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/resourcevar.h>
89 #include <sys/sched.h>
90 #include <sys/signalvar.h>
91 #include <sys/vnode.h>
92 #include <sys/vmmeter.h>
94 #include <sys/sysctl.h>
97 #include <vm/vm_param.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_map.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/swap_pager.h>
104 #include <vm/vm_extern.h>
107 #include <machine/mutex.h>
110 * System initialization
113 /* the kernel process "vm_pageout"*/
114 static void vm_pageout(void);
115 static int vm_pageout_clean(vm_page_t);
116 static void vm_pageout_scan(int pass);
118 struct proc *pageproc;
120 static struct kproc_desc page_kp = {
125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
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;
149 static int vm_max_launder = 32;
150 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
151 static int vm_pageout_full_stats_interval = 0;
152 static int vm_pageout_algorithm=0;
153 static int defer_swap_pageouts=0;
154 static int disable_swap_pageouts=0;
156 #if defined(NO_SWAPPING)
157 static int vm_swap_enabled=0;
158 static int vm_swap_idle_enabled=0;
160 static int vm_swap_enabled=1;
161 static int vm_swap_idle_enabled=0;
164 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
165 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
167 SYSCTL_INT(_vm, OID_AUTO, max_launder,
168 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
171 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
173 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
174 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
176 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
177 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
179 #if defined(NO_SWAPPING)
180 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
181 CTLFLAG_RD, &vm_swap_enabled, 0, "");
182 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
183 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
186 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
188 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
191 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
192 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
194 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
195 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
197 static int pageout_lock_miss;
198 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
199 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
201 #define VM_PAGEOUT_PAGE_COUNT 16
202 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
204 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
206 #if !defined(NO_SWAPPING)
207 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
208 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
209 static void vm_req_vmdaemon(void);
211 static void vm_pageout_page_stats(void);
214 * vm_pageout_fallback_object_lock:
216 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
217 * known to have failed and page queue must be either PQ_ACTIVE or
218 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
219 * while locking the vm object. Use marker page to detect page queue
220 * changes and maintain notion of next page on page queue. Return
221 * TRUE if no changes were detected, FALSE otherwise. vm object is
224 * This function depends on both the lock portion of struct vm_object
225 * and normal struct vm_page being type stable.
228 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
230 struct vm_page marker;
236 * Initialize our marker
238 bzero(&marker, sizeof(marker));
239 marker.flags = PG_FICTITIOUS | PG_MARKER;
240 marker.oflags = VPO_BUSY;
241 marker.queue = m->queue;
242 marker.wire_count = 1;
247 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
249 vm_page_unlock_queues();
250 VM_OBJECT_LOCK(object);
251 vm_page_lock_queues();
253 /* Page queue might have changed. */
254 *next = TAILQ_NEXT(&marker, pageq);
255 unchanged = (m->queue == queue &&
256 m->object == object &&
257 &marker == TAILQ_NEXT(m, pageq));
258 TAILQ_REMOVE(&vm_page_queues[queue].pl,
266 * Clean the page and remove it from the laundry.
268 * We set the busy bit to cause potential page faults on this page to
269 * block. Note the careful timing, however, the busy bit isn't set till
270 * late and we cannot do anything that will mess with the page.
277 vm_page_t mc[2*vm_pageout_page_count];
279 int ib, is, page_base;
280 vm_pindex_t pindex = m->pindex;
282 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
283 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
286 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
287 * with the new swapper, but we could have serious problems paging
288 * out other object types if there is insufficient memory.
290 * Unfortunately, checking free memory here is far too late, so the
291 * check has been moved up a procedural level.
295 * Don't mess with the page if it's busy, held, or special
297 if ((m->hold_count != 0) ||
298 ((m->busy != 0) || (m->oflags & VPO_BUSY) ||
299 (m->flags & PG_UNMANAGED))) {
303 mc[vm_pageout_page_count] = m;
305 page_base = vm_pageout_page_count;
310 * Scan object for clusterable pages.
312 * We can cluster ONLY if: ->> the page is NOT
313 * clean, wired, busy, held, or mapped into a
314 * buffer, and one of the following:
315 * 1) The page is inactive, or a seldom used
318 * 2) we force the issue.
320 * During heavy mmap/modification loads the pageout
321 * daemon can really fragment the underlying file
322 * due to flushing pages out of order and not trying
323 * align the clusters (which leave sporatic out-of-order
324 * holes). To solve this problem we do the reverse scan
325 * first and attempt to align our cluster, then do a
326 * forward scan if room remains.
330 while (ib && pageout_count < vm_pageout_page_count) {
338 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
342 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
343 (p->oflags & VPO_BUSY) || p->busy ||
344 (p->flags & PG_UNMANAGED)) {
348 vm_page_test_dirty(p);
349 if ((p->dirty & p->valid) == 0 ||
350 p->queue != PQ_INACTIVE ||
351 p->wire_count != 0 || /* may be held by buf cache */
352 p->hold_count != 0) { /* may be undergoing I/O */
360 * alignment boundry, stop here and switch directions. Do
363 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
367 while (pageout_count < vm_pageout_page_count &&
368 pindex + is < object->size) {
371 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
373 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
374 (p->oflags & VPO_BUSY) || p->busy ||
375 (p->flags & PG_UNMANAGED)) {
378 vm_page_test_dirty(p);
379 if ((p->dirty & p->valid) == 0 ||
380 p->queue != PQ_INACTIVE ||
381 p->wire_count != 0 || /* may be held by buf cache */
382 p->hold_count != 0) { /* may be undergoing I/O */
385 mc[page_base + pageout_count] = p;
391 * If we exhausted our forward scan, continue with the reverse scan
392 * when possible, even past a page boundry. This catches boundry
395 if (ib && pageout_count < vm_pageout_page_count)
399 * we allow reads during pageouts...
401 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
405 * vm_pageout_flush() - launder the given pages
407 * The given pages are laundered. Note that we setup for the start of
408 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
409 * reference count all in here rather then in the parent. If we want
410 * the parent to do more sophisticated things we may have to change
414 vm_pageout_flush(vm_page_t *mc, int count, int flags)
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,
444 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
447 vm_page_lock_queues();
448 for (i = 0; i < count; i++) {
449 vm_page_t mt = mc[i];
451 KASSERT((mt->flags & PG_WRITEABLE) == 0,
452 ("vm_pageout_flush: page %p is not write protected", mt));
453 switch (pageout_status[i]) {
460 * Page outside of range of object. Right now we
461 * essentially lose the changes by pretending it
464 pmap_clear_modify(mt);
470 * If page couldn't be paged out, then reactivate the
471 * page so it doesn't clog the inactive list. (We
472 * will try paging out it again later).
474 vm_page_activate(mt);
481 * If the operation is still going, leave the page busy to
482 * block all other accesses. Also, leave the paging in
483 * progress indicator set so that we don't attempt an object
486 if (pageout_status[i] != VM_PAGER_PEND) {
487 vm_object_pip_wakeup(object);
488 vm_page_io_finish(mt);
489 if (vm_page_count_severe())
490 vm_page_try_to_cache(mt);
496 #if !defined(NO_SWAPPING)
498 * vm_pageout_object_deactivate_pages
500 * deactivate enough pages to satisfy the inactive target
501 * requirements or if vm_page_proc_limit is set, then
502 * deactivate all of the pages in the object and its
505 * The object and map must be locked.
508 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
510 vm_object_t first_object;
513 vm_object_t backing_object, object;
515 int actcount, rcount, remove_mode;
517 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
518 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
520 for (object = first_object;; object = backing_object) {
521 if (pmap_resident_count(pmap) <= desired)
523 if (object->paging_in_progress)
527 if (object->shadow_count > 1)
530 * scan the objects entire memory queue
532 rcount = object->resident_page_count;
533 p = TAILQ_FIRST(&object->memq);
534 vm_page_lock_queues();
535 while (p && (rcount-- > 0)) {
536 if (pmap_resident_count(pmap) <= desired) {
537 vm_page_unlock_queues();
540 next = TAILQ_NEXT(p, listq);
542 if (p->wire_count != 0 ||
543 p->hold_count != 0 ||
545 (p->oflags & VPO_BUSY) ||
546 (p->flags & PG_UNMANAGED) ||
547 !pmap_page_exists_quick(pmap, p)) {
551 actcount = pmap_ts_referenced(p);
553 vm_page_flag_set(p, PG_REFERENCED);
554 } else if (p->flags & PG_REFERENCED) {
557 if ((p->queue != PQ_ACTIVE) &&
558 (p->flags & PG_REFERENCED)) {
560 p->act_count += actcount;
561 vm_page_flag_clear(p, PG_REFERENCED);
562 } else if (p->queue == PQ_ACTIVE) {
563 if ((p->flags & PG_REFERENCED) == 0) {
564 p->act_count -= min(p->act_count, ACT_DECLINE);
565 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
567 vm_page_deactivate(p);
573 vm_page_flag_clear(p, PG_REFERENCED);
574 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
575 p->act_count += ACT_ADVANCE;
578 } else if (p->queue == PQ_INACTIVE) {
583 vm_page_unlock_queues();
584 if ((backing_object = object->backing_object) == NULL)
586 VM_OBJECT_LOCK(backing_object);
587 if (object != first_object)
588 VM_OBJECT_UNLOCK(object);
591 if (object != first_object)
592 VM_OBJECT_UNLOCK(object);
596 * deactivate some number of pages in a map, try to do it fairly, but
597 * that is really hard to do.
600 vm_pageout_map_deactivate_pages(map, desired)
605 vm_object_t obj, bigobj;
608 if (!vm_map_trylock(map))
615 * first, search out the biggest object, and try to free pages from
618 tmpe = map->header.next;
619 while (tmpe != &map->header) {
620 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
621 obj = tmpe->object.vm_object;
622 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
623 if (obj->shadow_count <= 1 &&
625 bigobj->resident_page_count < obj->resident_page_count)) {
627 VM_OBJECT_UNLOCK(bigobj);
630 VM_OBJECT_UNLOCK(obj);
633 if (tmpe->wired_count > 0)
634 nothingwired = FALSE;
638 if (bigobj != NULL) {
639 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
640 VM_OBJECT_UNLOCK(bigobj);
643 * Next, hunt around for other pages to deactivate. We actually
644 * do this search sort of wrong -- .text first is not the best idea.
646 tmpe = map->header.next;
647 while (tmpe != &map->header) {
648 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
650 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
651 obj = tmpe->object.vm_object;
654 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
655 VM_OBJECT_UNLOCK(obj);
662 * Remove all mappings if a process is swapped out, this will free page
665 if (desired == 0 && nothingwired) {
666 pmap_remove(vm_map_pmap(map), vm_map_min(map),
671 #endif /* !defined(NO_SWAPPING) */
674 * vm_pageout_scan does the dirty work for the pageout daemon.
677 vm_pageout_scan(int pass)
680 struct vm_page marker;
681 int page_shortage, maxscan, pcount;
682 int addl_page_shortage, addl_page_shortage_init;
683 struct proc *p, *bigproc;
685 vm_offset_t size, bigsize;
687 int actcount, cache_cur, cache_first_failure;
688 static int cache_last_free;
689 int vnodes_skipped = 0;
694 * Decrease registered cache sizes.
696 EVENTHANDLER_INVOKE(vm_lowmem, 0);
698 * We do this explicitly after the caches have been drained above.
702 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
705 * Calculate the number of pages we want to either free or move
708 page_shortage = vm_paging_target() + addl_page_shortage_init;
711 * Initialize our marker
713 bzero(&marker, sizeof(marker));
714 marker.flags = PG_FICTITIOUS | PG_MARKER;
715 marker.oflags = VPO_BUSY;
716 marker.queue = PQ_INACTIVE;
717 marker.wire_count = 1;
720 * Start scanning the inactive queue for pages we can move to the
721 * cache or free. The scan will stop when the target is reached or
722 * we have scanned the entire inactive queue. Note that m->act_count
723 * is not used to form decisions for the inactive queue, only for the
726 * maxlaunder limits the number of dirty pages we flush per scan.
727 * For most systems a smaller value (16 or 32) is more robust under
728 * extreme memory and disk pressure because any unnecessary writes
729 * to disk can result in extreme performance degredation. However,
730 * systems with excessive dirty pages (especially when MAP_NOSYNC is
731 * used) will die horribly with limited laundering. If the pageout
732 * daemon cannot clean enough pages in the first pass, we let it go
733 * all out in succeeding passes.
735 if ((maxlaunder = vm_max_launder) <= 1)
739 vm_page_lock_queues();
741 addl_page_shortage = addl_page_shortage_init;
742 maxscan = cnt.v_inactive_count;
744 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
745 m != NULL && maxscan-- > 0 && page_shortage > 0;
750 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
754 next = TAILQ_NEXT(m, pageq);
760 if (m->flags & PG_MARKER)
764 * A held page may be undergoing I/O, so skip it.
768 addl_page_shortage++;
772 * Don't mess with busy pages, keep in the front of the
773 * queue, most likely are being paged out.
775 if (!VM_OBJECT_TRYLOCK(object) &&
776 (!vm_pageout_fallback_object_lock(m, &next) ||
777 m->hold_count != 0)) {
778 VM_OBJECT_UNLOCK(object);
779 addl_page_shortage++;
782 if (m->busy || (m->oflags & VPO_BUSY)) {
783 VM_OBJECT_UNLOCK(object);
784 addl_page_shortage++;
789 * If the object is not being used, we ignore previous
792 if (object->ref_count == 0) {
793 vm_page_flag_clear(m, PG_REFERENCED);
794 pmap_clear_reference(m);
797 * Otherwise, if the page has been referenced while in the
798 * inactive queue, we bump the "activation count" upwards,
799 * making it less likely that the page will be added back to
800 * the inactive queue prematurely again. Here we check the
801 * page tables (or emulated bits, if any), given the upper
802 * level VM system not knowing anything about existing
805 } else if (((m->flags & PG_REFERENCED) == 0) &&
806 (actcount = pmap_ts_referenced(m))) {
808 VM_OBJECT_UNLOCK(object);
809 m->act_count += (actcount + ACT_ADVANCE);
814 * If the upper level VM system knows about any page
815 * references, we activate the page. We also set the
816 * "activation count" higher than normal so that we will less
817 * likely place pages back onto the inactive queue again.
819 if ((m->flags & PG_REFERENCED) != 0) {
820 vm_page_flag_clear(m, PG_REFERENCED);
821 actcount = pmap_ts_referenced(m);
823 VM_OBJECT_UNLOCK(object);
824 m->act_count += (actcount + ACT_ADVANCE + 1);
829 * If the upper level VM system doesn't know anything about
830 * the page being dirty, we have to check for it again. As
831 * far as the VM code knows, any partially dirty pages are
834 if (m->dirty == 0 && !pmap_is_modified(m)) {
836 * Avoid a race condition: Unless write access is
837 * removed from the page, another processor could
838 * modify it before all access is removed by the call
839 * to vm_page_cache() below. If vm_page_cache() finds
840 * that the page has been modified when it removes all
841 * access, it panics because it cannot cache dirty
842 * pages. In principle, we could eliminate just write
843 * access here rather than all access. In the expected
844 * case, when there are no last instant modifications
845 * to the page, removing all access will be cheaper
848 if ((m->flags & PG_WRITEABLE) != 0)
856 * Invalid pages can be easily freed
861 } else if (m->dirty == 0) {
863 * Clean pages can be placed onto the cache queue.
864 * This effectively frees them.
868 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
870 * Dirty pages need to be paged out, but flushing
871 * a page is extremely expensive verses freeing
872 * a clean page. Rather then artificially limiting
873 * the number of pages we can flush, we instead give
874 * dirty pages extra priority on the inactive queue
875 * by forcing them to be cycled through the queue
876 * twice before being flushed, after which the
877 * (now clean) page will cycle through once more
878 * before being freed. This significantly extends
879 * the thrash point for a heavily loaded machine.
881 vm_page_flag_set(m, PG_WINATCFLS);
883 } else if (maxlaunder > 0) {
885 * We always want to try to flush some dirty pages if
886 * we encounter them, to keep the system stable.
887 * Normally this number is small, but under extreme
888 * pressure where there are insufficient clean pages
889 * on the inactive queue, we may have to go all out.
891 int swap_pageouts_ok;
892 struct vnode *vp = NULL;
895 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
896 swap_pageouts_ok = 1;
898 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
899 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
900 vm_page_count_min());
905 * We don't bother paging objects that are "dead".
906 * Those objects are in a "rundown" state.
908 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
909 VM_OBJECT_UNLOCK(object);
915 * Following operations may unlock
916 * vm_page_queue_mtx, invalidating the 'next'
917 * pointer. To prevent an inordinate number
918 * of restarts we use our marker to remember
922 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
925 * The object is already known NOT to be dead. It
926 * is possible for the vget() to block the whole
927 * pageout daemon, but the new low-memory handling
928 * code should prevent it.
930 * The previous code skipped locked vnodes and, worse,
931 * reordered pages in the queue. This results in
932 * completely non-deterministic operation and, on a
933 * busy system, can lead to extremely non-optimal
934 * pageouts. For example, it can cause clean pages
935 * to be freed and dirty pages to be moved to the end
936 * of the queue. Since dirty pages are also moved to
937 * the end of the queue once-cleaned, this gives
938 * way too large a weighting to defering the freeing
941 * We can't wait forever for the vnode lock, we might
942 * deadlock due to a vn_read() getting stuck in
943 * vm_wait while holding this vnode. We skip the
944 * vnode if we can't get it in a reasonable amount
947 if (object->type == OBJT_VNODE) {
950 if (vp->v_type == VREG &&
951 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
953 if (object->flags & OBJ_MIGHTBEDIRTY)
956 goto unlock_and_continue;
958 vm_page_unlock_queues();
960 VM_OBJECT_UNLOCK(object);
961 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK |
962 LK_TIMELOCK, curthread)) {
963 VM_OBJECT_LOCK(object);
964 vm_page_lock_queues();
966 vn_finished_write(mp);
967 if (object->flags & OBJ_MIGHTBEDIRTY)
970 goto unlock_and_continue;
972 VM_OBJECT_LOCK(object);
973 vm_page_lock_queues();
975 * The page might have been moved to another
976 * queue during potential blocking in vget()
977 * above. The page might have been freed and
978 * reused for another vnode. The object might
979 * have been reused for another vnode.
981 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
982 m->object != object ||
983 object->handle != vp ||
984 TAILQ_NEXT(m, pageq) != &marker) {
985 if (object->flags & OBJ_MIGHTBEDIRTY)
987 goto unlock_and_continue;
991 * The page may have been busied during the
992 * blocking in vput(); We don't move the
993 * page back onto the end of the queue so that
994 * statistics are more correct if we don't.
996 if (m->busy || (m->oflags & VPO_BUSY)) {
997 goto unlock_and_continue;
1001 * If the page has become held it might
1002 * be undergoing I/O, so skip it
1004 if (m->hold_count) {
1005 vm_pageq_requeue(m);
1006 if (object->flags & OBJ_MIGHTBEDIRTY)
1008 goto unlock_and_continue;
1013 * If a page is dirty, then it is either being washed
1014 * (but not yet cleaned) or it is still in the
1015 * laundry. If it is still in the laundry, then we
1016 * start the cleaning operation.
1018 * decrement page_shortage on success to account for
1019 * the (future) cleaned page. Otherwise we could wind
1020 * up laundering or cleaning too many pages.
1022 if (vm_pageout_clean(m) != 0) {
1026 unlock_and_continue:
1027 VM_OBJECT_UNLOCK(object);
1029 vm_page_unlock_queues();
1031 vn_finished_write(mp);
1032 vm_page_lock_queues();
1034 next = TAILQ_NEXT(&marker, pageq);
1035 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1039 VM_OBJECT_UNLOCK(object);
1043 * Compute the number of pages we want to try to move from the
1044 * active queue to the inactive queue.
1046 page_shortage = vm_paging_target() +
1047 cnt.v_inactive_target - cnt.v_inactive_count;
1048 page_shortage += addl_page_shortage;
1051 * Scan the active queue for things we can deactivate. We nominally
1052 * track the per-page activity counter and use it to locate
1053 * deactivation candidates.
1055 pcount = cnt.v_active_count;
1056 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1058 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1060 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1061 ("vm_pageout_scan: page %p isn't active", m));
1063 next = TAILQ_NEXT(m, pageq);
1065 if ((m->flags & PG_MARKER) != 0) {
1069 if (!VM_OBJECT_TRYLOCK(object) &&
1070 !vm_pageout_fallback_object_lock(m, &next)) {
1071 VM_OBJECT_UNLOCK(object);
1077 * Don't deactivate pages that are busy.
1079 if ((m->busy != 0) ||
1080 (m->oflags & VPO_BUSY) ||
1081 (m->hold_count != 0)) {
1082 VM_OBJECT_UNLOCK(object);
1083 vm_pageq_requeue(m);
1089 * The count for pagedaemon pages is done after checking the
1090 * page for eligibility...
1095 * Check to see "how much" the page has been used.
1098 if (object->ref_count != 0) {
1099 if (m->flags & PG_REFERENCED) {
1102 actcount += pmap_ts_referenced(m);
1104 m->act_count += ACT_ADVANCE + actcount;
1105 if (m->act_count > ACT_MAX)
1106 m->act_count = ACT_MAX;
1111 * Since we have "tested" this bit, we need to clear it now.
1113 vm_page_flag_clear(m, PG_REFERENCED);
1116 * Only if an object is currently being used, do we use the
1117 * page activation count stats.
1119 if (actcount && (object->ref_count != 0)) {
1120 vm_pageq_requeue(m);
1122 m->act_count -= min(m->act_count, ACT_DECLINE);
1123 if (vm_pageout_algorithm ||
1124 object->ref_count == 0 ||
1125 m->act_count == 0) {
1127 if (object->ref_count == 0) {
1132 vm_page_deactivate(m);
1134 vm_page_deactivate(m);
1137 vm_pageq_requeue(m);
1140 VM_OBJECT_UNLOCK(object);
1145 * We try to maintain some *really* free pages, this allows interrupt
1146 * code to be guaranteed space. Since both cache and free queues
1147 * are considered basically 'free', moving pages from cache to free
1148 * does not effect other calculations.
1150 cache_cur = cache_last_free;
1151 cache_first_failure = -1;
1152 while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur =
1153 (cache_cur + PQ_PRIME2) & PQ_COLORMASK) != cache_first_failure) {
1154 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl,
1156 KASSERT(m->dirty == 0,
1157 ("Found dirty cache page %p", m));
1158 KASSERT(!pmap_page_is_mapped(m),
1159 ("Found mapped cache page %p", m));
1160 KASSERT((m->flags & PG_UNMANAGED) == 0,
1161 ("Found unmanaged cache page %p", m));
1162 KASSERT(m->wire_count == 0,
1163 ("Found wired cache page %p", m));
1164 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1166 KASSERT((m->oflags & VPO_BUSY) == 0 &&
1167 m->busy == 0, ("Found busy cache page %p",
1170 VM_OBJECT_UNLOCK(object);
1172 cache_last_free = cache_cur;
1173 cache_first_failure = -1;
1177 if (m == NULL && cache_first_failure == -1)
1178 cache_first_failure = cache_cur;
1180 vm_page_unlock_queues();
1181 #if !defined(NO_SWAPPING)
1183 * Idle process swapout -- run once per second.
1185 if (vm_swap_idle_enabled) {
1187 if (time_second != lsec) {
1188 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1196 * If we didn't get enough free pages, and we have skipped a vnode
1197 * in a writeable object, wakeup the sync daemon. And kick swapout
1198 * if we did not get enough free pages.
1200 if (vm_paging_target() > 0) {
1201 if (vnodes_skipped && vm_page_count_min())
1202 (void) speedup_syncer();
1203 #if !defined(NO_SWAPPING)
1204 if (vm_swap_enabled && vm_page_count_target()) {
1206 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1212 * If we are critically low on one of RAM or swap and low on
1213 * the other, kill the largest process. However, we avoid
1214 * doing this on the first pass in order to give ourselves a
1215 * chance to flush out dirty vnode-backed pages and to allow
1216 * active pages to be moved to the inactive queue and reclaimed.
1218 * We keep the process bigproc locked once we find it to keep anyone
1219 * from messing with it; however, there is a possibility of
1220 * deadlock if process B is bigproc and one of it's child processes
1221 * attempts to propagate a signal to B while we are waiting for A's
1222 * lock while walking this list. To avoid this, we don't block on
1223 * the process lock but just skip a process if it is already locked.
1226 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1227 (swap_pager_full && vm_paging_target() > 0))) {
1230 sx_slock(&allproc_lock);
1231 FOREACH_PROC_IN_SYSTEM(p) {
1234 if (PROC_TRYLOCK(p) == 0)
1237 * If this is a system or protected process, skip it.
1239 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1240 (p->p_flag & P_PROTECTED) ||
1241 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1246 * If the process is in a non-running type state,
1247 * don't touch it. Check all the threads individually.
1249 mtx_lock_spin(&sched_lock);
1251 FOREACH_THREAD_IN_PROC(p, td) {
1252 if (!TD_ON_RUNQ(td) &&
1253 !TD_IS_RUNNING(td) &&
1254 !TD_IS_SLEEPING(td)) {
1260 mtx_unlock_spin(&sched_lock);
1264 mtx_unlock_spin(&sched_lock);
1266 * get the process size
1268 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1272 size = vmspace_swap_count(p->p_vmspace);
1273 vm_map_unlock_read(&p->p_vmspace->vm_map);
1274 size += vmspace_resident_count(p->p_vmspace);
1276 * if the this process is bigger than the biggest one
1279 if (size > bigsize) {
1280 if (bigproc != NULL)
1281 PROC_UNLOCK(bigproc);
1287 sx_sunlock(&allproc_lock);
1288 if (bigproc != NULL) {
1289 killproc(bigproc, "out of swap space");
1290 mtx_lock_spin(&sched_lock);
1291 sched_nice(bigproc, PRIO_MIN);
1292 mtx_unlock_spin(&sched_lock);
1293 PROC_UNLOCK(bigproc);
1294 wakeup(&cnt.v_free_count);
1301 * This routine tries to maintain the pseudo LRU active queue,
1302 * so that during long periods of time where there is no paging,
1303 * that some statistic accumulation still occurs. This code
1304 * helps the situation where paging just starts to occur.
1307 vm_pageout_page_stats()
1311 int pcount,tpcount; /* Number of pages to check */
1312 static int fullintervalcount = 0;
1315 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1317 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1318 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1320 if (page_shortage <= 0)
1323 pcount = cnt.v_active_count;
1324 fullintervalcount += vm_pageout_stats_interval;
1325 if (fullintervalcount < vm_pageout_full_stats_interval) {
1326 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1327 if (pcount > tpcount)
1330 fullintervalcount = 0;
1333 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1334 while ((m != NULL) && (pcount-- > 0)) {
1337 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1338 ("vm_pageout_page_stats: page %p isn't active", m));
1340 next = TAILQ_NEXT(m, pageq);
1343 if ((m->flags & PG_MARKER) != 0) {
1347 if (!VM_OBJECT_TRYLOCK(object) &&
1348 !vm_pageout_fallback_object_lock(m, &next)) {
1349 VM_OBJECT_UNLOCK(object);
1355 * Don't deactivate pages that are busy.
1357 if ((m->busy != 0) ||
1358 (m->oflags & VPO_BUSY) ||
1359 (m->hold_count != 0)) {
1360 VM_OBJECT_UNLOCK(object);
1361 vm_pageq_requeue(m);
1367 if (m->flags & PG_REFERENCED) {
1368 vm_page_flag_clear(m, PG_REFERENCED);
1372 actcount += pmap_ts_referenced(m);
1374 m->act_count += ACT_ADVANCE + actcount;
1375 if (m->act_count > ACT_MAX)
1376 m->act_count = ACT_MAX;
1377 vm_pageq_requeue(m);
1379 if (m->act_count == 0) {
1381 * We turn off page access, so that we have
1382 * more accurate RSS stats. We don't do this
1383 * in the normal page deactivation when the
1384 * system is loaded VM wise, because the
1385 * cost of the large number of page protect
1386 * operations would be higher than the value
1387 * of doing the operation.
1390 vm_page_deactivate(m);
1392 m->act_count -= min(m->act_count, ACT_DECLINE);
1393 vm_pageq_requeue(m);
1396 VM_OBJECT_UNLOCK(object);
1402 * vm_pageout is the high level pageout daemon.
1410 * Initialize some paging parameters.
1412 cnt.v_interrupt_free_min = 2;
1413 if (cnt.v_page_count < 2000)
1414 vm_pageout_page_count = 8;
1417 * v_free_reserved needs to include enough for the largest
1418 * swap pager structures plus enough for any pv_entry structs
1421 if (cnt.v_page_count > 1024)
1422 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1425 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1426 cnt.v_interrupt_free_min;
1427 cnt.v_free_reserved = vm_pageout_page_count +
1428 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_NUMCOLORS;
1429 cnt.v_free_severe = cnt.v_free_min / 2;
1430 cnt.v_free_min += cnt.v_free_reserved;
1431 cnt.v_free_severe += cnt.v_free_reserved;
1434 * v_free_target and v_cache_min control pageout hysteresis. Note
1435 * that these are more a measure of the VM cache queue hysteresis
1436 * then the VM free queue. Specifically, v_free_target is the
1437 * high water mark (free+cache pages).
1439 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1440 * low water mark, while v_free_min is the stop. v_cache_min must
1441 * be big enough to handle memory needs while the pageout daemon
1442 * is signalled and run to free more pages.
1444 if (cnt.v_free_count > 6144)
1445 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1447 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1449 if (cnt.v_free_count > 2048) {
1450 cnt.v_cache_min = cnt.v_free_target;
1451 cnt.v_cache_max = 2 * cnt.v_cache_min;
1452 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1454 cnt.v_cache_min = 0;
1455 cnt.v_cache_max = 0;
1456 cnt.v_inactive_target = cnt.v_free_count / 4;
1458 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1459 cnt.v_inactive_target = cnt.v_free_count / 3;
1461 /* XXX does not really belong here */
1462 if (vm_page_max_wired == 0)
1463 vm_page_max_wired = cnt.v_free_count / 3;
1465 if (vm_pageout_stats_max == 0)
1466 vm_pageout_stats_max = cnt.v_free_target;
1469 * Set interval in seconds for stats scan.
1471 if (vm_pageout_stats_interval == 0)
1472 vm_pageout_stats_interval = 5;
1473 if (vm_pageout_full_stats_interval == 0)
1474 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1476 swap_pager_swap_init();
1479 * The pageout daemon is never done, so loop forever.
1482 vm_page_lock_queues();
1484 * If we have enough free memory, wakeup waiters. Do
1485 * not clear vm_pages_needed until we reach our target,
1486 * otherwise we may be woken up over and over again and
1487 * waste a lot of cpu.
1489 if (vm_pages_needed && !vm_page_count_min()) {
1490 if (!vm_paging_needed())
1491 vm_pages_needed = 0;
1492 wakeup(&cnt.v_free_count);
1494 if (vm_pages_needed) {
1496 * Still not done, take a second pass without waiting
1497 * (unlimited dirty cleaning), otherwise sleep a bit
1502 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1506 * Good enough, sleep & handle stats. Prime the pass
1513 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1514 "psleep", vm_pageout_stats_interval * hz);
1515 if (error && !vm_pages_needed) {
1517 vm_pageout_page_stats();
1518 vm_page_unlock_queues();
1522 if (vm_pages_needed)
1524 vm_page_unlock_queues();
1525 vm_pageout_scan(pass);
1530 * Unless the page queue lock is held by the caller, this function
1531 * should be regarded as advisory. Specifically, the caller should
1532 * not msleep() on &cnt.v_free_count following this function unless
1533 * the page queue lock is held until the msleep() is performed.
1539 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1540 vm_pages_needed = 1;
1541 wakeup(&vm_pages_needed);
1545 #if !defined(NO_SWAPPING)
1549 static int lastrun = 0;
1551 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1552 wakeup(&vm_daemon_needed);
1560 struct rlimit rsslim;
1567 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1568 if (vm_pageout_req_swapout) {
1569 swapout_procs(vm_pageout_req_swapout);
1570 vm_pageout_req_swapout = 0;
1573 * scan the processes for exceeding their rlimits or if
1574 * process is swapped out -- deactivate pages
1576 sx_slock(&allproc_lock);
1577 LIST_FOREACH(p, &allproc, p_list) {
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_flag & (P_SYSTEM | P_WEXIT)) {
1590 * if the process is in a non-running type state,
1593 mtx_lock_spin(&sched_lock);
1595 FOREACH_THREAD_IN_PROC(p, td) {
1596 if (!TD_ON_RUNQ(td) &&
1597 !TD_IS_RUNNING(td) &&
1598 !TD_IS_SLEEPING(td)) {
1603 mtx_unlock_spin(&sched_lock);
1611 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1613 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1616 * let processes that are swapped out really be
1617 * swapped out set the limit to nothing (will force a
1620 if ((p->p_sflag & PS_INMEM) == 0)
1621 limit = 0; /* XXX */
1624 size = vmspace_resident_count(p->p_vmspace);
1625 if (limit >= 0 && size >= limit) {
1626 vm_pageout_map_deactivate_pages(
1627 &p->p_vmspace->vm_map, limit);
1630 sx_sunlock(&allproc_lock);
1633 #endif /* !defined(NO_SWAPPING) */