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 * Can't clean the page if it's busy or held.
297 if ((m->hold_count != 0) ||
298 ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
302 mc[vm_pageout_page_count] = m;
304 page_base = vm_pageout_page_count;
309 * Scan object for clusterable pages.
311 * We can cluster ONLY if: ->> the page is NOT
312 * clean, wired, busy, held, or mapped into a
313 * buffer, and one of the following:
314 * 1) The page is inactive, or a seldom used
317 * 2) we force the issue.
319 * During heavy mmap/modification loads the pageout
320 * daemon can really fragment the underlying file
321 * due to flushing pages out of order and not trying
322 * align the clusters (which leave sporatic out-of-order
323 * holes). To solve this problem we do the reverse scan
324 * first and attempt to align our cluster, then do a
325 * forward scan if room remains.
329 while (ib && pageout_count < vm_pageout_page_count) {
337 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
341 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
342 (p->oflags & VPO_BUSY) || p->busy) {
346 vm_page_test_dirty(p);
347 if ((p->dirty & p->valid) == 0 ||
348 p->queue != PQ_INACTIVE ||
349 p->wire_count != 0 || /* may be held by buf cache */
350 p->hold_count != 0) { /* may be undergoing I/O */
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_lookup(object, pindex + is)) == NULL)
371 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
372 (p->oflags & VPO_BUSY) || p->busy) {
375 vm_page_test_dirty(p);
376 if ((p->dirty & p->valid) == 0 ||
377 p->queue != PQ_INACTIVE ||
378 p->wire_count != 0 || /* may be held by buf cache */
379 p->hold_count != 0) { /* may be undergoing I/O */
382 mc[page_base + pageout_count] = p;
388 * If we exhausted our forward scan, continue with the reverse scan
389 * when possible, even past a page boundry. This catches boundry
392 if (ib && pageout_count < vm_pageout_page_count)
396 * we allow reads during pageouts...
398 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
402 * vm_pageout_flush() - launder the given pages
404 * The given pages are laundered. Note that we setup for the start of
405 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
406 * reference count all in here rather then in the parent. If we want
407 * the parent to do more sophisticated things we may have to change
411 vm_pageout_flush(vm_page_t *mc, int count, int flags)
413 vm_object_t object = mc[0]->object;
414 int pageout_status[count];
418 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
419 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
421 * Initiate I/O. Bump the vm_page_t->busy counter and
422 * mark the pages read-only.
424 * We do not have to fixup the clean/dirty bits here... we can
425 * allow the pager to do it after the I/O completes.
427 * NOTE! mc[i]->dirty may be partial or fragmented due to an
428 * edge case with file fragments.
430 for (i = 0; i < count; i++) {
431 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
432 ("vm_pageout_flush: partially invalid page %p index %d/%d",
434 vm_page_io_start(mc[i]);
435 pmap_remove_write(mc[i]);
437 vm_page_unlock_queues();
438 vm_object_pip_add(object, count);
440 vm_pager_put_pages(object, mc, count, flags, pageout_status);
442 vm_page_lock_queues();
443 for (i = 0; i < count; i++) {
444 vm_page_t mt = mc[i];
446 KASSERT((mt->flags & PG_WRITEABLE) == 0,
447 ("vm_pageout_flush: page %p is not write protected", mt));
448 switch (pageout_status[i]) {
455 * Page outside of range of object. Right now we
456 * essentially lose the changes by pretending it
459 pmap_clear_modify(mt);
465 * If page couldn't be paged out, then reactivate the
466 * page so it doesn't clog the inactive list. (We
467 * will try paging out it again later).
469 vm_page_activate(mt);
476 * If the operation is still going, leave the page busy to
477 * block all other accesses. Also, leave the paging in
478 * progress indicator set so that we don't attempt an object
481 if (pageout_status[i] != VM_PAGER_PEND) {
482 vm_object_pip_wakeup(object);
483 vm_page_io_finish(mt);
484 if (vm_page_count_severe())
485 vm_page_try_to_cache(mt);
491 #if !defined(NO_SWAPPING)
493 * vm_pageout_object_deactivate_pages
495 * deactivate enough pages to satisfy the inactive target
496 * requirements or if vm_page_proc_limit is set, then
497 * deactivate all of the pages in the object and its
500 * The object and map must be locked.
503 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
505 vm_object_t first_object;
508 vm_object_t backing_object, object;
510 int actcount, rcount, remove_mode;
512 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
513 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
515 for (object = first_object;; object = backing_object) {
516 if (pmap_resident_count(pmap) <= desired)
518 if (object->paging_in_progress)
522 if (object->shadow_count > 1)
525 * scan the objects entire memory queue
527 rcount = object->resident_page_count;
528 p = TAILQ_FIRST(&object->memq);
529 vm_page_lock_queues();
530 while (p && (rcount-- > 0)) {
531 if (pmap_resident_count(pmap) <= desired) {
532 vm_page_unlock_queues();
535 next = TAILQ_NEXT(p, listq);
537 if (p->wire_count != 0 ||
538 p->hold_count != 0 ||
540 (p->oflags & VPO_BUSY) ||
541 (p->flags & 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;
683 int vnodes_skipped = 0;
688 * Decrease registered cache sizes.
690 EVENTHANDLER_INVOKE(vm_lowmem, 0);
692 * We do this explicitly after the caches have been drained above.
696 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
699 * Calculate the number of pages we want to either free or move
702 page_shortage = vm_paging_target() + addl_page_shortage_init;
705 * Initialize our marker
707 bzero(&marker, sizeof(marker));
708 marker.flags = PG_FICTITIOUS | PG_MARKER;
709 marker.oflags = VPO_BUSY;
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->oflags & VPO_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 * Following operations may unlock
910 * vm_page_queue_mtx, invalidating the 'next'
911 * pointer. To prevent an inordinate number
912 * of restarts we use our marker to remember
916 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
919 * The object is already known NOT to be dead. It
920 * is possible for the vget() to block the whole
921 * pageout daemon, but the new low-memory handling
922 * code should prevent it.
924 * The previous code skipped locked vnodes and, worse,
925 * reordered pages in the queue. This results in
926 * completely non-deterministic operation and, on a
927 * busy system, can lead to extremely non-optimal
928 * pageouts. For example, it can cause clean pages
929 * to be freed and dirty pages to be moved to the end
930 * of the queue. Since dirty pages are also moved to
931 * the end of the queue once-cleaned, this gives
932 * way too large a weighting to defering the freeing
935 * We can't wait forever for the vnode lock, we might
936 * deadlock due to a vn_read() getting stuck in
937 * vm_wait while holding this vnode. We skip the
938 * vnode if we can't get it in a reasonable amount
941 if (object->type == OBJT_VNODE) {
944 if (vp->v_type == VREG &&
945 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
947 if (object->flags & OBJ_MIGHTBEDIRTY)
950 goto unlock_and_continue;
952 vm_page_unlock_queues();
954 VM_OBJECT_UNLOCK(object);
955 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK |
956 LK_TIMELOCK, curthread)) {
957 VM_OBJECT_LOCK(object);
958 vm_page_lock_queues();
960 vn_finished_write(mp);
961 if (object->flags & OBJ_MIGHTBEDIRTY)
964 goto unlock_and_continue;
966 VM_OBJECT_LOCK(object);
967 vm_page_lock_queues();
969 * The page might have been moved to another
970 * queue during potential blocking in vget()
971 * above. The page might have been freed and
972 * reused for another vnode. The object might
973 * have been reused for another vnode.
975 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
976 m->object != object ||
977 object->handle != vp ||
978 TAILQ_NEXT(m, pageq) != &marker) {
979 if (object->flags & OBJ_MIGHTBEDIRTY)
981 goto unlock_and_continue;
985 * The page may have been busied during the
986 * blocking in vput(); We don't move the
987 * page back onto the end of the queue so that
988 * statistics are more correct if we don't.
990 if (m->busy || (m->oflags & VPO_BUSY)) {
991 goto unlock_and_continue;
995 * If the page has become held it might
996 * be undergoing I/O, so skip it
1000 if (object->flags & OBJ_MIGHTBEDIRTY)
1002 goto unlock_and_continue;
1007 * If a page is dirty, then it is either being washed
1008 * (but not yet cleaned) or it is still in the
1009 * laundry. If it is still in the laundry, then we
1010 * start the cleaning operation.
1012 * decrement page_shortage on success to account for
1013 * the (future) cleaned page. Otherwise we could wind
1014 * up laundering or cleaning too many pages.
1016 if (vm_pageout_clean(m) != 0) {
1020 unlock_and_continue:
1021 VM_OBJECT_UNLOCK(object);
1023 vm_page_unlock_queues();
1025 vn_finished_write(mp);
1026 vm_page_lock_queues();
1028 next = TAILQ_NEXT(&marker, pageq);
1029 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1033 VM_OBJECT_UNLOCK(object);
1037 * Compute the number of pages we want to try to move from the
1038 * active queue to the inactive queue.
1040 page_shortage = vm_paging_target() +
1041 cnt.v_inactive_target - cnt.v_inactive_count;
1042 page_shortage += addl_page_shortage;
1045 * Scan the active queue for things we can deactivate. We nominally
1046 * track the per-page activity counter and use it to locate
1047 * deactivation candidates.
1049 pcount = cnt.v_active_count;
1050 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1052 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1054 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1055 ("vm_pageout_scan: page %p isn't active", m));
1057 next = TAILQ_NEXT(m, pageq);
1059 if ((m->flags & PG_MARKER) != 0) {
1063 if (!VM_OBJECT_TRYLOCK(object) &&
1064 !vm_pageout_fallback_object_lock(m, &next)) {
1065 VM_OBJECT_UNLOCK(object);
1071 * Don't deactivate pages that are busy.
1073 if ((m->busy != 0) ||
1074 (m->oflags & VPO_BUSY) ||
1075 (m->hold_count != 0)) {
1076 VM_OBJECT_UNLOCK(object);
1077 vm_pageq_requeue(m);
1083 * The count for pagedaemon pages is done after checking the
1084 * page for eligibility...
1089 * Check to see "how much" the page has been used.
1092 if (object->ref_count != 0) {
1093 if (m->flags & PG_REFERENCED) {
1096 actcount += pmap_ts_referenced(m);
1098 m->act_count += ACT_ADVANCE + actcount;
1099 if (m->act_count > ACT_MAX)
1100 m->act_count = ACT_MAX;
1105 * Since we have "tested" this bit, we need to clear it now.
1107 vm_page_flag_clear(m, PG_REFERENCED);
1110 * Only if an object is currently being used, do we use the
1111 * page activation count stats.
1113 if (actcount && (object->ref_count != 0)) {
1114 vm_pageq_requeue(m);
1116 m->act_count -= min(m->act_count, ACT_DECLINE);
1117 if (vm_pageout_algorithm ||
1118 object->ref_count == 0 ||
1119 m->act_count == 0) {
1121 if (object->ref_count == 0) {
1126 vm_page_deactivate(m);
1128 vm_page_deactivate(m);
1131 vm_pageq_requeue(m);
1134 VM_OBJECT_UNLOCK(object);
1139 * We try to maintain some *really* free pages, this allows interrupt
1140 * code to be guaranteed space. Since both cache and free queues
1141 * are considered basically 'free', moving pages from cache to free
1142 * does not effect other calculations.
1144 while (cnt.v_free_count < cnt.v_free_reserved) {
1145 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE].pl, pageq) {
1146 KASSERT(m->dirty == 0,
1147 ("Found dirty cache page %p", m));
1148 KASSERT(!pmap_page_is_mapped(m),
1149 ("Found mapped cache page %p", m));
1150 KASSERT((m->flags & PG_UNMANAGED) == 0,
1151 ("Found unmanaged cache page %p", m));
1152 KASSERT(m->wire_count == 0,
1153 ("Found wired cache page %p", m));
1154 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1156 KASSERT((m->oflags & VPO_BUSY) == 0 &&
1157 m->busy == 0, ("Found busy cache page %p",
1160 VM_OBJECT_UNLOCK(object);
1168 vm_page_unlock_queues();
1169 #if !defined(NO_SWAPPING)
1171 * Idle process swapout -- run once per second.
1173 if (vm_swap_idle_enabled) {
1175 if (time_second != lsec) {
1176 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1184 * If we didn't get enough free pages, and we have skipped a vnode
1185 * in a writeable object, wakeup the sync daemon. And kick swapout
1186 * if we did not get enough free pages.
1188 if (vm_paging_target() > 0) {
1189 if (vnodes_skipped && vm_page_count_min())
1190 (void) speedup_syncer();
1191 #if !defined(NO_SWAPPING)
1192 if (vm_swap_enabled && vm_page_count_target()) {
1194 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1200 * If we are critically low on one of RAM or swap and low on
1201 * the other, kill the largest process. However, we avoid
1202 * doing this on the first pass in order to give ourselves a
1203 * chance to flush out dirty vnode-backed pages and to allow
1204 * active pages to be moved to the inactive queue and reclaimed.
1206 * We keep the process bigproc locked once we find it to keep anyone
1207 * from messing with it; however, there is a possibility of
1208 * deadlock if process B is bigproc and one of it's child processes
1209 * attempts to propagate a signal to B while we are waiting for A's
1210 * lock while walking this list. To avoid this, we don't block on
1211 * the process lock but just skip a process if it is already locked.
1214 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1215 (swap_pager_full && vm_paging_target() > 0))) {
1218 sx_slock(&allproc_lock);
1219 FOREACH_PROC_IN_SYSTEM(p) {
1222 if (PROC_TRYLOCK(p) == 0)
1225 * If this is a system or protected process, skip it.
1227 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1228 (p->p_flag & P_PROTECTED) ||
1229 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1234 * If the process is in a non-running type state,
1235 * don't touch it. Check all the threads individually.
1239 FOREACH_THREAD_IN_PROC(p, td) {
1241 if (!TD_ON_RUNQ(td) &&
1242 !TD_IS_RUNNING(td) &&
1243 !TD_IS_SLEEPING(td)) {
1256 * get the process size
1258 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1262 size = vmspace_swap_count(p->p_vmspace);
1263 vm_map_unlock_read(&p->p_vmspace->vm_map);
1264 size += vmspace_resident_count(p->p_vmspace);
1266 * if the this process is bigger than the biggest one
1269 if (size > bigsize) {
1270 if (bigproc != NULL)
1271 PROC_UNLOCK(bigproc);
1277 sx_sunlock(&allproc_lock);
1278 if (bigproc != NULL) {
1279 killproc(bigproc, "out of swap space");
1280 PROC_SLOCK(bigproc);
1281 sched_nice(bigproc, PRIO_MIN);
1282 PROC_SUNLOCK(bigproc);
1283 PROC_UNLOCK(bigproc);
1284 wakeup(&cnt.v_free_count);
1291 * This routine tries to maintain the pseudo LRU active queue,
1292 * so that during long periods of time where there is no paging,
1293 * that some statistic accumulation still occurs. This code
1294 * helps the situation where paging just starts to occur.
1297 vm_pageout_page_stats()
1301 int pcount,tpcount; /* Number of pages to check */
1302 static int fullintervalcount = 0;
1305 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1307 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1308 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1310 if (page_shortage <= 0)
1313 pcount = cnt.v_active_count;
1314 fullintervalcount += vm_pageout_stats_interval;
1315 if (fullintervalcount < vm_pageout_full_stats_interval) {
1316 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_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);
1351 vm_pageq_requeue(m);
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;
1367 vm_pageq_requeue(m);
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);
1383 vm_pageq_requeue(m);
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)
1543 static int lastrun = 0;
1545 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1546 wakeup(&vm_daemon_needed);
1554 struct rlimit rsslim;
1561 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1562 if (vm_pageout_req_swapout) {
1563 swapout_procs(vm_pageout_req_swapout);
1564 vm_pageout_req_swapout = 0;
1567 * scan the processes for exceeding their rlimits or if
1568 * process is swapped out -- deactivate pages
1570 sx_slock(&allproc_lock);
1571 FOREACH_PROC_IN_SYSTEM(p) {
1572 vm_pindex_t limit, size;
1575 * if this is a system process or if we have already
1576 * looked at this process, skip it.
1579 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1584 * if the process is in a non-running type state,
1589 FOREACH_THREAD_IN_PROC(p, td) {
1591 if (!TD_ON_RUNQ(td) &&
1592 !TD_IS_RUNNING(td) &&
1593 !TD_IS_SLEEPING(td)) {
1608 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1610 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1613 * let processes that are swapped out really be
1614 * swapped out set the limit to nothing (will force a
1617 if ((p->p_sflag & PS_INMEM) == 0)
1618 limit = 0; /* XXX */
1621 size = vmspace_resident_count(p->p_vmspace);
1622 if (limit >= 0 && size >= limit) {
1623 vm_pageout_map_deactivate_pages(
1624 &p->p_vmspace->vm_map, limit);
1627 sx_sunlock(&allproc_lock);
1630 #endif /* !defined(NO_SWAPPING) */