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_BUSY | PG_FICTITIOUS | PG_MARKER;
240 marker.queue = m->queue;
241 marker.wire_count = 1;
246 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
248 vm_page_unlock_queues();
249 VM_OBJECT_LOCK(object);
250 vm_page_lock_queues();
252 /* Page queue might have changed. */
253 *next = TAILQ_NEXT(&marker, pageq);
254 unchanged = (m->queue == queue &&
255 m->object == object &&
256 &marker == TAILQ_NEXT(m, pageq));
257 TAILQ_REMOVE(&vm_page_queues[queue].pl,
265 * Clean the page and remove it from the laundry.
267 * We set the busy bit to cause potential page faults on this page to
268 * block. Note the careful timing, however, the busy bit isn't set till
269 * late and we cannot do anything that will mess with the page.
276 vm_page_t mc[2*vm_pageout_page_count];
278 int ib, is, page_base;
279 vm_pindex_t pindex = m->pindex;
281 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
282 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
285 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
286 * with the new swapper, but we could have serious problems paging
287 * out other object types if there is insufficient memory.
289 * Unfortunately, checking free memory here is far too late, so the
290 * check has been moved up a procedural level.
294 * Don't mess with the page if it's busy, held, or special
296 if ((m->hold_count != 0) ||
297 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
301 mc[vm_pageout_page_count] = m;
303 page_base = vm_pageout_page_count;
308 * Scan object for clusterable pages.
310 * We can cluster ONLY if: ->> the page is NOT
311 * clean, wired, busy, held, or mapped into a
312 * buffer, and one of the following:
313 * 1) The page is inactive, or a seldom used
316 * 2) we force the issue.
318 * During heavy mmap/modification loads the pageout
319 * daemon can really fragment the underlying file
320 * due to flushing pages out of order and not trying
321 * align the clusters (which leave sporatic out-of-order
322 * holes). To solve this problem we do the reverse scan
323 * first and attempt to align our cluster, then do a
324 * forward scan if room remains.
328 while (ib && pageout_count < vm_pageout_page_count) {
336 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
340 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
341 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
345 vm_page_test_dirty(p);
346 if ((p->dirty & p->valid) == 0 ||
347 p->queue != PQ_INACTIVE ||
348 p->wire_count != 0 || /* may be held by buf cache */
349 p->hold_count != 0) { /* may be undergoing I/O */
357 * alignment boundry, stop here and switch directions. Do
360 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
364 while (pageout_count < vm_pageout_page_count &&
365 pindex + is < object->size) {
368 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
370 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
371 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
374 vm_page_test_dirty(p);
375 if ((p->dirty & p->valid) == 0 ||
376 p->queue != PQ_INACTIVE ||
377 p->wire_count != 0 || /* may be held by buf cache */
378 p->hold_count != 0) { /* may be undergoing I/O */
381 mc[page_base + pageout_count] = p;
387 * If we exhausted our forward scan, continue with the reverse scan
388 * when possible, even past a page boundry. This catches boundry
391 if (ib && pageout_count < vm_pageout_page_count)
395 * we allow reads during pageouts...
397 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
401 * vm_pageout_flush() - launder the given pages
403 * The given pages are laundered. Note that we setup for the start of
404 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
405 * reference count all in here rather then in the parent. If we want
406 * the parent to do more sophisticated things we may have to change
410 vm_pageout_flush(vm_page_t *mc, int count, int flags)
412 vm_object_t object = mc[0]->object;
413 int pageout_status[count];
417 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
418 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
420 * Initiate I/O. Bump the vm_page_t->busy counter and
421 * mark the pages read-only.
423 * We do not have to fixup the clean/dirty bits here... we can
424 * allow the pager to do it after the I/O completes.
426 * NOTE! mc[i]->dirty may be partial or fragmented due to an
427 * edge case with file fragments.
429 for (i = 0; i < count; i++) {
430 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
431 ("vm_pageout_flush: partially invalid page %p index %d/%d",
433 vm_page_io_start(mc[i]);
434 pmap_page_protect(mc[i], VM_PROT_READ);
436 vm_page_unlock_queues();
437 vm_object_pip_add(object, count);
439 vm_pager_put_pages(object, mc, count,
440 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
443 vm_page_lock_queues();
444 for (i = 0; i < count; i++) {
445 vm_page_t mt = mc[i];
447 KASSERT((mt->flags & PG_WRITEABLE) == 0,
448 ("vm_pageout_flush: page %p is not write protected", mt));
449 switch (pageout_status[i]) {
456 * Page outside of range of object. Right now we
457 * essentially lose the changes by pretending it
460 pmap_clear_modify(mt);
466 * If page couldn't be paged out, then reactivate the
467 * page so it doesn't clog the inactive list. (We
468 * will try paging out it again later).
470 vm_page_activate(mt);
477 * If the operation is still going, leave the page busy to
478 * block all other accesses. Also, leave the paging in
479 * progress indicator set so that we don't attempt an object
482 if (pageout_status[i] != VM_PAGER_PEND) {
483 vm_object_pip_wakeup(object);
484 vm_page_io_finish(mt);
485 if (vm_page_count_severe())
486 vm_page_try_to_cache(mt);
492 #if !defined(NO_SWAPPING)
494 * vm_pageout_object_deactivate_pages
496 * deactivate enough pages to satisfy the inactive target
497 * requirements or if vm_page_proc_limit is set, then
498 * deactivate all of the pages in the object and its
501 * The object and map must be locked.
504 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
506 vm_object_t first_object;
509 vm_object_t backing_object, object;
511 int actcount, rcount, remove_mode;
513 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
514 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
516 for (object = first_object;; object = backing_object) {
517 if (pmap_resident_count(pmap) <= desired)
519 if (object->paging_in_progress)
523 if (object->shadow_count > 1)
526 * scan the objects entire memory queue
528 rcount = object->resident_page_count;
529 p = TAILQ_FIRST(&object->memq);
530 vm_page_lock_queues();
531 while (p && (rcount-- > 0)) {
532 if (pmap_resident_count(pmap) <= desired) {
533 vm_page_unlock_queues();
536 next = TAILQ_NEXT(p, listq);
538 if (p->wire_count != 0 ||
539 p->hold_count != 0 ||
541 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
542 !pmap_page_exists_quick(pmap, p)) {
546 actcount = pmap_ts_referenced(p);
548 vm_page_flag_set(p, PG_REFERENCED);
549 } else if (p->flags & PG_REFERENCED) {
552 if ((p->queue != PQ_ACTIVE) &&
553 (p->flags & PG_REFERENCED)) {
555 p->act_count += actcount;
556 vm_page_flag_clear(p, PG_REFERENCED);
557 } else if (p->queue == PQ_ACTIVE) {
558 if ((p->flags & PG_REFERENCED) == 0) {
559 p->act_count -= min(p->act_count, ACT_DECLINE);
560 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
562 vm_page_deactivate(p);
568 vm_page_flag_clear(p, PG_REFERENCED);
569 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
570 p->act_count += ACT_ADVANCE;
573 } else if (p->queue == PQ_INACTIVE) {
578 vm_page_unlock_queues();
579 if ((backing_object = object->backing_object) == NULL)
581 VM_OBJECT_LOCK(backing_object);
582 if (object != first_object)
583 VM_OBJECT_UNLOCK(object);
586 if (object != first_object)
587 VM_OBJECT_UNLOCK(object);
591 * deactivate some number of pages in a map, try to do it fairly, but
592 * that is really hard to do.
595 vm_pageout_map_deactivate_pages(map, desired)
600 vm_object_t obj, bigobj;
603 if (!vm_map_trylock(map))
610 * first, search out the biggest object, and try to free pages from
613 tmpe = map->header.next;
614 while (tmpe != &map->header) {
615 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
616 obj = tmpe->object.vm_object;
617 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
618 if (obj->shadow_count <= 1 &&
620 bigobj->resident_page_count < obj->resident_page_count)) {
622 VM_OBJECT_UNLOCK(bigobj);
625 VM_OBJECT_UNLOCK(obj);
628 if (tmpe->wired_count > 0)
629 nothingwired = FALSE;
633 if (bigobj != NULL) {
634 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
635 VM_OBJECT_UNLOCK(bigobj);
638 * Next, hunt around for other pages to deactivate. We actually
639 * do this search sort of wrong -- .text first is not the best idea.
641 tmpe = map->header.next;
642 while (tmpe != &map->header) {
643 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
645 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
646 obj = tmpe->object.vm_object;
649 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
650 VM_OBJECT_UNLOCK(obj);
657 * Remove all mappings if a process is swapped out, this will free page
660 if (desired == 0 && nothingwired) {
661 pmap_remove(vm_map_pmap(map), vm_map_min(map),
666 #endif /* !defined(NO_SWAPPING) */
669 * vm_pageout_scan does the dirty work for the pageout daemon.
672 vm_pageout_scan(int pass)
675 struct vm_page marker;
676 int page_shortage, maxscan, pcount;
677 int addl_page_shortage, addl_page_shortage_init;
678 struct proc *p, *bigproc;
680 vm_offset_t size, bigsize;
682 int actcount, cache_cur, cache_first_failure;
683 static int cache_last_free;
684 int vnodes_skipped = 0;
689 * Decrease registered cache sizes.
691 EVENTHANDLER_INVOKE(vm_lowmem, 0);
693 * We do this explicitly after the caches have been drained above.
697 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
700 * Calculate the number of pages we want to either free or move
703 page_shortage = vm_paging_target() + addl_page_shortage_init;
706 * Initialize our marker
708 bzero(&marker, sizeof(marker));
709 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
710 marker.queue = PQ_INACTIVE;
711 marker.wire_count = 1;
714 * Start scanning the inactive queue for pages we can move to the
715 * cache or free. The scan will stop when the target is reached or
716 * we have scanned the entire inactive queue. Note that m->act_count
717 * is not used to form decisions for the inactive queue, only for the
720 * maxlaunder limits the number of dirty pages we flush per scan.
721 * For most systems a smaller value (16 or 32) is more robust under
722 * extreme memory and disk pressure because any unnecessary writes
723 * to disk can result in extreme performance degredation. However,
724 * systems with excessive dirty pages (especially when MAP_NOSYNC is
725 * used) will die horribly with limited laundering. If the pageout
726 * daemon cannot clean enough pages in the first pass, we let it go
727 * all out in succeeding passes.
729 if ((maxlaunder = vm_max_launder) <= 1)
733 vm_page_lock_queues();
735 addl_page_shortage = addl_page_shortage_init;
736 maxscan = cnt.v_inactive_count;
738 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
739 m != NULL && maxscan-- > 0 && page_shortage > 0;
744 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
748 next = TAILQ_NEXT(m, pageq);
754 if (m->flags & PG_MARKER)
758 * A held page may be undergoing I/O, so skip it.
762 addl_page_shortage++;
766 * Don't mess with busy pages, keep in the front of the
767 * queue, most likely are being paged out.
769 if (!VM_OBJECT_TRYLOCK(object) &&
770 (!vm_pageout_fallback_object_lock(m, &next) ||
771 m->hold_count != 0)) {
772 VM_OBJECT_UNLOCK(object);
773 addl_page_shortage++;
776 if (m->busy || (m->flags & PG_BUSY)) {
777 VM_OBJECT_UNLOCK(object);
778 addl_page_shortage++;
783 * If the object is not being used, we ignore previous
786 if (object->ref_count == 0) {
787 vm_page_flag_clear(m, PG_REFERENCED);
788 pmap_clear_reference(m);
791 * Otherwise, if the page has been referenced while in the
792 * inactive queue, we bump the "activation count" upwards,
793 * making it less likely that the page will be added back to
794 * the inactive queue prematurely again. Here we check the
795 * page tables (or emulated bits, if any), given the upper
796 * level VM system not knowing anything about existing
799 } else if (((m->flags & PG_REFERENCED) == 0) &&
800 (actcount = pmap_ts_referenced(m))) {
802 VM_OBJECT_UNLOCK(object);
803 m->act_count += (actcount + ACT_ADVANCE);
808 * If the upper level VM system knows about any page
809 * references, we activate the page. We also set the
810 * "activation count" higher than normal so that we will less
811 * likely place pages back onto the inactive queue again.
813 if ((m->flags & PG_REFERENCED) != 0) {
814 vm_page_flag_clear(m, PG_REFERENCED);
815 actcount = pmap_ts_referenced(m);
817 VM_OBJECT_UNLOCK(object);
818 m->act_count += (actcount + ACT_ADVANCE + 1);
823 * If the upper level VM system doesn't know anything about
824 * the page being dirty, we have to check for it again. As
825 * far as the VM code knows, any partially dirty pages are
828 if (m->dirty == 0 && !pmap_is_modified(m)) {
830 * Avoid a race condition: Unless write access is
831 * removed from the page, another processor could
832 * modify it before all access is removed by the call
833 * to vm_page_cache() below. If vm_page_cache() finds
834 * that the page has been modified when it removes all
835 * access, it panics because it cannot cache dirty
836 * pages. In principle, we could eliminate just write
837 * access here rather than all access. In the expected
838 * case, when there are no last instant modifications
839 * to the page, removing all access will be cheaper
842 if ((m->flags & PG_WRITEABLE) != 0)
850 * Invalid pages can be easily freed
855 } else if (m->dirty == 0) {
857 * Clean pages can be placed onto the cache queue.
858 * This effectively frees them.
862 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
864 * Dirty pages need to be paged out, but flushing
865 * a page is extremely expensive verses freeing
866 * a clean page. Rather then artificially limiting
867 * the number of pages we can flush, we instead give
868 * dirty pages extra priority on the inactive queue
869 * by forcing them to be cycled through the queue
870 * twice before being flushed, after which the
871 * (now clean) page will cycle through once more
872 * before being freed. This significantly extends
873 * the thrash point for a heavily loaded machine.
875 vm_page_flag_set(m, PG_WINATCFLS);
877 } else if (maxlaunder > 0) {
879 * We always want to try to flush some dirty pages if
880 * we encounter them, to keep the system stable.
881 * Normally this number is small, but under extreme
882 * pressure where there are insufficient clean pages
883 * on the inactive queue, we may have to go all out.
885 int swap_pageouts_ok;
886 struct vnode *vp = NULL;
889 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
890 swap_pageouts_ok = 1;
892 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
893 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
894 vm_page_count_min());
899 * We don't bother paging objects that are "dead".
900 * Those objects are in a "rundown" state.
902 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
903 VM_OBJECT_UNLOCK(object);
909 * The object is already known NOT to be dead. It
910 * is possible for the vget() to block the whole
911 * pageout daemon, but the new low-memory handling
912 * code should prevent it.
914 * The previous code skipped locked vnodes and, worse,
915 * reordered pages in the queue. This results in
916 * completely non-deterministic operation and, on a
917 * busy system, can lead to extremely non-optimal
918 * pageouts. For example, it can cause clean pages
919 * to be freed and dirty pages to be moved to the end
920 * of the queue. Since dirty pages are also moved to
921 * the end of the queue once-cleaned, this gives
922 * way too large a weighting to defering the freeing
925 * We can't wait forever for the vnode lock, we might
926 * deadlock due to a vn_read() getting stuck in
927 * vm_wait while holding this vnode. We skip the
928 * vnode if we can't get it in a reasonable amount
931 if (object->type == OBJT_VNODE) {
934 if (vp->v_type == VREG)
935 vn_start_write(vp, &mp, V_NOWAIT);
936 vm_page_unlock_queues();
938 VM_OBJECT_UNLOCK(object);
939 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK |
940 LK_TIMELOCK, curthread)) {
941 VM_OBJECT_LOCK(object);
942 vm_page_lock_queues();
944 vn_finished_write(mp);
945 if (object->flags & OBJ_MIGHTBEDIRTY)
947 VM_OBJECT_UNLOCK(object);
950 VM_OBJECT_LOCK(object);
951 vm_page_lock_queues();
953 * The page might have been moved to another
954 * queue during potential blocking in vget()
955 * above. The page might have been freed and
956 * reused for another vnode. The object might
957 * have been reused for another vnode.
959 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
960 m->object != object ||
961 object->handle != vp) {
962 if (object->flags & OBJ_MIGHTBEDIRTY)
964 goto unlock_and_continue;
968 * The page may have been busied during the
969 * blocking in vput(); We don't move the
970 * page back onto the end of the queue so that
971 * statistics are more correct if we don't.
973 if (m->busy || (m->flags & PG_BUSY)) {
974 goto unlock_and_continue;
978 * If the page has become held it might
979 * be undergoing I/O, so skip it
983 if (object->flags & OBJ_MIGHTBEDIRTY)
985 goto unlock_and_continue;
990 * If a page is dirty, then it is either being washed
991 * (but not yet cleaned) or it is still in the
992 * laundry. If it is still in the laundry, then we
993 * start the cleaning operation.
995 * This operation may cluster, invalidating the 'next'
996 * pointer. To prevent an inordinate number of
997 * restarts we use our marker to remember our place.
999 * decrement page_shortage on success to account for
1000 * the (future) cleaned page. Otherwise we could wind
1001 * up laundering or cleaning too many pages.
1003 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1004 if (vm_pageout_clean(m) != 0) {
1008 next = TAILQ_NEXT(&marker, pageq);
1009 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1010 unlock_and_continue:
1011 VM_OBJECT_UNLOCK(object);
1013 vm_page_unlock_queues();
1015 vn_finished_write(mp);
1016 vm_page_lock_queues();
1020 VM_OBJECT_UNLOCK(object);
1024 * Compute the number of pages we want to try to move from the
1025 * active queue to the inactive queue.
1027 page_shortage = vm_paging_target() +
1028 cnt.v_inactive_target - cnt.v_inactive_count;
1029 page_shortage += addl_page_shortage;
1032 * Scan the active queue for things we can deactivate. We nominally
1033 * track the per-page activity counter and use it to locate
1034 * deactivation candidates.
1036 pcount = cnt.v_active_count;
1037 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1039 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1041 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1042 ("vm_pageout_scan: page %p isn't active", m));
1044 next = TAILQ_NEXT(m, pageq);
1046 if ((m->flags & PG_MARKER) != 0) {
1050 if (!VM_OBJECT_TRYLOCK(object) &&
1051 !vm_pageout_fallback_object_lock(m, &next)) {
1052 VM_OBJECT_UNLOCK(object);
1058 * Don't deactivate pages that are busy.
1060 if ((m->busy != 0) ||
1061 (m->flags & PG_BUSY) ||
1062 (m->hold_count != 0)) {
1063 VM_OBJECT_UNLOCK(object);
1064 vm_pageq_requeue(m);
1070 * The count for pagedaemon pages is done after checking the
1071 * page for eligibility...
1076 * Check to see "how much" the page has been used.
1079 if (object->ref_count != 0) {
1080 if (m->flags & PG_REFERENCED) {
1083 actcount += pmap_ts_referenced(m);
1085 m->act_count += ACT_ADVANCE + actcount;
1086 if (m->act_count > ACT_MAX)
1087 m->act_count = ACT_MAX;
1092 * Since we have "tested" this bit, we need to clear it now.
1094 vm_page_flag_clear(m, PG_REFERENCED);
1097 * Only if an object is currently being used, do we use the
1098 * page activation count stats.
1100 if (actcount && (object->ref_count != 0)) {
1101 vm_pageq_requeue(m);
1103 m->act_count -= min(m->act_count, ACT_DECLINE);
1104 if (vm_pageout_algorithm ||
1105 object->ref_count == 0 ||
1106 m->act_count == 0) {
1108 if (object->ref_count == 0) {
1113 vm_page_deactivate(m);
1115 vm_page_deactivate(m);
1118 vm_pageq_requeue(m);
1121 VM_OBJECT_UNLOCK(object);
1126 * We try to maintain some *really* free pages, this allows interrupt
1127 * code to be guaranteed space. Since both cache and free queues
1128 * are considered basically 'free', moving pages from cache to free
1129 * does not effect other calculations.
1131 cache_cur = cache_last_free;
1132 cache_first_failure = -1;
1133 while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur =
1134 (cache_cur + PQ_PRIME2) & PQ_COLORMASK) != cache_first_failure) {
1135 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl,
1137 KASSERT(m->dirty == 0,
1138 ("Found dirty cache page %p", m));
1139 KASSERT(!pmap_page_is_mapped(m),
1140 ("Found mapped cache page %p", m));
1141 KASSERT((m->flags & PG_UNMANAGED) == 0,
1142 ("Found unmanaged cache page %p", m));
1143 KASSERT(m->wire_count == 0,
1144 ("Found wired cache page %p", m));
1145 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1147 KASSERT((m->flags & PG_BUSY) == 0 &&
1148 m->busy == 0, ("Found busy cache page %p",
1151 VM_OBJECT_UNLOCK(object);
1153 cache_last_free = cache_cur;
1154 cache_first_failure = -1;
1158 if (m == NULL && cache_first_failure == -1)
1159 cache_first_failure = cache_cur;
1161 vm_page_unlock_queues();
1162 #if !defined(NO_SWAPPING)
1164 * Idle process swapout -- run once per second.
1166 if (vm_swap_idle_enabled) {
1168 if (time_second != lsec) {
1169 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1177 * If we didn't get enough free pages, and we have skipped a vnode
1178 * in a writeable object, wakeup the sync daemon. And kick swapout
1179 * if we did not get enough free pages.
1181 if (vm_paging_target() > 0) {
1182 if (vnodes_skipped && vm_page_count_min())
1183 (void) speedup_syncer();
1184 #if !defined(NO_SWAPPING)
1185 if (vm_swap_enabled && vm_page_count_target()) {
1187 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1193 * If we are critically low on one of RAM or swap and low on
1194 * the other, kill the largest process. However, we avoid
1195 * doing this on the first pass in order to give ourselves a
1196 * chance to flush out dirty vnode-backed pages and to allow
1197 * active pages to be moved to the inactive queue and reclaimed.
1199 * We keep the process bigproc locked once we find it to keep anyone
1200 * from messing with it; however, there is a possibility of
1201 * deadlock if process B is bigproc and one of it's child processes
1202 * attempts to propagate a signal to B while we are waiting for A's
1203 * lock while walking this list. To avoid this, we don't block on
1204 * the process lock but just skip a process if it is already locked.
1207 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1208 (swap_pager_full && vm_paging_target() > 0))) {
1211 sx_slock(&allproc_lock);
1212 FOREACH_PROC_IN_SYSTEM(p) {
1215 if (PROC_TRYLOCK(p) == 0)
1218 * If this is a system or protected process, skip it.
1220 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1221 (p->p_flag & P_PROTECTED) ||
1222 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1227 * If the process is in a non-running type state,
1228 * don't touch it. Check all the threads individually.
1230 mtx_lock_spin(&sched_lock);
1232 FOREACH_THREAD_IN_PROC(p, td) {
1233 if (!TD_ON_RUNQ(td) &&
1234 !TD_IS_RUNNING(td) &&
1235 !TD_IS_SLEEPING(td)) {
1241 mtx_unlock_spin(&sched_lock);
1245 mtx_unlock_spin(&sched_lock);
1247 * get the process size
1249 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1253 size = vmspace_swap_count(p->p_vmspace);
1254 vm_map_unlock_read(&p->p_vmspace->vm_map);
1255 size += vmspace_resident_count(p->p_vmspace);
1257 * if the this process is bigger than the biggest one
1260 if (size > bigsize) {
1261 if (bigproc != NULL)
1262 PROC_UNLOCK(bigproc);
1268 sx_sunlock(&allproc_lock);
1269 if (bigproc != NULL) {
1270 killproc(bigproc, "out of swap space");
1271 mtx_lock_spin(&sched_lock);
1272 sched_nice(bigproc, PRIO_MIN);
1273 mtx_unlock_spin(&sched_lock);
1274 PROC_UNLOCK(bigproc);
1275 wakeup(&cnt.v_free_count);
1282 * This routine tries to maintain the pseudo LRU active queue,
1283 * so that during long periods of time where there is no paging,
1284 * that some statistic accumulation still occurs. This code
1285 * helps the situation where paging just starts to occur.
1288 vm_pageout_page_stats()
1292 int pcount,tpcount; /* Number of pages to check */
1293 static int fullintervalcount = 0;
1296 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1298 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1299 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1301 if (page_shortage <= 0)
1304 pcount = cnt.v_active_count;
1305 fullintervalcount += vm_pageout_stats_interval;
1306 if (fullintervalcount < vm_pageout_full_stats_interval) {
1307 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1308 if (pcount > tpcount)
1311 fullintervalcount = 0;
1314 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1315 while ((m != NULL) && (pcount-- > 0)) {
1318 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1319 ("vm_pageout_page_stats: page %p isn't active", m));
1321 next = TAILQ_NEXT(m, pageq);
1324 if ((m->flags & PG_MARKER) != 0) {
1328 if (!VM_OBJECT_TRYLOCK(object) &&
1329 !vm_pageout_fallback_object_lock(m, &next)) {
1330 VM_OBJECT_UNLOCK(object);
1336 * Don't deactivate pages that are busy.
1338 if ((m->busy != 0) ||
1339 (m->flags & PG_BUSY) ||
1340 (m->hold_count != 0)) {
1341 VM_OBJECT_UNLOCK(object);
1342 vm_pageq_requeue(m);
1348 if (m->flags & PG_REFERENCED) {
1349 vm_page_flag_clear(m, PG_REFERENCED);
1353 actcount += pmap_ts_referenced(m);
1355 m->act_count += ACT_ADVANCE + actcount;
1356 if (m->act_count > ACT_MAX)
1357 m->act_count = ACT_MAX;
1358 vm_pageq_requeue(m);
1360 if (m->act_count == 0) {
1362 * We turn off page access, so that we have
1363 * more accurate RSS stats. We don't do this
1364 * in the normal page deactivation when the
1365 * system is loaded VM wise, because the
1366 * cost of the large number of page protect
1367 * operations would be higher than the value
1368 * of doing the operation.
1371 vm_page_deactivate(m);
1373 m->act_count -= min(m->act_count, ACT_DECLINE);
1374 vm_pageq_requeue(m);
1377 VM_OBJECT_UNLOCK(object);
1383 * vm_pageout is the high level pageout daemon.
1391 * Initialize some paging parameters.
1393 cnt.v_interrupt_free_min = 2;
1394 if (cnt.v_page_count < 2000)
1395 vm_pageout_page_count = 8;
1398 * v_free_reserved needs to include enough for the largest
1399 * swap pager structures plus enough for any pv_entry structs
1402 if (cnt.v_page_count > 1024)
1403 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1406 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1407 cnt.v_interrupt_free_min;
1408 cnt.v_free_reserved = vm_pageout_page_count +
1409 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_NUMCOLORS;
1410 cnt.v_free_severe = cnt.v_free_min / 2;
1411 cnt.v_free_min += cnt.v_free_reserved;
1412 cnt.v_free_severe += cnt.v_free_reserved;
1415 * v_free_target and v_cache_min control pageout hysteresis. Note
1416 * that these are more a measure of the VM cache queue hysteresis
1417 * then the VM free queue. Specifically, v_free_target is the
1418 * high water mark (free+cache pages).
1420 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1421 * low water mark, while v_free_min is the stop. v_cache_min must
1422 * be big enough to handle memory needs while the pageout daemon
1423 * is signalled and run to free more pages.
1425 if (cnt.v_free_count > 6144)
1426 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1428 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1430 if (cnt.v_free_count > 2048) {
1431 cnt.v_cache_min = cnt.v_free_target;
1432 cnt.v_cache_max = 2 * cnt.v_cache_min;
1433 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1435 cnt.v_cache_min = 0;
1436 cnt.v_cache_max = 0;
1437 cnt.v_inactive_target = cnt.v_free_count / 4;
1439 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1440 cnt.v_inactive_target = cnt.v_free_count / 3;
1442 /* XXX does not really belong here */
1443 if (vm_page_max_wired == 0)
1444 vm_page_max_wired = cnt.v_free_count / 3;
1446 if (vm_pageout_stats_max == 0)
1447 vm_pageout_stats_max = cnt.v_free_target;
1450 * Set interval in seconds for stats scan.
1452 if (vm_pageout_stats_interval == 0)
1453 vm_pageout_stats_interval = 5;
1454 if (vm_pageout_full_stats_interval == 0)
1455 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1457 swap_pager_swap_init();
1460 * The pageout daemon is never done, so loop forever.
1463 vm_page_lock_queues();
1465 * If we have enough free memory, wakeup waiters. Do
1466 * not clear vm_pages_needed until we reach our target,
1467 * otherwise we may be woken up over and over again and
1468 * waste a lot of cpu.
1470 if (vm_pages_needed && !vm_page_count_min()) {
1471 if (!vm_paging_needed())
1472 vm_pages_needed = 0;
1473 wakeup(&cnt.v_free_count);
1475 if (vm_pages_needed) {
1477 * Still not done, take a second pass without waiting
1478 * (unlimited dirty cleaning), otherwise sleep a bit
1483 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1487 * Good enough, sleep & handle stats. Prime the pass
1494 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1495 "psleep", vm_pageout_stats_interval * hz);
1496 if (error && !vm_pages_needed) {
1498 vm_pageout_page_stats();
1499 vm_page_unlock_queues();
1503 if (vm_pages_needed)
1505 vm_page_unlock_queues();
1506 vm_pageout_scan(pass);
1511 * Unless the page queue lock is held by the caller, this function
1512 * should be regarded as advisory. Specifically, the caller should
1513 * not msleep() on &cnt.v_free_count following this function unless
1514 * the page queue lock is held until the msleep() is performed.
1520 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1521 vm_pages_needed = 1;
1522 wakeup(&vm_pages_needed);
1526 #if !defined(NO_SWAPPING)
1530 static int lastrun = 0;
1532 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1533 wakeup(&vm_daemon_needed);
1541 struct rlimit rsslim;
1548 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1549 if (vm_pageout_req_swapout) {
1550 swapout_procs(vm_pageout_req_swapout);
1551 vm_pageout_req_swapout = 0;
1554 * scan the processes for exceeding their rlimits or if
1555 * process is swapped out -- deactivate pages
1557 sx_slock(&allproc_lock);
1558 LIST_FOREACH(p, &allproc, p_list) {
1559 vm_pindex_t limit, size;
1562 * if this is a system process or if we have already
1563 * looked at this process, skip it.
1566 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1571 * if the process is in a non-running type state,
1574 mtx_lock_spin(&sched_lock);
1576 FOREACH_THREAD_IN_PROC(p, td) {
1577 if (!TD_ON_RUNQ(td) &&
1578 !TD_IS_RUNNING(td) &&
1579 !TD_IS_SLEEPING(td)) {
1584 mtx_unlock_spin(&sched_lock);
1592 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1594 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1597 * let processes that are swapped out really be
1598 * swapped out set the limit to nothing (will force a
1601 if ((p->p_sflag & PS_INMEM) == 0)
1602 limit = 0; /* XXX */
1605 size = vmspace_resident_count(p->p_vmspace);
1606 if (limit >= 0 && size >= limit) {
1607 vm_pageout_map_deactivate_pages(
1608 &p->p_vmspace->vm_map, limit);
1611 sx_sunlock(&allproc_lock);
1614 #endif /* !defined(NO_SWAPPING) */