2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
10 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * The Mach Operating System project at Carnegie-Mellon University.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
44 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
47 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
48 * All rights reserved.
50 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
52 * Permission to use, copy, modify and distribute this software and
53 * its documentation is hereby granted, provided that both the copyright
54 * notice and this permission notice appear in all copies of the
55 * software, derivative works or modified versions, and any portions
56 * thereof, and that both notices appear in supporting documentation.
58 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
59 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
60 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
62 * Carnegie Mellon requests users of this software to return to
64 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
65 * School of Computer Science
66 * Carnegie Mellon University
67 * Pittsburgh PA 15213-3890
69 * any improvements or extensions that they make and grant Carnegie the
70 * rights to redistribute these changes.
74 * The proverbial page-out daemon.
77 #include <sys/cdefs.h>
78 __FBSDID("$FreeBSD$");
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
85 #include <sys/eventhandler.h>
87 #include <sys/mutex.h>
89 #include <sys/kthread.h>
91 #include <sys/mount.h>
92 #include <sys/racct.h>
93 #include <sys/resourcevar.h>
94 #include <sys/sched.h>
96 #include <sys/signalvar.h>
99 #include <sys/vnode.h>
100 #include <sys/vmmeter.h>
101 #include <sys/rwlock.h>
103 #include <sys/sysctl.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/vm_pagequeue.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
119 * System initialization
122 /* the kernel process "vm_pageout"*/
123 static void vm_pageout(void);
124 static void vm_pageout_init(void);
125 static int vm_pageout_clean(vm_page_t m, int *numpagedout);
126 static int vm_pageout_cluster(vm_page_t m);
127 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
128 int starting_page_shortage);
130 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
133 struct proc *pageproc;
135 static struct kproc_desc page_kp = {
140 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
143 SDT_PROVIDER_DEFINE(vm);
144 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
146 /* Pagedaemon activity rates, in subdivisions of one second. */
147 #define VM_LAUNDER_RATE 10
148 #define VM_INACT_SCAN_RATE 10
150 static int vm_pageout_oom_seq = 12;
152 static int vm_pageout_update_period;
153 static int disable_swap_pageouts;
154 static int lowmem_period = 10;
155 static int swapdev_enabled;
157 static int vm_panic_on_oom = 0;
159 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
160 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
161 "panic on out of memory instead of killing the largest process");
163 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
164 CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
165 "Maximum active LRU update period");
167 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
168 "Low memory callback period");
170 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
171 CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
173 static int pageout_lock_miss;
174 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
175 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
177 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
178 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
179 "back-to-back calls to oom detector to start OOM");
181 static int act_scan_laundry_weight = 3;
182 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
183 &act_scan_laundry_weight, 0,
184 "weight given to clean vs. dirty pages in active queue scans");
186 static u_int vm_background_launder_rate = 4096;
187 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
188 &vm_background_launder_rate, 0,
189 "background laundering rate, in kilobytes per second");
191 static u_int vm_background_launder_max = 20 * 1024;
192 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
193 &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
195 int vm_pageout_page_count = 32;
197 u_long vm_page_max_user_wired;
198 SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW,
199 &vm_page_max_user_wired, 0,
200 "system-wide limit to user-wired page count");
202 static u_int isqrt(u_int num);
203 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
205 static void vm_pageout_laundry_worker(void *arg);
208 struct vm_batchqueue bq;
209 struct vm_pagequeue *pq;
216 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
217 vm_page_t marker, vm_page_t after, int maxscan)
220 vm_pagequeue_assert_locked(pq);
221 KASSERT((marker->aflags & PGA_ENQUEUED) == 0,
222 ("marker %p already enqueued", marker));
225 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
227 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
228 vm_page_aflag_set(marker, PGA_ENQUEUED);
230 vm_batchqueue_init(&ss->bq);
233 ss->maxscan = maxscan;
235 vm_pagequeue_unlock(pq);
239 vm_pageout_end_scan(struct scan_state *ss)
241 struct vm_pagequeue *pq;
244 vm_pagequeue_assert_locked(pq);
245 KASSERT((ss->marker->aflags & PGA_ENQUEUED) != 0,
246 ("marker %p not enqueued", ss->marker));
248 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
249 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
250 pq->pq_pdpages += ss->scanned;
254 * Add a small number of queued pages to a batch queue for later processing
255 * without the corresponding queue lock held. The caller must have enqueued a
256 * marker page at the desired start point for the scan. Pages will be
257 * physically dequeued if the caller so requests. Otherwise, the returned
258 * batch may contain marker pages, and it is up to the caller to handle them.
260 * When processing the batch queue, vm_page_queue() must be used to
261 * determine whether the page has been logically dequeued by another thread.
262 * Once this check is performed, the page lock guarantees that the page will
263 * not be disassociated from the queue.
265 static __always_inline void
266 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
268 struct vm_pagequeue *pq;
274 KASSERT((marker->aflags & PGA_ENQUEUED) != 0,
275 ("marker %p not enqueued", ss->marker));
277 vm_pagequeue_lock(pq);
278 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
279 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
280 m = TAILQ_NEXT(m, plinks.q), ss->scanned++) {
281 if ((m->flags & PG_MARKER) == 0) {
282 KASSERT((m->aflags & PGA_ENQUEUED) != 0,
283 ("page %p not enqueued", m));
284 KASSERT((m->flags & PG_FICTITIOUS) == 0,
285 ("Fictitious page %p cannot be in page queue", m));
286 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
287 ("Unmanaged page %p cannot be in page queue", m));
291 (void)vm_batchqueue_insert(&ss->bq, m);
293 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
294 vm_page_aflag_clear(m, PGA_ENQUEUED);
297 TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
298 if (__predict_true(m != NULL))
299 TAILQ_INSERT_BEFORE(m, marker, plinks.q);
301 TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
303 vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
304 vm_pagequeue_unlock(pq);
307 /* Return the next page to be scanned, or NULL if the scan is complete. */
308 static __always_inline vm_page_t
309 vm_pageout_next(struct scan_state *ss, const bool dequeue)
312 if (ss->bq.bq_cnt == 0)
313 vm_pageout_collect_batch(ss, dequeue);
314 return (vm_batchqueue_pop(&ss->bq));
318 * Scan for pages at adjacent offsets within the given page's object that are
319 * eligible for laundering, form a cluster of these pages and the given page,
320 * and launder that cluster.
323 vm_pageout_cluster(vm_page_t m)
326 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
328 int ib, is, page_base, pageout_count;
330 vm_page_assert_locked(m);
332 VM_OBJECT_ASSERT_WLOCKED(object);
335 vm_page_assert_unbusied(m);
336 KASSERT(!vm_page_held(m), ("page %p is held", m));
338 pmap_remove_write(m);
341 mc[vm_pageout_page_count] = pb = ps = m;
343 page_base = vm_pageout_page_count;
348 * We can cluster only if the page is not clean, busy, or held, and
349 * the page is in the laundry queue.
351 * During heavy mmap/modification loads the pageout
352 * daemon can really fragment the underlying file
353 * due to flushing pages out of order and not trying to
354 * align the clusters (which leaves sporadic out-of-order
355 * holes). To solve this problem we do the reverse scan
356 * first and attempt to align our cluster, then do a
357 * forward scan if room remains.
360 while (ib != 0 && pageout_count < vm_pageout_page_count) {
365 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
369 vm_page_test_dirty(p);
375 if (vm_page_held(p) || !vm_page_in_laundry(p)) {
380 pmap_remove_write(p);
382 mc[--page_base] = pb = p;
387 * We are at an alignment boundary. Stop here, and switch
388 * directions. Do not clear ib.
390 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
393 while (pageout_count < vm_pageout_page_count &&
394 pindex + is < object->size) {
395 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
397 vm_page_test_dirty(p);
401 if (vm_page_held(p) || !vm_page_in_laundry(p)) {
405 pmap_remove_write(p);
407 mc[page_base + pageout_count] = ps = p;
413 * If we exhausted our forward scan, continue with the reverse scan
414 * when possible, even past an alignment boundary. This catches
415 * boundary conditions.
417 if (ib != 0 && pageout_count < vm_pageout_page_count)
420 return (vm_pageout_flush(&mc[page_base], pageout_count,
421 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
425 * vm_pageout_flush() - launder the given pages
427 * The given pages are laundered. Note that we setup for the start of
428 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
429 * reference count all in here rather then in the parent. If we want
430 * the parent to do more sophisticated things we may have to change
433 * Returned runlen is the count of pages between mreq and first
434 * page after mreq with status VM_PAGER_AGAIN.
435 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
436 * for any page in runlen set.
439 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
442 vm_object_t object = mc[0]->object;
443 int pageout_status[count];
447 VM_OBJECT_ASSERT_WLOCKED(object);
450 * Initiate I/O. Mark the pages busy and verify that they're valid
453 * We do not have to fixup the clean/dirty bits here... we can
454 * allow the pager to do it after the I/O completes.
456 * NOTE! mc[i]->dirty may be partial or fragmented due to an
457 * edge case with file fragments.
459 for (i = 0; i < count; i++) {
460 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
461 ("vm_pageout_flush: partially invalid page %p index %d/%d",
463 KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
464 ("vm_pageout_flush: writeable page %p", mc[i]));
465 vm_page_sbusy(mc[i]);
467 vm_object_pip_add(object, count);
469 vm_pager_put_pages(object, mc, count, flags, pageout_status);
471 runlen = count - mreq;
474 for (i = 0; i < count; i++) {
475 vm_page_t mt = mc[i];
477 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
478 !pmap_page_is_write_mapped(mt),
479 ("vm_pageout_flush: page %p is not write protected", mt));
480 switch (pageout_status[i]) {
483 if (vm_page_in_laundry(mt))
484 vm_page_deactivate_noreuse(mt);
492 * The page is outside the object's range. We pretend
493 * that the page out worked and clean the page, so the
494 * changes will be lost if the page is reclaimed by
499 if (vm_page_in_laundry(mt))
500 vm_page_deactivate_noreuse(mt);
506 * If the page couldn't be paged out to swap because the
507 * pager wasn't able to find space, place the page in
508 * the PQ_UNSWAPPABLE holding queue. This is an
509 * optimization that prevents the page daemon from
510 * wasting CPU cycles on pages that cannot be reclaimed
511 * becase no swap device is configured.
513 * Otherwise, reactivate the page so that it doesn't
514 * clog the laundry and inactive queues. (We will try
515 * paging it out again later.)
518 if (object->type == OBJT_SWAP &&
519 pageout_status[i] == VM_PAGER_FAIL) {
520 vm_page_unswappable(mt);
523 vm_page_activate(mt);
525 if (eio != NULL && i >= mreq && i - mreq < runlen)
529 if (i >= mreq && i - mreq < runlen)
535 * If the operation is still going, leave the page busy to
536 * block all other accesses. Also, leave the paging in
537 * progress indicator set so that we don't attempt an object
540 if (pageout_status[i] != VM_PAGER_PEND) {
541 vm_object_pip_wakeup(object);
547 return (numpagedout);
551 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
554 atomic_store_rel_int(&swapdev_enabled, 1);
558 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
561 if (swap_pager_nswapdev() == 1)
562 atomic_store_rel_int(&swapdev_enabled, 0);
566 * Attempt to acquire all of the necessary locks to launder a page and
567 * then call through the clustering layer to PUTPAGES. Wait a short
568 * time for a vnode lock.
570 * Requires the page and object lock on entry, releases both before return.
571 * Returns 0 on success and an errno otherwise.
574 vm_pageout_clean(vm_page_t m, int *numpagedout)
582 vm_page_assert_locked(m);
584 VM_OBJECT_ASSERT_WLOCKED(object);
590 * The object is already known NOT to be dead. It
591 * is possible for the vget() to block the whole
592 * pageout daemon, but the new low-memory handling
593 * code should prevent it.
595 * We can't wait forever for the vnode lock, we might
596 * deadlock due to a vn_read() getting stuck in
597 * vm_wait while holding this vnode. We skip the
598 * vnode if we can't get it in a reasonable amount
601 if (object->type == OBJT_VNODE) {
604 if (vp->v_type == VREG &&
605 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
611 ("vp %p with NULL v_mount", vp));
612 vm_object_reference_locked(object);
614 VM_OBJECT_WUNLOCK(object);
615 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
616 LK_SHARED : LK_EXCLUSIVE;
617 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
622 VM_OBJECT_WLOCK(object);
625 * Ensure that the object and vnode were not disassociated
626 * while locks were dropped.
628 if (vp->v_object != object) {
635 * While the object and page were unlocked, the page
637 * (1) moved to a different queue,
638 * (2) reallocated to a different object,
639 * (3) reallocated to a different offset, or
642 if (!vm_page_in_laundry(m) || m->object != object ||
643 m->pindex != pindex || m->dirty == 0) {
650 * The page may have been busied or referenced while the object
651 * and page locks were released.
653 if (vm_page_busied(m) || vm_page_held(m)) {
661 * If a page is dirty, then it is either being washed
662 * (but not yet cleaned) or it is still in the
663 * laundry. If it is still in the laundry, then we
664 * start the cleaning operation.
666 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
670 VM_OBJECT_WUNLOCK(object);
673 vm_page_lock_assert(m, MA_NOTOWNED);
677 vm_object_deallocate(object);
678 vn_finished_write(mp);
685 * Attempt to launder the specified number of pages.
687 * Returns the number of pages successfully laundered.
690 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
692 struct scan_state ss;
693 struct vm_pagequeue *pq;
697 int act_delta, error, numpagedout, queue, starting_target;
703 starting_target = launder;
707 * Scan the laundry queues for pages eligible to be laundered. We stop
708 * once the target number of dirty pages have been laundered, or once
709 * we've reached the end of the queue. A single iteration of this loop
710 * may cause more than one page to be laundered because of clustering.
712 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
713 * swap devices are configured.
715 if (atomic_load_acq_int(&swapdev_enabled))
716 queue = PQ_UNSWAPPABLE;
721 marker = &vmd->vmd_markers[queue];
722 pq = &vmd->vmd_pagequeues[queue];
723 vm_pagequeue_lock(pq);
724 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
725 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
726 if (__predict_false((m->flags & PG_MARKER) != 0))
729 vm_page_change_lock(m, &mtx);
733 * The page may have been disassociated from the queue
734 * while locks were dropped.
736 if (vm_page_queue(m) != queue)
740 * A requeue was requested, so this page gets a second
743 if ((m->aflags & PGA_REQUEUE) != 0) {
749 * Held pages are essentially stuck in the queue.
751 * Wired pages may not be freed. Complete their removal
752 * from the queue now to avoid needless revisits during
755 if (m->hold_count != 0)
757 if (m->wire_count != 0) {
758 vm_page_dequeue_deferred(m);
762 if (object != m->object) {
764 VM_OBJECT_WUNLOCK(object);
766 if (!VM_OBJECT_TRYWLOCK(object)) {
768 /* Depends on type-stability. */
769 VM_OBJECT_WLOCK(object);
775 if (vm_page_busied(m))
779 * Invalid pages can be easily freed. They cannot be
780 * mapped; vm_page_free() asserts this.
786 * If the page has been referenced and the object is not dead,
787 * reactivate or requeue the page depending on whether the
790 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
791 * that a reference from a concurrently destroyed mapping is
792 * observed here and now.
794 if (object->ref_count != 0)
795 act_delta = pmap_ts_referenced(m);
797 KASSERT(!pmap_page_is_mapped(m),
798 ("page %p is mapped", m));
801 if ((m->aflags & PGA_REFERENCED) != 0) {
802 vm_page_aflag_clear(m, PGA_REFERENCED);
805 if (act_delta != 0) {
806 if (object->ref_count != 0) {
807 VM_CNT_INC(v_reactivated);
811 * Increase the activation count if the page
812 * was referenced while in the laundry queue.
813 * This makes it less likely that the page will
814 * be returned prematurely to the inactive
817 m->act_count += act_delta + ACT_ADVANCE;
820 * If this was a background laundering, count
821 * activated pages towards our target. The
822 * purpose of background laundering is to ensure
823 * that pages are eventually cycled through the
824 * laundry queue, and an activation is a valid
830 } else if ((object->flags & OBJ_DEAD) == 0) {
837 * If the page appears to be clean at the machine-independent
838 * layer, then remove all of its mappings from the pmap in
839 * anticipation of freeing it. If, however, any of the page's
840 * mappings allow write access, then the page may still be
841 * modified until the last of those mappings are removed.
843 if (object->ref_count != 0) {
844 vm_page_test_dirty(m);
850 * Clean pages are freed, and dirty pages are paged out unless
851 * they belong to a dead object. Requeueing dirty pages from
852 * dead objects is pointless, as they are being paged out and
853 * freed by the thread that destroyed the object.
859 } else if ((object->flags & OBJ_DEAD) == 0) {
860 if (object->type != OBJT_SWAP &&
861 object->type != OBJT_DEFAULT)
863 else if (disable_swap_pageouts)
873 * Form a cluster with adjacent, dirty pages from the
874 * same object, and page out that entire cluster.
876 * The adjacent, dirty pages must also be in the
877 * laundry. However, their mappings are not checked
878 * for new references. Consequently, a recently
879 * referenced page may be paged out. However, that
880 * page will not be prematurely reclaimed. After page
881 * out, the page will be placed in the inactive queue,
882 * where any new references will be detected and the
885 error = vm_pageout_clean(m, &numpagedout);
887 launder -= numpagedout;
888 ss.scanned += numpagedout;
889 } else if (error == EDEADLK) {
901 if (object != NULL) {
902 VM_OBJECT_WUNLOCK(object);
905 vm_pagequeue_lock(pq);
906 vm_pageout_end_scan(&ss);
907 vm_pagequeue_unlock(pq);
909 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
915 * Wakeup the sync daemon if we skipped a vnode in a writeable object
916 * and we didn't launder enough pages.
918 if (vnodes_skipped > 0 && launder > 0)
919 (void)speedup_syncer();
921 return (starting_target - launder);
925 * Compute the integer square root.
930 u_int bit, root, tmp;
932 bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
947 * Perform the work of the laundry thread: periodically wake up and determine
948 * whether any pages need to be laundered. If so, determine the number of pages
949 * that need to be laundered, and launder them.
952 vm_pageout_laundry_worker(void *arg)
954 struct vm_domain *vmd;
955 struct vm_pagequeue *pq;
956 uint64_t nclean, ndirty, nfreed;
957 int domain, last_target, launder, shortfall, shortfall_cycle, target;
960 domain = (uintptr_t)arg;
961 vmd = VM_DOMAIN(domain);
962 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
963 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
966 in_shortfall = false;
968 last_target = target = 0;
972 * Calls to these handlers are serialized by the swap syscall lock.
974 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
975 EVENTHANDLER_PRI_ANY);
976 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
977 EVENTHANDLER_PRI_ANY);
980 * The pageout laundry worker is never done, so loop forever.
983 KASSERT(target >= 0, ("negative target %d", target));
984 KASSERT(shortfall_cycle >= 0,
985 ("negative cycle %d", shortfall_cycle));
989 * First determine whether we need to launder pages to meet a
990 * shortage of free pages.
994 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
996 } else if (!in_shortfall)
998 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1000 * We recently entered shortfall and began laundering
1001 * pages. If we have completed that laundering run
1002 * (and we are no longer in shortfall) or we have met
1003 * our laundry target through other activity, then we
1004 * can stop laundering pages.
1006 in_shortfall = false;
1010 launder = target / shortfall_cycle--;
1014 * There's no immediate need to launder any pages; see if we
1015 * meet the conditions to perform background laundering:
1017 * 1. The ratio of dirty to clean inactive pages exceeds the
1018 * background laundering threshold, or
1019 * 2. we haven't yet reached the target of the current
1020 * background laundering run.
1022 * The background laundering threshold is not a constant.
1023 * Instead, it is a slowly growing function of the number of
1024 * clean pages freed by the page daemon since the last
1025 * background laundering. Thus, as the ratio of dirty to
1026 * clean inactive pages grows, the amount of memory pressure
1027 * required to trigger laundering decreases. We ensure
1028 * that the threshold is non-zero after an inactive queue
1029 * scan, even if that scan failed to free a single clean page.
1032 nclean = vmd->vmd_free_count +
1033 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1034 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1035 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1036 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1037 target = vmd->vmd_background_launder_target;
1041 * We have a non-zero background laundering target. If we've
1042 * laundered up to our maximum without observing a page daemon
1043 * request, just stop. This is a safety belt that ensures we
1044 * don't launder an excessive amount if memory pressure is low
1045 * and the ratio of dirty to clean pages is large. Otherwise,
1046 * proceed at the background laundering rate.
1051 last_target = target;
1052 } else if (last_target - target >=
1053 vm_background_launder_max * PAGE_SIZE / 1024) {
1056 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1057 launder /= VM_LAUNDER_RATE;
1058 if (launder > target)
1065 * Because of I/O clustering, the number of laundered
1066 * pages could exceed "target" by the maximum size of
1067 * a cluster minus one.
1069 target -= min(vm_pageout_launder(vmd, launder,
1070 in_shortfall), target);
1071 pause("laundp", hz / VM_LAUNDER_RATE);
1075 * If we're not currently laundering pages and the page daemon
1076 * hasn't posted a new request, sleep until the page daemon
1079 vm_pagequeue_lock(pq);
1080 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1081 (void)mtx_sleep(&vmd->vmd_laundry_request,
1082 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1085 * If the pagedaemon has indicated that it's in shortfall, start
1086 * a shortfall laundering unless we're already in the middle of
1087 * one. This may preempt a background laundering.
1089 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1090 (!in_shortfall || shortfall_cycle == 0)) {
1091 shortfall = vm_laundry_target(vmd) +
1092 vmd->vmd_pageout_deficit;
1098 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1099 nfreed += vmd->vmd_clean_pages_freed;
1100 vmd->vmd_clean_pages_freed = 0;
1101 vm_pagequeue_unlock(pq);
1106 * Compute the number of pages we want to try to move from the
1107 * active queue to either the inactive or laundry queue.
1109 * When scanning active pages during a shortage, we make clean pages
1110 * count more heavily towards the page shortage than dirty pages.
1111 * This is because dirty pages must be laundered before they can be
1112 * reused and thus have less utility when attempting to quickly
1113 * alleviate a free page shortage. However, this weighting also
1114 * causes the scan to deactivate dirty pages more aggressively,
1115 * improving the effectiveness of clustering.
1118 vm_pageout_active_target(struct vm_domain *vmd)
1122 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1123 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1124 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1125 shortage *= act_scan_laundry_weight;
1130 * Scan the active queue. If there is no shortage of inactive pages, scan a
1131 * small portion of the queue in order to maintain quasi-LRU.
1134 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1136 struct scan_state ss;
1138 vm_page_t m, marker;
1139 struct vm_pagequeue *pq;
1141 int act_delta, max_scan, scan_tick;
1143 marker = &vmd->vmd_markers[PQ_ACTIVE];
1144 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1145 vm_pagequeue_lock(pq);
1148 * If we're just idle polling attempt to visit every
1149 * active page within 'update_period' seconds.
1152 if (vm_pageout_update_period != 0) {
1153 min_scan = pq->pq_cnt;
1154 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1155 min_scan /= hz * vm_pageout_update_period;
1158 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1159 vmd->vmd_last_active_scan = scan_tick;
1162 * Scan the active queue for pages that can be deactivated. Update
1163 * the per-page activity counter and use it to identify deactivation
1164 * candidates. Held pages may be deactivated.
1166 * To avoid requeuing each page that remains in the active queue, we
1167 * implement the CLOCK algorithm. To keep the implementation of the
1168 * enqueue operation consistent for all page queues, we use two hands,
1169 * represented by marker pages. Scans begin at the first hand, which
1170 * precedes the second hand in the queue. When the two hands meet,
1171 * they are moved back to the head and tail of the queue, respectively,
1172 * and scanning resumes.
1174 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1177 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1178 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1179 if (__predict_false(m == &vmd->vmd_clock[1])) {
1180 vm_pagequeue_lock(pq);
1181 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1182 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1183 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1185 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1187 max_scan -= ss.scanned;
1188 vm_pageout_end_scan(&ss);
1191 if (__predict_false((m->flags & PG_MARKER) != 0))
1194 vm_page_change_lock(m, &mtx);
1197 * The page may have been disassociated from the queue
1198 * while locks were dropped.
1200 if (vm_page_queue(m) != PQ_ACTIVE)
1204 * Wired pages are dequeued lazily.
1206 if (m->wire_count != 0) {
1207 vm_page_dequeue_deferred(m);
1212 * Check to see "how much" the page has been used.
1214 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1215 * that a reference from a concurrently destroyed mapping is
1216 * observed here and now.
1218 * Perform an unsynchronized object ref count check. While
1219 * the page lock ensures that the page is not reallocated to
1220 * another object, in particular, one with unmanaged mappings
1221 * that cannot support pmap_ts_referenced(), two races are,
1222 * nonetheless, possible:
1223 * 1) The count was transitioning to zero, but we saw a non-
1224 * zero value. pmap_ts_referenced() will return zero
1225 * because the page is not mapped.
1226 * 2) The count was transitioning to one, but we saw zero.
1227 * This race delays the detection of a new reference. At
1228 * worst, we will deactivate and reactivate the page.
1230 if (m->object->ref_count != 0)
1231 act_delta = pmap_ts_referenced(m);
1234 if ((m->aflags & PGA_REFERENCED) != 0) {
1235 vm_page_aflag_clear(m, PGA_REFERENCED);
1240 * Advance or decay the act_count based on recent usage.
1242 if (act_delta != 0) {
1243 m->act_count += ACT_ADVANCE + act_delta;
1244 if (m->act_count > ACT_MAX)
1245 m->act_count = ACT_MAX;
1247 m->act_count -= min(m->act_count, ACT_DECLINE);
1249 if (m->act_count == 0) {
1251 * When not short for inactive pages, let dirty pages go
1252 * through the inactive queue before moving to the
1253 * laundry queues. This gives them some extra time to
1254 * be reactivated, potentially avoiding an expensive
1255 * pageout. However, during a page shortage, the
1256 * inactive queue is necessarily small, and so dirty
1257 * pages would only spend a trivial amount of time in
1258 * the inactive queue. Therefore, we might as well
1259 * place them directly in the laundry queue to reduce
1262 if (page_shortage <= 0)
1263 vm_page_deactivate(m);
1266 * Calling vm_page_test_dirty() here would
1267 * require acquisition of the object's write
1268 * lock. However, during a page shortage,
1269 * directing dirty pages into the laundry
1270 * queue is only an optimization and not a
1271 * requirement. Therefore, we simply rely on
1272 * the opportunistic updates to the page's
1273 * dirty field by the pmap.
1275 if (m->dirty == 0) {
1276 vm_page_deactivate(m);
1278 act_scan_laundry_weight;
1290 vm_pagequeue_lock(pq);
1291 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1292 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1293 vm_pageout_end_scan(&ss);
1294 vm_pagequeue_unlock(pq);
1298 vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m)
1300 struct vm_domain *vmd;
1302 if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0)
1304 vm_page_aflag_set(m, PGA_ENQUEUED);
1305 if ((m->aflags & PGA_REQUEUE_HEAD) != 0) {
1306 vmd = vm_pagequeue_domain(m);
1307 TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
1308 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1309 } else if ((m->aflags & PGA_REQUEUE) != 0) {
1310 TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q);
1311 vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1313 TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q);
1318 * Re-add stuck pages to the inactive queue. We will examine them again
1319 * during the next scan. If the queue state of a page has changed since
1320 * it was physically removed from the page queue in
1321 * vm_pageout_collect_batch(), don't do anything with that page.
1324 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1327 struct vm_pagequeue *pq;
1334 if (vm_batchqueue_insert(bq, m))
1336 vm_pagequeue_lock(pq);
1337 delta += vm_pageout_reinsert_inactive_page(ss, m);
1339 vm_pagequeue_lock(pq);
1340 while ((m = vm_batchqueue_pop(bq)) != NULL)
1341 delta += vm_pageout_reinsert_inactive_page(ss, m);
1342 vm_pagequeue_cnt_add(pq, delta);
1343 vm_pagequeue_unlock(pq);
1344 vm_batchqueue_init(bq);
1348 * Attempt to reclaim the requested number of pages from the inactive queue.
1349 * Returns true if the shortage was addressed.
1352 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1355 struct scan_state ss;
1356 struct vm_batchqueue rq;
1358 vm_page_t m, marker;
1359 struct vm_pagequeue *pq;
1361 int act_delta, addl_page_shortage, deficit, page_shortage;
1362 int starting_page_shortage;
1365 * The addl_page_shortage is an estimate of the number of temporarily
1366 * stuck pages in the inactive queue. In other words, the
1367 * number of pages from the inactive count that should be
1368 * discounted in setting the target for the active queue scan.
1370 addl_page_shortage = 0;
1373 * vmd_pageout_deficit counts the number of pages requested in
1374 * allocations that failed because of a free page shortage. We assume
1375 * that the allocations will be reattempted and thus include the deficit
1376 * in our scan target.
1378 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1379 starting_page_shortage = page_shortage = shortage + deficit;
1383 vm_batchqueue_init(&rq);
1386 * Start scanning the inactive queue for pages that we can free. The
1387 * scan will stop when we reach the target or we have scanned the
1388 * entire queue. (Note that m->act_count is not used to make
1389 * decisions for the inactive queue, only for the active queue.)
1391 marker = &vmd->vmd_markers[PQ_INACTIVE];
1392 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1393 vm_pagequeue_lock(pq);
1394 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1395 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1396 KASSERT((m->flags & PG_MARKER) == 0,
1397 ("marker page %p was dequeued", m));
1399 vm_page_change_lock(m, &mtx);
1403 * The page may have been disassociated from the queue
1404 * while locks were dropped.
1406 if (vm_page_queue(m) != PQ_INACTIVE) {
1407 addl_page_shortage++;
1412 * The page was re-enqueued after the page queue lock was
1413 * dropped, or a requeue was requested. This page gets a second
1416 if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE |
1417 PGA_REQUEUE_HEAD)) != 0)
1421 * Held pages are essentially stuck in the queue. So,
1422 * they ought to be discounted from the inactive count.
1423 * See the description of addl_page_shortage above.
1425 * Wired pages may not be freed. Complete their removal
1426 * from the queue now to avoid needless revisits during
1429 if (m->hold_count != 0) {
1430 addl_page_shortage++;
1433 if (m->wire_count != 0) {
1434 vm_page_dequeue_deferred(m);
1438 if (object != m->object) {
1440 VM_OBJECT_WUNLOCK(object);
1442 if (!VM_OBJECT_TRYWLOCK(object)) {
1444 /* Depends on type-stability. */
1445 VM_OBJECT_WLOCK(object);
1451 if (vm_page_busied(m)) {
1453 * Don't mess with busy pages. Leave them at
1454 * the front of the queue. Most likely, they
1455 * are being paged out and will leave the
1456 * queue shortly after the scan finishes. So,
1457 * they ought to be discounted from the
1460 addl_page_shortage++;
1465 * Invalid pages can be easily freed. They cannot be
1466 * mapped, vm_page_free() asserts this.
1472 * If the page has been referenced and the object is not dead,
1473 * reactivate or requeue the page depending on whether the
1476 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1477 * that a reference from a concurrently destroyed mapping is
1478 * observed here and now.
1480 if (object->ref_count != 0)
1481 act_delta = pmap_ts_referenced(m);
1483 KASSERT(!pmap_page_is_mapped(m),
1484 ("page %p is mapped", m));
1487 if ((m->aflags & PGA_REFERENCED) != 0) {
1488 vm_page_aflag_clear(m, PGA_REFERENCED);
1491 if (act_delta != 0) {
1492 if (object->ref_count != 0) {
1493 VM_CNT_INC(v_reactivated);
1494 vm_page_activate(m);
1497 * Increase the activation count if the page
1498 * was referenced while in the inactive queue.
1499 * This makes it less likely that the page will
1500 * be returned prematurely to the inactive
1503 m->act_count += act_delta + ACT_ADVANCE;
1505 } else if ((object->flags & OBJ_DEAD) == 0) {
1506 vm_page_aflag_set(m, PGA_REQUEUE);
1512 * If the page appears to be clean at the machine-independent
1513 * layer, then remove all of its mappings from the pmap in
1514 * anticipation of freeing it. If, however, any of the page's
1515 * mappings allow write access, then the page may still be
1516 * modified until the last of those mappings are removed.
1518 if (object->ref_count != 0) {
1519 vm_page_test_dirty(m);
1525 * Clean pages can be freed, but dirty pages must be sent back
1526 * to the laundry, unless they belong to a dead object.
1527 * Requeueing dirty pages from dead objects is pointless, as
1528 * they are being paged out and freed by the thread that
1529 * destroyed the object.
1531 if (m->dirty == 0) {
1534 * Because we dequeued the page and have already
1535 * checked for concurrent dequeue and enqueue
1536 * requests, we can safely disassociate the page
1537 * from the inactive queue.
1539 KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
1540 ("page %p has queue state", m));
1544 } else if ((object->flags & OBJ_DEAD) == 0)
1548 vm_pageout_reinsert_inactive(&ss, &rq, m);
1553 VM_OBJECT_WUNLOCK(object);
1554 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1555 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1556 vm_pagequeue_lock(pq);
1557 vm_pageout_end_scan(&ss);
1558 vm_pagequeue_unlock(pq);
1560 VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1563 * Wake up the laundry thread so that it can perform any needed
1564 * laundering. If we didn't meet our target, we're in shortfall and
1565 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1566 * swap devices are configured, the laundry thread has no work to do, so
1567 * don't bother waking it up.
1569 * The laundry thread uses the number of inactive queue scans elapsed
1570 * since the last laundering to determine whether to launder again, so
1573 if (starting_page_shortage > 0) {
1574 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1575 vm_pagequeue_lock(pq);
1576 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1577 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1578 if (page_shortage > 0) {
1579 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1580 VM_CNT_INC(v_pdshortfalls);
1581 } else if (vmd->vmd_laundry_request !=
1582 VM_LAUNDRY_SHORTFALL)
1583 vmd->vmd_laundry_request =
1584 VM_LAUNDRY_BACKGROUND;
1585 wakeup(&vmd->vmd_laundry_request);
1587 vmd->vmd_clean_pages_freed +=
1588 starting_page_shortage - page_shortage;
1589 vm_pagequeue_unlock(pq);
1593 * Wakeup the swapout daemon if we didn't free the targeted number of
1596 if (page_shortage > 0)
1600 * If the inactive queue scan fails repeatedly to meet its
1601 * target, kill the largest process.
1603 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1606 * Reclaim pages by swapping out idle processes, if configured to do so.
1608 vm_swapout_run_idle();
1611 * See the description of addl_page_shortage above.
1613 *addl_shortage = addl_page_shortage + deficit;
1615 return (page_shortage <= 0);
1618 static int vm_pageout_oom_vote;
1621 * The pagedaemon threads randlomly select one to perform the
1622 * OOM. Trying to kill processes before all pagedaemons
1623 * failed to reach free target is premature.
1626 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1627 int starting_page_shortage)
1631 if (starting_page_shortage <= 0 || starting_page_shortage !=
1633 vmd->vmd_oom_seq = 0;
1636 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1638 vmd->vmd_oom = FALSE;
1639 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1645 * Do not follow the call sequence until OOM condition is
1648 vmd->vmd_oom_seq = 0;
1653 vmd->vmd_oom = TRUE;
1654 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1655 if (old_vote != vm_ndomains - 1)
1659 * The current pagedaemon thread is the last in the quorum to
1660 * start OOM. Initiate the selection and signaling of the
1663 vm_pageout_oom(VM_OOM_MEM);
1666 * After one round of OOM terror, recall our vote. On the
1667 * next pass, current pagedaemon would vote again if the low
1668 * memory condition is still there, due to vmd_oom being
1671 vmd->vmd_oom = FALSE;
1672 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1676 * The OOM killer is the page daemon's action of last resort when
1677 * memory allocation requests have been stalled for a prolonged period
1678 * of time because it cannot reclaim memory. This function computes
1679 * the approximate number of physical pages that could be reclaimed if
1680 * the specified address space is destroyed.
1682 * Private, anonymous memory owned by the address space is the
1683 * principal resource that we expect to recover after an OOM kill.
1684 * Since the physical pages mapped by the address space's COW entries
1685 * are typically shared pages, they are unlikely to be released and so
1686 * they are not counted.
1688 * To get to the point where the page daemon runs the OOM killer, its
1689 * efforts to write-back vnode-backed pages may have stalled. This
1690 * could be caused by a memory allocation deadlock in the write path
1691 * that might be resolved by an OOM kill. Therefore, physical pages
1692 * belonging to vnode-backed objects are counted, because they might
1693 * be freed without being written out first if the address space holds
1694 * the last reference to an unlinked vnode.
1696 * Similarly, physical pages belonging to OBJT_PHYS objects are
1697 * counted because the address space might hold the last reference to
1701 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1704 vm_map_entry_t entry;
1708 map = &vmspace->vm_map;
1709 KASSERT(!map->system_map, ("system map"));
1710 sx_assert(&map->lock, SA_LOCKED);
1712 for (entry = map->header.next; entry != &map->header;
1713 entry = entry->next) {
1714 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1716 obj = entry->object.vm_object;
1719 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1720 obj->ref_count != 1)
1722 switch (obj->type) {
1727 res += obj->resident_page_count;
1735 vm_pageout_oom(int shortage)
1737 struct proc *p, *bigproc;
1738 vm_offset_t size, bigsize;
1744 * We keep the process bigproc locked once we find it to keep anyone
1745 * from messing with it; however, there is a possibility of
1746 * deadlock if process B is bigproc and one of its child processes
1747 * attempts to propagate a signal to B while we are waiting for A's
1748 * lock while walking this list. To avoid this, we don't block on
1749 * the process lock but just skip a process if it is already locked.
1753 sx_slock(&allproc_lock);
1754 FOREACH_PROC_IN_SYSTEM(p) {
1758 * If this is a system, protected or killed process, skip it.
1760 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1761 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1762 p->p_pid == 1 || P_KILLED(p) ||
1763 (p->p_pid < 48 && swap_pager_avail != 0)) {
1768 * If the process is in a non-running type state,
1769 * don't touch it. Check all the threads individually.
1772 FOREACH_THREAD_IN_PROC(p, td) {
1774 if (!TD_ON_RUNQ(td) &&
1775 !TD_IS_RUNNING(td) &&
1776 !TD_IS_SLEEPING(td) &&
1777 !TD_IS_SUSPENDED(td) &&
1778 !TD_IS_SWAPPED(td)) {
1790 * get the process size
1792 vm = vmspace_acquire_ref(p);
1799 sx_sunlock(&allproc_lock);
1800 if (!vm_map_trylock_read(&vm->vm_map)) {
1802 sx_slock(&allproc_lock);
1806 size = vmspace_swap_count(vm);
1807 if (shortage == VM_OOM_MEM)
1808 size += vm_pageout_oom_pagecount(vm);
1809 vm_map_unlock_read(&vm->vm_map);
1811 sx_slock(&allproc_lock);
1814 * If this process is bigger than the biggest one,
1817 if (size > bigsize) {
1818 if (bigproc != NULL)
1826 sx_sunlock(&allproc_lock);
1827 if (bigproc != NULL) {
1828 if (vm_panic_on_oom != 0)
1829 panic("out of swap space");
1831 killproc(bigproc, "out of swap space");
1832 sched_nice(bigproc, PRIO_MIN);
1834 PROC_UNLOCK(bigproc);
1839 vm_pageout_lowmem(void)
1841 static int lowmem_ticks = 0;
1844 last = atomic_load_int(&lowmem_ticks);
1845 while ((u_int)(ticks - last) / hz >= lowmem_period) {
1846 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
1850 * Decrease registered cache sizes.
1852 SDT_PROBE0(vm, , , vm__lowmem_scan);
1853 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1856 * We do this explicitly after the caches have been
1866 vm_pageout_worker(void *arg)
1868 struct vm_domain *vmd;
1870 int addl_shortage, domain, shortage;
1873 domain = (uintptr_t)arg;
1874 vmd = VM_DOMAIN(domain);
1879 * XXXKIB It could be useful to bind pageout daemon threads to
1880 * the cores belonging to the domain, from which vm_page_array
1884 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1885 vmd->vmd_last_active_scan = ticks;
1888 * The pageout daemon worker is never done, so loop forever.
1891 vm_domain_pageout_lock(vmd);
1894 * We need to clear wanted before we check the limits. This
1895 * prevents races with wakers who will check wanted after they
1898 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
1901 * Might the page daemon need to run again?
1903 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
1905 * Yes. If the scan failed to produce enough free
1906 * pages, sleep uninterruptibly for some time in the
1907 * hope that the laundry thread will clean some pages.
1909 vm_domain_pageout_unlock(vmd);
1911 pause("pwait", hz / VM_INACT_SCAN_RATE);
1914 * No, sleep until the next wakeup or until pages
1915 * need to have their reference stats updated.
1917 if (mtx_sleep(&vmd->vmd_pageout_wanted,
1918 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
1919 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
1920 VM_CNT_INC(v_pdwakeups);
1923 /* Prevent spurious wakeups by ensuring that wanted is set. */
1924 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1927 * Use the controller to calculate how many pages to free in
1928 * this interval, and scan the inactive queue. If the lowmem
1929 * handlers appear to have freed up some pages, subtract the
1930 * difference from the inactive queue scan target.
1932 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
1934 ofree = vmd->vmd_free_count;
1935 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
1936 shortage -= min(vmd->vmd_free_count - ofree,
1938 target_met = vm_pageout_scan_inactive(vmd, shortage,
1944 * Scan the active queue. A positive value for shortage
1945 * indicates that we must aggressively deactivate pages to avoid
1948 shortage = vm_pageout_active_target(vmd) + addl_shortage;
1949 vm_pageout_scan_active(vmd, shortage);
1954 * vm_pageout_init initialises basic pageout daemon settings.
1957 vm_pageout_init_domain(int domain)
1959 struct vm_domain *vmd;
1960 struct sysctl_oid *oid;
1962 vmd = VM_DOMAIN(domain);
1963 vmd->vmd_interrupt_free_min = 2;
1966 * v_free_reserved needs to include enough for the largest
1967 * swap pager structures plus enough for any pv_entry structs
1970 if (vmd->vmd_page_count > 1024)
1971 vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200;
1973 vmd->vmd_free_min = 4;
1974 vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1975 vmd->vmd_interrupt_free_min;
1976 vmd->vmd_free_reserved = vm_pageout_page_count +
1977 vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768);
1978 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
1979 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
1980 vmd->vmd_free_min += vmd->vmd_free_reserved;
1981 vmd->vmd_free_severe += vmd->vmd_free_reserved;
1982 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
1983 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
1984 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
1987 * Set the default wakeup threshold to be 10% below the paging
1988 * target. This keeps the steady state out of shortfall.
1990 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
1993 * Target amount of memory to move out of the laundry queue during a
1994 * background laundering. This is proportional to the amount of system
1997 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
1998 vmd->vmd_free_min) / 10;
2000 /* Initialize the pageout daemon pid controller. */
2001 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2002 vmd->vmd_free_target, PIDCTRL_BOUND,
2003 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2004 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2005 "pidctrl", CTLFLAG_RD, NULL, "");
2006 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2010 vm_pageout_init(void)
2016 * Initialize some paging parameters.
2018 if (vm_cnt.v_page_count < 2000)
2019 vm_pageout_page_count = 8;
2022 for (i = 0; i < vm_ndomains; i++) {
2023 struct vm_domain *vmd;
2025 vm_pageout_init_domain(i);
2027 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2028 vm_cnt.v_free_target += vmd->vmd_free_target;
2029 vm_cnt.v_free_min += vmd->vmd_free_min;
2030 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2031 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2032 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2033 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2034 freecount += vmd->vmd_free_count;
2038 * Set interval in seconds for active scan. We want to visit each
2039 * page at least once every ten minutes. This is to prevent worst
2040 * case paging behaviors with stale active LRU.
2042 if (vm_pageout_update_period == 0)
2043 vm_pageout_update_period = 600;
2045 if (vm_page_max_user_wired == 0)
2046 vm_page_max_user_wired = freecount / 3;
2050 * vm_pageout is the high level pageout daemon.
2057 int error, first, i;
2062 swap_pager_swap_init();
2063 for (first = -1, i = 0; i < vm_ndomains; i++) {
2064 if (VM_DOMAIN_EMPTY(i)) {
2066 printf("domain %d empty; skipping pageout\n",
2073 error = kthread_add(vm_pageout_worker,
2074 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2076 panic("starting pageout for domain %d: %d\n",
2079 error = kthread_add(vm_pageout_laundry_worker,
2080 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2082 panic("starting laundry for domain %d: %d", i, error);
2084 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2086 panic("starting uma_reclaim helper, error %d\n", error);
2088 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2089 vm_pageout_worker((void *)(uintptr_t)first);
2093 * Perform an advisory wakeup of the page daemon.
2096 pagedaemon_wakeup(int domain)
2098 struct vm_domain *vmd;
2100 vmd = VM_DOMAIN(domain);
2101 vm_domain_pageout_assert_unlocked(vmd);
2102 if (curproc == pageproc)
2105 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2106 vm_domain_pageout_lock(vmd);
2107 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2108 wakeup(&vmd->vmd_pageout_wanted);
2109 vm_domain_pageout_unlock(vmd);