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();
255 VM_OBJECT_LOCK(object);
256 vm_page_lock_queues();
258 /* Page queue might have changed. */
259 *next = TAILQ_NEXT(&marker, pageq);
260 unchanged = (m->queue == queue &&
261 m->object == object &&
262 &marker == TAILQ_NEXT(m, pageq));
263 TAILQ_REMOVE(&vm_page_queues[queue].pl,
271 * Clean the page and remove it from the laundry.
273 * We set the busy bit to cause potential page faults on this page to
274 * block. Note the careful timing, however, the busy bit isn't set till
275 * late and we cannot do anything that will mess with the page.
282 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
284 int ib, is, page_base;
285 vm_pindex_t pindex = m->pindex;
287 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
288 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
291 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
292 * with the new swapper, but we could have serious problems paging
293 * out other object types if there is insufficient memory.
295 * Unfortunately, checking free memory here is far too late, so the
296 * check has been moved up a procedural level.
300 * Can't clean the page if it's busy or held.
302 if ((m->hold_count != 0) ||
303 ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
307 mc[vm_pageout_page_count] = pb = ps = m;
309 page_base = vm_pageout_page_count;
314 * Scan object for clusterable pages.
316 * We can cluster ONLY if: ->> the page is NOT
317 * clean, wired, busy, held, or mapped into a
318 * buffer, and one of the following:
319 * 1) The page is inactive, or a seldom used
322 * 2) we force the issue.
324 * During heavy mmap/modification loads the pageout
325 * daemon can really fragment the underlying file
326 * due to flushing pages out of order and not trying
327 * align the clusters (which leave sporatic out-of-order
328 * holes). To solve this problem we do the reverse scan
329 * first and attempt to align our cluster, then do a
330 * forward scan if room remains.
334 while (ib && pageout_count < vm_pageout_page_count) {
342 if ((p = vm_page_prev(pb)) == NULL ||
343 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
347 vm_page_test_dirty(p);
349 p->queue != PQ_INACTIVE ||
350 p->hold_count != 0) { /* may be undergoing I/O */
354 mc[--page_base] = pb = p;
358 * alignment boundry, stop here and switch directions. Do
361 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
365 while (pageout_count < vm_pageout_page_count &&
366 pindex + is < object->size) {
369 if ((p = vm_page_next(ps)) == NULL ||
370 (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
372 vm_page_test_dirty(p);
374 p->queue != PQ_INACTIVE ||
375 p->hold_count != 0) { /* may be undergoing I/O */
378 mc[page_base + pageout_count] = ps = p;
384 * If we exhausted our forward scan, continue with the reverse scan
385 * when possible, even past a page boundry. This catches boundry
388 if (ib && pageout_count < vm_pageout_page_count)
392 * we allow reads during pageouts...
394 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
398 * vm_pageout_flush() - launder the given pages
400 * The given pages are laundered. Note that we setup for the start of
401 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
402 * reference count all in here rather then in the parent. If we want
403 * the parent to do more sophisticated things we may have to change
407 vm_pageout_flush(vm_page_t *mc, int count, int flags)
409 vm_object_t object = mc[0]->object;
410 int pageout_status[count];
414 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
415 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
417 * Initiate I/O. Bump the vm_page_t->busy counter and
418 * mark the pages read-only.
420 * We do not have to fixup the clean/dirty bits here... we can
421 * allow the pager to do it after the I/O completes.
423 * NOTE! mc[i]->dirty may be partial or fragmented due to an
424 * edge case with file fragments.
426 for (i = 0; i < count; i++) {
427 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
428 ("vm_pageout_flush: partially invalid page %p index %d/%d",
430 vm_page_io_start(mc[i]);
431 pmap_remove_write(mc[i]);
433 vm_page_unlock_queues();
434 vm_object_pip_add(object, count);
436 vm_pager_put_pages(object, mc, count, flags, pageout_status);
438 vm_page_lock_queues();
439 for (i = 0; i < count; i++) {
440 vm_page_t mt = mc[i];
442 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
443 (mt->flags & PG_WRITEABLE) == 0,
444 ("vm_pageout_flush: page %p is not write protected", mt));
445 switch (pageout_status[i]) {
452 * Page outside of range of object. Right now we
453 * essentially lose the changes by pretending it
461 * If page couldn't be paged out, then reactivate the
462 * page so it doesn't clog the inactive list. (We
463 * will try paging out it again later).
465 vm_page_activate(mt);
472 * If the operation is still going, leave the page busy to
473 * block all other accesses. Also, leave the paging in
474 * progress indicator set so that we don't attempt an object
477 if (pageout_status[i] != VM_PAGER_PEND) {
478 vm_object_pip_wakeup(object);
479 vm_page_io_finish(mt);
480 if (vm_page_count_severe())
481 vm_page_try_to_cache(mt);
487 #if !defined(NO_SWAPPING)
489 * vm_pageout_object_deactivate_pages
491 * deactivate enough pages to satisfy the inactive target
492 * requirements or if vm_page_proc_limit is set, then
493 * deactivate all of the pages in the object and its
496 * The object and map must be locked.
499 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
501 vm_object_t first_object;
504 vm_object_t backing_object, object;
506 int actcount, rcount, remove_mode;
508 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
509 if (first_object->type == OBJT_DEVICE ||
510 first_object->type == OBJT_SG ||
511 first_object->type == OBJT_PHYS)
513 for (object = first_object;; object = backing_object) {
514 if (pmap_resident_count(pmap) <= desired)
516 if (object->paging_in_progress)
520 if (object->shadow_count > 1)
523 * scan the objects entire memory queue
525 rcount = object->resident_page_count;
526 p = TAILQ_FIRST(&object->memq);
527 vm_page_lock_queues();
528 while (p && (rcount-- > 0)) {
529 if (pmap_resident_count(pmap) <= desired) {
530 vm_page_unlock_queues();
533 next = TAILQ_NEXT(p, listq);
535 if (p->wire_count != 0 ||
536 p->hold_count != 0 ||
538 (p->oflags & VPO_BUSY) ||
539 (p->flags & PG_UNMANAGED) ||
540 !pmap_page_exists_quick(pmap, p)) {
544 actcount = pmap_ts_referenced(p);
546 vm_page_flag_set(p, PG_REFERENCED);
547 } else if (p->flags & PG_REFERENCED) {
550 if ((p->queue != PQ_ACTIVE) &&
551 (p->flags & PG_REFERENCED)) {
553 p->act_count += actcount;
554 vm_page_flag_clear(p, PG_REFERENCED);
555 } else if (p->queue == PQ_ACTIVE) {
556 if ((p->flags & PG_REFERENCED) == 0) {
557 p->act_count -= min(p->act_count, ACT_DECLINE);
558 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
560 vm_page_deactivate(p);
566 vm_page_flag_clear(p, PG_REFERENCED);
567 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
568 p->act_count += ACT_ADVANCE;
571 } else if (p->queue == PQ_INACTIVE) {
576 vm_page_unlock_queues();
577 if ((backing_object = object->backing_object) == NULL)
579 VM_OBJECT_LOCK(backing_object);
580 if (object != first_object)
581 VM_OBJECT_UNLOCK(object);
584 if (object != first_object)
585 VM_OBJECT_UNLOCK(object);
589 * deactivate some number of pages in a map, try to do it fairly, but
590 * that is really hard to do.
593 vm_pageout_map_deactivate_pages(map, desired)
598 vm_object_t obj, bigobj;
601 if (!vm_map_trylock(map))
608 * first, search out the biggest object, and try to free pages from
611 tmpe = map->header.next;
612 while (tmpe != &map->header) {
613 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
614 obj = tmpe->object.vm_object;
615 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
616 if (obj->shadow_count <= 1 &&
618 bigobj->resident_page_count < obj->resident_page_count)) {
620 VM_OBJECT_UNLOCK(bigobj);
623 VM_OBJECT_UNLOCK(obj);
626 if (tmpe->wired_count > 0)
627 nothingwired = FALSE;
631 if (bigobj != NULL) {
632 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
633 VM_OBJECT_UNLOCK(bigobj);
636 * Next, hunt around for other pages to deactivate. We actually
637 * do this search sort of wrong -- .text first is not the best idea.
639 tmpe = map->header.next;
640 while (tmpe != &map->header) {
641 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
643 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
644 obj = tmpe->object.vm_object;
647 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
648 VM_OBJECT_UNLOCK(obj);
655 * Remove all mappings if a process is swapped out, this will free page
658 if (desired == 0 && nothingwired) {
659 tmpe = map->header.next;
660 while (tmpe != &map->header) {
661 pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end);
667 #endif /* !defined(NO_SWAPPING) */
670 * vm_pageout_scan does the dirty work for the pageout daemon.
673 vm_pageout_scan(int pass)
676 struct vm_page marker;
677 int page_shortage, maxscan, pcount;
678 int addl_page_shortage, addl_page_shortage_init;
681 int vnodes_skipped = 0;
685 * Decrease registered cache sizes.
687 EVENTHANDLER_INVOKE(vm_lowmem, 0);
689 * We do this explicitly after the caches have been drained above.
693 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
696 * Calculate the number of pages we want to either free or move
699 page_shortage = vm_paging_target() + addl_page_shortage_init;
702 * Initialize our marker
704 bzero(&marker, sizeof(marker));
705 marker.flags = PG_FICTITIOUS | PG_MARKER;
706 marker.oflags = VPO_BUSY;
707 marker.queue = PQ_INACTIVE;
708 marker.wire_count = 1;
711 * Start scanning the inactive queue for pages we can move to the
712 * cache or free. The scan will stop when the target is reached or
713 * we have scanned the entire inactive queue. Note that m->act_count
714 * is not used to form decisions for the inactive queue, only for the
717 * maxlaunder limits the number of dirty pages we flush per scan.
718 * For most systems a smaller value (16 or 32) is more robust under
719 * extreme memory and disk pressure because any unnecessary writes
720 * to disk can result in extreme performance degredation. However,
721 * systems with excessive dirty pages (especially when MAP_NOSYNC is
722 * used) will die horribly with limited laundering. If the pageout
723 * daemon cannot clean enough pages in the first pass, we let it go
724 * all out in succeeding passes.
726 if ((maxlaunder = vm_max_launder) <= 1)
730 vm_page_lock_queues();
732 addl_page_shortage = addl_page_shortage_init;
733 maxscan = cnt.v_inactive_count;
735 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
736 m != NULL && maxscan-- > 0 && page_shortage > 0;
741 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
745 next = TAILQ_NEXT(m, pageq);
751 if (m->flags & PG_MARKER)
755 * A held page may be undergoing I/O, so skip it.
759 addl_page_shortage++;
763 * Don't mess with busy pages, keep in the front of the
764 * queue, most likely are being paged out.
766 if (!VM_OBJECT_TRYLOCK(object) &&
767 (!vm_pageout_fallback_object_lock(m, &next) ||
768 m->hold_count != 0)) {
769 VM_OBJECT_UNLOCK(object);
770 addl_page_shortage++;
773 if (m->busy || (m->oflags & VPO_BUSY)) {
774 VM_OBJECT_UNLOCK(object);
775 addl_page_shortage++;
780 * If the object is not being used, we ignore previous
783 if (object->ref_count == 0) {
784 vm_page_flag_clear(m, PG_REFERENCED);
785 KASSERT(!pmap_page_is_mapped(m),
786 ("vm_pageout_scan: page %p is mapped", m));
789 * Otherwise, if the page has been referenced while in the
790 * inactive queue, we bump the "activation count" upwards,
791 * making it less likely that the page will be added back to
792 * the inactive queue prematurely again. Here we check the
793 * page tables (or emulated bits, if any), given the upper
794 * level VM system not knowing anything about existing
797 } else if (((m->flags & PG_REFERENCED) == 0) &&
798 (actcount = pmap_ts_referenced(m))) {
800 VM_OBJECT_UNLOCK(object);
801 m->act_count += (actcount + ACT_ADVANCE);
806 * If the upper level VM system knows about any page
807 * references, we activate the page. We also set the
808 * "activation count" higher than normal so that we will less
809 * likely place pages back onto the inactive queue again.
811 if ((m->flags & PG_REFERENCED) != 0) {
812 vm_page_flag_clear(m, PG_REFERENCED);
813 actcount = pmap_ts_referenced(m);
815 VM_OBJECT_UNLOCK(object);
816 m->act_count += (actcount + ACT_ADVANCE + 1);
821 * If the upper level VM system does not believe that the page
822 * is fully dirty, but it is mapped for write access, then we
823 * consult the pmap to see if the page's dirty status should
826 if (m->dirty != VM_PAGE_BITS_ALL &&
827 (m->flags & PG_WRITEABLE) != 0) {
829 * Avoid a race condition: Unless write access is
830 * removed from the page, another processor could
831 * modify it before all access is removed by the call
832 * to vm_page_cache() below. If vm_page_cache() finds
833 * that the page has been modified when it removes all
834 * access, it panics because it cannot cache dirty
835 * pages. In principle, we could eliminate just write
836 * access here rather than all access. In the expected
837 * case, when there are no last instant modifications
838 * to the page, removing all access will be cheaper
841 if (pmap_is_modified(m))
843 else if (m->dirty == 0)
849 * Invalid pages can be easily freed
854 } else if (m->dirty == 0) {
856 * Clean pages can be placed onto the cache queue.
857 * This effectively frees them.
861 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
863 * Dirty pages need to be paged out, but flushing
864 * a page is extremely expensive verses freeing
865 * a clean page. Rather then artificially limiting
866 * the number of pages we can flush, we instead give
867 * dirty pages extra priority on the inactive queue
868 * by forcing them to be cycled through the queue
869 * twice before being flushed, after which the
870 * (now clean) page will cycle through once more
871 * before being freed. This significantly extends
872 * the thrash point for a heavily loaded machine.
874 vm_page_flag_set(m, PG_WINATCFLS);
876 } else if (maxlaunder > 0) {
878 * We always want to try to flush some dirty pages if
879 * we encounter them, to keep the system stable.
880 * Normally this number is small, but under extreme
881 * pressure where there are insufficient clean pages
882 * on the inactive queue, we may have to go all out.
884 int swap_pageouts_ok, vfslocked = 0;
885 struct vnode *vp = NULL;
886 struct mount *mp = NULL;
888 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
889 swap_pageouts_ok = 1;
891 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
892 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
893 vm_page_count_min());
898 * We don't bother paging objects that are "dead".
899 * Those objects are in a "rundown" state.
901 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
902 VM_OBJECT_UNLOCK(object);
908 * Following operations may unlock
909 * vm_page_queue_mtx, invalidating the 'next'
910 * pointer. To prevent an inordinate number
911 * of restarts we use our marker to remember
915 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
918 * The object is already known NOT to be dead. It
919 * is possible for the vget() to block the whole
920 * pageout daemon, but the new low-memory handling
921 * code should prevent it.
923 * The previous code skipped locked vnodes and, worse,
924 * reordered pages in the queue. This results in
925 * completely non-deterministic operation and, on a
926 * busy system, can lead to extremely non-optimal
927 * pageouts. For example, it can cause clean pages
928 * to be freed and dirty pages to be moved to the end
929 * of the queue. Since dirty pages are also moved to
930 * the end of the queue once-cleaned, this gives
931 * way too large a weighting to defering the freeing
934 * We can't wait forever for the vnode lock, we might
935 * deadlock due to a vn_read() getting stuck in
936 * vm_wait while holding this vnode. We skip the
937 * vnode if we can't get it in a reasonable amount
940 if (object->type == OBJT_VNODE) {
942 if (vp->v_type == VREG &&
943 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
946 if (object->flags & OBJ_MIGHTBEDIRTY)
948 goto unlock_and_continue;
951 ("vp %p with NULL v_mount", vp));
952 vm_page_unlock_queues();
953 vm_object_reference_locked(object);
954 VM_OBJECT_UNLOCK(object);
955 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
956 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
958 VM_OBJECT_LOCK(object);
959 vm_page_lock_queues();
961 if (object->flags & OBJ_MIGHTBEDIRTY)
964 goto unlock_and_continue;
966 VM_OBJECT_LOCK(object);
967 vm_page_lock_queues();
969 * The page might have been moved to another
970 * queue during potential blocking in vget()
971 * above. The page might have been freed and
972 * reused for another vnode.
974 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
975 m->object != object ||
976 TAILQ_NEXT(m, pageq) != &marker) {
977 if (object->flags & OBJ_MIGHTBEDIRTY)
979 goto unlock_and_continue;
983 * The page may have been busied during the
984 * blocking in vget(). We don't move the
985 * page back onto the end of the queue so that
986 * statistics are more correct if we don't.
988 if (m->busy || (m->oflags & VPO_BUSY)) {
989 goto unlock_and_continue;
993 * If the page has become held it might
994 * be undergoing I/O, so skip it
998 if (object->flags & OBJ_MIGHTBEDIRTY)
1000 goto unlock_and_continue;
1005 * If a page is dirty, then it is either being washed
1006 * (but not yet cleaned) or it is still in the
1007 * laundry. If it is still in the laundry, then we
1008 * start the cleaning operation.
1010 * decrement page_shortage on success to account for
1011 * the (future) cleaned page. Otherwise we could wind
1012 * up laundering or cleaning too many pages.
1014 if (vm_pageout_clean(m) != 0) {
1018 unlock_and_continue:
1019 VM_OBJECT_UNLOCK(object);
1021 vm_page_unlock_queues();
1024 VFS_UNLOCK_GIANT(vfslocked);
1025 vm_object_deallocate(object);
1026 vn_finished_write(mp);
1027 vm_page_lock_queues();
1029 next = TAILQ_NEXT(&marker, pageq);
1030 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1034 VM_OBJECT_UNLOCK(object);
1038 * Compute the number of pages we want to try to move from the
1039 * active queue to the inactive queue.
1041 page_shortage = vm_paging_target() +
1042 cnt.v_inactive_target - cnt.v_inactive_count;
1043 page_shortage += addl_page_shortage;
1046 * Scan the active queue for things we can deactivate. We nominally
1047 * track the per-page activity counter and use it to locate
1048 * deactivation candidates.
1050 pcount = cnt.v_active_count;
1051 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1053 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1055 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1056 ("vm_pageout_scan: page %p isn't active", m));
1058 next = TAILQ_NEXT(m, pageq);
1060 if ((m->flags & PG_MARKER) != 0) {
1064 if (!VM_OBJECT_TRYLOCK(object) &&
1065 !vm_pageout_fallback_object_lock(m, &next)) {
1066 VM_OBJECT_UNLOCK(object);
1072 * Don't deactivate pages that are busy.
1074 if ((m->busy != 0) ||
1075 (m->oflags & VPO_BUSY) ||
1076 (m->hold_count != 0)) {
1077 VM_OBJECT_UNLOCK(object);
1084 * The count for pagedaemon pages is done after checking the
1085 * page for eligibility...
1090 * Check to see "how much" the page has been used.
1093 if (object->ref_count != 0) {
1094 if (m->flags & PG_REFERENCED) {
1097 actcount += pmap_ts_referenced(m);
1099 m->act_count += ACT_ADVANCE + actcount;
1100 if (m->act_count > ACT_MAX)
1101 m->act_count = ACT_MAX;
1106 * Since we have "tested" this bit, we need to clear it now.
1108 vm_page_flag_clear(m, PG_REFERENCED);
1111 * Only if an object is currently being used, do we use the
1112 * page activation count stats.
1114 if (actcount && (object->ref_count != 0)) {
1117 m->act_count -= min(m->act_count, ACT_DECLINE);
1118 if (vm_pageout_algorithm ||
1119 object->ref_count == 0 ||
1120 m->act_count == 0) {
1122 if (object->ref_count == 0) {
1127 vm_page_deactivate(m);
1129 vm_page_deactivate(m);
1135 VM_OBJECT_UNLOCK(object);
1138 vm_page_unlock_queues();
1139 #if !defined(NO_SWAPPING)
1141 * Idle process swapout -- run once per second.
1143 if (vm_swap_idle_enabled) {
1145 if (time_second != lsec) {
1146 vm_req_vmdaemon(VM_SWAP_IDLE);
1153 * If we didn't get enough free pages, and we have skipped a vnode
1154 * in a writeable object, wakeup the sync daemon. And kick swapout
1155 * if we did not get enough free pages.
1157 if (vm_paging_target() > 0) {
1158 if (vnodes_skipped && vm_page_count_min())
1159 (void) speedup_syncer();
1160 #if !defined(NO_SWAPPING)
1161 if (vm_swap_enabled && vm_page_count_target())
1162 vm_req_vmdaemon(VM_SWAP_NORMAL);
1167 * If we are critically low on one of RAM or swap and low on
1168 * the other, kill the largest process. However, we avoid
1169 * doing this on the first pass in order to give ourselves a
1170 * chance to flush out dirty vnode-backed pages and to allow
1171 * active pages to be moved to the inactive queue and reclaimed.
1174 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1175 (swap_pager_full && vm_paging_target() > 0)))
1176 vm_pageout_oom(VM_OOM_MEM);
1181 vm_pageout_oom(int shortage)
1183 struct proc *p, *bigproc;
1184 vm_offset_t size, bigsize;
1189 * We keep the process bigproc locked once we find it to keep anyone
1190 * from messing with it; however, there is a possibility of
1191 * deadlock if process B is bigproc and one of it's child processes
1192 * attempts to propagate a signal to B while we are waiting for A's
1193 * lock while walking this list. To avoid this, we don't block on
1194 * the process lock but just skip a process if it is already locked.
1198 sx_slock(&allproc_lock);
1199 FOREACH_PROC_IN_SYSTEM(p) {
1202 if (PROC_TRYLOCK(p) == 0)
1205 * If this is a system, protected or killed process, skip it.
1207 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1208 (p->p_pid == 1) || P_KILLED(p) ||
1209 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1214 * If the process is in a non-running type state,
1215 * don't touch it. Check all the threads individually.
1218 FOREACH_THREAD_IN_PROC(p, td) {
1220 if (!TD_ON_RUNQ(td) &&
1221 !TD_IS_RUNNING(td) &&
1222 !TD_IS_SLEEPING(td)) {
1234 * get the process size
1236 vm = vmspace_acquire_ref(p);
1241 if (!vm_map_trylock_read(&vm->vm_map)) {
1246 size = vmspace_swap_count(vm);
1247 vm_map_unlock_read(&vm->vm_map);
1248 if (shortage == VM_OOM_MEM)
1249 size += vmspace_resident_count(vm);
1252 * if the this process is bigger than the biggest one
1255 if (size > bigsize) {
1256 if (bigproc != NULL)
1257 PROC_UNLOCK(bigproc);
1263 sx_sunlock(&allproc_lock);
1264 if (bigproc != NULL) {
1265 killproc(bigproc, "out of swap space");
1266 sched_nice(bigproc, PRIO_MIN);
1267 PROC_UNLOCK(bigproc);
1268 wakeup(&cnt.v_free_count);
1273 * This routine tries to maintain the pseudo LRU active queue,
1274 * so that during long periods of time where there is no paging,
1275 * that some statistic accumulation still occurs. This code
1276 * helps the situation where paging just starts to occur.
1279 vm_pageout_page_stats()
1283 int pcount,tpcount; /* Number of pages to check */
1284 static int fullintervalcount = 0;
1287 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1289 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1290 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1292 if (page_shortage <= 0)
1295 pcount = cnt.v_active_count;
1296 fullintervalcount += vm_pageout_stats_interval;
1297 if (fullintervalcount < vm_pageout_full_stats_interval) {
1298 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1300 if (pcount > tpcount)
1303 fullintervalcount = 0;
1306 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1307 while ((m != NULL) && (pcount-- > 0)) {
1310 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1311 ("vm_pageout_page_stats: page %p isn't active", m));
1313 next = TAILQ_NEXT(m, pageq);
1316 if ((m->flags & PG_MARKER) != 0) {
1320 if (!VM_OBJECT_TRYLOCK(object) &&
1321 !vm_pageout_fallback_object_lock(m, &next)) {
1322 VM_OBJECT_UNLOCK(object);
1328 * Don't deactivate pages that are busy.
1330 if ((m->busy != 0) ||
1331 (m->oflags & VPO_BUSY) ||
1332 (m->hold_count != 0)) {
1333 VM_OBJECT_UNLOCK(object);
1340 if (m->flags & PG_REFERENCED) {
1341 vm_page_flag_clear(m, PG_REFERENCED);
1345 actcount += pmap_ts_referenced(m);
1347 m->act_count += ACT_ADVANCE + actcount;
1348 if (m->act_count > ACT_MAX)
1349 m->act_count = ACT_MAX;
1352 if (m->act_count == 0) {
1354 * We turn off page access, so that we have
1355 * more accurate RSS stats. We don't do this
1356 * in the normal page deactivation when the
1357 * system is loaded VM wise, because the
1358 * cost of the large number of page protect
1359 * operations would be higher than the value
1360 * of doing the operation.
1363 vm_page_deactivate(m);
1365 m->act_count -= min(m->act_count, ACT_DECLINE);
1369 VM_OBJECT_UNLOCK(object);
1375 * vm_pageout is the high level pageout daemon.
1383 * Initialize some paging parameters.
1385 cnt.v_interrupt_free_min = 2;
1386 if (cnt.v_page_count < 2000)
1387 vm_pageout_page_count = 8;
1390 * v_free_reserved needs to include enough for the largest
1391 * swap pager structures plus enough for any pv_entry structs
1394 if (cnt.v_page_count > 1024)
1395 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1398 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1399 cnt.v_interrupt_free_min;
1400 cnt.v_free_reserved = vm_pageout_page_count +
1401 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1402 cnt.v_free_severe = cnt.v_free_min / 2;
1403 cnt.v_free_min += cnt.v_free_reserved;
1404 cnt.v_free_severe += cnt.v_free_reserved;
1407 * v_free_target and v_cache_min control pageout hysteresis. Note
1408 * that these are more a measure of the VM cache queue hysteresis
1409 * then the VM free queue. Specifically, v_free_target is the
1410 * high water mark (free+cache pages).
1412 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1413 * low water mark, while v_free_min is the stop. v_cache_min must
1414 * be big enough to handle memory needs while the pageout daemon
1415 * is signalled and run to free more pages.
1417 if (cnt.v_free_count > 6144)
1418 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1420 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1422 if (cnt.v_free_count > 2048) {
1423 cnt.v_cache_min = cnt.v_free_target;
1424 cnt.v_cache_max = 2 * cnt.v_cache_min;
1425 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1427 cnt.v_cache_min = 0;
1428 cnt.v_cache_max = 0;
1429 cnt.v_inactive_target = cnt.v_free_count / 4;
1431 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1432 cnt.v_inactive_target = cnt.v_free_count / 3;
1434 /* XXX does not really belong here */
1435 if (vm_page_max_wired == 0)
1436 vm_page_max_wired = cnt.v_free_count / 3;
1438 if (vm_pageout_stats_max == 0)
1439 vm_pageout_stats_max = cnt.v_free_target;
1442 * Set interval in seconds for stats scan.
1444 if (vm_pageout_stats_interval == 0)
1445 vm_pageout_stats_interval = 5;
1446 if (vm_pageout_full_stats_interval == 0)
1447 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1449 swap_pager_swap_init();
1452 * The pageout daemon is never done, so loop forever.
1456 * If we have enough free memory, wakeup waiters. Do
1457 * not clear vm_pages_needed until we reach our target,
1458 * otherwise we may be woken up over and over again and
1459 * waste a lot of cpu.
1461 mtx_lock(&vm_page_queue_free_mtx);
1462 if (vm_pages_needed && !vm_page_count_min()) {
1463 if (!vm_paging_needed())
1464 vm_pages_needed = 0;
1465 wakeup(&cnt.v_free_count);
1467 if (vm_pages_needed) {
1469 * Still not done, take a second pass without waiting
1470 * (unlimited dirty cleaning), otherwise sleep a bit
1475 msleep(&vm_pages_needed,
1476 &vm_page_queue_free_mtx, PVM, "psleep",
1480 * Good enough, sleep & handle stats. Prime the pass
1487 error = msleep(&vm_pages_needed,
1488 &vm_page_queue_free_mtx, PVM, "psleep",
1489 vm_pageout_stats_interval * hz);
1490 if (error && !vm_pages_needed) {
1491 mtx_unlock(&vm_page_queue_free_mtx);
1493 vm_page_lock_queues();
1494 vm_pageout_page_stats();
1495 vm_page_unlock_queues();
1499 if (vm_pages_needed)
1501 mtx_unlock(&vm_page_queue_free_mtx);
1502 vm_pageout_scan(pass);
1507 * Unless the free page queue lock is held by the caller, this function
1508 * should be regarded as advisory. Specifically, the caller should
1509 * not msleep() on &cnt.v_free_count following this function unless
1510 * the free page queue lock is held until the msleep() is performed.
1516 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1517 vm_pages_needed = 1;
1518 wakeup(&vm_pages_needed);
1522 #if !defined(NO_SWAPPING)
1524 vm_req_vmdaemon(int req)
1526 static int lastrun = 0;
1528 mtx_lock(&vm_daemon_mtx);
1529 vm_pageout_req_swapout |= req;
1530 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1531 wakeup(&vm_daemon_needed);
1534 mtx_unlock(&vm_daemon_mtx);
1540 struct rlimit rsslim;
1544 int breakout, swapout_flags;
1547 mtx_lock(&vm_daemon_mtx);
1548 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1549 swapout_flags = vm_pageout_req_swapout;
1550 vm_pageout_req_swapout = 0;
1551 mtx_unlock(&vm_daemon_mtx);
1553 swapout_procs(swapout_flags);
1556 * scan the processes for exceeding their rlimits or if
1557 * process is swapped out -- deactivate pages
1559 sx_slock(&allproc_lock);
1560 FOREACH_PROC_IN_SYSTEM(p) {
1561 vm_pindex_t limit, size;
1564 * if this is a system process or if we have already
1565 * looked at this process, skip it.
1568 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1573 * if the process is in a non-running type state,
1577 FOREACH_THREAD_IN_PROC(p, td) {
1579 if (!TD_ON_RUNQ(td) &&
1580 !TD_IS_RUNNING(td) &&
1581 !TD_IS_SLEEPING(td)) {
1595 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1597 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1600 * let processes that are swapped out really be
1601 * swapped out set the limit to nothing (will force a
1604 if ((p->p_flag & P_INMEM) == 0)
1605 limit = 0; /* XXX */
1606 vm = vmspace_acquire_ref(p);
1611 size = vmspace_resident_count(vm);
1612 if (limit >= 0 && size >= limit) {
1613 vm_pageout_map_deactivate_pages(
1614 &vm->vm_map, limit);
1618 sx_sunlock(&allproc_lock);
1621 #endif /* !defined(NO_SWAPPING) */