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, 0, NULL,
399 * vm_pageout_flush() - launder the given pages
401 * The given pages are laundered. Note that we setup for the start of
402 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
403 * reference count all in here rather then in the parent. If we want
404 * the parent to do more sophisticated things we may have to change
407 * Returned runlen is the count of pages between mreq and first
408 * page after mreq with status VM_PAGER_AGAIN.
409 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
410 * for any page in runlen set.
413 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
416 vm_object_t object = mc[0]->object;
417 int pageout_status[count];
421 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
422 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
424 * Initiate I/O. Bump the vm_page_t->busy counter and
425 * mark the pages read-only.
427 * We do not have to fixup the clean/dirty bits here... we can
428 * allow the pager to do it after the I/O completes.
430 * NOTE! mc[i]->dirty may be partial or fragmented due to an
431 * edge case with file fragments.
433 for (i = 0; i < count; i++) {
434 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
435 ("vm_pageout_flush: partially invalid page %p index %d/%d",
437 vm_page_io_start(mc[i]);
438 pmap_remove_write(mc[i]);
440 vm_page_unlock_queues();
441 vm_object_pip_add(object, count);
443 vm_pager_put_pages(object, mc, count, flags, pageout_status);
445 runlen = count - mreq;
448 vm_page_lock_queues();
449 for (i = 0; i < count; i++) {
450 vm_page_t mt = mc[i];
452 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
453 (mt->flags & PG_WRITEABLE) == 0,
454 ("vm_pageout_flush: page %p is not write protected", mt));
455 switch (pageout_status[i]) {
462 * Page outside of range of object. Right now we
463 * essentially lose the changes by pretending it
471 * If page couldn't be paged out, then reactivate the
472 * page so it doesn't clog the inactive list. (We
473 * will try paging out it again later).
475 vm_page_activate(mt);
476 if (eio != NULL && i >= mreq && i - mreq < runlen)
480 if (i >= mreq && i - mreq < runlen)
486 * If the operation is still going, leave the page busy to
487 * block all other accesses. Also, leave the paging in
488 * progress indicator set so that we don't attempt an object
491 if (pageout_status[i] != VM_PAGER_PEND) {
492 vm_object_pip_wakeup(object);
493 vm_page_io_finish(mt);
494 if (vm_page_count_severe())
495 vm_page_try_to_cache(mt);
500 return (numpagedout);
503 #if !defined(NO_SWAPPING)
505 * vm_pageout_object_deactivate_pages
507 * deactivate enough pages to satisfy the inactive target
508 * requirements or if vm_page_proc_limit is set, then
509 * deactivate all of the pages in the object and its
512 * The object and map must be locked.
515 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
517 vm_object_t first_object;
520 vm_object_t backing_object, object;
522 int actcount, rcount, remove_mode;
524 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
525 if (first_object->type == OBJT_DEVICE ||
526 first_object->type == OBJT_SG ||
527 first_object->type == OBJT_PHYS)
529 for (object = first_object;; object = backing_object) {
530 if (pmap_resident_count(pmap) <= desired)
532 if (object->paging_in_progress)
536 if (object->shadow_count > 1)
539 * scan the objects entire memory queue
541 rcount = object->resident_page_count;
542 p = TAILQ_FIRST(&object->memq);
543 vm_page_lock_queues();
544 while (p && (rcount-- > 0)) {
545 if (pmap_resident_count(pmap) <= desired) {
546 vm_page_unlock_queues();
549 next = TAILQ_NEXT(p, listq);
551 if (p->wire_count != 0 ||
552 p->hold_count != 0 ||
554 (p->oflags & VPO_BUSY) ||
555 (p->flags & PG_UNMANAGED) ||
556 !pmap_page_exists_quick(pmap, p)) {
560 actcount = pmap_ts_referenced(p);
562 vm_page_flag_set(p, PG_REFERENCED);
563 } else if (p->flags & PG_REFERENCED) {
566 if ((p->queue != PQ_ACTIVE) &&
567 (p->flags & PG_REFERENCED)) {
569 p->act_count += actcount;
570 vm_page_flag_clear(p, PG_REFERENCED);
571 } else if (p->queue == PQ_ACTIVE) {
572 if ((p->flags & PG_REFERENCED) == 0) {
573 p->act_count -= min(p->act_count, ACT_DECLINE);
574 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
576 vm_page_deactivate(p);
582 vm_page_flag_clear(p, PG_REFERENCED);
583 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
584 p->act_count += ACT_ADVANCE;
587 } else if (p->queue == PQ_INACTIVE) {
592 vm_page_unlock_queues();
593 if ((backing_object = object->backing_object) == NULL)
595 VM_OBJECT_LOCK(backing_object);
596 if (object != first_object)
597 VM_OBJECT_UNLOCK(object);
600 if (object != first_object)
601 VM_OBJECT_UNLOCK(object);
605 * deactivate some number of pages in a map, try to do it fairly, but
606 * that is really hard to do.
609 vm_pageout_map_deactivate_pages(map, desired)
614 vm_object_t obj, bigobj;
617 if (!vm_map_trylock(map))
624 * first, search out the biggest object, and try to free pages from
627 tmpe = map->header.next;
628 while (tmpe != &map->header) {
629 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
630 obj = tmpe->object.vm_object;
631 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
632 if (obj->shadow_count <= 1 &&
634 bigobj->resident_page_count < obj->resident_page_count)) {
636 VM_OBJECT_UNLOCK(bigobj);
639 VM_OBJECT_UNLOCK(obj);
642 if (tmpe->wired_count > 0)
643 nothingwired = FALSE;
647 if (bigobj != NULL) {
648 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
649 VM_OBJECT_UNLOCK(bigobj);
652 * Next, hunt around for other pages to deactivate. We actually
653 * do this search sort of wrong -- .text first is not the best idea.
655 tmpe = map->header.next;
656 while (tmpe != &map->header) {
657 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
659 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
660 obj = tmpe->object.vm_object;
663 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
664 VM_OBJECT_UNLOCK(obj);
671 * Remove all mappings if a process is swapped out, this will free page
674 if (desired == 0 && nothingwired) {
675 tmpe = map->header.next;
676 while (tmpe != &map->header) {
677 pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end);
683 #endif /* !defined(NO_SWAPPING) */
686 * vm_pageout_scan does the dirty work for the pageout daemon.
689 vm_pageout_scan(int pass)
692 struct vm_page marker;
693 int page_shortage, maxscan, pcount;
694 int addl_page_shortage, addl_page_shortage_init;
697 int vnodes_skipped = 0;
701 * Decrease registered cache sizes.
703 EVENTHANDLER_INVOKE(vm_lowmem, 0);
705 * We do this explicitly after the caches have been drained above.
709 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
712 * Calculate the number of pages we want to either free or move
715 page_shortage = vm_paging_target() + addl_page_shortage_init;
718 * Initialize our marker
720 bzero(&marker, sizeof(marker));
721 marker.flags = PG_FICTITIOUS | PG_MARKER;
722 marker.oflags = VPO_BUSY;
723 marker.queue = PQ_INACTIVE;
724 marker.wire_count = 1;
727 * Start scanning the inactive queue for pages we can move to the
728 * cache or free. The scan will stop when the target is reached or
729 * we have scanned the entire inactive queue. Note that m->act_count
730 * is not used to form decisions for the inactive queue, only for the
733 * maxlaunder limits the number of dirty pages we flush per scan.
734 * For most systems a smaller value (16 or 32) is more robust under
735 * extreme memory and disk pressure because any unnecessary writes
736 * to disk can result in extreme performance degredation. However,
737 * systems with excessive dirty pages (especially when MAP_NOSYNC is
738 * used) will die horribly with limited laundering. If the pageout
739 * daemon cannot clean enough pages in the first pass, we let it go
740 * all out in succeeding passes.
742 if ((maxlaunder = vm_max_launder) <= 1)
746 vm_page_lock_queues();
748 addl_page_shortage = addl_page_shortage_init;
749 maxscan = cnt.v_inactive_count;
751 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
752 m != NULL && maxscan-- > 0 && page_shortage > 0;
757 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
761 next = TAILQ_NEXT(m, pageq);
767 if (m->flags & PG_MARKER)
771 * A held page may be undergoing I/O, so skip it.
775 addl_page_shortage++;
779 * Don't mess with busy pages, keep in the front of the
780 * queue, most likely are being paged out.
782 if (!VM_OBJECT_TRYLOCK(object) &&
783 (!vm_pageout_fallback_object_lock(m, &next) ||
784 m->hold_count != 0)) {
785 VM_OBJECT_UNLOCK(object);
786 addl_page_shortage++;
789 if (m->busy || (m->oflags & VPO_BUSY)) {
790 VM_OBJECT_UNLOCK(object);
791 addl_page_shortage++;
796 * If the object is not being used, we ignore previous
799 if (object->ref_count == 0) {
800 vm_page_flag_clear(m, PG_REFERENCED);
801 KASSERT(!pmap_page_is_mapped(m),
802 ("vm_pageout_scan: page %p is mapped", m));
805 * Otherwise, if the page has been referenced while in the
806 * inactive queue, we bump the "activation count" upwards,
807 * making it less likely that the page will be added back to
808 * the inactive queue prematurely again. Here we check the
809 * page tables (or emulated bits, if any), given the upper
810 * level VM system not knowing anything about existing
813 } else if (((m->flags & PG_REFERENCED) == 0) &&
814 (actcount = pmap_ts_referenced(m))) {
816 VM_OBJECT_UNLOCK(object);
817 m->act_count += (actcount + ACT_ADVANCE);
822 * If the upper level VM system knows about any page
823 * references, we activate the page. We also set the
824 * "activation count" higher than normal so that we will less
825 * likely place pages back onto the inactive queue again.
827 if ((m->flags & PG_REFERENCED) != 0) {
828 vm_page_flag_clear(m, PG_REFERENCED);
829 actcount = pmap_ts_referenced(m);
831 VM_OBJECT_UNLOCK(object);
832 m->act_count += (actcount + ACT_ADVANCE + 1);
837 * If the upper level VM system does not believe that the page
838 * is fully dirty, but it is mapped for write access, then we
839 * consult the pmap to see if the page's dirty status should
842 if (m->dirty != VM_PAGE_BITS_ALL &&
843 (m->flags & PG_WRITEABLE) != 0) {
845 * Avoid a race condition: Unless write access is
846 * removed from the page, another processor could
847 * modify it before all access is removed by the call
848 * to vm_page_cache() below. If vm_page_cache() finds
849 * that the page has been modified when it removes all
850 * access, it panics because it cannot cache dirty
851 * pages. In principle, we could eliminate just write
852 * access here rather than all access. In the expected
853 * case, when there are no last instant modifications
854 * to the page, removing all access will be cheaper
857 if (pmap_is_modified(m))
859 else if (m->dirty == 0)
865 * Invalid pages can be easily freed
870 } else if (m->dirty == 0) {
872 * Clean pages can be placed onto the cache queue.
873 * This effectively frees them.
877 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
879 * Dirty pages need to be paged out, but flushing
880 * a page is extremely expensive verses freeing
881 * a clean page. Rather then artificially limiting
882 * the number of pages we can flush, we instead give
883 * dirty pages extra priority on the inactive queue
884 * by forcing them to be cycled through the queue
885 * twice before being flushed, after which the
886 * (now clean) page will cycle through once more
887 * before being freed. This significantly extends
888 * the thrash point for a heavily loaded machine.
890 vm_page_flag_set(m, PG_WINATCFLS);
892 } else if (maxlaunder > 0) {
894 * We always want to try to flush some dirty pages if
895 * we encounter them, to keep the system stable.
896 * Normally this number is small, but under extreme
897 * pressure where there are insufficient clean pages
898 * on the inactive queue, we may have to go all out.
900 int swap_pageouts_ok, vfslocked = 0;
901 struct vnode *vp = NULL;
902 struct mount *mp = NULL;
904 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
905 swap_pageouts_ok = 1;
907 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
908 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
909 vm_page_count_min());
914 * We don't bother paging objects that are "dead".
915 * Those objects are in a "rundown" state.
917 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
918 VM_OBJECT_UNLOCK(object);
924 * Following operations may unlock
925 * vm_page_queue_mtx, invalidating the 'next'
926 * pointer. To prevent an inordinate number
927 * of restarts we use our marker to remember
931 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
934 * The object is already known NOT to be dead. It
935 * is possible for the vget() to block the whole
936 * pageout daemon, but the new low-memory handling
937 * code should prevent it.
939 * The previous code skipped locked vnodes and, worse,
940 * reordered pages in the queue. This results in
941 * completely non-deterministic operation and, on a
942 * busy system, can lead to extremely non-optimal
943 * pageouts. For example, it can cause clean pages
944 * to be freed and dirty pages to be moved to the end
945 * of the queue. Since dirty pages are also moved to
946 * the end of the queue once-cleaned, this gives
947 * way too large a weighting to defering the freeing
950 * We can't wait forever for the vnode lock, we might
951 * deadlock due to a vn_read() getting stuck in
952 * vm_wait while holding this vnode. We skip the
953 * vnode if we can't get it in a reasonable amount
956 if (object->type == OBJT_VNODE) {
958 if (vp->v_type == VREG &&
959 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
962 if (object->flags & OBJ_MIGHTBEDIRTY)
964 goto unlock_and_continue;
967 ("vp %p with NULL v_mount", vp));
968 vm_page_unlock_queues();
969 vm_object_reference_locked(object);
970 VM_OBJECT_UNLOCK(object);
971 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
972 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
974 VM_OBJECT_LOCK(object);
975 vm_page_lock_queues();
977 if (object->flags & OBJ_MIGHTBEDIRTY)
980 goto unlock_and_continue;
982 VM_OBJECT_LOCK(object);
983 vm_page_lock_queues();
985 * The page might have been moved to another
986 * queue during potential blocking in vget()
987 * above. The page might have been freed and
988 * reused for another vnode.
990 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
991 m->object != object ||
992 TAILQ_NEXT(m, pageq) != &marker) {
993 if (object->flags & OBJ_MIGHTBEDIRTY)
995 goto unlock_and_continue;
999 * The page may have been busied during the
1000 * blocking in vget(). We don't move the
1001 * page back onto the end of the queue so that
1002 * statistics are more correct if we don't.
1004 if (m->busy || (m->oflags & VPO_BUSY)) {
1005 goto unlock_and_continue;
1009 * If the page has become held it might
1010 * be undergoing I/O, so skip it
1012 if (m->hold_count) {
1014 if (object->flags & OBJ_MIGHTBEDIRTY)
1016 goto unlock_and_continue;
1021 * If a page is dirty, then it is either being washed
1022 * (but not yet cleaned) or it is still in the
1023 * laundry. If it is still in the laundry, then we
1024 * start the cleaning operation.
1026 * decrement page_shortage on success to account for
1027 * the (future) cleaned page. Otherwise we could wind
1028 * up laundering or cleaning too many pages.
1030 if (vm_pageout_clean(m) != 0) {
1034 unlock_and_continue:
1035 VM_OBJECT_UNLOCK(object);
1037 vm_page_unlock_queues();
1040 VFS_UNLOCK_GIANT(vfslocked);
1041 vm_object_deallocate(object);
1042 vn_finished_write(mp);
1043 vm_page_lock_queues();
1045 next = TAILQ_NEXT(&marker, pageq);
1046 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1050 VM_OBJECT_UNLOCK(object);
1054 * Compute the number of pages we want to try to move from the
1055 * active queue to the inactive queue.
1057 page_shortage = vm_paging_target() +
1058 cnt.v_inactive_target - cnt.v_inactive_count;
1059 page_shortage += addl_page_shortage;
1062 * Scan the active queue for things we can deactivate. We nominally
1063 * track the per-page activity counter and use it to locate
1064 * deactivation candidates.
1066 pcount = cnt.v_active_count;
1067 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1069 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1071 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1072 ("vm_pageout_scan: page %p isn't active", m));
1074 next = TAILQ_NEXT(m, pageq);
1076 if ((m->flags & PG_MARKER) != 0) {
1080 if (!VM_OBJECT_TRYLOCK(object) &&
1081 !vm_pageout_fallback_object_lock(m, &next)) {
1082 VM_OBJECT_UNLOCK(object);
1088 * Don't deactivate pages that are busy.
1090 if ((m->busy != 0) ||
1091 (m->oflags & VPO_BUSY) ||
1092 (m->hold_count != 0)) {
1093 VM_OBJECT_UNLOCK(object);
1100 * The count for pagedaemon pages is done after checking the
1101 * page for eligibility...
1106 * Check to see "how much" the page has been used.
1109 if (object->ref_count != 0) {
1110 if (m->flags & PG_REFERENCED) {
1113 actcount += pmap_ts_referenced(m);
1115 m->act_count += ACT_ADVANCE + actcount;
1116 if (m->act_count > ACT_MAX)
1117 m->act_count = ACT_MAX;
1122 * Since we have "tested" this bit, we need to clear it now.
1124 vm_page_flag_clear(m, PG_REFERENCED);
1127 * Only if an object is currently being used, do we use the
1128 * page activation count stats.
1130 if (actcount && (object->ref_count != 0)) {
1133 m->act_count -= min(m->act_count, ACT_DECLINE);
1134 if (vm_pageout_algorithm ||
1135 object->ref_count == 0 ||
1136 m->act_count == 0) {
1138 if (object->ref_count == 0) {
1143 vm_page_deactivate(m);
1145 vm_page_deactivate(m);
1151 VM_OBJECT_UNLOCK(object);
1154 vm_page_unlock_queues();
1155 #if !defined(NO_SWAPPING)
1157 * Idle process swapout -- run once per second.
1159 if (vm_swap_idle_enabled) {
1161 if (time_second != lsec) {
1162 vm_req_vmdaemon(VM_SWAP_IDLE);
1169 * If we didn't get enough free pages, and we have skipped a vnode
1170 * in a writeable object, wakeup the sync daemon. And kick swapout
1171 * if we did not get enough free pages.
1173 if (vm_paging_target() > 0) {
1174 if (vnodes_skipped && vm_page_count_min())
1175 (void) speedup_syncer();
1176 #if !defined(NO_SWAPPING)
1177 if (vm_swap_enabled && vm_page_count_target())
1178 vm_req_vmdaemon(VM_SWAP_NORMAL);
1183 * If we are critically low on one of RAM or swap and low on
1184 * the other, kill the largest process. However, we avoid
1185 * doing this on the first pass in order to give ourselves a
1186 * chance to flush out dirty vnode-backed pages and to allow
1187 * active pages to be moved to the inactive queue and reclaimed.
1190 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1191 (swap_pager_full && vm_paging_target() > 0)))
1192 vm_pageout_oom(VM_OOM_MEM);
1197 vm_pageout_oom(int shortage)
1199 struct proc *p, *bigproc;
1200 vm_offset_t size, bigsize;
1205 * We keep the process bigproc locked once we find it to keep anyone
1206 * from messing with it; however, there is a possibility of
1207 * deadlock if process B is bigproc and one of it's child processes
1208 * attempts to propagate a signal to B while we are waiting for A's
1209 * lock while walking this list. To avoid this, we don't block on
1210 * the process lock but just skip a process if it is already locked.
1214 sx_slock(&allproc_lock);
1215 FOREACH_PROC_IN_SYSTEM(p) {
1218 if (PROC_TRYLOCK(p) == 0)
1221 * If this is a system, protected or killed process, skip it.
1223 if (p->p_state != PRS_NORMAL ||
1224 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1225 (p->p_pid == 1) || P_KILLED(p) ||
1226 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1231 * If the process is in a non-running type state,
1232 * don't touch it. Check all the threads individually.
1235 FOREACH_THREAD_IN_PROC(p, td) {
1237 if (!TD_ON_RUNQ(td) &&
1238 !TD_IS_RUNNING(td) &&
1239 !TD_IS_SLEEPING(td)) {
1251 * get the process size
1253 vm = vmspace_acquire_ref(p);
1258 if (!vm_map_trylock_read(&vm->vm_map)) {
1263 size = vmspace_swap_count(vm);
1264 vm_map_unlock_read(&vm->vm_map);
1265 if (shortage == VM_OOM_MEM)
1266 size += vmspace_resident_count(vm);
1269 * if the this process is bigger than the biggest one
1272 if (size > bigsize) {
1273 if (bigproc != NULL)
1274 PROC_UNLOCK(bigproc);
1280 sx_sunlock(&allproc_lock);
1281 if (bigproc != NULL) {
1282 killproc(bigproc, "out of swap space");
1283 sched_nice(bigproc, PRIO_MIN);
1284 PROC_UNLOCK(bigproc);
1285 wakeup(&cnt.v_free_count);
1290 * This routine tries to maintain the pseudo LRU active queue,
1291 * so that during long periods of time where there is no paging,
1292 * that some statistic accumulation still occurs. This code
1293 * helps the situation where paging just starts to occur.
1296 vm_pageout_page_stats()
1300 int pcount,tpcount; /* Number of pages to check */
1301 static int fullintervalcount = 0;
1304 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1306 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1307 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1309 if (page_shortage <= 0)
1312 pcount = cnt.v_active_count;
1313 fullintervalcount += vm_pageout_stats_interval;
1314 if (fullintervalcount < vm_pageout_full_stats_interval) {
1315 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1317 if (pcount > tpcount)
1320 fullintervalcount = 0;
1323 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1324 while ((m != NULL) && (pcount-- > 0)) {
1327 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1328 ("vm_pageout_page_stats: page %p isn't active", m));
1330 next = TAILQ_NEXT(m, pageq);
1333 if ((m->flags & PG_MARKER) != 0) {
1337 if (!VM_OBJECT_TRYLOCK(object) &&
1338 !vm_pageout_fallback_object_lock(m, &next)) {
1339 VM_OBJECT_UNLOCK(object);
1345 * Don't deactivate pages that are busy.
1347 if ((m->busy != 0) ||
1348 (m->oflags & VPO_BUSY) ||
1349 (m->hold_count != 0)) {
1350 VM_OBJECT_UNLOCK(object);
1357 if (m->flags & PG_REFERENCED) {
1358 vm_page_flag_clear(m, PG_REFERENCED);
1362 actcount += pmap_ts_referenced(m);
1364 m->act_count += ACT_ADVANCE + actcount;
1365 if (m->act_count > ACT_MAX)
1366 m->act_count = ACT_MAX;
1369 if (m->act_count == 0) {
1371 * We turn off page access, so that we have
1372 * more accurate RSS stats. We don't do this
1373 * in the normal page deactivation when the
1374 * system is loaded VM wise, because the
1375 * cost of the large number of page protect
1376 * operations would be higher than the value
1377 * of doing the operation.
1380 vm_page_deactivate(m);
1382 m->act_count -= min(m->act_count, ACT_DECLINE);
1386 VM_OBJECT_UNLOCK(object);
1392 * vm_pageout is the high level pageout daemon.
1400 * Initialize some paging parameters.
1402 cnt.v_interrupt_free_min = 2;
1403 if (cnt.v_page_count < 2000)
1404 vm_pageout_page_count = 8;
1407 * v_free_reserved needs to include enough for the largest
1408 * swap pager structures plus enough for any pv_entry structs
1411 if (cnt.v_page_count > 1024)
1412 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1415 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1416 cnt.v_interrupt_free_min;
1417 cnt.v_free_reserved = vm_pageout_page_count +
1418 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1419 cnt.v_free_severe = cnt.v_free_min / 2;
1420 cnt.v_free_min += cnt.v_free_reserved;
1421 cnt.v_free_severe += cnt.v_free_reserved;
1424 * v_free_target and v_cache_min control pageout hysteresis. Note
1425 * that these are more a measure of the VM cache queue hysteresis
1426 * then the VM free queue. Specifically, v_free_target is the
1427 * high water mark (free+cache pages).
1429 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1430 * low water mark, while v_free_min is the stop. v_cache_min must
1431 * be big enough to handle memory needs while the pageout daemon
1432 * is signalled and run to free more pages.
1434 if (cnt.v_free_count > 6144)
1435 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1437 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1439 if (cnt.v_free_count > 2048) {
1440 cnt.v_cache_min = cnt.v_free_target;
1441 cnt.v_cache_max = 2 * cnt.v_cache_min;
1442 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1444 cnt.v_cache_min = 0;
1445 cnt.v_cache_max = 0;
1446 cnt.v_inactive_target = cnt.v_free_count / 4;
1448 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1449 cnt.v_inactive_target = cnt.v_free_count / 3;
1451 /* XXX does not really belong here */
1452 if (vm_page_max_wired == 0)
1453 vm_page_max_wired = cnt.v_free_count / 3;
1455 if (vm_pageout_stats_max == 0)
1456 vm_pageout_stats_max = cnt.v_free_target;
1459 * Set interval in seconds for stats scan.
1461 if (vm_pageout_stats_interval == 0)
1462 vm_pageout_stats_interval = 5;
1463 if (vm_pageout_full_stats_interval == 0)
1464 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1466 swap_pager_swap_init();
1469 * The pageout daemon is never done, so loop forever.
1473 * If we have enough free memory, wakeup waiters. Do
1474 * not clear vm_pages_needed until we reach our target,
1475 * otherwise we may be woken up over and over again and
1476 * waste a lot of cpu.
1478 mtx_lock(&vm_page_queue_free_mtx);
1479 if (vm_pages_needed && !vm_page_count_min()) {
1480 if (!vm_paging_needed())
1481 vm_pages_needed = 0;
1482 wakeup(&cnt.v_free_count);
1484 if (vm_pages_needed) {
1486 * Still not done, take a second pass without waiting
1487 * (unlimited dirty cleaning), otherwise sleep a bit
1492 msleep(&vm_pages_needed,
1493 &vm_page_queue_free_mtx, PVM, "psleep",
1497 * Good enough, sleep & handle stats. Prime the pass
1504 error = msleep(&vm_pages_needed,
1505 &vm_page_queue_free_mtx, PVM, "psleep",
1506 vm_pageout_stats_interval * hz);
1507 if (error && !vm_pages_needed) {
1508 mtx_unlock(&vm_page_queue_free_mtx);
1510 vm_page_lock_queues();
1511 vm_pageout_page_stats();
1512 vm_page_unlock_queues();
1516 if (vm_pages_needed)
1518 mtx_unlock(&vm_page_queue_free_mtx);
1519 vm_pageout_scan(pass);
1524 * Unless the free page queue lock is held by the caller, this function
1525 * should be regarded as advisory. Specifically, the caller should
1526 * not msleep() on &cnt.v_free_count following this function unless
1527 * the free page queue lock is held until the msleep() is performed.
1533 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1534 vm_pages_needed = 1;
1535 wakeup(&vm_pages_needed);
1539 #if !defined(NO_SWAPPING)
1541 vm_req_vmdaemon(int req)
1543 static int lastrun = 0;
1545 mtx_lock(&vm_daemon_mtx);
1546 vm_pageout_req_swapout |= req;
1547 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1548 wakeup(&vm_daemon_needed);
1551 mtx_unlock(&vm_daemon_mtx);
1557 struct rlimit rsslim;
1561 int breakout, swapout_flags;
1564 mtx_lock(&vm_daemon_mtx);
1565 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1566 swapout_flags = vm_pageout_req_swapout;
1567 vm_pageout_req_swapout = 0;
1568 mtx_unlock(&vm_daemon_mtx);
1570 swapout_procs(swapout_flags);
1573 * scan the processes for exceeding their rlimits or if
1574 * process is swapped out -- deactivate pages
1576 sx_slock(&allproc_lock);
1577 FOREACH_PROC_IN_SYSTEM(p) {
1578 vm_pindex_t limit, size;
1581 * if this is a system process or if we have already
1582 * looked at this process, skip it.
1585 if (p->p_state != PRS_NORMAL ||
1586 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1591 * if the process is in a non-running type state,
1595 FOREACH_THREAD_IN_PROC(p, td) {
1597 if (!TD_ON_RUNQ(td) &&
1598 !TD_IS_RUNNING(td) &&
1599 !TD_IS_SLEEPING(td)) {
1613 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1615 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1618 * let processes that are swapped out really be
1619 * swapped out set the limit to nothing (will force a
1622 if ((p->p_flag & P_INMEM) == 0)
1623 limit = 0; /* XXX */
1624 vm = vmspace_acquire_ref(p);
1629 size = vmspace_resident_count(vm);
1630 if (limit >= 0 && size >= limit) {
1631 vm_pageout_map_deactivate_pages(
1632 &vm->vm_map, limit);
1636 sx_sunlock(&allproc_lock);
1639 #endif /* !defined(NO_SWAPPING) */