vmd_free_reserved = vm_pageout_page_count +
2254 vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768;
2255 vmd->vmd_free_min = vmd->vmd_page_count / 200;
2256 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2257 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2258 vmd->vmd_free_min += vmd->vmd_free_reserved;
2259 vmd->vmd_free_severe += vmd->vmd_free_reserved;
2260 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2261 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2262 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2265 * Set the default wakeup threshold to be 10% below the paging
2266 * target. This keeps the steady state out of shortfall.
2268 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2271 * Target amount of memory to move out of the laundry queue during a
2272 * background laundering. This is proportional to the amount of system
2275 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2276 vmd->vmd_free_min) / 10;
2278 /* Initialize the pageout daemon pid controller. */
2279 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2280 vmd->vmd_free_target, PIDCTRL_BOUND,
2281 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2282 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2283 "pidctrl", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2284 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2286 vmd->vmd_inactive_threads = get_pageout_threads_per_domain(vmd);
2290 vm_pageout_init(void)
2296 * Initialize some paging parameters.
2298 if (vm_cnt.v_page_count < 2000)
2299 vm_pageout_page_count = 8;
2302 for (i = 0; i < vm_ndomains; i++) {
2303 struct vm_domain *vmd;
2305 vm_pageout_init_domain(i);
2307 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2308 vm_cnt.v_free_target += vmd->vmd_free_target;
2309 vm_cnt.v_free_min += vmd->vmd_free_min;
2310 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2311 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2312 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2313 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2314 freecount += vmd->vmd_free_count;
2318 * Set interval in seconds for active scan. We want to visit each
2319 * page at least once every ten minutes. This is to prevent worst
2320 * case paging behaviors with stale active LRU.
2322 if (vm_pageout_update_period == 0)
2323 vm_pageout_update_period = 600;
2326 * Set the maximum number of user-wired virtual pages. Historically the
2327 * main source of such pages was mlock(2) and mlockall(2). Hypervisors
2328 * may also request user-wired memory.
2330 if (vm_page_max_user_wired == 0)
2331 vm_page_max_user_wired = 4 * freecount / 5;
2335 * vm_pageout is the high level pageout daemon.
2342 int error, first, i, j, pageout_threads;
2347 mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2348 swap_pager_swap_init();
2349 for (first = -1, i = 0; i < vm_ndomains; i++) {
2350 if (VM_DOMAIN_EMPTY(i)) {
2352 printf("domain %d empty; skipping pageout\n",
2359 error = kthread_add(vm_pageout_worker,
2360 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2362 panic("starting pageout for domain %d: %d\n",
2365 pageout_threads = VM_DOMAIN(i)->vmd_inactive_threads;
2366 for (j = 0; j < pageout_threads - 1; j++) {
2367 error = kthread_add(vm_pageout_helper,
2368 (void *)(uintptr_t)i, p, NULL, 0, 0,
2369 "dom%d helper%d", i, j);
2371 panic("starting pageout helper %d for domain "
2372 "%d: %d\n", j, i, error);
2374 error = kthread_add(vm_pageout_laundry_worker,
2375 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2377 panic("starting laundry for domain %d: %d", i, error);
2379 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2381 panic("starting uma_reclaim helper, error %d\n", error);
2383 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2384 vm_pageout_worker((void *)(uintptr_t)first);
2388 * Perform an advisory wakeup of the page daemon.
2391 pagedaemon_wakeup(int domain)
2393 struct vm_domain *vmd;
2395 vmd = VM_DOMAIN(domain);
2396 vm_domain_pageout_assert_unlocked(vmd);
2397 if (curproc == pageproc)
2400 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2401 vm_domain_pageout_lock(vmd);
2402 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2403 wakeup(&vmd->vmd_pageout_wanted);
2404 vm_domain_pageout_unlock(vmd);