2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * This code is derived from software contributed to Berkeley by
12 * The Mach Operating System project at Carnegie-Mellon University.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
50 * Permission to use, copy, modify and distribute this software and
51 * its documentation is hereby granted, provided that both the copyright
52 * notice and this permission notice appear in all copies of the
53 * software, derivative works or modified versions, and any portions
54 * thereof, and that both notices appear in supporting documentation.
56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
60 * Carnegie Mellon requests users of this software to return to
62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
63 * School of Computer Science
64 * Carnegie Mellon University
65 * Pittsburgh PA 15213-3890
67 * any improvements or extensions that they make and grant Carnegie the
68 * rights to redistribute these changes.
72 * The proverbial page-out daemon.
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
84 #include <sys/mutex.h>
86 #include <sys/kthread.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
95 #include <sys/sysctl.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
108 #include <machine/mutex.h>
111 * System initialization
114 /* the kernel process "vm_pageout"*/
115 static void vm_pageout(void);
116 static int vm_pageout_clean(vm_page_t);
117 static void vm_pageout_pmap_collect(void);
118 static void vm_pageout_scan(int pass);
120 struct proc *pageproc;
122 static struct kproc_desc page_kp = {
127 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
129 #if !defined(NO_SWAPPING)
130 /* the kernel process "vm_daemon"*/
131 static void vm_daemon(void);
132 static struct proc *vmproc;
134 static struct kproc_desc vm_kp = {
139 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
143 int vm_pages_needed; /* Event on which pageout daemon sleeps */
144 int vm_pageout_deficit; /* Estimated number of pages deficit */
145 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
147 #if !defined(NO_SWAPPING)
148 static int vm_pageout_req_swapout; /* XXX */
149 static int vm_daemon_needed;
150 static struct mtx vm_daemon_mtx;
151 /* Allow for use by vm_pageout before vm_daemon is initialized. */
152 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
154 static int vm_max_launder = 32;
155 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
156 static int vm_pageout_full_stats_interval = 0;
157 static int vm_pageout_algorithm=0;
158 static int defer_swap_pageouts=0;
159 static int disable_swap_pageouts=0;
161 #if defined(NO_SWAPPING)
162 static int vm_swap_enabled=0;
163 static int vm_swap_idle_enabled=0;
165 static int vm_swap_enabled=1;
166 static int vm_swap_idle_enabled=0;
169 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
170 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
172 SYSCTL_INT(_vm, OID_AUTO, max_launder,
173 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
176 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
178 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
179 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
181 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
182 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
184 #if defined(NO_SWAPPING)
185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
186 CTLFLAG_RD, &vm_swap_enabled, 0, "");
187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
188 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
190 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
191 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
192 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
193 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
196 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
197 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
199 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
200 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
202 static int pageout_lock_miss;
203 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
204 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
206 #define VM_PAGEOUT_PAGE_COUNT 16
207 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
209 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
210 SYSCTL_INT(_vm, OID_AUTO, max_wired,
211 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
213 #if !defined(NO_SWAPPING)
214 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
215 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
216 static void vm_req_vmdaemon(int req);
218 static void vm_pageout_page_stats(void);
221 * vm_pageout_fallback_object_lock:
223 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
224 * known to have failed and page queue must be either PQ_ACTIVE or
225 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
226 * while locking the vm object. Use marker page to detect page queue
227 * changes and maintain notion of next page on page queue. Return
228 * TRUE if no changes were detected, FALSE otherwise. vm object is
231 * This function depends on both the lock portion of struct vm_object
232 * and normal struct vm_page being type stable.
235 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
237 struct vm_page marker;
243 * Initialize our marker
245 bzero(&marker, sizeof(marker));
246 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
247 marker.queue = m->queue;
248 marker.wire_count = 1;
253 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
255 vm_page_unlock_queues();
256 VM_OBJECT_LOCK(object);
257 vm_page_lock_queues();
259 /* Page queue might have changed. */
260 *next = TAILQ_NEXT(&marker, pageq);
261 unchanged = (m->queue == queue &&
262 m->object == object &&
263 &marker == TAILQ_NEXT(m, pageq));
264 TAILQ_REMOVE(&vm_page_queues[queue].pl,
272 * Clean the page and remove it from the laundry.
274 * We set the busy bit to cause potential page faults on this page to
275 * block. Note the careful timing, however, the busy bit isn't set till
276 * late and we cannot do anything that will mess with the page.
283 vm_page_t mc[2*vm_pageout_page_count];
285 int ib, is, page_base;
286 vm_pindex_t pindex = m->pindex;
288 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
289 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
292 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
293 * with the new swapper, but we could have serious problems paging
294 * out other object types if there is insufficient memory.
296 * Unfortunately, checking free memory here is far too late, so the
297 * check has been moved up a procedural level.
301 * Don't mess with the page if it's busy, held, or special
303 if ((m->hold_count != 0) ||
304 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
308 mc[vm_pageout_page_count] = m;
310 page_base = vm_pageout_page_count;
315 * Scan object for clusterable pages.
317 * We can cluster ONLY if: ->> the page is NOT
318 * clean, wired, busy, held, or mapped into a
319 * buffer, and one of the following:
320 * 1) The page is inactive, or a seldom used
323 * 2) we force the issue.
325 * During heavy mmap/modification loads the pageout
326 * daemon can really fragment the underlying file
327 * due to flushing pages out of order and not trying
328 * align the clusters (which leave sporatic out-of-order
329 * holes). To solve this problem we do the reverse scan
330 * first and attempt to align our cluster, then do a
331 * forward scan if room remains.
335 while (ib && pageout_count < vm_pageout_page_count) {
343 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
347 if (((p->queue - p->pc) == PQ_CACHE) ||
348 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
352 vm_page_test_dirty(p);
353 if ((p->dirty & p->valid) == 0 ||
354 p->queue != PQ_INACTIVE ||
355 p->wire_count != 0 || /* may be held by buf cache */
356 p->hold_count != 0) { /* may be undergoing I/O */
364 * alignment boundry, stop here and switch directions. Do
367 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
371 while (pageout_count < vm_pageout_page_count &&
372 pindex + is < object->size) {
375 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
377 if (((p->queue - p->pc) == PQ_CACHE) ||
378 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
381 vm_page_test_dirty(p);
382 if ((p->dirty & p->valid) == 0 ||
383 p->queue != PQ_INACTIVE ||
384 p->wire_count != 0 || /* may be held by buf cache */
385 p->hold_count != 0) { /* may be undergoing I/O */
388 mc[page_base + pageout_count] = p;
394 * If we exhausted our forward scan, continue with the reverse scan
395 * when possible, even past a page boundry. This catches boundry
398 if (ib && pageout_count < vm_pageout_page_count)
402 * we allow reads during pageouts...
404 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
408 * vm_pageout_flush() - launder the given pages
410 * The given pages are laundered. Note that we setup for the start of
411 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
412 * reference count all in here rather then in the parent. If we want
413 * the parent to do more sophisticated things we may have to change
417 vm_pageout_flush(vm_page_t *mc, int count, int flags)
419 vm_object_t object = mc[0]->object;
420 int pageout_status[count];
424 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
425 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
427 * Initiate I/O. Bump the vm_page_t->busy counter and
428 * mark the pages read-only.
430 * We do not have to fixup the clean/dirty bits here... we can
431 * allow the pager to do it after the I/O completes.
433 * NOTE! mc[i]->dirty may be partial or fragmented due to an
434 * edge case with file fragments.
436 for (i = 0; i < count; i++) {
437 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
438 ("vm_pageout_flush: partially invalid page %p index %d/%d",
440 vm_page_io_start(mc[i]);
441 pmap_page_protect(mc[i], VM_PROT_READ);
443 vm_page_unlock_queues();
444 vm_object_pip_add(object, count);
446 vm_pager_put_pages(object, mc, count,
447 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
450 vm_page_lock_queues();
451 for (i = 0; i < count; i++) {
452 vm_page_t mt = mc[i];
454 KASSERT((mt->flags & PG_WRITEABLE) == 0,
455 ("vm_pageout_flush: page %p is not write protected", mt));
456 switch (pageout_status[i]) {
463 * Page outside of range of object. Right now we
464 * essentially lose the changes by pretending it
467 pmap_clear_modify(mt);
473 * If page couldn't be paged out, then reactivate the
474 * page so it doesn't clog the inactive list. (We
475 * will try paging out it again later).
477 vm_page_activate(mt);
484 * If the operation is still going, leave the page busy to
485 * block all other accesses. Also, leave the paging in
486 * progress indicator set so that we don't attempt an object
489 if (pageout_status[i] != VM_PAGER_PEND) {
490 vm_object_pip_wakeup(object);
491 vm_page_io_finish(mt);
492 if (vm_page_count_severe())
493 vm_page_try_to_cache(mt);
499 #if !defined(NO_SWAPPING)
501 * vm_pageout_object_deactivate_pages
503 * deactivate enough pages to satisfy the inactive target
504 * requirements or if vm_page_proc_limit is set, then
505 * deactivate all of the pages in the object and its
508 * The object and map must be locked.
511 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
513 vm_object_t first_object;
516 vm_object_t backing_object, object;
518 int actcount, rcount, remove_mode;
520 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
521 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
523 for (object = first_object;; object = backing_object) {
524 if (pmap_resident_count(pmap) <= desired)
526 if (object->paging_in_progress)
530 if (object->shadow_count > 1)
533 * scan the objects entire memory queue
535 rcount = object->resident_page_count;
536 p = TAILQ_FIRST(&object->memq);
537 vm_page_lock_queues();
538 while (p && (rcount-- > 0)) {
539 if (pmap_resident_count(pmap) <= desired) {
540 vm_page_unlock_queues();
543 next = TAILQ_NEXT(p, listq);
545 if (p->wire_count != 0 ||
546 p->hold_count != 0 ||
548 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
549 !pmap_page_exists_quick(pmap, p)) {
553 actcount = pmap_ts_referenced(p);
555 vm_page_flag_set(p, PG_REFERENCED);
556 } else if (p->flags & PG_REFERENCED) {
559 if ((p->queue != PQ_ACTIVE) &&
560 (p->flags & PG_REFERENCED)) {
562 p->act_count += actcount;
563 vm_page_flag_clear(p, PG_REFERENCED);
564 } else if (p->queue == PQ_ACTIVE) {
565 if ((p->flags & PG_REFERENCED) == 0) {
566 p->act_count -= min(p->act_count, ACT_DECLINE);
567 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
569 vm_page_deactivate(p);
575 vm_page_flag_clear(p, PG_REFERENCED);
576 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
577 p->act_count += ACT_ADVANCE;
580 } else if (p->queue == PQ_INACTIVE) {
585 vm_page_unlock_queues();
586 if ((backing_object = object->backing_object) == NULL)
588 VM_OBJECT_LOCK(backing_object);
589 if (object != first_object)
590 VM_OBJECT_UNLOCK(object);
593 if (object != first_object)
594 VM_OBJECT_UNLOCK(object);
598 * deactivate some number of pages in a map, try to do it fairly, but
599 * that is really hard to do.
602 vm_pageout_map_deactivate_pages(map, desired)
607 vm_object_t obj, bigobj;
610 if (!vm_map_trylock(map))
617 * first, search out the biggest object, and try to free pages from
620 tmpe = map->header.next;
621 while (tmpe != &map->header) {
622 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
623 obj = tmpe->object.vm_object;
624 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
625 if (obj->shadow_count <= 1 &&
627 bigobj->resident_page_count < obj->resident_page_count)) {
629 VM_OBJECT_UNLOCK(bigobj);
632 VM_OBJECT_UNLOCK(obj);
635 if (tmpe->wired_count > 0)
636 nothingwired = FALSE;
640 if (bigobj != NULL) {
641 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
642 VM_OBJECT_UNLOCK(bigobj);
645 * Next, hunt around for other pages to deactivate. We actually
646 * do this search sort of wrong -- .text first is not the best idea.
648 tmpe = map->header.next;
649 while (tmpe != &map->header) {
650 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
652 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
653 obj = tmpe->object.vm_object;
656 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
657 VM_OBJECT_UNLOCK(obj);
664 * Remove all mappings if a process is swapped out, this will free page
667 if (desired == 0 && nothingwired) {
668 pmap_remove(vm_map_pmap(map), vm_map_min(map),
673 #endif /* !defined(NO_SWAPPING) */
676 * This routine is very drastic, but can save the system
680 vm_pageout_pmap_collect(void)
684 static int warningdone;
686 if (pmap_pagedaemon_waken == 0)
688 if (warningdone < 5) {
689 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
692 vm_page_lock_queues();
693 for (i = 0; i < vm_page_array_size; i++) {
694 m = &vm_page_array[i];
695 if (m->wire_count || m->hold_count || m->busy ||
696 (m->flags & (PG_BUSY | PG_UNMANAGED)))
700 vm_page_unlock_queues();
701 pmap_pagedaemon_waken = 0;
705 * vm_pageout_scan does the dirty work for the pageout daemon.
708 vm_pageout_scan(int pass)
711 struct vm_page marker;
712 int page_shortage, maxscan, pcount;
713 int addl_page_shortage, addl_page_shortage_init;
714 struct proc *p, *bigproc;
716 vm_offset_t size, bigsize;
718 int actcount, cache_cur, cache_first_failure;
719 static int cache_last_free;
720 int vnodes_skipped = 0;
724 * Decrease registered cache sizes.
726 EVENTHANDLER_INVOKE(vm_lowmem, 0);
728 * We do this explicitly after the caches have been drained above.
732 * Do whatever cleanup that the pmap code can.
734 vm_pageout_pmap_collect();
736 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
739 * Calculate the number of pages we want to either free or move
742 page_shortage = vm_paging_target() + addl_page_shortage_init;
745 * Initialize our marker
747 bzero(&marker, sizeof(marker));
748 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
749 marker.queue = PQ_INACTIVE;
750 marker.wire_count = 1;
753 * Start scanning the inactive queue for pages we can move to the
754 * cache or free. The scan will stop when the target is reached or
755 * we have scanned the entire inactive queue. Note that m->act_count
756 * is not used to form decisions for the inactive queue, only for the
759 * maxlaunder limits the number of dirty pages we flush per scan.
760 * For most systems a smaller value (16 or 32) is more robust under
761 * extreme memory and disk pressure because any unnecessary writes
762 * to disk can result in extreme performance degredation. However,
763 * systems with excessive dirty pages (especially when MAP_NOSYNC is
764 * used) will die horribly with limited laundering. If the pageout
765 * daemon cannot clean enough pages in the first pass, we let it go
766 * all out in succeeding passes.
768 if ((maxlaunder = vm_max_launder) <= 1)
772 vm_page_lock_queues();
774 addl_page_shortage = addl_page_shortage_init;
775 maxscan = cnt.v_inactive_count;
777 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
778 m != NULL && maxscan-- > 0 && page_shortage > 0;
783 if (m->queue != PQ_INACTIVE) {
787 next = TAILQ_NEXT(m, pageq);
793 if (m->flags & PG_MARKER)
797 * A held page may be undergoing I/O, so skip it.
801 addl_page_shortage++;
805 * Don't mess with busy pages, keep in the front of the
806 * queue, most likely are being paged out.
808 if (!VM_OBJECT_TRYLOCK(object) &&
809 (!vm_pageout_fallback_object_lock(m, &next) ||
810 m->hold_count != 0)) {
811 VM_OBJECT_UNLOCK(object);
812 addl_page_shortage++;
815 if (m->busy || (m->flags & PG_BUSY)) {
816 VM_OBJECT_UNLOCK(object);
817 addl_page_shortage++;
822 * If the object is not being used, we ignore previous
825 if (object->ref_count == 0) {
826 vm_page_flag_clear(m, PG_REFERENCED);
827 pmap_clear_reference(m);
830 * Otherwise, if the page has been referenced while in the
831 * inactive queue, we bump the "activation count" upwards,
832 * making it less likely that the page will be added back to
833 * the inactive queue prematurely again. Here we check the
834 * page tables (or emulated bits, if any), given the upper
835 * level VM system not knowing anything about existing
838 } else if (((m->flags & PG_REFERENCED) == 0) &&
839 (actcount = pmap_ts_referenced(m))) {
841 VM_OBJECT_UNLOCK(object);
842 m->act_count += (actcount + ACT_ADVANCE);
847 * If the upper level VM system knows about any page
848 * references, we activate the page. We also set the
849 * "activation count" higher than normal so that we will less
850 * likely place pages back onto the inactive queue again.
852 if ((m->flags & PG_REFERENCED) != 0) {
853 vm_page_flag_clear(m, PG_REFERENCED);
854 actcount = pmap_ts_referenced(m);
856 VM_OBJECT_UNLOCK(object);
857 m->act_count += (actcount + ACT_ADVANCE + 1);
862 * If the upper level VM system doesn't know anything about
863 * the page being dirty, we have to check for it again. As
864 * far as the VM code knows, any partially dirty pages are
867 if (m->dirty == 0 && !pmap_is_modified(m)) {
869 * Avoid a race condition: Unless write access is
870 * removed from the page, another processor could
871 * modify it before all access is removed by the call
872 * to vm_page_cache() below. If vm_page_cache() finds
873 * that the page has been modified when it removes all
874 * access, it panics because it cannot cache dirty
875 * pages. In principle, we could eliminate just write
876 * access here rather than all access. In the expected
877 * case, when there are no last instant modifications
878 * to the page, removing all access will be cheaper
881 if ((m->flags & PG_WRITEABLE) != 0)
889 * Invalid pages can be easily freed
895 } else if (m->dirty == 0) {
897 * Clean pages can be placed onto the cache queue.
898 * This effectively frees them.
902 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
904 * Dirty pages need to be paged out, but flushing
905 * a page is extremely expensive verses freeing
906 * a clean page. Rather then artificially limiting
907 * the number of pages we can flush, we instead give
908 * dirty pages extra priority on the inactive queue
909 * by forcing them to be cycled through the queue
910 * twice before being flushed, after which the
911 * (now clean) page will cycle through once more
912 * before being freed. This significantly extends
913 * the thrash point for a heavily loaded machine.
915 vm_page_flag_set(m, PG_WINATCFLS);
917 } else if (maxlaunder > 0) {
919 * We always want to try to flush some dirty pages if
920 * we encounter them, to keep the system stable.
921 * Normally this number is small, but under extreme
922 * pressure where there are insufficient clean pages
923 * on the inactive queue, we may have to go all out.
925 int swap_pageouts_ok, vfslocked = 0;
926 struct vnode *vp = NULL;
927 struct mount *mp = NULL;
929 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
930 swap_pageouts_ok = 1;
932 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
933 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
934 vm_page_count_min());
939 * We don't bother paging objects that are "dead".
940 * Those objects are in a "rundown" state.
942 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
943 VM_OBJECT_UNLOCK(object);
949 * Following operations may unlock
950 * vm_page_queue_mtx, invalidating the 'next'
951 * pointer. To prevent an inordinate number
952 * of restarts we use our marker to remember
956 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
959 * The object is already known NOT to be dead. It
960 * is possible for the vget() to block the whole
961 * pageout daemon, but the new low-memory handling
962 * code should prevent it.
964 * The previous code skipped locked vnodes and, worse,
965 * reordered pages in the queue. This results in
966 * completely non-deterministic operation and, on a
967 * busy system, can lead to extremely non-optimal
968 * pageouts. For example, it can cause clean pages
969 * to be freed and dirty pages to be moved to the end
970 * of the queue. Since dirty pages are also moved to
971 * the end of the queue once-cleaned, this gives
972 * way too large a weighting to defering the freeing
975 * We can't wait forever for the vnode lock, we might
976 * deadlock due to a vn_read() getting stuck in
977 * vm_wait while holding this vnode. We skip the
978 * vnode if we can't get it in a reasonable amount
981 if (object->type == OBJT_VNODE) {
983 if (vp->v_type == VREG &&
984 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
986 ("vm_pageout_scan: mp != NULL"));
988 if (object->flags & OBJ_MIGHTBEDIRTY)
990 goto unlock_and_continue;
992 vm_page_unlock_queues();
993 vm_object_reference_locked(object);
994 VM_OBJECT_UNLOCK(object);
995 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
996 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
998 VM_OBJECT_LOCK(object);
999 vm_page_lock_queues();
1000 ++pageout_lock_miss;
1001 if (object->flags & OBJ_MIGHTBEDIRTY)
1004 goto unlock_and_continue;
1006 VM_OBJECT_LOCK(object);
1007 vm_page_lock_queues();
1009 * The page might have been moved to another
1010 * queue during potential blocking in vget()
1011 * above. The page might have been freed and
1012 * reused for another vnode.
1014 if (m->queue != PQ_INACTIVE ||
1015 m->object != object ||
1016 TAILQ_NEXT(m, pageq) != &marker) {
1017 if (object->flags & OBJ_MIGHTBEDIRTY)
1019 goto unlock_and_continue;
1023 * The page may have been busied during the
1024 * blocking in vget(). We don't move the
1025 * page back onto the end of the queue so that
1026 * statistics are more correct if we don't.
1028 if (m->busy || (m->flags & PG_BUSY)) {
1029 goto unlock_and_continue;
1033 * If the page has become held it might
1034 * be undergoing I/O, so skip it
1036 if (m->hold_count) {
1037 vm_pageq_requeue(m);
1038 if (object->flags & OBJ_MIGHTBEDIRTY)
1040 goto unlock_and_continue;
1045 * If a page is dirty, then it is either being washed
1046 * (but not yet cleaned) or it is still in the
1047 * laundry. If it is still in the laundry, then we
1048 * start the cleaning operation.
1050 * decrement page_shortage on success to account for
1051 * the (future) cleaned page. Otherwise we could wind
1052 * up laundering or cleaning too many pages.
1054 if (vm_pageout_clean(m) != 0) {
1058 unlock_and_continue:
1059 VM_OBJECT_UNLOCK(object);
1061 vm_page_unlock_queues();
1064 VFS_UNLOCK_GIANT(vfslocked);
1065 vm_object_deallocate(object);
1066 vn_finished_write(mp);
1067 vm_page_lock_queues();
1069 next = TAILQ_NEXT(&marker, pageq);
1070 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1074 VM_OBJECT_UNLOCK(object);
1078 * Compute the number of pages we want to try to move from the
1079 * active queue to the inactive queue.
1081 page_shortage = vm_paging_target() +
1082 cnt.v_inactive_target - cnt.v_inactive_count;
1083 page_shortage += addl_page_shortage;
1086 * Scan the active queue for things we can deactivate. We nominally
1087 * track the per-page activity counter and use it to locate
1088 * deactivation candidates.
1090 pcount = cnt.v_active_count;
1091 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1093 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1095 KASSERT(m->queue == PQ_ACTIVE,
1096 ("vm_pageout_scan: page %p isn't active", m));
1098 next = TAILQ_NEXT(m, pageq);
1100 if ((m->flags & PG_MARKER) != 0) {
1104 if (!VM_OBJECT_TRYLOCK(object) &&
1105 !vm_pageout_fallback_object_lock(m, &next)) {
1106 VM_OBJECT_UNLOCK(object);
1112 * Don't deactivate pages that are busy.
1114 if ((m->busy != 0) ||
1115 (m->flags & PG_BUSY) ||
1116 (m->hold_count != 0)) {
1117 VM_OBJECT_UNLOCK(object);
1118 vm_pageq_requeue(m);
1124 * The count for pagedaemon pages is done after checking the
1125 * page for eligibility...
1130 * Check to see "how much" the page has been used.
1133 if (object->ref_count != 0) {
1134 if (m->flags & PG_REFERENCED) {
1137 actcount += pmap_ts_referenced(m);
1139 m->act_count += ACT_ADVANCE + actcount;
1140 if (m->act_count > ACT_MAX)
1141 m->act_count = ACT_MAX;
1146 * Since we have "tested" this bit, we need to clear it now.
1148 vm_page_flag_clear(m, PG_REFERENCED);
1151 * Only if an object is currently being used, do we use the
1152 * page activation count stats.
1154 if (actcount && (object->ref_count != 0)) {
1155 vm_pageq_requeue(m);
1157 m->act_count -= min(m->act_count, ACT_DECLINE);
1158 if (vm_pageout_algorithm ||
1159 object->ref_count == 0 ||
1160 m->act_count == 0) {
1162 if (object->ref_count == 0) {
1167 vm_page_deactivate(m);
1169 vm_page_deactivate(m);
1172 vm_pageq_requeue(m);
1175 VM_OBJECT_UNLOCK(object);
1180 * We try to maintain some *really* free pages, this allows interrupt
1181 * code to be guaranteed space. Since both cache and free queues
1182 * are considered basically 'free', moving pages from cache to free
1183 * does not effect other calculations.
1185 cache_cur = cache_last_free;
1186 cache_first_failure = -1;
1187 while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur =
1188 (cache_cur + PQ_PRIME2) & PQ_L2_MASK) != cache_first_failure) {
1189 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl,
1191 KASSERT(m->dirty == 0,
1192 ("Found dirty cache page %p", m));
1193 KASSERT(!pmap_page_is_mapped(m),
1194 ("Found mapped cache page %p", m));
1195 KASSERT((m->flags & PG_UNMANAGED) == 0,
1196 ("Found unmanaged cache page %p", m));
1197 KASSERT(m->wire_count == 0,
1198 ("Found wired cache page %p", m));
1199 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1201 KASSERT((m->flags & PG_BUSY) == 0 &&
1202 m->busy == 0, ("Found busy cache page %p",
1205 VM_OBJECT_UNLOCK(object);
1207 cache_last_free = cache_cur;
1208 cache_first_failure = -1;
1212 if (m == NULL && cache_first_failure == -1)
1213 cache_first_failure = cache_cur;
1215 vm_page_unlock_queues();
1216 #if !defined(NO_SWAPPING)
1218 * Idle process swapout -- run once per second.
1220 if (vm_swap_idle_enabled) {
1222 if (time_second != lsec) {
1223 vm_req_vmdaemon(VM_SWAP_IDLE);
1230 * If we didn't get enough free pages, and we have skipped a vnode
1231 * in a writeable object, wakeup the sync daemon. And kick swapout
1232 * if we did not get enough free pages.
1234 if (vm_paging_target() > 0) {
1235 if (vnodes_skipped && vm_page_count_min())
1236 (void) speedup_syncer();
1237 #if !defined(NO_SWAPPING)
1238 if (vm_swap_enabled && vm_page_count_target())
1239 vm_req_vmdaemon(VM_SWAP_NORMAL);
1244 * If we are critically low on one of RAM or swap and low on
1245 * the other, kill the largest process. However, we avoid
1246 * doing this on the first pass in order to give ourselves a
1247 * chance to flush out dirty vnode-backed pages and to allow
1248 * active pages to be moved to the inactive queue and reclaimed.
1250 * We keep the process bigproc locked once we find it to keep anyone
1251 * from messing with it; however, there is a possibility of
1252 * deadlock if process B is bigproc and one of it's child processes
1253 * attempts to propagate a signal to B while we are waiting for A's
1254 * lock while walking this list. To avoid this, we don't block on
1255 * the process lock but just skip a process if it is already locked.
1258 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1259 (swap_pager_full && vm_paging_target() > 0))) {
1262 sx_slock(&allproc_lock);
1263 FOREACH_PROC_IN_SYSTEM(p) {
1266 if (PROC_TRYLOCK(p) == 0)
1269 * If this is a system or protected process, skip it.
1271 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1272 (p->p_flag & P_PROTECTED) ||
1273 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1278 * If the process is in a non-running type state,
1279 * don't touch it. Check all the threads individually.
1281 mtx_lock_spin(&sched_lock);
1283 FOREACH_THREAD_IN_PROC(p, td) {
1284 if (!TD_ON_RUNQ(td) &&
1285 !TD_IS_RUNNING(td) &&
1286 !TD_IS_SLEEPING(td)) {
1292 mtx_unlock_spin(&sched_lock);
1296 mtx_unlock_spin(&sched_lock);
1298 * get the process size
1300 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1304 size = vmspace_swap_count(p->p_vmspace);
1305 vm_map_unlock_read(&p->p_vmspace->vm_map);
1306 size += vmspace_resident_count(p->p_vmspace);
1308 * if the this process is bigger than the biggest one
1311 if (size > bigsize) {
1312 if (bigproc != NULL)
1313 PROC_UNLOCK(bigproc);
1319 sx_sunlock(&allproc_lock);
1320 if (bigproc != NULL) {
1321 killproc(bigproc, "out of swap space");
1322 mtx_lock_spin(&sched_lock);
1323 sched_nice(bigproc, PRIO_MIN);
1324 mtx_unlock_spin(&sched_lock);
1325 PROC_UNLOCK(bigproc);
1326 wakeup(&cnt.v_free_count);
1332 * This routine tries to maintain the pseudo LRU active queue,
1333 * so that during long periods of time where there is no paging,
1334 * that some statistic accumulation still occurs. This code
1335 * helps the situation where paging just starts to occur.
1338 vm_pageout_page_stats()
1342 int pcount,tpcount; /* Number of pages to check */
1343 static int fullintervalcount = 0;
1346 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1348 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1349 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1351 if (page_shortage <= 0)
1354 pcount = cnt.v_active_count;
1355 fullintervalcount += vm_pageout_stats_interval;
1356 if (fullintervalcount < vm_pageout_full_stats_interval) {
1357 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1359 if (pcount > tpcount)
1362 fullintervalcount = 0;
1365 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1366 while ((m != NULL) && (pcount-- > 0)) {
1369 KASSERT(m->queue == PQ_ACTIVE,
1370 ("vm_pageout_page_stats: page %p isn't active", m));
1372 next = TAILQ_NEXT(m, pageq);
1375 if ((m->flags & PG_MARKER) != 0) {
1379 if (!VM_OBJECT_TRYLOCK(object) &&
1380 !vm_pageout_fallback_object_lock(m, &next)) {
1381 VM_OBJECT_UNLOCK(object);
1387 * Don't deactivate pages that are busy.
1389 if ((m->busy != 0) ||
1390 (m->flags & PG_BUSY) ||
1391 (m->hold_count != 0)) {
1392 VM_OBJECT_UNLOCK(object);
1393 vm_pageq_requeue(m);
1399 if (m->flags & PG_REFERENCED) {
1400 vm_page_flag_clear(m, PG_REFERENCED);
1404 actcount += pmap_ts_referenced(m);
1406 m->act_count += ACT_ADVANCE + actcount;
1407 if (m->act_count > ACT_MAX)
1408 m->act_count = ACT_MAX;
1409 vm_pageq_requeue(m);
1411 if (m->act_count == 0) {
1413 * We turn off page access, so that we have
1414 * more accurate RSS stats. We don't do this
1415 * in the normal page deactivation when the
1416 * system is loaded VM wise, because the
1417 * cost of the large number of page protect
1418 * operations would be higher than the value
1419 * of doing the operation.
1422 vm_page_deactivate(m);
1424 m->act_count -= min(m->act_count, ACT_DECLINE);
1425 vm_pageq_requeue(m);
1428 VM_OBJECT_UNLOCK(object);
1434 * vm_pageout is the high level pageout daemon.
1442 * Initialize some paging parameters.
1444 cnt.v_interrupt_free_min = 2;
1445 if (cnt.v_page_count < 2000)
1446 vm_pageout_page_count = 8;
1449 * v_free_reserved needs to include enough for the largest
1450 * swap pager structures plus enough for any pv_entry structs
1453 if (cnt.v_page_count > 1024)
1454 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1457 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1458 cnt.v_interrupt_free_min;
1459 cnt.v_free_reserved = vm_pageout_page_count +
1460 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_L2_SIZE;
1461 cnt.v_free_severe = cnt.v_free_min / 2;
1462 cnt.v_free_min += cnt.v_free_reserved;
1463 cnt.v_free_severe += cnt.v_free_reserved;
1466 * v_free_target and v_cache_min control pageout hysteresis. Note
1467 * that these are more a measure of the VM cache queue hysteresis
1468 * then the VM free queue. Specifically, v_free_target is the
1469 * high water mark (free+cache pages).
1471 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1472 * low water mark, while v_free_min is the stop. v_cache_min must
1473 * be big enough to handle memory needs while the pageout daemon
1474 * is signalled and run to free more pages.
1476 if (cnt.v_free_count > 6144)
1477 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1479 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1481 if (cnt.v_free_count > 2048) {
1482 cnt.v_cache_min = cnt.v_free_target;
1483 cnt.v_cache_max = 2 * cnt.v_cache_min;
1484 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1486 cnt.v_cache_min = 0;
1487 cnt.v_cache_max = 0;
1488 cnt.v_inactive_target = cnt.v_free_count / 4;
1490 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1491 cnt.v_inactive_target = cnt.v_free_count / 3;
1493 /* XXX does not really belong here */
1494 if (vm_page_max_wired == 0)
1495 vm_page_max_wired = cnt.v_free_count / 3;
1497 if (vm_pageout_stats_max == 0)
1498 vm_pageout_stats_max = cnt.v_free_target;
1501 * Set interval in seconds for stats scan.
1503 if (vm_pageout_stats_interval == 0)
1504 vm_pageout_stats_interval = 5;
1505 if (vm_pageout_full_stats_interval == 0)
1506 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1508 swap_pager_swap_init();
1511 * The pageout daemon is never done, so loop forever.
1514 vm_page_lock_queues();
1516 * If we have enough free memory, wakeup waiters. Do
1517 * not clear vm_pages_needed until we reach our target,
1518 * otherwise we may be woken up over and over again and
1519 * waste a lot of cpu.
1521 if (vm_pages_needed && !vm_page_count_min()) {
1522 if (!vm_paging_needed())
1523 vm_pages_needed = 0;
1524 wakeup(&cnt.v_free_count);
1526 if (vm_pages_needed) {
1528 * Still not done, take a second pass without waiting
1529 * (unlimited dirty cleaning), otherwise sleep a bit
1534 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1538 * Good enough, sleep & handle stats. Prime the pass
1545 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1546 "psleep", vm_pageout_stats_interval * hz);
1547 if (error && !vm_pages_needed) {
1549 vm_pageout_page_stats();
1550 vm_page_unlock_queues();
1554 if (vm_pages_needed)
1556 vm_page_unlock_queues();
1557 vm_pageout_scan(pass);
1562 * Unless the page queue lock is held by the caller, this function
1563 * should be regarded as advisory. Specifically, the caller should
1564 * not msleep() on &cnt.v_free_count following this function unless
1565 * the page queue lock is held until the msleep() is performed.
1571 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1572 vm_pages_needed = 1;
1573 wakeup(&vm_pages_needed);
1577 #if !defined(NO_SWAPPING)
1579 vm_req_vmdaemon(int req)
1581 static int lastrun = 0;
1583 mtx_lock(&vm_daemon_mtx);
1584 vm_pageout_req_swapout |= req;
1585 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1586 wakeup(&vm_daemon_needed);
1589 mtx_unlock(&vm_daemon_mtx);
1595 struct rlimit rsslim;
1598 int breakout, swapout_flags;
1601 mtx_lock(&vm_daemon_mtx);
1602 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1603 swapout_flags = vm_pageout_req_swapout;
1604 vm_pageout_req_swapout = 0;
1605 mtx_unlock(&vm_daemon_mtx);
1607 swapout_procs(swapout_flags);
1610 * scan the processes for exceeding their rlimits or if
1611 * process is swapped out -- deactivate pages
1613 sx_slock(&allproc_lock);
1614 LIST_FOREACH(p, &allproc, p_list) {
1615 vm_pindex_t limit, size;
1618 * if this is a system process or if we have already
1619 * looked at this process, skip it.
1622 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1627 * if the process is in a non-running type state,
1630 mtx_lock_spin(&sched_lock);
1632 FOREACH_THREAD_IN_PROC(p, td) {
1633 if (!TD_ON_RUNQ(td) &&
1634 !TD_IS_RUNNING(td) &&
1635 !TD_IS_SLEEPING(td)) {
1640 mtx_unlock_spin(&sched_lock);
1648 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1650 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1653 * let processes that are swapped out really be
1654 * swapped out set the limit to nothing (will force a
1657 if ((p->p_sflag & PS_INMEM) == 0)
1658 limit = 0; /* XXX */
1661 size = vmspace_resident_count(p->p_vmspace);
1662 if (limit >= 0 && size >= limit) {
1663 vm_pageout_map_deactivate_pages(
1664 &p->p_vmspace->vm_map, limit);
1667 sx_sunlock(&allproc_lock);
1670 #endif /* !defined(NO_SWAPPING) */