2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * This code is derived from software contributed to Berkeley by
12 * The Mach Operating System project at Carnegie-Mellon University.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
50 * Permission to use, copy, modify and distribute this software and
51 * its documentation is hereby granted, provided that both the copyright
52 * notice and this permission notice appear in all copies of the
53 * software, derivative works or modified versions, and any portions
54 * thereof, and that both notices appear in supporting documentation.
56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
60 * Carnegie Mellon requests users of this software to return to
62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
63 * School of Computer Science
64 * Carnegie Mellon University
65 * Pittsburgh PA 15213-3890
67 * any improvements or extensions that they make and grant Carnegie the
68 * rights to redistribute these changes.
72 * The proverbial page-out daemon.
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
83 #include <sys/eventhandler.h>
85 #include <sys/mutex.h>
87 #include <sys/kthread.h>
89 #include <sys/mount.h>
90 #include <sys/racct.h>
91 #include <sys/resourcevar.h>
92 #include <sys/sched.h>
94 #include <sys/signalvar.h>
97 #include <sys/vnode.h>
98 #include <sys/vmmeter.h>
99 #include <sys/rwlock.h>
101 #include <sys/sysctl.h>
104 #include <vm/vm_param.h>
105 #include <vm/vm_object.h>
106 #include <vm/vm_page.h>
107 #include <vm/vm_map.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_pager.h>
110 #include <vm/vm_phys.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
116 * System initialization
119 /* the kernel process "vm_pageout"*/
120 static void vm_pageout(void);
121 static void vm_pageout_init(void);
122 static int vm_pageout_clean(vm_page_t m);
123 static int vm_pageout_cluster(vm_page_t m);
124 static void vm_pageout_scan(struct vm_domain *vmd, int pass);
125 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
126 int starting_page_shortage);
128 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
131 struct proc *pageproc;
133 static struct kproc_desc page_kp = {
138 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
141 SDT_PROVIDER_DEFINE(vm);
142 SDT_PROBE_DEFINE(vm, , , vm__lowmem_cache);
143 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
145 #if !defined(NO_SWAPPING)
146 /* the kernel process "vm_daemon"*/
147 static void vm_daemon(void);
148 static struct proc *vmproc;
150 static struct kproc_desc vm_kp = {
155 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
159 int vm_pages_needed; /* Event on which pageout daemon sleeps */
160 int vm_pageout_deficit; /* Estimated number of pages deficit */
161 int vm_pageout_wakeup_thresh;
162 static int vm_pageout_oom_seq = 12;
164 #if !defined(NO_SWAPPING)
165 static int vm_pageout_req_swapout; /* XXX */
166 static int vm_daemon_needed;
167 static struct mtx vm_daemon_mtx;
168 /* Allow for use by vm_pageout before vm_daemon is initialized. */
169 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
171 static int vm_max_launder = 32;
172 static int vm_pageout_update_period;
173 static int defer_swap_pageouts;
174 static int disable_swap_pageouts;
175 static int lowmem_period = 10;
176 static time_t lowmem_uptime;
178 #if defined(NO_SWAPPING)
179 static int vm_swap_enabled = 0;
180 static int vm_swap_idle_enabled = 0;
182 static int vm_swap_enabled = 1;
183 static int vm_swap_idle_enabled = 0;
186 static int vm_panic_on_oom = 0;
188 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
189 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
190 "panic on out of memory instead of killing the largest process");
192 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
193 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
194 "free page threshold for waking up the pageout daemon");
196 SYSCTL_INT(_vm, OID_AUTO, max_launder,
197 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
199 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
200 CTLFLAG_RW, &vm_pageout_update_period, 0,
201 "Maximum active LRU update period");
203 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
204 "Low memory callback period");
206 #if defined(NO_SWAPPING)
207 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
208 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
209 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
210 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
212 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
213 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
214 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
215 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
218 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
219 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
221 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
222 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
224 static int pageout_lock_miss;
225 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
226 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
228 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
229 CTLFLAG_RW, &vm_pageout_oom_seq, 0,
230 "back-to-back calls to oom detector to start OOM");
232 #define VM_PAGEOUT_PAGE_COUNT 16
233 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
235 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
236 SYSCTL_INT(_vm, OID_AUTO, max_wired,
237 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
239 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
240 #if !defined(NO_SWAPPING)
241 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
242 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
243 static void vm_req_vmdaemon(int req);
245 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
248 * Initialize a dummy page for marking the caller's place in the specified
249 * paging queue. In principle, this function only needs to set the flag
250 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count
251 * to one as safety precautions.
254 vm_pageout_init_marker(vm_page_t marker, u_short queue)
257 bzero(marker, sizeof(*marker));
258 marker->flags = PG_MARKER;
259 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
260 marker->queue = queue;
261 marker->hold_count = 1;
265 * vm_pageout_fallback_object_lock:
267 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
268 * known to have failed and page queue must be either PQ_ACTIVE or
269 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
270 * while locking the vm object. Use marker page to detect page queue
271 * changes and maintain notion of next page on page queue. Return
272 * TRUE if no changes were detected, FALSE otherwise. vm object is
275 * This function depends on both the lock portion of struct vm_object
276 * and normal struct vm_page being type stable.
279 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
281 struct vm_page marker;
282 struct vm_pagequeue *pq;
288 vm_pageout_init_marker(&marker, queue);
289 pq = vm_page_pagequeue(m);
292 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
293 vm_pagequeue_unlock(pq);
295 VM_OBJECT_WLOCK(object);
297 vm_pagequeue_lock(pq);
300 * The page's object might have changed, and/or the page might
301 * have moved from its original position in the queue. If the
302 * page's object has changed, then the caller should abandon
303 * processing the page because the wrong object lock was
304 * acquired. Use the marker's plinks.q, not the page's, to
305 * determine if the page has been moved. The state of the
306 * page's plinks.q can be indeterminate; whereas, the marker's
307 * plinks.q must be valid.
309 *next = TAILQ_NEXT(&marker, plinks.q);
310 unchanged = m->object == object &&
311 m == TAILQ_PREV(&marker, pglist, plinks.q);
312 KASSERT(!unchanged || m->queue == queue,
313 ("page %p queue %d %d", m, queue, m->queue));
314 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
319 * Lock the page while holding the page queue lock. Use marker page
320 * to detect page queue changes and maintain notion of next page on
321 * page queue. Return TRUE if no changes were detected, FALSE
322 * otherwise. The page is locked on return. The page queue lock might
323 * be dropped and reacquired.
325 * This function depends on normal struct vm_page being type stable.
328 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
330 struct vm_page marker;
331 struct vm_pagequeue *pq;
335 vm_page_lock_assert(m, MA_NOTOWNED);
336 if (vm_page_trylock(m))
340 vm_pageout_init_marker(&marker, queue);
341 pq = vm_page_pagequeue(m);
343 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
344 vm_pagequeue_unlock(pq);
346 vm_pagequeue_lock(pq);
348 /* Page queue might have changed. */
349 *next = TAILQ_NEXT(&marker, plinks.q);
350 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
351 KASSERT(!unchanged || m->queue == queue,
352 ("page %p queue %d %d", m, queue, m->queue));
353 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
360 * Clean the page and remove it from the laundry.
362 * We set the busy bit to cause potential page faults on this page to
363 * block. Note the careful timing, however, the busy bit isn't set till
364 * late and we cannot do anything that will mess with the page.
367 vm_pageout_cluster(vm_page_t m)
370 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
372 int ib, is, page_base;
373 vm_pindex_t pindex = m->pindex;
375 vm_page_lock_assert(m, MA_OWNED);
377 VM_OBJECT_ASSERT_WLOCKED(object);
380 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
381 * with the new swapper, but we could have serious problems paging
382 * out other object types if there is insufficient memory.
384 * Unfortunately, checking free memory here is far too late, so the
385 * check has been moved up a procedural level.
389 * Can't clean the page if it's busy or held.
391 vm_page_assert_unbusied(m);
392 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
395 mc[vm_pageout_page_count] = pb = ps = m;
397 page_base = vm_pageout_page_count;
402 * Scan object for clusterable pages.
404 * We can cluster ONLY if: ->> the page is NOT
405 * clean, wired, busy, held, or mapped into a
406 * buffer, and one of the following:
407 * 1) The page is inactive, or a seldom used
410 * 2) we force the issue.
412 * During heavy mmap/modification loads the pageout
413 * daemon can really fragment the underlying file
414 * due to flushing pages out of order and not trying
415 * align the clusters (which leave sporatic out-of-order
416 * holes). To solve this problem we do the reverse scan
417 * first and attempt to align our cluster, then do a
418 * forward scan if room remains.
421 while (ib && pageout_count < vm_pageout_page_count) {
429 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
433 vm_page_test_dirty(p);
439 if (p->queue != PQ_INACTIVE ||
440 p->hold_count != 0) { /* may be undergoing I/O */
446 mc[--page_base] = pb = p;
450 * alignment boundry, stop here and switch directions. Do
453 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
457 while (pageout_count < vm_pageout_page_count &&
458 pindex + is < object->size) {
461 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
463 vm_page_test_dirty(p);
467 if (p->queue != PQ_INACTIVE ||
468 p->hold_count != 0) { /* may be undergoing I/O */
473 mc[page_base + pageout_count] = ps = p;
479 * If we exhausted our forward scan, continue with the reverse scan
480 * when possible, even past a page boundry. This catches boundry
483 if (ib && pageout_count < vm_pageout_page_count)
487 * we allow reads during pageouts...
489 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
494 * vm_pageout_flush() - launder the given pages
496 * The given pages are laundered. Note that we setup for the start of
497 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
498 * reference count all in here rather then in the parent. If we want
499 * the parent to do more sophisticated things we may have to change
502 * Returned runlen is the count of pages between mreq and first
503 * page after mreq with status VM_PAGER_AGAIN.
504 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
505 * for any page in runlen set.
508 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
511 vm_object_t object = mc[0]->object;
512 int pageout_status[count];
516 VM_OBJECT_ASSERT_WLOCKED(object);
519 * Initiate I/O. Bump the vm_page_t->busy counter and
520 * mark the pages read-only.
522 * We do not have to fixup the clean/dirty bits here... we can
523 * allow the pager to do it after the I/O completes.
525 * NOTE! mc[i]->dirty may be partial or fragmented due to an
526 * edge case with file fragments.
528 for (i = 0; i < count; i++) {
529 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
530 ("vm_pageout_flush: partially invalid page %p index %d/%d",
532 vm_page_sbusy(mc[i]);
533 pmap_remove_write(mc[i]);
535 vm_object_pip_add(object, count);
537 vm_pager_put_pages(object, mc, count, flags, pageout_status);
539 runlen = count - mreq;
542 for (i = 0; i < count; i++) {
543 vm_page_t mt = mc[i];
545 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
546 !pmap_page_is_write_mapped(mt),
547 ("vm_pageout_flush: page %p is not write protected", mt));
548 switch (pageout_status[i]) {
555 * Page outside of range of object. Right now we
556 * essentially lose the changes by pretending it
564 * If page couldn't be paged out, then reactivate the
565 * page so it doesn't clog the inactive list. (We
566 * will try paging out it again later).
569 vm_page_activate(mt);
571 if (eio != NULL && i >= mreq && i - mreq < runlen)
575 if (i >= mreq && i - mreq < runlen)
581 * If the operation is still going, leave the page busy to
582 * block all other accesses. Also, leave the paging in
583 * progress indicator set so that we don't attempt an object
586 if (pageout_status[i] != VM_PAGER_PEND) {
587 vm_object_pip_wakeup(object);
593 return (numpagedout);
596 #if !defined(NO_SWAPPING)
598 * vm_pageout_object_deactivate_pages
600 * Deactivate enough pages to satisfy the inactive target
603 * The object and map must be locked.
606 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
609 vm_object_t backing_object, object;
611 int act_delta, remove_mode;
613 VM_OBJECT_ASSERT_LOCKED(first_object);
614 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
616 for (object = first_object;; object = backing_object) {
617 if (pmap_resident_count(pmap) <= desired)
619 VM_OBJECT_ASSERT_LOCKED(object);
620 if ((object->flags & OBJ_UNMANAGED) != 0 ||
621 object->paging_in_progress != 0)
625 if (object->shadow_count > 1)
628 * Scan the object's entire memory queue.
630 TAILQ_FOREACH(p, &object->memq, listq) {
631 if (pmap_resident_count(pmap) <= desired)
633 if (vm_page_busied(p))
635 PCPU_INC(cnt.v_pdpages);
637 if (p->wire_count != 0 || p->hold_count != 0 ||
638 !pmap_page_exists_quick(pmap, p)) {
642 act_delta = pmap_ts_referenced(p);
643 if ((p->aflags & PGA_REFERENCED) != 0) {
646 vm_page_aflag_clear(p, PGA_REFERENCED);
648 if (p->queue != PQ_ACTIVE && act_delta != 0) {
650 p->act_count += act_delta;
651 } else if (p->queue == PQ_ACTIVE) {
652 if (act_delta == 0) {
653 p->act_count -= min(p->act_count,
655 if (!remove_mode && p->act_count == 0) {
657 vm_page_deactivate(p);
662 if (p->act_count < ACT_MAX -
664 p->act_count += ACT_ADVANCE;
667 } else if (p->queue == PQ_INACTIVE)
671 if ((backing_object = object->backing_object) == NULL)
673 VM_OBJECT_RLOCK(backing_object);
674 if (object != first_object)
675 VM_OBJECT_RUNLOCK(object);
678 if (object != first_object)
679 VM_OBJECT_RUNLOCK(object);
683 * deactivate some number of pages in a map, try to do it fairly, but
684 * that is really hard to do.
687 vm_pageout_map_deactivate_pages(map, desired)
692 vm_object_t obj, bigobj;
695 if (!vm_map_trylock(map))
702 * first, search out the biggest object, and try to free pages from
705 tmpe = map->header.next;
706 while (tmpe != &map->header) {
707 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
708 obj = tmpe->object.vm_object;
709 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
710 if (obj->shadow_count <= 1 &&
712 bigobj->resident_page_count < obj->resident_page_count)) {
714 VM_OBJECT_RUNLOCK(bigobj);
717 VM_OBJECT_RUNLOCK(obj);
720 if (tmpe->wired_count > 0)
721 nothingwired = FALSE;
725 if (bigobj != NULL) {
726 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
727 VM_OBJECT_RUNLOCK(bigobj);
730 * Next, hunt around for other pages to deactivate. We actually
731 * do this search sort of wrong -- .text first is not the best idea.
733 tmpe = map->header.next;
734 while (tmpe != &map->header) {
735 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
737 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
738 obj = tmpe->object.vm_object;
740 VM_OBJECT_RLOCK(obj);
741 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
742 VM_OBJECT_RUNLOCK(obj);
749 * Remove all mappings if a process is swapped out, this will free page
752 if (desired == 0 && nothingwired) {
753 pmap_remove(vm_map_pmap(map), vm_map_min(map),
759 #endif /* !defined(NO_SWAPPING) */
762 * Attempt to acquire all of the necessary locks to launder a page and
763 * then call through the clustering layer to PUTPAGES. Wait a short
764 * time for a vnode lock.
766 * Requires the page and object lock on entry, releases both before return.
767 * Returns 0 on success and an errno otherwise.
770 vm_pageout_clean(vm_page_t m)
778 vm_page_assert_locked(m);
780 VM_OBJECT_ASSERT_WLOCKED(object);
786 * The object is already known NOT to be dead. It
787 * is possible for the vget() to block the whole
788 * pageout daemon, but the new low-memory handling
789 * code should prevent it.
791 * We can't wait forever for the vnode lock, we might
792 * deadlock due to a vn_read() getting stuck in
793 * vm_wait while holding this vnode. We skip the
794 * vnode if we can't get it in a reasonable amount
797 if (object->type == OBJT_VNODE) {
800 if (vp->v_type == VREG &&
801 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
807 ("vp %p with NULL v_mount", vp));
808 vm_object_reference_locked(object);
810 VM_OBJECT_WUNLOCK(object);
811 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
812 LK_SHARED : LK_EXCLUSIVE;
813 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
818 VM_OBJECT_WLOCK(object);
821 * While the object and page were unlocked, the page
823 * (1) moved to a different queue,
824 * (2) reallocated to a different object,
825 * (3) reallocated to a different offset, or
828 if (m->queue != PQ_INACTIVE || m->object != object ||
829 m->pindex != pindex || m->dirty == 0) {
836 * The page may have been busied or held while the object
837 * and page locks were released.
839 if (vm_page_busied(m) || m->hold_count != 0) {
847 * If a page is dirty, then it is either being washed
848 * (but not yet cleaned) or it is still in the
849 * laundry. If it is still in the laundry, then we
850 * start the cleaning operation.
852 if (vm_pageout_cluster(m) == 0)
856 VM_OBJECT_WUNLOCK(object);
859 vm_page_lock_assert(m, MA_NOTOWNED);
863 vm_object_deallocate(object);
864 vn_finished_write(mp);
871 * vm_pageout_scan does the dirty work for the pageout daemon.
873 * pass 0 - Update active LRU/deactivate pages
874 * pass 1 - Move inactive to cache or free
875 * pass 2 - Launder dirty pages
878 vm_pageout_scan(struct vm_domain *vmd, int pass)
881 struct vm_pagequeue *pq;
884 int act_delta, addl_page_shortage, deficit, error, maxlaunder, maxscan;
885 int page_shortage, scan_tick, scanned, starting_page_shortage;
887 boolean_t pageout_ok, queues_locked;
890 * If we need to reclaim memory ask kernel caches to return
891 * some. We rate limit to avoid thrashing.
893 if (vmd == &vm_dom[0] && pass > 0 &&
894 (time_uptime - lowmem_uptime) >= lowmem_period) {
896 * Decrease registered cache sizes.
898 SDT_PROBE0(vm, , , vm__lowmem_scan);
899 EVENTHANDLER_INVOKE(vm_lowmem, 0);
901 * We do this explicitly after the caches have been
905 lowmem_uptime = time_uptime;
909 * The addl_page_shortage is the number of temporarily
910 * stuck pages in the inactive queue. In other words, the
911 * number of pages from the inactive count that should be
912 * discounted in setting the target for the active queue scan.
914 addl_page_shortage = 0;
917 * Calculate the number of pages we want to either free or move
921 deficit = atomic_readandclear_int(&vm_pageout_deficit);
922 page_shortage = vm_paging_target() + deficit;
924 page_shortage = deficit = 0;
925 starting_page_shortage = page_shortage;
928 * maxlaunder limits the number of dirty pages we flush per scan.
929 * For most systems a smaller value (16 or 32) is more robust under
930 * extreme memory and disk pressure because any unnecessary writes
931 * to disk can result in extreme performance degredation. However,
932 * systems with excessive dirty pages (especially when MAP_NOSYNC is
933 * used) will die horribly with limited laundering. If the pageout
934 * daemon cannot clean enough pages in the first pass, we let it go
935 * all out in succeeding passes.
937 if ((maxlaunder = vm_max_launder) <= 1)
945 * Start scanning the inactive queue for pages we can move to the
946 * cache or free. The scan will stop when the target is reached or
947 * we have scanned the entire inactive queue. Note that m->act_count
948 * is not used to form decisions for the inactive queue, only for the
951 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
952 maxscan = pq->pq_cnt;
953 vm_pagequeue_lock(pq);
954 queues_locked = TRUE;
955 for (m = TAILQ_FIRST(&pq->pq_pl);
956 m != NULL && maxscan-- > 0 && page_shortage > 0;
958 vm_pagequeue_assert_locked(pq);
959 KASSERT(queues_locked, ("unlocked queues"));
960 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
962 PCPU_INC(cnt.v_pdpages);
963 next = TAILQ_NEXT(m, plinks.q);
968 if (m->flags & PG_MARKER)
971 KASSERT((m->flags & PG_FICTITIOUS) == 0,
972 ("Fictitious page %p cannot be in inactive queue", m));
973 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
974 ("Unmanaged page %p cannot be in inactive queue", m));
977 * The page or object lock acquisitions fail if the
978 * page was removed from the queue or moved to a
979 * different position within the queue. In either
980 * case, addl_page_shortage should not be incremented.
982 if (!vm_pageout_page_lock(m, &next))
984 else if (m->hold_count != 0) {
986 * Held pages are essentially stuck in the
987 * queue. So, they ought to be discounted
988 * from the inactive count. See the
989 * calculation of the page_shortage for the
990 * loop over the active queue below.
992 addl_page_shortage++;
996 if (!VM_OBJECT_TRYWLOCK(object)) {
997 if (!vm_pageout_fallback_object_lock(m, &next))
999 else if (m->hold_count != 0) {
1000 addl_page_shortage++;
1004 if (vm_page_busied(m)) {
1006 * Don't mess with busy pages. Leave them at
1007 * the front of the queue. Most likely, they
1008 * are being paged out and will leave the
1009 * queue shortly after the scan finishes. So,
1010 * they ought to be discounted from the
1013 addl_page_shortage++;
1015 VM_OBJECT_WUNLOCK(object);
1020 KASSERT(m->hold_count == 0, ("Held page %p", m));
1023 * We unlock the inactive page queue, invalidating the
1024 * 'next' pointer. Use our marker to remember our
1027 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1028 vm_pagequeue_unlock(pq);
1029 queues_locked = FALSE;
1032 * Invalid pages can be easily freed. They cannot be
1033 * mapped, vm_page_free() asserts this.
1039 * If the page has been referenced and the object is not dead,
1040 * reactivate or requeue the page depending on whether the
1043 if ((m->aflags & PGA_REFERENCED) != 0) {
1044 vm_page_aflag_clear(m, PGA_REFERENCED);
1048 if (object->ref_count != 0) {
1049 act_delta += pmap_ts_referenced(m);
1051 KASSERT(!pmap_page_is_mapped(m),
1052 ("vm_pageout_scan: page %p is mapped", m));
1054 if (act_delta != 0) {
1055 if (object->ref_count != 0) {
1056 vm_page_activate(m);
1059 * Increase the activation count if the page
1060 * was referenced while in the inactive queue.
1061 * This makes it less likely that the page will
1062 * be returned prematurely to the inactive
1065 m->act_count += act_delta + ACT_ADVANCE;
1067 } else if ((object->flags & OBJ_DEAD) == 0)
1072 * If the page appears to be clean at the machine-independent
1073 * layer, then remove all of its mappings from the pmap in
1074 * anticipation of placing it onto the cache queue. If,
1075 * however, any of the page's mappings allow write access,
1076 * then the page may still be modified until the last of those
1077 * mappings are removed.
1079 if (object->ref_count != 0) {
1080 vm_page_test_dirty(m);
1085 if (m->dirty == 0) {
1087 * Clean pages can be freed.
1091 PCPU_INC(cnt.v_dfree);
1093 } else if ((object->flags & OBJ_DEAD) != 0) {
1095 * Leave dirty pages from dead objects at the front of
1096 * the queue. They are being paged out and freed by
1097 * the thread that destroyed the object. They will
1098 * leave the queue shortly after the scan finishes, so
1099 * they should be discounted from the inactive count.
1101 addl_page_shortage++;
1102 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1104 * Dirty pages need to be paged out, but flushing
1105 * a page is extremely expensive versus freeing
1106 * a clean page. Rather then artificially limiting
1107 * the number of pages we can flush, we instead give
1108 * dirty pages extra priority on the inactive queue
1109 * by forcing them to be cycled through the queue
1110 * twice before being flushed, after which the
1111 * (now clean) page will cycle through once more
1112 * before being freed. This significantly extends
1113 * the thrash point for a heavily loaded machine.
1115 m->flags |= PG_WINATCFLS;
1117 vm_pagequeue_lock(pq);
1118 queues_locked = TRUE;
1119 vm_page_requeue_locked(m);
1120 } else if (maxlaunder > 0) {
1122 * We always want to try to flush some dirty pages if
1123 * we encounter them, to keep the system stable.
1124 * Normally this number is small, but under extreme
1125 * pressure where there are insufficient clean pages
1126 * on the inactive queue, we may have to go all out.
1129 if (object->type != OBJT_SWAP &&
1130 object->type != OBJT_DEFAULT)
1132 else if (disable_swap_pageouts)
1134 else if (defer_swap_pageouts)
1135 pageout_ok = vm_page_count_min();
1140 error = vm_pageout_clean(m);
1142 * Decrement page_shortage on success to account for
1143 * the (future) cleaned page. Otherwise we could wind
1144 * up laundering or cleaning too many pages.
1149 } else if (error == EDEADLK) {
1150 pageout_lock_miss++;
1152 } else if (error == EBUSY) {
1153 addl_page_shortage++;
1155 vm_page_lock_assert(m, MA_NOTOWNED);
1160 VM_OBJECT_WUNLOCK(object);
1162 if (!queues_locked) {
1163 vm_pagequeue_lock(pq);
1164 queues_locked = TRUE;
1166 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1167 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1169 vm_pagequeue_unlock(pq);
1171 #if !defined(NO_SWAPPING)
1173 * Wakeup the swapout daemon if we didn't cache or free the targeted
1176 if (vm_swap_enabled && page_shortage > 0)
1177 vm_req_vmdaemon(VM_SWAP_NORMAL);
1181 * Wakeup the sync daemon if we skipped a vnode in a writeable object
1182 * and we didn't cache or free enough pages.
1184 if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target -
1186 (void)speedup_syncer();
1189 * If the inactive queue scan fails repeatedly to meet its
1190 * target, kill the largest process.
1192 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1195 * Compute the number of pages we want to try to move from the
1196 * active queue to the inactive queue.
1198 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count +
1199 vm_paging_target() + deficit + addl_page_shortage;
1201 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1202 vm_pagequeue_lock(pq);
1203 maxscan = pq->pq_cnt;
1206 * If we're just idle polling attempt to visit every
1207 * active page within 'update_period' seconds.
1210 if (vm_pageout_update_period != 0) {
1211 min_scan = pq->pq_cnt;
1212 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1213 min_scan /= hz * vm_pageout_update_period;
1216 if (min_scan > 0 || (page_shortage > 0 && maxscan > 0))
1217 vmd->vmd_last_active_scan = scan_tick;
1220 * Scan the active queue for pages that can be deactivated. Update
1221 * the per-page activity counter and use it to identify deactivation
1224 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1225 min_scan || (page_shortage > 0 && scanned < maxscan)); m = next,
1228 KASSERT(m->queue == PQ_ACTIVE,
1229 ("vm_pageout_scan: page %p isn't active", m));
1231 next = TAILQ_NEXT(m, plinks.q);
1232 if ((m->flags & PG_MARKER) != 0)
1234 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1235 ("Fictitious page %p cannot be in active queue", m));
1236 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1237 ("Unmanaged page %p cannot be in active queue", m));
1238 if (!vm_pageout_page_lock(m, &next)) {
1244 * The count for pagedaemon pages is done after checking the
1245 * page for eligibility...
1247 PCPU_INC(cnt.v_pdpages);
1250 * Check to see "how much" the page has been used.
1252 if ((m->aflags & PGA_REFERENCED) != 0) {
1253 vm_page_aflag_clear(m, PGA_REFERENCED);
1259 * Unlocked object ref count check. Two races are possible.
1260 * 1) The ref was transitioning to zero and we saw non-zero,
1261 * the pmap bits will be checked unnecessarily.
1262 * 2) The ref was transitioning to one and we saw zero.
1263 * The page lock prevents a new reference to this page so
1264 * we need not check the reference bits.
1266 if (m->object->ref_count != 0)
1267 act_delta += pmap_ts_referenced(m);
1270 * Advance or decay the act_count based on recent usage.
1272 if (act_delta != 0) {
1273 m->act_count += ACT_ADVANCE + act_delta;
1274 if (m->act_count > ACT_MAX)
1275 m->act_count = ACT_MAX;
1277 m->act_count -= min(m->act_count, ACT_DECLINE);
1280 * Move this page to the tail of the active or inactive
1281 * queue depending on usage.
1283 if (m->act_count == 0) {
1284 /* Dequeue to avoid later lock recursion. */
1285 vm_page_dequeue_locked(m);
1286 vm_page_deactivate(m);
1289 vm_page_requeue_locked(m);
1292 vm_pagequeue_unlock(pq);
1293 #if !defined(NO_SWAPPING)
1295 * Idle process swapout -- run once per second.
1297 if (vm_swap_idle_enabled) {
1299 if (time_second != lsec) {
1300 vm_req_vmdaemon(VM_SWAP_IDLE);
1307 static int vm_pageout_oom_vote;
1310 * The pagedaemon threads randlomly select one to perform the
1311 * OOM. Trying to kill processes before all pagedaemons
1312 * failed to reach free target is premature.
1315 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1316 int starting_page_shortage)
1320 if (starting_page_shortage <= 0 || starting_page_shortage !=
1322 vmd->vmd_oom_seq = 0;
1325 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1327 vmd->vmd_oom = FALSE;
1328 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1334 * Do not follow the call sequence until OOM condition is
1337 vmd->vmd_oom_seq = 0;
1342 vmd->vmd_oom = TRUE;
1343 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1344 if (old_vote != vm_ndomains - 1)
1348 * The current pagedaemon thread is the last in the quorum to
1349 * start OOM. Initiate the selection and signaling of the
1352 vm_pageout_oom(VM_OOM_MEM);
1355 * After one round of OOM terror, recall our vote. On the
1356 * next pass, current pagedaemon would vote again if the low
1357 * memory condition is still there, due to vmd_oom being
1360 vmd->vmd_oom = FALSE;
1361 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1365 * The OOM killer is the page daemon's action of last resort when
1366 * memory allocation requests have been stalled for a prolonged period
1367 * of time because it cannot reclaim memory. This function computes
1368 * the approximate number of physical pages that could be reclaimed if
1369 * the specified address space is destroyed.
1371 * Private, anonymous memory owned by the address space is the
1372 * principal resource that we expect to recover after an OOM kill.
1373 * Since the physical pages mapped by the address space's COW entries
1374 * are typically shared pages, they are unlikely to be released and so
1375 * they are not counted.
1377 * To get to the point where the page daemon runs the OOM killer, its
1378 * efforts to write-back vnode-backed pages may have stalled. This
1379 * could be caused by a memory allocation deadlock in the write path
1380 * that might be resolved by an OOM kill. Therefore, physical pages
1381 * belonging to vnode-backed objects are counted, because they might
1382 * be freed without being written out first if the address space holds
1383 * the last reference to an unlinked vnode.
1385 * Similarly, physical pages belonging to OBJT_PHYS objects are
1386 * counted because the address space might hold the last reference to
1390 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1393 vm_map_entry_t entry;
1397 map = &vmspace->vm_map;
1398 KASSERT(!map->system_map, ("system map"));
1399 sx_assert(&map->lock, SA_LOCKED);
1401 for (entry = map->header.next; entry != &map->header;
1402 entry = entry->next) {
1403 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1405 obj = entry->object.vm_object;
1408 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1409 obj->ref_count != 1)
1411 switch (obj->type) {
1416 res += obj->resident_page_count;
1424 vm_pageout_oom(int shortage)
1426 struct proc *p, *bigproc;
1427 vm_offset_t size, bigsize;
1432 * We keep the process bigproc locked once we find it to keep anyone
1433 * from messing with it; however, there is a possibility of
1434 * deadlock if process B is bigproc and one of it's child processes
1435 * attempts to propagate a signal to B while we are waiting for A's
1436 * lock while walking this list. To avoid this, we don't block on
1437 * the process lock but just skip a process if it is already locked.
1441 sx_slock(&allproc_lock);
1442 FOREACH_PROC_IN_SYSTEM(p) {
1448 * If this is a system, protected or killed process, skip it.
1450 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1451 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1452 p->p_pid == 1 || P_KILLED(p) ||
1453 (p->p_pid < 48 && swap_pager_avail != 0)) {
1458 * If the process is in a non-running type state,
1459 * don't touch it. Check all the threads individually.
1462 FOREACH_THREAD_IN_PROC(p, td) {
1464 if (!TD_ON_RUNQ(td) &&
1465 !TD_IS_RUNNING(td) &&
1466 !TD_IS_SLEEPING(td) &&
1467 !TD_IS_SUSPENDED(td) &&
1468 !TD_IS_SWAPPED(td)) {
1480 * get the process size
1482 vm = vmspace_acquire_ref(p);
1488 if (!vm_map_trylock_read(&vm->vm_map)) {
1495 size = vmspace_swap_count(vm);
1496 if (shortage == VM_OOM_MEM)
1497 size += vm_pageout_oom_pagecount(vm);
1498 vm_map_unlock_read(&vm->vm_map);
1502 * If this process is bigger than the biggest one,
1505 if (size > bigsize) {
1506 if (bigproc != NULL)
1514 sx_sunlock(&allproc_lock);
1515 if (bigproc != NULL) {
1516 if (vm_panic_on_oom != 0)
1517 panic("out of swap space");
1519 killproc(bigproc, "out of swap space");
1520 sched_nice(bigproc, PRIO_MIN);
1522 PROC_UNLOCK(bigproc);
1523 wakeup(&vm_cnt.v_free_count);
1528 vm_pageout_worker(void *arg)
1530 struct vm_domain *domain;
1533 domidx = (uintptr_t)arg;
1534 domain = &vm_dom[domidx];
1537 * XXXKIB It could be useful to bind pageout daemon threads to
1538 * the cores belonging to the domain, from which vm_page_array
1542 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1543 domain->vmd_last_active_scan = ticks;
1544 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1545 vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
1546 TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1547 &domain->vmd_inacthead, plinks.q);
1550 * The pageout daemon worker is never done, so loop forever.
1554 * If we have enough free memory, wakeup waiters. Do
1555 * not clear vm_pages_needed until we reach our target,
1556 * otherwise we may be woken up over and over again and
1557 * waste a lot of cpu.
1559 mtx_lock(&vm_page_queue_free_mtx);
1560 if (vm_pages_needed && !vm_page_count_min()) {
1561 if (!vm_paging_needed())
1562 vm_pages_needed = 0;
1563 wakeup(&vm_cnt.v_free_count);
1565 if (vm_pages_needed) {
1567 * We're still not done. Either vm_pages_needed was
1568 * set by another thread during the previous scan
1569 * (typically, this happens during a level 0 scan) or
1570 * vm_pages_needed was already set and the scan failed
1571 * to free enough pages. If we haven't yet performed
1572 * a level >= 2 scan (unlimited dirty cleaning), then
1573 * upgrade the level and scan again now. Otherwise,
1574 * sleep a bit and try again later. While sleeping,
1575 * vm_pages_needed can be cleared.
1577 if (domain->vmd_pass > 1)
1578 msleep(&vm_pages_needed,
1579 &vm_page_queue_free_mtx, PVM, "psleep",
1583 * Good enough, sleep until required to refresh
1586 msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1589 if (vm_pages_needed) {
1590 vm_cnt.v_pdwakeups++;
1593 domain->vmd_pass = 0;
1594 mtx_unlock(&vm_page_queue_free_mtx);
1595 vm_pageout_scan(domain, domain->vmd_pass);
1600 * vm_pageout_init initialises basic pageout daemon settings.
1603 vm_pageout_init(void)
1606 * Initialize some paging parameters.
1608 vm_cnt.v_interrupt_free_min = 2;
1609 if (vm_cnt.v_page_count < 2000)
1610 vm_pageout_page_count = 8;
1613 * v_free_reserved needs to include enough for the largest
1614 * swap pager structures plus enough for any pv_entry structs
1617 if (vm_cnt.v_page_count > 1024)
1618 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1620 vm_cnt.v_free_min = 4;
1621 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1622 vm_cnt.v_interrupt_free_min;
1623 vm_cnt.v_free_reserved = vm_pageout_page_count +
1624 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1625 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1626 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1627 vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1628 vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1629 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1630 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1631 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1634 * Set the default wakeup threshold to be 10% above the minimum
1635 * page limit. This keeps the steady state out of shortfall.
1637 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1640 * Set interval in seconds for active scan. We want to visit each
1641 * page at least once every ten minutes. This is to prevent worst
1642 * case paging behaviors with stale active LRU.
1644 if (vm_pageout_update_period == 0)
1645 vm_pageout_update_period = 600;
1647 /* XXX does not really belong here */
1648 if (vm_page_max_wired == 0)
1649 vm_page_max_wired = vm_cnt.v_free_count / 3;
1653 * vm_pageout is the high level pageout daemon.
1663 swap_pager_swap_init();
1665 for (i = 1; i < vm_ndomains; i++) {
1666 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1667 curproc, NULL, 0, 0, "dom%d", i);
1669 panic("starting pageout for domain %d, error %d\n",
1674 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1677 panic("starting uma_reclaim helper, error %d\n", error);
1678 vm_pageout_worker((void *)(uintptr_t)0);
1682 * Unless the free page queue lock is held by the caller, this function
1683 * should be regarded as advisory. Specifically, the caller should
1684 * not msleep() on &vm_cnt.v_free_count following this function unless
1685 * the free page queue lock is held until the msleep() is performed.
1688 pagedaemon_wakeup(void)
1691 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1692 vm_pages_needed = 1;
1693 wakeup(&vm_pages_needed);
1697 #if !defined(NO_SWAPPING)
1699 vm_req_vmdaemon(int req)
1701 static int lastrun = 0;
1703 mtx_lock(&vm_daemon_mtx);
1704 vm_pageout_req_swapout |= req;
1705 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1706 wakeup(&vm_daemon_needed);
1709 mtx_unlock(&vm_daemon_mtx);
1715 struct rlimit rsslim;
1719 int breakout, swapout_flags, tryagain, attempts;
1721 uint64_t rsize, ravailable;
1725 mtx_lock(&vm_daemon_mtx);
1726 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
1728 racct_enable ? hz : 0
1733 swapout_flags = vm_pageout_req_swapout;
1734 vm_pageout_req_swapout = 0;
1735 mtx_unlock(&vm_daemon_mtx);
1737 swapout_procs(swapout_flags);
1740 * scan the processes for exceeding their rlimits or if
1741 * process is swapped out -- deactivate pages
1747 sx_slock(&allproc_lock);
1748 FOREACH_PROC_IN_SYSTEM(p) {
1749 vm_pindex_t limit, size;
1752 * if this is a system process or if we have already
1753 * looked at this process, skip it.
1756 if (p->p_state != PRS_NORMAL ||
1757 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1762 * if the process is in a non-running type state,
1766 FOREACH_THREAD_IN_PROC(p, td) {
1768 if (!TD_ON_RUNQ(td) &&
1769 !TD_IS_RUNNING(td) &&
1770 !TD_IS_SLEEPING(td) &&
1771 !TD_IS_SUSPENDED(td)) {
1785 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
1787 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1790 * let processes that are swapped out really be
1791 * swapped out set the limit to nothing (will force a
1794 if ((p->p_flag & P_INMEM) == 0)
1795 limit = 0; /* XXX */
1796 vm = vmspace_acquire_ref(p);
1801 size = vmspace_resident_count(vm);
1802 if (size >= limit) {
1803 vm_pageout_map_deactivate_pages(
1804 &vm->vm_map, limit);
1808 rsize = IDX_TO_OFF(size);
1810 racct_set(p, RACCT_RSS, rsize);
1811 ravailable = racct_get_available(p, RACCT_RSS);
1813 if (rsize > ravailable) {
1815 * Don't be overly aggressive; this
1816 * might be an innocent process,
1817 * and the limit could've been exceeded
1818 * by some memory hog. Don't try
1819 * to deactivate more than 1/4th
1820 * of process' resident set size.
1822 if (attempts <= 8) {
1823 if (ravailable < rsize -
1825 ravailable = rsize -
1829 vm_pageout_map_deactivate_pages(
1831 OFF_TO_IDX(ravailable));
1832 /* Update RSS usage after paging out. */
1833 size = vmspace_resident_count(vm);
1834 rsize = IDX_TO_OFF(size);
1836 racct_set(p, RACCT_RSS, rsize);
1838 if (rsize > ravailable)
1845 sx_sunlock(&allproc_lock);
1846 if (tryagain != 0 && attempts <= 10)
1850 #endif /* !defined(NO_SWAPPING) */