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/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
95 #include <sys/sysctl.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
109 * System initialization
112 /* the kernel process "vm_pageout"*/
113 static void vm_pageout(void);
114 static int vm_pageout_clean(vm_page_t);
115 static void vm_pageout_scan(int pass);
117 struct proc *pageproc;
119 static struct kproc_desc page_kp = {
124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
127 #if !defined(NO_SWAPPING)
128 /* the kernel process "vm_daemon"*/
129 static void vm_daemon(void);
130 static struct proc *vmproc;
132 static struct kproc_desc vm_kp = {
137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
141 int vm_pages_needed; /* Event on which pageout daemon sleeps */
142 int vm_pageout_deficit; /* Estimated number of pages deficit */
143 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
145 #if !defined(NO_SWAPPING)
146 static int vm_pageout_req_swapout; /* XXX */
147 static int vm_daemon_needed;
148 static struct mtx vm_daemon_mtx;
149 /* Allow for use by vm_pageout before vm_daemon is initialized. */
150 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
152 static int vm_max_launder = 32;
153 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
154 static int vm_pageout_full_stats_interval = 0;
155 static int vm_pageout_algorithm=0;
156 static int defer_swap_pageouts=0;
157 static int disable_swap_pageouts=0;
159 #if defined(NO_SWAPPING)
160 static int vm_swap_enabled=0;
161 static int vm_swap_idle_enabled=0;
163 static int vm_swap_enabled=1;
164 static int vm_swap_idle_enabled=0;
167 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
168 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
170 SYSCTL_INT(_vm, OID_AUTO, max_launder,
171 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
174 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
176 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
177 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
180 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
182 #if defined(NO_SWAPPING)
183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
184 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
186 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
194 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
195 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
197 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
198 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
200 static int pageout_lock_miss;
201 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
202 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
204 #define VM_PAGEOUT_PAGE_COUNT 16
205 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
207 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
208 SYSCTL_INT(_vm, OID_AUTO, max_wired,
209 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
211 #if !defined(NO_SWAPPING)
212 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
213 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
214 static void vm_req_vmdaemon(int req);
216 static void vm_pageout_page_stats(void);
219 * vm_pageout_fallback_object_lock:
221 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
222 * known to have failed and page queue must be either PQ_ACTIVE or
223 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
224 * while locking the vm object. Use marker page to detect page queue
225 * changes and maintain notion of next page on page queue. Return
226 * TRUE if no changes were detected, FALSE otherwise. vm object is
229 * This function depends on both the lock portion of struct vm_object
230 * and normal struct vm_page being type stable.
233 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
235 struct vm_page marker;
241 * Initialize our marker
243 bzero(&marker, sizeof(marker));
244 marker.flags = PG_FICTITIOUS | PG_MARKER;
245 marker.oflags = VPO_BUSY;
246 marker.queue = m->queue;
247 marker.wire_count = 1;
252 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
254 vm_page_unlock_queues();
256 VM_OBJECT_LOCK(object);
258 vm_page_lock_queues();
260 /* Page queue might have changed. */
261 *next = TAILQ_NEXT(&marker, pageq);
262 unchanged = (m->queue == queue &&
263 m->object == object &&
264 &marker == TAILQ_NEXT(m, pageq));
265 TAILQ_REMOVE(&vm_page_queues[queue].pl,
273 * Clean the page and remove it from the laundry.
275 * We set the busy bit to cause potential page faults on this page to
276 * block. Note the careful timing, however, the busy bit isn't set till
277 * late and we cannot do anything that will mess with the page.
280 vm_pageout_clean(vm_page_t m)
283 vm_page_t mc[2*vm_pageout_page_count];
285 int ib, is, page_base;
286 vm_pindex_t pindex = m->pindex;
288 vm_page_lock_assert(m, MA_NOTOWNED);
290 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
293 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
294 * with the new swapper, but we could have serious problems paging
295 * out other object types if there is insufficient memory.
297 * Unfortunately, checking free memory here is far too late, so the
298 * check has been moved up a procedural level.
302 * Can't clean the page if it's busy or held.
304 if ((m->hold_count != 0) ||
305 ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
310 mc[vm_pageout_page_count] = m;
312 page_base = vm_pageout_page_count;
317 * Scan object for clusterable pages.
319 * We can cluster ONLY if: ->> the page is NOT
320 * clean, wired, busy, held, or mapped into a
321 * buffer, and one of the following:
322 * 1) The page is inactive, or a seldom used
325 * 2) we force the issue.
327 * During heavy mmap/modification loads the pageout
328 * daemon can really fragment the underlying file
329 * due to flushing pages out of order and not trying
330 * align the clusters (which leave sporatic out-of-order
331 * holes). To solve this problem we do the reverse scan
332 * first and attempt to align our cluster, then do a
333 * forward scan if room remains.
337 while (ib && pageout_count < vm_pageout_page_count) {
345 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
349 if ((p->oflags & VPO_BUSY) || p->busy) {
354 vm_page_lock_queues();
355 vm_page_test_dirty(p);
357 p->queue != PQ_INACTIVE ||
358 p->hold_count != 0) { /* may be undergoing I/O */
360 vm_page_unlock_queues();
364 vm_page_unlock_queues();
370 * alignment boundry, stop here and switch directions. Do
373 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
377 while (pageout_count < vm_pageout_page_count &&
378 pindex + is < object->size) {
381 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
383 if ((p->oflags & VPO_BUSY) || p->busy) {
387 vm_page_lock_queues();
388 vm_page_test_dirty(p);
390 p->queue != PQ_INACTIVE ||
391 p->hold_count != 0) { /* may be undergoing I/O */
392 vm_page_unlock_queues();
396 vm_page_unlock_queues();
398 mc[page_base + pageout_count] = p;
404 * If we exhausted our forward scan, continue with the reverse scan
405 * when possible, even past a page boundry. This catches boundry
408 if (ib && pageout_count < vm_pageout_page_count)
413 * we allow reads during pageouts...
415 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
419 * vm_pageout_flush() - launder the given pages
421 * The given pages are laundered. Note that we setup for the start of
422 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
423 * reference count all in here rather then in the parent. If we want
424 * the parent to do more sophisticated things we may have to change
428 vm_pageout_flush(vm_page_t *mc, int count, int flags)
430 vm_object_t object = mc[0]->object;
431 int pageout_status[count];
435 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
436 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
439 * Initiate I/O. Bump the vm_page_t->busy counter and
440 * mark the pages read-only.
442 * We do not have to fixup the clean/dirty bits here... we can
443 * allow the pager to do it after the I/O completes.
445 * NOTE! mc[i]->dirty may be partial or fragmented due to an
446 * edge case with file fragments.
448 for (i = 0; i < count; i++) {
449 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
450 ("vm_pageout_flush: partially invalid page %p index %d/%d",
452 vm_page_io_start(mc[i]);
454 vm_page_lock_queues();
455 pmap_remove_write(mc[i]);
456 vm_page_unlock(mc[i]);
457 vm_page_unlock_queues();
459 vm_object_pip_add(object, count);
461 vm_pager_put_pages(object, mc, count, flags, pageout_status);
463 for (i = 0; i < count; i++) {
464 vm_page_t mt = mc[i];
467 vm_page_lock_queues();
468 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
469 (mt->flags & PG_WRITEABLE) == 0,
470 ("vm_pageout_flush: page %p is not write protected", mt));
471 switch (pageout_status[i]) {
478 * Page outside of range of object. Right now we
479 * essentially lose the changes by pretending it
487 * If page couldn't be paged out, then reactivate the
488 * page so it doesn't clog the inactive list. (We
489 * will try paging out it again later).
491 vm_page_activate(mt);
498 * If the operation is still going, leave the page busy to
499 * block all other accesses. Also, leave the paging in
500 * progress indicator set so that we don't attempt an object
503 if (pageout_status[i] != VM_PAGER_PEND) {
504 vm_object_pip_wakeup(object);
505 vm_page_io_finish(mt);
506 if (vm_page_count_severe())
507 vm_page_try_to_cache(mt);
509 vm_page_unlock_queues();
515 #if !defined(NO_SWAPPING)
517 * vm_pageout_object_deactivate_pages
519 * deactivate enough pages to satisfy the inactive target
520 * requirements or if vm_page_proc_limit is set, then
521 * deactivate all of the pages in the object and its
524 * The object and map must be locked.
527 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
529 vm_object_t first_object;
532 vm_object_t backing_object, object;
534 int actcount, remove_mode;
536 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
537 if (first_object->type == OBJT_DEVICE ||
538 first_object->type == OBJT_SG)
540 for (object = first_object;; object = backing_object) {
541 if (pmap_resident_count(pmap) <= desired)
543 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
544 if (object->type == OBJT_PHYS || object->paging_in_progress)
548 if (object->shadow_count > 1)
551 * scan the objects entire memory queue
553 p = TAILQ_FIRST(&object->memq);
555 if (pmap_resident_count(pmap) <= desired)
557 next = TAILQ_NEXT(p, listq);
558 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
563 vm_page_lock_queues();
565 if (p->wire_count != 0 ||
566 p->hold_count != 0 ||
567 !pmap_page_exists_quick(pmap, p)) {
568 vm_page_unlock_queues();
573 actcount = pmap_ts_referenced(p);
575 vm_page_flag_set(p, PG_REFERENCED);
576 } else if (p->flags & PG_REFERENCED) {
579 if ((p->queue != PQ_ACTIVE) &&
580 (p->flags & PG_REFERENCED)) {
582 p->act_count += actcount;
583 vm_page_flag_clear(p, PG_REFERENCED);
584 } else if (p->queue == PQ_ACTIVE) {
585 if ((p->flags & PG_REFERENCED) == 0) {
586 p->act_count -= min(p->act_count, ACT_DECLINE);
587 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
589 vm_page_deactivate(p);
595 vm_page_flag_clear(p, PG_REFERENCED);
596 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
597 p->act_count += ACT_ADVANCE;
600 } else if (p->queue == PQ_INACTIVE) {
603 vm_page_unlock_queues();
607 if ((backing_object = object->backing_object) == NULL)
609 VM_OBJECT_LOCK(backing_object);
610 if (object != first_object)
611 VM_OBJECT_UNLOCK(object);
614 if (object != first_object)
615 VM_OBJECT_UNLOCK(object);
619 * deactivate some number of pages in a map, try to do it fairly, but
620 * that is really hard to do.
623 vm_pageout_map_deactivate_pages(map, desired)
628 vm_object_t obj, bigobj;
631 if (!vm_map_trylock(map))
638 * first, search out the biggest object, and try to free pages from
641 tmpe = map->header.next;
642 while (tmpe != &map->header) {
643 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
644 obj = tmpe->object.vm_object;
645 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
646 if (obj->shadow_count <= 1 &&
648 bigobj->resident_page_count < obj->resident_page_count)) {
650 VM_OBJECT_UNLOCK(bigobj);
653 VM_OBJECT_UNLOCK(obj);
656 if (tmpe->wired_count > 0)
657 nothingwired = FALSE;
661 if (bigobj != NULL) {
662 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
663 VM_OBJECT_UNLOCK(bigobj);
666 * Next, hunt around for other pages to deactivate. We actually
667 * do this search sort of wrong -- .text first is not the best idea.
669 tmpe = map->header.next;
670 while (tmpe != &map->header) {
671 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
673 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
674 obj = tmpe->object.vm_object;
677 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
678 VM_OBJECT_UNLOCK(obj);
685 * Remove all mappings if a process is swapped out, this will free page
688 if (desired == 0 && nothingwired) {
689 pmap_remove(vm_map_pmap(map), vm_map_min(map),
694 #endif /* !defined(NO_SWAPPING) */
697 * vm_pageout_scan does the dirty work for the pageout daemon.
700 vm_pageout_scan(int pass)
703 struct vm_page marker;
704 int page_shortage, maxscan, pcount;
705 int addl_page_shortage, addl_page_shortage_init;
708 int vnodes_skipped = 0;
712 * Decrease registered cache sizes.
714 EVENTHANDLER_INVOKE(vm_lowmem, 0);
716 * We do this explicitly after the caches have been drained above.
720 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
723 * Calculate the number of pages we want to either free or move
726 page_shortage = vm_paging_target() + addl_page_shortage_init;
729 * Initialize our marker
731 bzero(&marker, sizeof(marker));
732 marker.flags = PG_FICTITIOUS | PG_MARKER;
733 marker.oflags = VPO_BUSY;
734 marker.queue = PQ_INACTIVE;
735 marker.wire_count = 1;
738 * Start scanning the inactive queue for pages we can move to the
739 * cache or free. The scan will stop when the target is reached or
740 * we have scanned the entire inactive queue. Note that m->act_count
741 * is not used to form decisions for the inactive queue, only for the
744 * maxlaunder limits the number of dirty pages we flush per scan.
745 * For most systems a smaller value (16 or 32) is more robust under
746 * extreme memory and disk pressure because any unnecessary writes
747 * to disk can result in extreme performance degredation. However,
748 * systems with excessive dirty pages (especially when MAP_NOSYNC is
749 * used) will die horribly with limited laundering. If the pageout
750 * daemon cannot clean enough pages in the first pass, we let it go
751 * all out in succeeding passes.
753 if ((maxlaunder = vm_max_launder) <= 1)
757 vm_page_lock_queues();
759 addl_page_shortage = addl_page_shortage_init;
760 maxscan = cnt.v_inactive_count;
762 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
763 m != NULL && maxscan-- > 0 && page_shortage > 0;
768 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
772 next = TAILQ_NEXT(m, pageq);
777 if (m->flags & PG_MARKER)
780 if (!vm_page_trylock(m)) {
781 addl_page_shortage++;
786 * A held page may be undergoing I/O, so skip it.
788 if (m->hold_count || (object = m->object) == NULL) {
791 addl_page_shortage++;
796 * Don't mess with busy pages, keep in the front of the
797 * queue, most likely are being paged out.
799 if (!VM_OBJECT_TRYLOCK(object) &&
800 (!vm_pageout_fallback_object_lock(m, &next) ||
801 m->hold_count != 0)) {
802 VM_OBJECT_UNLOCK(object);
804 addl_page_shortage++;
807 if (m->busy || (m->oflags & VPO_BUSY)) {
809 VM_OBJECT_UNLOCK(object);
810 addl_page_shortage++;
815 * If the object is not being used, we ignore previous
818 if (object->ref_count == 0) {
819 vm_page_flag_clear(m, PG_REFERENCED);
820 KASSERT(!pmap_page_is_mapped(m),
821 ("vm_pageout_scan: page %p is mapped", m));
824 * Otherwise, if the page has been referenced while in the
825 * inactive queue, we bump the "activation count" upwards,
826 * making it less likely that the page will be added back to
827 * the inactive queue prematurely again. Here we check the
828 * page tables (or emulated bits, if any), given the upper
829 * level VM system not knowing anything about existing
832 } else if (((m->flags & PG_REFERENCED) == 0) &&
833 (actcount = pmap_ts_referenced(m))) {
835 VM_OBJECT_UNLOCK(object);
836 m->act_count += (actcount + ACT_ADVANCE);
842 * If the upper level VM system knows about any page
843 * references, we activate the page. We also set the
844 * "activation count" higher than normal so that we will less
845 * likely place pages back onto the inactive queue again.
847 if ((m->flags & PG_REFERENCED) != 0) {
848 vm_page_flag_clear(m, PG_REFERENCED);
849 actcount = pmap_ts_referenced(m);
851 VM_OBJECT_UNLOCK(object);
852 m->act_count += (actcount + ACT_ADVANCE + 1);
858 * If the upper level VM system does not believe that the page
859 * is fully dirty, but it is mapped for write access, then we
860 * consult the pmap to see if the page's dirty status should
863 if (m->dirty != VM_PAGE_BITS_ALL &&
864 (m->flags & PG_WRITEABLE) != 0) {
866 * Avoid a race condition: Unless write access is
867 * removed from the page, another processor could
868 * modify it before all access is removed by the call
869 * to vm_page_cache() below. If vm_page_cache() finds
870 * that the page has been modified when it removes all
871 * access, it panics because it cannot cache dirty
872 * pages. In principle, we could eliminate just write
873 * access here rather than all access. In the expected
874 * case, when there are no last instant modifications
875 * to the page, removing all access will be cheaper
878 if (pmap_is_modified(m))
880 else if (m->dirty == 0)
886 * Invalid pages can be easily freed
891 } else if (m->dirty == 0) {
893 * Clean pages can be placed onto the cache queue.
894 * This effectively frees them.
898 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
900 * Dirty pages need to be paged out, but flushing
901 * a page is extremely expensive verses freeing
902 * a clean page. Rather then artificially limiting
903 * the number of pages we can flush, we instead give
904 * dirty pages extra priority on the inactive queue
905 * by forcing them to be cycled through the queue
906 * twice before being flushed, after which the
907 * (now clean) page will cycle through once more
908 * before being freed. This significantly extends
909 * the thrash point for a heavily loaded machine.
911 vm_page_flag_set(m, PG_WINATCFLS);
913 } else if (maxlaunder > 0) {
915 * We always want to try to flush some dirty pages if
916 * we encounter them, to keep the system stable.
917 * Normally this number is small, but under extreme
918 * pressure where there are insufficient clean pages
919 * on the inactive queue, we may have to go all out.
921 int swap_pageouts_ok, vfslocked = 0;
922 struct vnode *vp = NULL;
923 struct mount *mp = NULL;
925 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
926 swap_pageouts_ok = 1;
928 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
929 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
930 vm_page_count_min());
935 * We don't bother paging objects that are "dead".
936 * Those objects are in a "rundown" state.
938 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
940 VM_OBJECT_UNLOCK(object);
946 * Following operations may unlock
947 * vm_page_queue_mtx, invalidating the 'next'
948 * pointer. To prevent an inordinate number
949 * of restarts we use our marker to remember
953 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
956 * The object is already known NOT to be dead. It
957 * is possible for the vget() to block the whole
958 * pageout daemon, but the new low-memory handling
959 * code should prevent it.
961 * The previous code skipped locked vnodes and, worse,
962 * reordered pages in the queue. This results in
963 * completely non-deterministic operation and, on a
964 * busy system, can lead to extremely non-optimal
965 * pageouts. For example, it can cause clean pages
966 * to be freed and dirty pages to be moved to the end
967 * of the queue. Since dirty pages are also moved to
968 * the end of the queue once-cleaned, this gives
969 * way too large a weighting to defering the freeing
972 * We can't wait forever for the vnode lock, we might
973 * deadlock due to a vn_read() getting stuck in
974 * vm_wait while holding this vnode. We skip the
975 * vnode if we can't get it in a reasonable amount
978 if (object->type == OBJT_VNODE) {
979 vm_page_unlock_queues();
982 if (vp->v_type == VREG &&
983 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
986 if (object->flags & OBJ_MIGHTBEDIRTY)
988 vm_page_lock_queues();
989 goto unlock_and_continue;
992 ("vp %p with NULL v_mount", vp));
993 vm_object_reference_locked(object);
994 VM_OBJECT_UNLOCK(object);
995 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
996 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
998 VM_OBJECT_LOCK(object);
999 vm_page_lock_queues();
1000 ++pageout_lock_miss;
1001 if (object->flags & OBJ_MIGHTBEDIRTY)
1004 goto unlock_and_continue;
1006 VM_OBJECT_LOCK(object);
1008 vm_page_lock_queues();
1010 * The page might have been moved to another
1011 * queue during potential blocking in vget()
1012 * above. The page might have been freed and
1013 * reused for another vnode.
1015 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
1016 m->object != object ||
1017 TAILQ_NEXT(m, pageq) != &marker) {
1019 if (object->flags & OBJ_MIGHTBEDIRTY)
1021 goto unlock_and_continue;
1025 * The page may have been busied during the
1026 * blocking in vget(). We don't move the
1027 * page back onto the end of the queue so that
1028 * statistics are more correct if we don't.
1030 if (m->busy || (m->oflags & VPO_BUSY)) {
1032 goto unlock_and_continue;
1036 * If the page has become held it might
1037 * be undergoing I/O, so skip it
1039 if (m->hold_count) {
1042 if (object->flags & OBJ_MIGHTBEDIRTY)
1044 goto unlock_and_continue;
1050 * If a page is dirty, then it is either being washed
1051 * (but not yet cleaned) or it is still in the
1052 * laundry. If it is still in the laundry, then we
1053 * start the cleaning operation.
1055 * decrement page_shortage on success to account for
1056 * the (future) cleaned page. Otherwise we could wind
1057 * up laundering or cleaning too many pages.
1059 vm_page_unlock_queues();
1060 if (vm_pageout_clean(m) != 0) {
1064 vm_page_lock_queues();
1065 unlock_and_continue:
1066 vm_page_lock_assert(m, MA_NOTOWNED);
1067 VM_OBJECT_UNLOCK(object);
1069 vm_page_unlock_queues();
1072 VFS_UNLOCK_GIANT(vfslocked);
1073 vm_object_deallocate(object);
1074 vn_finished_write(mp);
1075 vm_page_lock_queues();
1077 next = TAILQ_NEXT(&marker, pageq);
1078 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1080 vm_page_lock_assert(m, MA_NOTOWNED);
1084 VM_OBJECT_UNLOCK(object);
1088 * Compute the number of pages we want to try to move from the
1089 * active queue to the inactive queue.
1091 page_shortage = vm_paging_target() +
1092 cnt.v_inactive_target - cnt.v_inactive_count;
1093 page_shortage += addl_page_shortage;
1096 * Scan the active queue for things we can deactivate. We nominally
1097 * track the per-page activity counter and use it to locate
1098 * deactivation candidates.
1100 pcount = cnt.v_active_count;
1101 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1102 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1104 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1106 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1107 ("vm_pageout_scan: page %p isn't active", m));
1109 next = TAILQ_NEXT(m, pageq);
1111 if ((m->flags & PG_MARKER) != 0) {
1115 if (!vm_page_trylock(m) || (object = m->object) == NULL) {
1119 if (!VM_OBJECT_TRYLOCK(object) &&
1120 !vm_pageout_fallback_object_lock(m, &next)) {
1121 VM_OBJECT_UNLOCK(object);
1128 * Don't deactivate pages that are busy.
1130 if ((m->busy != 0) ||
1131 (m->oflags & VPO_BUSY) ||
1132 (m->hold_count != 0)) {
1134 VM_OBJECT_UNLOCK(object);
1141 * The count for pagedaemon pages is done after checking the
1142 * page for eligibility...
1147 * Check to see "how much" the page has been used.
1150 if (object->ref_count != 0) {
1151 if (m->flags & PG_REFERENCED) {
1154 actcount += pmap_ts_referenced(m);
1156 m->act_count += ACT_ADVANCE + actcount;
1157 if (m->act_count > ACT_MAX)
1158 m->act_count = ACT_MAX;
1163 * Since we have "tested" this bit, we need to clear it now.
1165 vm_page_flag_clear(m, PG_REFERENCED);
1168 * Only if an object is currently being used, do we use the
1169 * page activation count stats.
1171 if (actcount && (object->ref_count != 0)) {
1174 m->act_count -= min(m->act_count, ACT_DECLINE);
1175 if (vm_pageout_algorithm ||
1176 object->ref_count == 0 ||
1177 m->act_count == 0) {
1179 if (object->ref_count == 0) {
1180 KASSERT(!pmap_page_is_mapped(m),
1181 ("vm_pageout_scan: page %p is mapped", m));
1185 vm_page_deactivate(m);
1187 vm_page_deactivate(m);
1194 VM_OBJECT_UNLOCK(object);
1197 vm_page_unlock_queues();
1198 #if !defined(NO_SWAPPING)
1200 * Idle process swapout -- run once per second.
1202 if (vm_swap_idle_enabled) {
1204 if (time_second != lsec) {
1205 vm_req_vmdaemon(VM_SWAP_IDLE);
1212 * If we didn't get enough free pages, and we have skipped a vnode
1213 * in a writeable object, wakeup the sync daemon. And kick swapout
1214 * if we did not get enough free pages.
1216 if (vm_paging_target() > 0) {
1217 if (vnodes_skipped && vm_page_count_min())
1218 (void) speedup_syncer();
1219 #if !defined(NO_SWAPPING)
1220 if (vm_swap_enabled && vm_page_count_target())
1221 vm_req_vmdaemon(VM_SWAP_NORMAL);
1226 * If we are critically low on one of RAM or swap and low on
1227 * the other, kill the largest process. However, we avoid
1228 * doing this on the first pass in order to give ourselves a
1229 * chance to flush out dirty vnode-backed pages and to allow
1230 * active pages to be moved to the inactive queue and reclaimed.
1233 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1234 (swap_pager_full && vm_paging_target() > 0)))
1235 vm_pageout_oom(VM_OOM_MEM);
1240 vm_pageout_oom(int shortage)
1242 struct proc *p, *bigproc;
1243 vm_offset_t size, bigsize;
1248 * We keep the process bigproc locked once we find it to keep anyone
1249 * from messing with it; however, there is a possibility of
1250 * deadlock if process B is bigproc and one of it's child processes
1251 * attempts to propagate a signal to B while we are waiting for A's
1252 * lock while walking this list. To avoid this, we don't block on
1253 * the process lock but just skip a process if it is already locked.
1257 sx_slock(&allproc_lock);
1258 FOREACH_PROC_IN_SYSTEM(p) {
1261 if (PROC_TRYLOCK(p) == 0)
1264 * If this is a system, protected or killed process, skip it.
1266 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1267 (p->p_pid == 1) || P_KILLED(p) ||
1268 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1273 * If the process is in a non-running type state,
1274 * don't touch it. Check all the threads individually.
1277 FOREACH_THREAD_IN_PROC(p, td) {
1279 if (!TD_ON_RUNQ(td) &&
1280 !TD_IS_RUNNING(td) &&
1281 !TD_IS_SLEEPING(td)) {
1293 * get the process size
1295 vm = vmspace_acquire_ref(p);
1300 if (!vm_map_trylock_read(&vm->vm_map)) {
1305 size = vmspace_swap_count(vm);
1306 vm_map_unlock_read(&vm->vm_map);
1307 if (shortage == VM_OOM_MEM)
1308 size += vmspace_resident_count(vm);
1311 * if the this process is bigger than the biggest one
1314 if (size > bigsize) {
1315 if (bigproc != NULL)
1316 PROC_UNLOCK(bigproc);
1322 sx_sunlock(&allproc_lock);
1323 if (bigproc != NULL) {
1324 killproc(bigproc, "out of swap space");
1325 sched_nice(bigproc, PRIO_MIN);
1326 PROC_UNLOCK(bigproc);
1327 wakeup(&cnt.v_free_count);
1332 * This routine tries to maintain the pseudo LRU active queue,
1333 * so that during long periods of time where there is no paging,
1334 * that some statistic accumulation still occurs. This code
1335 * helps the situation where paging just starts to occur.
1338 vm_pageout_page_stats()
1342 int pcount,tpcount; /* Number of pages to check */
1343 static int fullintervalcount = 0;
1346 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1348 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1349 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1351 if (page_shortage <= 0)
1354 pcount = cnt.v_active_count;
1355 fullintervalcount += vm_pageout_stats_interval;
1356 if (fullintervalcount < vm_pageout_full_stats_interval) {
1357 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1359 if (pcount > tpcount)
1362 fullintervalcount = 0;
1365 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1366 while ((m != NULL) && (pcount-- > 0)) {
1369 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1370 ("vm_pageout_page_stats: page %p isn't active", m));
1372 next = TAILQ_NEXT(m, pageq);
1375 if ((m->flags & PG_MARKER) != 0) {
1379 vm_page_lock_assert(m, MA_NOTOWNED);
1380 if (vm_page_trylock(m) == 0 || (object = m->object) == NULL) {
1384 if (!VM_OBJECT_TRYLOCK(object) &&
1385 !vm_pageout_fallback_object_lock(m, &next)) {
1386 VM_OBJECT_UNLOCK(object);
1393 * Don't deactivate pages that are busy.
1395 if ((m->busy != 0) ||
1396 (m->oflags & VPO_BUSY) ||
1397 (m->hold_count != 0)) {
1399 VM_OBJECT_UNLOCK(object);
1406 if (m->flags & PG_REFERENCED) {
1407 vm_page_flag_clear(m, PG_REFERENCED);
1411 actcount += pmap_ts_referenced(m);
1413 m->act_count += ACT_ADVANCE + actcount;
1414 if (m->act_count > ACT_MAX)
1415 m->act_count = ACT_MAX;
1418 if (m->act_count == 0) {
1420 * We turn off page access, so that we have
1421 * more accurate RSS stats. We don't do this
1422 * in the normal page deactivation when the
1423 * system is loaded VM wise, because the
1424 * cost of the large number of page protect
1425 * operations would be higher than the value
1426 * of doing the operation.
1429 vm_page_deactivate(m);
1431 m->act_count -= min(m->act_count, ACT_DECLINE);
1436 VM_OBJECT_UNLOCK(object);
1442 * vm_pageout is the high level pageout daemon.
1450 * Initialize some paging parameters.
1452 cnt.v_interrupt_free_min = 2;
1453 if (cnt.v_page_count < 2000)
1454 vm_pageout_page_count = 8;
1457 * v_free_reserved needs to include enough for the largest
1458 * swap pager structures plus enough for any pv_entry structs
1461 if (cnt.v_page_count > 1024)
1462 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1465 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1466 cnt.v_interrupt_free_min;
1467 cnt.v_free_reserved = vm_pageout_page_count +
1468 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1469 cnt.v_free_severe = cnt.v_free_min / 2;
1470 cnt.v_free_min += cnt.v_free_reserved;
1471 cnt.v_free_severe += cnt.v_free_reserved;
1474 * v_free_target and v_cache_min control pageout hysteresis. Note
1475 * that these are more a measure of the VM cache queue hysteresis
1476 * then the VM free queue. Specifically, v_free_target is the
1477 * high water mark (free+cache pages).
1479 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1480 * low water mark, while v_free_min is the stop. v_cache_min must
1481 * be big enough to handle memory needs while the pageout daemon
1482 * is signalled and run to free more pages.
1484 if (cnt.v_free_count > 6144)
1485 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1487 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1489 if (cnt.v_free_count > 2048) {
1490 cnt.v_cache_min = cnt.v_free_target;
1491 cnt.v_cache_max = 2 * cnt.v_cache_min;
1492 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1494 cnt.v_cache_min = 0;
1495 cnt.v_cache_max = 0;
1496 cnt.v_inactive_target = cnt.v_free_count / 4;
1498 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1499 cnt.v_inactive_target = cnt.v_free_count / 3;
1501 /* XXX does not really belong here */
1502 if (vm_page_max_wired == 0)
1503 vm_page_max_wired = cnt.v_free_count / 3;
1505 if (vm_pageout_stats_max == 0)
1506 vm_pageout_stats_max = cnt.v_free_target;
1509 * Set interval in seconds for stats scan.
1511 if (vm_pageout_stats_interval == 0)
1512 vm_pageout_stats_interval = 5;
1513 if (vm_pageout_full_stats_interval == 0)
1514 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1516 swap_pager_swap_init();
1519 * The pageout daemon is never done, so loop forever.
1523 * If we have enough free memory, wakeup waiters. Do
1524 * not clear vm_pages_needed until we reach our target,
1525 * otherwise we may be woken up over and over again and
1526 * waste a lot of cpu.
1528 mtx_lock(&vm_page_queue_free_mtx);
1529 if (vm_pages_needed && !vm_page_count_min()) {
1530 if (!vm_paging_needed())
1531 vm_pages_needed = 0;
1532 wakeup(&cnt.v_free_count);
1534 if (vm_pages_needed) {
1536 * Still not done, take a second pass without waiting
1537 * (unlimited dirty cleaning), otherwise sleep a bit
1542 msleep(&vm_pages_needed,
1543 &vm_page_queue_free_mtx, PVM, "psleep",
1547 * Good enough, sleep & handle stats. Prime the pass
1554 error = msleep(&vm_pages_needed,
1555 &vm_page_queue_free_mtx, PVM, "psleep",
1556 vm_pageout_stats_interval * hz);
1557 if (error && !vm_pages_needed) {
1558 mtx_unlock(&vm_page_queue_free_mtx);
1560 vm_page_lock_queues();
1561 vm_pageout_page_stats();
1562 vm_page_unlock_queues();
1566 if (vm_pages_needed)
1568 mtx_unlock(&vm_page_queue_free_mtx);
1569 vm_pageout_scan(pass);
1574 * Unless the free page queue lock is held by the caller, this function
1575 * should be regarded as advisory. Specifically, the caller should
1576 * not msleep() on &cnt.v_free_count following this function unless
1577 * the free page queue lock is held until the msleep() is performed.
1583 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1584 vm_pages_needed = 1;
1585 wakeup(&vm_pages_needed);
1589 #if !defined(NO_SWAPPING)
1591 vm_req_vmdaemon(int req)
1593 static int lastrun = 0;
1595 mtx_lock(&vm_daemon_mtx);
1596 vm_pageout_req_swapout |= req;
1597 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1598 wakeup(&vm_daemon_needed);
1601 mtx_unlock(&vm_daemon_mtx);
1607 struct rlimit rsslim;
1611 int breakout, swapout_flags;
1614 mtx_lock(&vm_daemon_mtx);
1615 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1616 swapout_flags = vm_pageout_req_swapout;
1617 vm_pageout_req_swapout = 0;
1618 mtx_unlock(&vm_daemon_mtx);
1620 swapout_procs(swapout_flags);
1623 * scan the processes for exceeding their rlimits or if
1624 * process is swapped out -- deactivate pages
1626 sx_slock(&allproc_lock);
1627 FOREACH_PROC_IN_SYSTEM(p) {
1628 vm_pindex_t limit, size;
1631 * if this is a system process or if we have already
1632 * looked at this process, skip it.
1635 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1640 * if the process is in a non-running type state,
1644 FOREACH_THREAD_IN_PROC(p, td) {
1646 if (!TD_ON_RUNQ(td) &&
1647 !TD_IS_RUNNING(td) &&
1648 !TD_IS_SLEEPING(td)) {
1662 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1664 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1667 * let processes that are swapped out really be
1668 * swapped out set the limit to nothing (will force a
1671 if ((p->p_flag & P_INMEM) == 0)
1672 limit = 0; /* XXX */
1673 vm = vmspace_acquire_ref(p);
1678 size = vmspace_resident_count(vm);
1679 if (limit >= 0 && size >= limit) {
1680 vm_pageout_map_deactivate_pages(
1681 &vm->vm_map, limit);
1685 sx_sunlock(&allproc_lock);
1688 #endif /* !defined(NO_SWAPPING) */