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("$FreeBSD$");
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
84 #include <sys/mutex.h>
86 #include <sys/kthread.h>
88 #include <sys/mount.h>
89 #include <sys/racct.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sched.h>
92 #include <sys/signalvar.h>
93 #include <sys/vnode.h>
94 #include <sys/vmmeter.h>
96 #include <sys/sysctl.h>
99 #include <vm/vm_param.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_map.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_pager.h>
105 #include <vm/swap_pager.h>
106 #include <vm/vm_extern.h>
110 * System initialization
113 /* the kernel process "vm_pageout"*/
114 static void vm_pageout(void);
115 static int vm_pageout_clean(vm_page_t);
116 static void vm_pageout_scan(int pass);
118 struct proc *pageproc;
120 static struct kproc_desc page_kp = {
125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
128 #if !defined(NO_SWAPPING)
129 /* the kernel process "vm_daemon"*/
130 static void vm_daemon(void);
131 static struct proc *vmproc;
133 static struct kproc_desc vm_kp = {
138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
142 int vm_pages_needed; /* Event on which pageout daemon sleeps */
143 int vm_pageout_deficit; /* Estimated number of pages deficit */
144 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
146 #if !defined(NO_SWAPPING)
147 static int vm_pageout_req_swapout; /* XXX */
148 static int vm_daemon_needed;
149 static struct mtx vm_daemon_mtx;
150 /* Allow for use by vm_pageout before vm_daemon is initialized. */
151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
153 static int vm_max_launder = 32;
154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
155 static int vm_pageout_full_stats_interval = 0;
156 static int vm_pageout_algorithm=0;
157 static int defer_swap_pageouts=0;
158 static int disable_swap_pageouts=0;
160 #if defined(NO_SWAPPING)
161 static int vm_swap_enabled=0;
162 static int vm_swap_idle_enabled=0;
164 static int vm_swap_enabled=1;
165 static int vm_swap_idle_enabled=0;
168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
169 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
171 SYSCTL_INT(_vm, OID_AUTO, max_launder,
172 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
175 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
178 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
181 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
183 #if defined(NO_SWAPPING)
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
190 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
192 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
196 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
199 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
201 static int pageout_lock_miss;
202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
203 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
205 #define VM_PAGEOUT_PAGE_COUNT 16
206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
208 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
209 SYSCTL_INT(_vm, OID_AUTO, max_wired,
210 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
212 #if !defined(NO_SWAPPING)
213 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
214 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
215 static void vm_req_vmdaemon(int req);
217 static void vm_pageout_page_stats(void);
220 * Initialize a dummy page for marking the caller's place in the specified
221 * paging queue. In principle, this function only needs to set the flag
222 * PG_MARKER. Nonetheless, it sets the flag VPO_BUSY and initializes the hold
223 * count to one as safety precautions.
226 vm_pageout_init_marker(vm_page_t marker, u_short queue)
229 bzero(marker, sizeof(*marker));
230 marker->flags = PG_MARKER;
231 marker->oflags = VPO_BUSY;
232 marker->queue = queue;
233 marker->hold_count = 1;
237 * vm_pageout_fallback_object_lock:
239 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
240 * known to have failed and page queue must be either PQ_ACTIVE or
241 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
242 * while locking the vm object. Use marker page to detect page queue
243 * changes and maintain notion of next page on page queue. Return
244 * TRUE if no changes were detected, FALSE otherwise. vm object is
247 * This function depends on both the lock portion of struct vm_object
248 * and normal struct vm_page being type stable.
251 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
253 struct vm_page marker;
259 vm_pageout_init_marker(&marker, queue);
262 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
264 vm_page_unlock_queues();
266 VM_OBJECT_LOCK(object);
268 vm_page_lock_queues();
270 /* Page queue might have changed. */
271 *next = TAILQ_NEXT(&marker, pageq);
272 unchanged = (m->queue == queue &&
273 m->object == object &&
274 &marker == TAILQ_NEXT(m, pageq));
275 TAILQ_REMOVE(&vm_page_queues[queue].pl,
281 * Lock the page while holding the page queue lock. Use marker page
282 * to detect page queue changes and maintain notion of next page on
283 * page queue. Return TRUE if no changes were detected, FALSE
284 * otherwise. The page is locked on return. The page queue lock might
285 * be dropped and reacquired.
287 * This function depends on normal struct vm_page being type stable.
290 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
292 struct vm_page marker;
296 vm_page_lock_assert(m, MA_NOTOWNED);
297 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
299 if (vm_page_trylock(m))
303 vm_pageout_init_marker(&marker, queue);
305 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq);
306 vm_page_unlock_queues();
308 vm_page_lock_queues();
310 /* Page queue might have changed. */
311 *next = TAILQ_NEXT(&marker, pageq);
312 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
313 TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq);
320 * Clean the page and remove it from the laundry.
322 * We set the busy bit to cause potential page faults on this page to
323 * block. Note the careful timing, however, the busy bit isn't set till
324 * late and we cannot do anything that will mess with the page.
327 vm_pageout_clean(vm_page_t m)
330 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
332 int ib, is, page_base;
333 vm_pindex_t pindex = m->pindex;
335 vm_page_lock_assert(m, MA_OWNED);
337 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
340 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
341 * with the new swapper, but we could have serious problems paging
342 * out other object types if there is insufficient memory.
344 * Unfortunately, checking free memory here is far too late, so the
345 * check has been moved up a procedural level.
349 * Can't clean the page if it's busy or held.
351 KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
352 ("vm_pageout_clean: page %p is busy", m));
353 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
356 mc[vm_pageout_page_count] = pb = ps = m;
358 page_base = vm_pageout_page_count;
363 * Scan object for clusterable pages.
365 * We can cluster ONLY if: ->> the page is NOT
366 * clean, wired, busy, held, or mapped into a
367 * buffer, and one of the following:
368 * 1) The page is inactive, or a seldom used
371 * 2) we force the issue.
373 * During heavy mmap/modification loads the pageout
374 * daemon can really fragment the underlying file
375 * due to flushing pages out of order and not trying
376 * align the clusters (which leave sporatic out-of-order
377 * holes). To solve this problem we do the reverse scan
378 * first and attempt to align our cluster, then do a
379 * forward scan if room remains.
382 while (ib && pageout_count < vm_pageout_page_count) {
390 if ((p = vm_page_prev(pb)) == NULL ||
391 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
396 vm_page_test_dirty(p);
398 p->queue != PQ_INACTIVE ||
399 p->hold_count != 0) { /* may be undergoing I/O */
405 mc[--page_base] = pb = p;
409 * alignment boundry, stop here and switch directions. Do
412 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
416 while (pageout_count < vm_pageout_page_count &&
417 pindex + is < object->size) {
420 if ((p = vm_page_next(ps)) == NULL ||
421 (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
424 vm_page_test_dirty(p);
426 p->queue != PQ_INACTIVE ||
427 p->hold_count != 0) { /* may be undergoing I/O */
432 mc[page_base + pageout_count] = ps = p;
438 * If we exhausted our forward scan, continue with the reverse scan
439 * when possible, even past a page boundry. This catches boundry
442 if (ib && pageout_count < vm_pageout_page_count)
446 * we allow reads during pageouts...
448 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL));
452 * vm_pageout_flush() - launder the given pages
454 * The given pages are laundered. Note that we setup for the start of
455 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
456 * reference count all in here rather then in the parent. If we want
457 * the parent to do more sophisticated things we may have to change
460 * Returned runlen is the count of pages between mreq and first
461 * page after mreq with status VM_PAGER_AGAIN.
464 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen)
466 vm_object_t object = mc[0]->object;
467 int pageout_status[count];
471 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
472 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
475 * Initiate I/O. Bump the vm_page_t->busy counter and
476 * mark the pages read-only.
478 * We do not have to fixup the clean/dirty bits here... we can
479 * allow the pager to do it after the I/O completes.
481 * NOTE! mc[i]->dirty may be partial or fragmented due to an
482 * edge case with file fragments.
484 for (i = 0; i < count; i++) {
485 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
486 ("vm_pageout_flush: partially invalid page %p index %d/%d",
488 vm_page_io_start(mc[i]);
489 pmap_remove_write(mc[i]);
491 vm_object_pip_add(object, count);
493 vm_pager_put_pages(object, mc, count, flags, pageout_status);
495 runlen = count - mreq;
496 for (i = 0; i < count; i++) {
497 vm_page_t mt = mc[i];
499 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
500 (mt->flags & PG_WRITEABLE) == 0,
501 ("vm_pageout_flush: page %p is not write protected", mt));
502 switch (pageout_status[i]) {
509 * Page outside of range of object. Right now we
510 * essentially lose the changes by pretending it
518 * If page couldn't be paged out, then reactivate the
519 * page so it doesn't clog the inactive list. (We
520 * will try paging out it again later).
523 vm_page_activate(mt);
527 if (i >= mreq && i - mreq < runlen)
533 * If the operation is still going, leave the page busy to
534 * block all other accesses. Also, leave the paging in
535 * progress indicator set so that we don't attempt an object
538 if (pageout_status[i] != VM_PAGER_PEND) {
539 vm_object_pip_wakeup(object);
540 vm_page_io_finish(mt);
541 if (vm_page_count_severe()) {
543 vm_page_try_to_cache(mt);
550 return (numpagedout);
553 #if !defined(NO_SWAPPING)
555 * vm_pageout_object_deactivate_pages
557 * Deactivate enough pages to satisfy the inactive target
560 * The object and map must be locked.
563 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
566 vm_object_t backing_object, object;
568 int actcount, remove_mode;
570 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
571 if (first_object->type == OBJT_DEVICE ||
572 first_object->type == OBJT_SG)
574 for (object = first_object;; object = backing_object) {
575 if (pmap_resident_count(pmap) <= desired)
577 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
578 if (object->type == OBJT_PHYS || object->paging_in_progress)
582 if (object->shadow_count > 1)
585 * Scan the object's entire memory queue.
587 TAILQ_FOREACH(p, &object->memq, listq) {
588 if (pmap_resident_count(pmap) <= desired)
590 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
592 PCPU_INC(cnt.v_pdpages);
594 if (p->wire_count != 0 || p->hold_count != 0 ||
595 !pmap_page_exists_quick(pmap, p)) {
599 actcount = pmap_ts_referenced(p);
600 if ((p->flags & PG_REFERENCED) != 0) {
603 vm_page_lock_queues();
604 vm_page_flag_clear(p, PG_REFERENCED);
605 vm_page_unlock_queues();
607 if (p->queue != PQ_ACTIVE && actcount != 0) {
609 p->act_count += actcount;
610 } else if (p->queue == PQ_ACTIVE) {
612 p->act_count -= min(p->act_count,
615 (vm_pageout_algorithm ||
616 p->act_count == 0)) {
618 vm_page_deactivate(p);
620 vm_page_lock_queues();
622 vm_page_unlock_queues();
626 if (p->act_count < ACT_MAX -
628 p->act_count += ACT_ADVANCE;
629 vm_page_lock_queues();
631 vm_page_unlock_queues();
633 } else if (p->queue == PQ_INACTIVE)
637 if ((backing_object = object->backing_object) == NULL)
639 VM_OBJECT_LOCK(backing_object);
640 if (object != first_object)
641 VM_OBJECT_UNLOCK(object);
644 if (object != first_object)
645 VM_OBJECT_UNLOCK(object);
649 * deactivate some number of pages in a map, try to do it fairly, but
650 * that is really hard to do.
653 vm_pageout_map_deactivate_pages(map, desired)
658 vm_object_t obj, bigobj;
661 if (!vm_map_trylock(map))
668 * first, search out the biggest object, and try to free pages from
671 tmpe = map->header.next;
672 while (tmpe != &map->header) {
673 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
674 obj = tmpe->object.vm_object;
675 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
676 if (obj->shadow_count <= 1 &&
678 bigobj->resident_page_count < obj->resident_page_count)) {
680 VM_OBJECT_UNLOCK(bigobj);
683 VM_OBJECT_UNLOCK(obj);
686 if (tmpe->wired_count > 0)
687 nothingwired = FALSE;
691 if (bigobj != NULL) {
692 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
693 VM_OBJECT_UNLOCK(bigobj);
696 * Next, hunt around for other pages to deactivate. We actually
697 * do this search sort of wrong -- .text first is not the best idea.
699 tmpe = map->header.next;
700 while (tmpe != &map->header) {
701 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
703 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
704 obj = tmpe->object.vm_object;
707 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
708 VM_OBJECT_UNLOCK(obj);
715 * Remove all mappings if a process is swapped out, this will free page
718 if (desired == 0 && nothingwired) {
719 tmpe = map->header.next;
720 while (tmpe != &map->header) {
721 pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end);
727 #endif /* !defined(NO_SWAPPING) */
730 * vm_pageout_scan does the dirty work for the pageout daemon.
733 vm_pageout_scan(int pass)
736 struct vm_page marker;
737 int page_shortage, maxscan, pcount;
738 int addl_page_shortage, addl_page_shortage_init;
741 int vnodes_skipped = 0;
745 * Decrease registered cache sizes.
747 EVENTHANDLER_INVOKE(vm_lowmem, 0);
749 * We do this explicitly after the caches have been drained above.
753 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
756 * Calculate the number of pages we want to either free or move
759 page_shortage = vm_paging_target() + addl_page_shortage_init;
761 vm_pageout_init_marker(&marker, PQ_INACTIVE);
764 * Start scanning the inactive queue for pages we can move to the
765 * cache or free. The scan will stop when the target is reached or
766 * we have scanned the entire inactive queue. Note that m->act_count
767 * is not used to form decisions for the inactive queue, only for the
770 * maxlaunder limits the number of dirty pages we flush per scan.
771 * For most systems a smaller value (16 or 32) is more robust under
772 * extreme memory and disk pressure because any unnecessary writes
773 * to disk can result in extreme performance degredation. However,
774 * systems with excessive dirty pages (especially when MAP_NOSYNC is
775 * used) will die horribly with limited laundering. If the pageout
776 * daemon cannot clean enough pages in the first pass, we let it go
777 * all out in succeeding passes.
779 if ((maxlaunder = vm_max_launder) <= 1)
783 vm_page_lock_queues();
785 addl_page_shortage = addl_page_shortage_init;
786 maxscan = cnt.v_inactive_count;
788 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
789 m != NULL && maxscan-- > 0 && page_shortage > 0;
794 if (m->queue != PQ_INACTIVE)
797 next = TAILQ_NEXT(m, pageq);
802 if (m->flags & PG_MARKER)
808 if (!vm_pageout_page_lock(m, &next)) {
810 addl_page_shortage++;
815 * A held page may be undergoing I/O, so skip it.
820 addl_page_shortage++;
825 * Don't mess with busy pages, keep in the front of the
826 * queue, most likely are being paged out.
829 if (!VM_OBJECT_TRYLOCK(object) &&
830 (!vm_pageout_fallback_object_lock(m, &next) ||
831 m->hold_count != 0)) {
832 VM_OBJECT_UNLOCK(object);
834 addl_page_shortage++;
837 if (m->busy || (m->oflags & VPO_BUSY)) {
839 VM_OBJECT_UNLOCK(object);
840 addl_page_shortage++;
845 * If the object is not being used, we ignore previous
848 if (object->ref_count == 0) {
849 vm_page_flag_clear(m, PG_REFERENCED);
850 KASSERT(!pmap_page_is_mapped(m),
851 ("vm_pageout_scan: page %p is mapped", m));
854 * Otherwise, if the page has been referenced while in the
855 * inactive queue, we bump the "activation count" upwards,
856 * making it less likely that the page will be added back to
857 * the inactive queue prematurely again. Here we check the
858 * page tables (or emulated bits, if any), given the upper
859 * level VM system not knowing anything about existing
862 } else if (((m->flags & PG_REFERENCED) == 0) &&
863 (actcount = pmap_ts_referenced(m))) {
866 m->act_count += actcount + ACT_ADVANCE;
867 VM_OBJECT_UNLOCK(object);
872 * If the upper level VM system knows about any page
873 * references, we activate the page. We also set the
874 * "activation count" higher than normal so that we will less
875 * likely place pages back onto the inactive queue again.
877 if ((m->flags & PG_REFERENCED) != 0) {
878 vm_page_flag_clear(m, PG_REFERENCED);
879 actcount = pmap_ts_referenced(m);
882 m->act_count += actcount + ACT_ADVANCE + 1;
883 VM_OBJECT_UNLOCK(object);
888 * If the upper level VM system does not believe that the page
889 * is fully dirty, but it is mapped for write access, then we
890 * consult the pmap to see if the page's dirty status should
893 if (m->dirty != VM_PAGE_BITS_ALL &&
894 (m->flags & PG_WRITEABLE) != 0) {
896 * Avoid a race condition: Unless write access is
897 * removed from the page, another processor could
898 * modify it before all access is removed by the call
899 * to vm_page_cache() below. If vm_page_cache() finds
900 * that the page has been modified when it removes all
901 * access, it panics because it cannot cache dirty
902 * pages. In principle, we could eliminate just write
903 * access here rather than all access. In the expected
904 * case, when there are no last instant modifications
905 * to the page, removing all access will be cheaper
908 if (pmap_is_modified(m))
910 else if (m->dirty == 0)
916 * Invalid pages can be easily freed
921 } else if (m->dirty == 0) {
923 * Clean pages can be placed onto the cache queue.
924 * This effectively frees them.
928 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
930 * Dirty pages need to be paged out, but flushing
931 * a page is extremely expensive verses freeing
932 * a clean page. Rather then artificially limiting
933 * the number of pages we can flush, we instead give
934 * dirty pages extra priority on the inactive queue
935 * by forcing them to be cycled through the queue
936 * twice before being flushed, after which the
937 * (now clean) page will cycle through once more
938 * before being freed. This significantly extends
939 * the thrash point for a heavily loaded machine.
941 vm_page_flag_set(m, PG_WINATCFLS);
943 } else if (maxlaunder > 0) {
945 * We always want to try to flush some dirty pages if
946 * we encounter them, to keep the system stable.
947 * Normally this number is small, but under extreme
948 * pressure where there are insufficient clean pages
949 * on the inactive queue, we may have to go all out.
951 int swap_pageouts_ok, vfslocked = 0;
952 struct vnode *vp = NULL;
953 struct mount *mp = NULL;
955 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
956 swap_pageouts_ok = 1;
958 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
959 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
960 vm_page_count_min());
965 * We don't bother paging objects that are "dead".
966 * Those objects are in a "rundown" state.
968 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
970 VM_OBJECT_UNLOCK(object);
976 * Following operations may unlock
977 * vm_page_queue_mtx, invalidating the 'next'
978 * pointer. To prevent an inordinate number
979 * of restarts we use our marker to remember
983 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
986 * The object is already known NOT to be dead. It
987 * is possible for the vget() to block the whole
988 * pageout daemon, but the new low-memory handling
989 * code should prevent it.
991 * The previous code skipped locked vnodes and, worse,
992 * reordered pages in the queue. This results in
993 * completely non-deterministic operation and, on a
994 * busy system, can lead to extremely non-optimal
995 * pageouts. For example, it can cause clean pages
996 * to be freed and dirty pages to be moved to the end
997 * of the queue. Since dirty pages are also moved to
998 * the end of the queue once-cleaned, this gives
999 * way too large a weighting to defering the freeing
1002 * We can't wait forever for the vnode lock, we might
1003 * deadlock due to a vn_read() getting stuck in
1004 * vm_wait while holding this vnode. We skip the
1005 * vnode if we can't get it in a reasonable amount
1008 if (object->type == OBJT_VNODE) {
1009 vm_page_unlock_queues();
1011 vp = object->handle;
1012 if (vp->v_type == VREG &&
1013 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1015 ++pageout_lock_miss;
1016 if (object->flags & OBJ_MIGHTBEDIRTY)
1018 vm_page_lock_queues();
1019 goto unlock_and_continue;
1022 ("vp %p with NULL v_mount", vp));
1023 vm_object_reference_locked(object);
1024 VM_OBJECT_UNLOCK(object);
1025 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1026 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1028 VM_OBJECT_LOCK(object);
1029 vm_page_lock_queues();
1030 ++pageout_lock_miss;
1031 if (object->flags & OBJ_MIGHTBEDIRTY)
1034 goto unlock_and_continue;
1036 VM_OBJECT_LOCK(object);
1038 vm_page_lock_queues();
1040 * The page might have been moved to another
1041 * queue during potential blocking in vget()
1042 * above. The page might have been freed and
1043 * reused for another vnode.
1045 if (m->queue != PQ_INACTIVE ||
1046 m->object != object ||
1047 TAILQ_NEXT(m, pageq) != &marker) {
1049 if (object->flags & OBJ_MIGHTBEDIRTY)
1051 goto unlock_and_continue;
1055 * The page may have been busied during the
1056 * blocking in vget(). We don't move the
1057 * page back onto the end of the queue so that
1058 * statistics are more correct if we don't.
1060 if (m->busy || (m->oflags & VPO_BUSY)) {
1062 goto unlock_and_continue;
1066 * If the page has become held it might
1067 * be undergoing I/O, so skip it
1069 if (m->hold_count) {
1072 if (object->flags & OBJ_MIGHTBEDIRTY)
1074 goto unlock_and_continue;
1079 * If a page is dirty, then it is either being washed
1080 * (but not yet cleaned) or it is still in the
1081 * laundry. If it is still in the laundry, then we
1082 * start the cleaning operation.
1084 * decrement page_shortage on success to account for
1085 * the (future) cleaned page. Otherwise we could wind
1086 * up laundering or cleaning too many pages.
1088 vm_page_unlock_queues();
1089 if (vm_pageout_clean(m) != 0) {
1093 vm_page_lock_queues();
1094 unlock_and_continue:
1095 vm_page_lock_assert(m, MA_NOTOWNED);
1096 VM_OBJECT_UNLOCK(object);
1098 vm_page_unlock_queues();
1101 VFS_UNLOCK_GIANT(vfslocked);
1102 vm_object_deallocate(object);
1103 vn_finished_write(mp);
1104 vm_page_lock_queues();
1106 next = TAILQ_NEXT(&marker, pageq);
1107 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1109 vm_page_lock_assert(m, MA_NOTOWNED);
1113 VM_OBJECT_UNLOCK(object);
1117 * Compute the number of pages we want to try to move from the
1118 * active queue to the inactive queue.
1120 page_shortage = vm_paging_target() +
1121 cnt.v_inactive_target - cnt.v_inactive_count;
1122 page_shortage += addl_page_shortage;
1125 * Scan the active queue for things we can deactivate. We nominally
1126 * track the per-page activity counter and use it to locate
1127 * deactivation candidates.
1129 pcount = cnt.v_active_count;
1130 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1131 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1133 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1135 KASSERT(m->queue == PQ_ACTIVE,
1136 ("vm_pageout_scan: page %p isn't active", m));
1138 next = TAILQ_NEXT(m, pageq);
1139 if ((m->flags & PG_MARKER) != 0) {
1143 if (!vm_pageout_page_lock(m, &next)) {
1149 if (!VM_OBJECT_TRYLOCK(object) &&
1150 !vm_pageout_fallback_object_lock(m, &next)) {
1151 VM_OBJECT_UNLOCK(object);
1158 * Don't deactivate pages that are busy.
1160 if ((m->busy != 0) ||
1161 (m->oflags & VPO_BUSY) ||
1162 (m->hold_count != 0)) {
1164 VM_OBJECT_UNLOCK(object);
1171 * The count for pagedaemon pages is done after checking the
1172 * page for eligibility...
1177 * Check to see "how much" the page has been used.
1180 if (object->ref_count != 0) {
1181 if (m->flags & PG_REFERENCED) {
1184 actcount += pmap_ts_referenced(m);
1186 m->act_count += ACT_ADVANCE + actcount;
1187 if (m->act_count > ACT_MAX)
1188 m->act_count = ACT_MAX;
1193 * Since we have "tested" this bit, we need to clear it now.
1195 vm_page_flag_clear(m, PG_REFERENCED);
1198 * Only if an object is currently being used, do we use the
1199 * page activation count stats.
1201 if (actcount && (object->ref_count != 0)) {
1204 m->act_count -= min(m->act_count, ACT_DECLINE);
1205 if (vm_pageout_algorithm ||
1206 object->ref_count == 0 ||
1207 m->act_count == 0) {
1209 if (object->ref_count == 0) {
1210 KASSERT(!pmap_page_is_mapped(m),
1211 ("vm_pageout_scan: page %p is mapped", m));
1215 vm_page_deactivate(m);
1217 vm_page_deactivate(m);
1224 VM_OBJECT_UNLOCK(object);
1227 vm_page_unlock_queues();
1228 #if !defined(NO_SWAPPING)
1230 * Idle process swapout -- run once per second.
1232 if (vm_swap_idle_enabled) {
1234 if (time_second != lsec) {
1235 vm_req_vmdaemon(VM_SWAP_IDLE);
1242 * If we didn't get enough free pages, and we have skipped a vnode
1243 * in a writeable object, wakeup the sync daemon. And kick swapout
1244 * if we did not get enough free pages.
1246 if (vm_paging_target() > 0) {
1247 if (vnodes_skipped && vm_page_count_min())
1248 (void) speedup_syncer();
1249 #if !defined(NO_SWAPPING)
1250 if (vm_swap_enabled && vm_page_count_target())
1251 vm_req_vmdaemon(VM_SWAP_NORMAL);
1256 * If we are critically low on one of RAM or swap and low on
1257 * the other, kill the largest process. However, we avoid
1258 * doing this on the first pass in order to give ourselves a
1259 * chance to flush out dirty vnode-backed pages and to allow
1260 * active pages to be moved to the inactive queue and reclaimed.
1263 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1264 (swap_pager_full && vm_paging_target() > 0)))
1265 vm_pageout_oom(VM_OOM_MEM);
1270 vm_pageout_oom(int shortage)
1272 struct proc *p, *bigproc;
1273 vm_offset_t size, bigsize;
1278 * We keep the process bigproc locked once we find it to keep anyone
1279 * from messing with it; however, there is a possibility of
1280 * deadlock if process B is bigproc and one of it's child processes
1281 * attempts to propagate a signal to B while we are waiting for A's
1282 * lock while walking this list. To avoid this, we don't block on
1283 * the process lock but just skip a process if it is already locked.
1287 sx_slock(&allproc_lock);
1288 FOREACH_PROC_IN_SYSTEM(p) {
1291 if (PROC_TRYLOCK(p) == 0)
1294 * If this is a system, protected or killed process, skip it.
1296 if (p->p_state != PRS_NORMAL ||
1297 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1298 (p->p_pid == 1) || P_KILLED(p) ||
1299 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1304 * If the process is in a non-running type state,
1305 * don't touch it. Check all the threads individually.
1308 FOREACH_THREAD_IN_PROC(p, td) {
1310 if (!TD_ON_RUNQ(td) &&
1311 !TD_IS_RUNNING(td) &&
1312 !TD_IS_SLEEPING(td) &&
1313 !TD_IS_SUSPENDED(td)) {
1325 * get the process size
1327 vm = vmspace_acquire_ref(p);
1332 if (!vm_map_trylock_read(&vm->vm_map)) {
1337 size = vmspace_swap_count(vm);
1338 vm_map_unlock_read(&vm->vm_map);
1339 if (shortage == VM_OOM_MEM)
1340 size += vmspace_resident_count(vm);
1343 * if the this process is bigger than the biggest one
1346 if (size > bigsize) {
1347 if (bigproc != NULL)
1348 PROC_UNLOCK(bigproc);
1354 sx_sunlock(&allproc_lock);
1355 if (bigproc != NULL) {
1356 killproc(bigproc, "out of swap space");
1357 sched_nice(bigproc, PRIO_MIN);
1358 PROC_UNLOCK(bigproc);
1359 wakeup(&cnt.v_free_count);
1364 * This routine tries to maintain the pseudo LRU active queue,
1365 * so that during long periods of time where there is no paging,
1366 * that some statistic accumulation still occurs. This code
1367 * helps the situation where paging just starts to occur.
1370 vm_pageout_page_stats()
1374 int pcount,tpcount; /* Number of pages to check */
1375 static int fullintervalcount = 0;
1379 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1380 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1382 if (page_shortage <= 0)
1385 vm_page_lock_queues();
1386 pcount = cnt.v_active_count;
1387 fullintervalcount += vm_pageout_stats_interval;
1388 if (fullintervalcount < vm_pageout_full_stats_interval) {
1389 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1391 if (pcount > tpcount)
1394 fullintervalcount = 0;
1397 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1398 while ((m != NULL) && (pcount-- > 0)) {
1401 KASSERT(m->queue == PQ_ACTIVE,
1402 ("vm_pageout_page_stats: page %p isn't active", m));
1404 next = TAILQ_NEXT(m, pageq);
1405 if ((m->flags & PG_MARKER) != 0) {
1409 vm_page_lock_assert(m, MA_NOTOWNED);
1410 if (!vm_pageout_page_lock(m, &next)) {
1416 if (!VM_OBJECT_TRYLOCK(object) &&
1417 !vm_pageout_fallback_object_lock(m, &next)) {
1418 VM_OBJECT_UNLOCK(object);
1425 * Don't deactivate pages that are busy.
1427 if ((m->busy != 0) ||
1428 (m->oflags & VPO_BUSY) ||
1429 (m->hold_count != 0)) {
1431 VM_OBJECT_UNLOCK(object);
1438 if (m->flags & PG_REFERENCED) {
1439 vm_page_flag_clear(m, PG_REFERENCED);
1443 actcount += pmap_ts_referenced(m);
1445 m->act_count += ACT_ADVANCE + actcount;
1446 if (m->act_count > ACT_MAX)
1447 m->act_count = ACT_MAX;
1450 if (m->act_count == 0) {
1452 * We turn off page access, so that we have
1453 * more accurate RSS stats. We don't do this
1454 * in the normal page deactivation when the
1455 * system is loaded VM wise, because the
1456 * cost of the large number of page protect
1457 * operations would be higher than the value
1458 * of doing the operation.
1461 vm_page_deactivate(m);
1463 m->act_count -= min(m->act_count, ACT_DECLINE);
1468 VM_OBJECT_UNLOCK(object);
1471 vm_page_unlock_queues();
1475 * vm_pageout is the high level pageout daemon.
1483 * Initialize some paging parameters.
1485 cnt.v_interrupt_free_min = 2;
1486 if (cnt.v_page_count < 2000)
1487 vm_pageout_page_count = 8;
1490 * v_free_reserved needs to include enough for the largest
1491 * swap pager structures plus enough for any pv_entry structs
1494 if (cnt.v_page_count > 1024)
1495 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1498 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1499 cnt.v_interrupt_free_min;
1500 cnt.v_free_reserved = vm_pageout_page_count +
1501 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1502 cnt.v_free_severe = cnt.v_free_min / 2;
1503 cnt.v_free_min += cnt.v_free_reserved;
1504 cnt.v_free_severe += cnt.v_free_reserved;
1507 * v_free_target and v_cache_min control pageout hysteresis. Note
1508 * that these are more a measure of the VM cache queue hysteresis
1509 * then the VM free queue. Specifically, v_free_target is the
1510 * high water mark (free+cache pages).
1512 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1513 * low water mark, while v_free_min is the stop. v_cache_min must
1514 * be big enough to handle memory needs while the pageout daemon
1515 * is signalled and run to free more pages.
1517 if (cnt.v_free_count > 6144)
1518 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1520 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1522 if (cnt.v_free_count > 2048) {
1523 cnt.v_cache_min = cnt.v_free_target;
1524 cnt.v_cache_max = 2 * cnt.v_cache_min;
1525 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1527 cnt.v_cache_min = 0;
1528 cnt.v_cache_max = 0;
1529 cnt.v_inactive_target = cnt.v_free_count / 4;
1531 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1532 cnt.v_inactive_target = cnt.v_free_count / 3;
1534 /* XXX does not really belong here */
1535 if (vm_page_max_wired == 0)
1536 vm_page_max_wired = cnt.v_free_count / 3;
1538 if (vm_pageout_stats_max == 0)
1539 vm_pageout_stats_max = cnt.v_free_target;
1542 * Set interval in seconds for stats scan.
1544 if (vm_pageout_stats_interval == 0)
1545 vm_pageout_stats_interval = 5;
1546 if (vm_pageout_full_stats_interval == 0)
1547 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1549 swap_pager_swap_init();
1552 * The pageout daemon is never done, so loop forever.
1556 * If we have enough free memory, wakeup waiters. Do
1557 * not clear vm_pages_needed until we reach our target,
1558 * otherwise we may be woken up over and over again and
1559 * waste a lot of cpu.
1561 mtx_lock(&vm_page_queue_free_mtx);
1562 if (vm_pages_needed && !vm_page_count_min()) {
1563 if (!vm_paging_needed())
1564 vm_pages_needed = 0;
1565 wakeup(&cnt.v_free_count);
1567 if (vm_pages_needed) {
1569 * Still not done, take a second pass without waiting
1570 * (unlimited dirty cleaning), otherwise sleep a bit
1575 msleep(&vm_pages_needed,
1576 &vm_page_queue_free_mtx, PVM, "psleep",
1580 * Good enough, sleep & handle stats. Prime the pass
1587 error = msleep(&vm_pages_needed,
1588 &vm_page_queue_free_mtx, PVM, "psleep",
1589 vm_pageout_stats_interval * hz);
1590 if (error && !vm_pages_needed) {
1591 mtx_unlock(&vm_page_queue_free_mtx);
1593 vm_pageout_page_stats();
1597 if (vm_pages_needed)
1599 mtx_unlock(&vm_page_queue_free_mtx);
1600 vm_pageout_scan(pass);
1605 * Unless the free page queue lock is held by the caller, this function
1606 * should be regarded as advisory. Specifically, the caller should
1607 * not msleep() on &cnt.v_free_count following this function unless
1608 * the free page queue lock is held until the msleep() is performed.
1614 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1615 vm_pages_needed = 1;
1616 wakeup(&vm_pages_needed);
1620 #if !defined(NO_SWAPPING)
1622 vm_req_vmdaemon(int req)
1624 static int lastrun = 0;
1626 mtx_lock(&vm_daemon_mtx);
1627 vm_pageout_req_swapout |= req;
1628 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1629 wakeup(&vm_daemon_needed);
1632 mtx_unlock(&vm_daemon_mtx);
1638 struct rlimit rsslim;
1642 int breakout, swapout_flags, tryagain, attempts;
1644 uint64_t rsize, ravailable;
1648 mtx_lock(&vm_daemon_mtx);
1650 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz);
1652 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1654 swapout_flags = vm_pageout_req_swapout;
1655 vm_pageout_req_swapout = 0;
1656 mtx_unlock(&vm_daemon_mtx);
1658 swapout_procs(swapout_flags);
1661 * scan the processes for exceeding their rlimits or if
1662 * process is swapped out -- deactivate pages
1668 sx_slock(&allproc_lock);
1669 FOREACH_PROC_IN_SYSTEM(p) {
1670 vm_pindex_t limit, size;
1673 * if this is a system process or if we have already
1674 * looked at this process, skip it.
1677 if (p->p_state != PRS_NORMAL ||
1678 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1683 * if the process is in a non-running type state,
1687 FOREACH_THREAD_IN_PROC(p, td) {
1689 if (!TD_ON_RUNQ(td) &&
1690 !TD_IS_RUNNING(td) &&
1691 !TD_IS_SLEEPING(td) &&
1692 !TD_IS_SUSPENDED(td)) {
1706 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1708 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1711 * let processes that are swapped out really be
1712 * swapped out set the limit to nothing (will force a
1715 if ((p->p_flag & P_INMEM) == 0)
1716 limit = 0; /* XXX */
1717 vm = vmspace_acquire_ref(p);
1722 size = vmspace_resident_count(vm);
1723 if (limit >= 0 && size >= limit) {
1724 vm_pageout_map_deactivate_pages(
1725 &vm->vm_map, limit);
1728 rsize = IDX_TO_OFF(size);
1730 racct_set(p, RACCT_RSS, rsize);
1731 ravailable = racct_get_available(p, RACCT_RSS);
1733 if (rsize > ravailable) {
1735 * Don't be overly aggressive; this might be
1736 * an innocent process, and the limit could've
1737 * been exceeded by some memory hog. Don't
1738 * try to deactivate more than 1/4th of process'
1739 * resident set size.
1741 if (attempts <= 8) {
1742 if (ravailable < rsize - (rsize / 4))
1743 ravailable = rsize - (rsize / 4);
1745 vm_pageout_map_deactivate_pages(
1746 &vm->vm_map, OFF_TO_IDX(ravailable));
1747 /* Update RSS usage after paging out. */
1748 size = vmspace_resident_count(vm);
1749 rsize = IDX_TO_OFF(size);
1751 racct_set(p, RACCT_RSS, rsize);
1753 if (rsize > ravailable)
1759 sx_sunlock(&allproc_lock);
1760 if (tryagain != 0 && attempts <= 10)
1764 #endif /* !defined(NO_SWAPPING) */