2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * This code is derived from software contributed to Berkeley by
12 * The Mach Operating System project at Carnegie-Mellon University.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
50 * Permission to use, copy, modify and distribute this software and
51 * its documentation is hereby granted, provided that both the copyright
52 * notice and this permission notice appear in all copies of the
53 * software, derivative works or modified versions, and any portions
54 * thereof, and that both notices appear in supporting documentation.
56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
60 * Carnegie Mellon requests users of this software to return to
62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
63 * School of Computer Science
64 * Carnegie Mellon University
65 * Pittsburgh PA 15213-3890
67 * any improvements or extensions that they make and grant Carnegie the
68 * rights to redistribute these changes.
72 * The proverbial page-out daemon.
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
84 #include <sys/mutex.h>
86 #include <sys/kthread.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
95 #include <sys/sysctl.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
109 * System initialization
112 /* the kernel process "vm_pageout"*/
113 static void vm_pageout(void);
114 static int vm_pageout_clean(vm_page_t);
115 static void vm_pageout_scan(int pass);
117 struct proc *pageproc;
119 static struct kproc_desc page_kp = {
124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
127 #if !defined(NO_SWAPPING)
128 /* the kernel process "vm_daemon"*/
129 static void vm_daemon(void);
130 static struct proc *vmproc;
132 static struct kproc_desc vm_kp = {
137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
141 int vm_pages_needed; /* Event on which pageout daemon sleeps */
142 int vm_pageout_deficit; /* Estimated number of pages deficit */
143 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
145 #if !defined(NO_SWAPPING)
146 static int vm_pageout_req_swapout; /* XXX */
147 static int vm_daemon_needed;
148 static struct mtx vm_daemon_mtx;
149 /* Allow for use by vm_pageout before vm_daemon is initialized. */
150 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
152 static int vm_max_launder = 32;
153 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
154 static int vm_pageout_full_stats_interval = 0;
155 static int vm_pageout_algorithm=0;
156 static int defer_swap_pageouts=0;
157 static int disable_swap_pageouts=0;
159 #if defined(NO_SWAPPING)
160 static int vm_swap_enabled=0;
161 static int vm_swap_idle_enabled=0;
163 static int vm_swap_enabled=1;
164 static int vm_swap_idle_enabled=0;
167 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
168 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
170 SYSCTL_INT(_vm, OID_AUTO, max_launder,
171 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
174 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
176 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
177 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
180 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
182 #if defined(NO_SWAPPING)
183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
184 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
186 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
194 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
195 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
197 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
198 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
200 static int pageout_lock_miss;
201 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
202 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
204 #define VM_PAGEOUT_PAGE_COUNT 16
205 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
207 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
208 SYSCTL_INT(_vm, OID_AUTO, max_wired,
209 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
211 #if !defined(NO_SWAPPING)
212 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
213 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
214 static void vm_req_vmdaemon(int req);
216 static void vm_pageout_page_stats(void);
219 * vm_pageout_fallback_object_lock:
221 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
222 * known to have failed and page queue must be either PQ_ACTIVE or
223 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
224 * while locking the vm object. Use marker page to detect page queue
225 * changes and maintain notion of next page on page queue. Return
226 * TRUE if no changes were detected, FALSE otherwise. vm object is
229 * This function depends on both the lock portion of struct vm_object
230 * and normal struct vm_page being type stable.
233 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
235 struct vm_page marker;
241 * Initialize our marker
243 bzero(&marker, sizeof(marker));
244 marker.flags = PG_FICTITIOUS | PG_MARKER;
245 marker.oflags = VPO_BUSY;
246 marker.queue = m->queue;
247 marker.wire_count = 1;
252 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
254 vm_page_unlock_queues();
255 VM_OBJECT_LOCK(object);
256 vm_page_lock_queues();
258 /* Page queue might have changed. */
259 *next = TAILQ_NEXT(&marker, pageq);
260 unchanged = (m->queue == queue &&
261 m->object == object &&
262 &marker == TAILQ_NEXT(m, pageq));
263 TAILQ_REMOVE(&vm_page_queues[queue].pl,
271 * Clean the page and remove it from the laundry.
273 * We set the busy bit to cause potential page faults on this page to
274 * block. Note the careful timing, however, the busy bit isn't set till
275 * late and we cannot do anything that will mess with the page.
282 vm_page_t mc[2*vm_pageout_page_count];
284 int ib, is, page_base;
285 vm_pindex_t pindex = m->pindex;
287 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
288 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
291 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
292 * with the new swapper, but we could have serious problems paging
293 * out other object types if there is insufficient memory.
295 * Unfortunately, checking free memory here is far too late, so the
296 * check has been moved up a procedural level.
300 * Can't clean the page if it's busy or held.
302 if ((m->hold_count != 0) ||
303 ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
307 mc[vm_pageout_page_count] = m;
309 page_base = vm_pageout_page_count;
314 * Scan object for clusterable pages.
316 * We can cluster ONLY if: ->> the page is NOT
317 * clean, wired, busy, held, or mapped into a
318 * buffer, and one of the following:
319 * 1) The page is inactive, or a seldom used
322 * 2) we force the issue.
324 * During heavy mmap/modification loads the pageout
325 * daemon can really fragment the underlying file
326 * due to flushing pages out of order and not trying
327 * align the clusters (which leave sporatic out-of-order
328 * holes). To solve this problem we do the reverse scan
329 * first and attempt to align our cluster, then do a
330 * forward scan if room remains.
334 while (ib && pageout_count < vm_pageout_page_count) {
342 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
346 if ((p->oflags & VPO_BUSY) || p->busy) {
350 vm_page_test_dirty(p);
352 p->queue != PQ_INACTIVE ||
353 p->wire_count != 0 || /* may be held by buf cache */
354 p->hold_count != 0) { /* may be undergoing I/O */
362 * alignment boundry, stop here and switch directions. Do
365 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
369 while (pageout_count < vm_pageout_page_count &&
370 pindex + is < object->size) {
373 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
375 if ((p->oflags & VPO_BUSY) || p->busy) {
378 vm_page_test_dirty(p);
380 p->queue != PQ_INACTIVE ||
381 p->wire_count != 0 || /* may be held by buf cache */
382 p->hold_count != 0) { /* may be undergoing I/O */
385 mc[page_base + pageout_count] = p;
391 * If we exhausted our forward scan, continue with the reverse scan
392 * when possible, even past a page boundry. This catches boundry
395 if (ib && pageout_count < vm_pageout_page_count)
399 * we allow reads during pageouts...
401 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
405 * vm_pageout_flush() - launder the given pages
407 * The given pages are laundered. Note that we setup for the start of
408 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
409 * reference count all in here rather then in the parent. If we want
410 * the parent to do more sophisticated things we may have to change
414 vm_pageout_flush(vm_page_t *mc, int count, int flags)
416 vm_object_t object = mc[0]->object;
417 int pageout_status[count];
421 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
422 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
424 * Initiate I/O. Bump the vm_page_t->busy counter and
425 * mark the pages read-only.
427 * We do not have to fixup the clean/dirty bits here... we can
428 * allow the pager to do it after the I/O completes.
430 * NOTE! mc[i]->dirty may be partial or fragmented due to an
431 * edge case with file fragments.
433 for (i = 0; i < count; i++) {
434 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
435 ("vm_pageout_flush: partially invalid page %p index %d/%d",
437 vm_page_io_start(mc[i]);
438 pmap_remove_write(mc[i]);
440 vm_page_unlock_queues();
441 vm_object_pip_add(object, count);
443 vm_pager_put_pages(object, mc, count, flags, pageout_status);
445 vm_page_lock_queues();
446 for (i = 0; i < count; i++) {
447 vm_page_t mt = mc[i];
449 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
450 (mt->flags & PG_WRITEABLE) == 0,
451 ("vm_pageout_flush: page %p is not write protected", mt));
452 switch (pageout_status[i]) {
459 * Page outside of range of object. Right now we
460 * essentially lose the changes by pretending it
468 * If page couldn't be paged out, then reactivate the
469 * page so it doesn't clog the inactive list. (We
470 * will try paging out it again later).
472 vm_page_activate(mt);
479 * If the operation is still going, leave the page busy to
480 * block all other accesses. Also, leave the paging in
481 * progress indicator set so that we don't attempt an object
484 if (pageout_status[i] != VM_PAGER_PEND) {
485 vm_object_pip_wakeup(object);
486 vm_page_io_finish(mt);
487 if (vm_page_count_severe())
488 vm_page_try_to_cache(mt);
494 #if !defined(NO_SWAPPING)
496 * vm_pageout_object_deactivate_pages
498 * deactivate enough pages to satisfy the inactive target
499 * requirements or if vm_page_proc_limit is set, then
500 * deactivate all of the pages in the object and its
503 * The object and map must be locked.
506 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
508 vm_object_t first_object;
511 vm_object_t backing_object, object;
513 int actcount, rcount, remove_mode;
515 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
516 if (first_object->type == OBJT_DEVICE ||
517 first_object->type == OBJT_SG ||
518 first_object->type == OBJT_PHYS)
520 for (object = first_object;; object = backing_object) {
521 if (pmap_resident_count(pmap) <= desired)
523 if (object->paging_in_progress)
527 if (object->shadow_count > 1)
530 * scan the objects entire memory queue
532 rcount = object->resident_page_count;
533 p = TAILQ_FIRST(&object->memq);
534 vm_page_lock_queues();
535 while (p && (rcount-- > 0)) {
536 if (pmap_resident_count(pmap) <= desired) {
537 vm_page_unlock_queues();
540 next = TAILQ_NEXT(p, listq);
542 if (p->wire_count != 0 ||
543 p->hold_count != 0 ||
545 (p->oflags & VPO_BUSY) ||
546 (p->flags & PG_UNMANAGED) ||
547 !pmap_page_exists_quick(pmap, p)) {
551 actcount = pmap_ts_referenced(p);
553 vm_page_flag_set(p, PG_REFERENCED);
554 } else if (p->flags & PG_REFERENCED) {
557 if ((p->queue != PQ_ACTIVE) &&
558 (p->flags & PG_REFERENCED)) {
560 p->act_count += actcount;
561 vm_page_flag_clear(p, PG_REFERENCED);
562 } else if (p->queue == PQ_ACTIVE) {
563 if ((p->flags & PG_REFERENCED) == 0) {
564 p->act_count -= min(p->act_count, ACT_DECLINE);
565 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
567 vm_page_deactivate(p);
573 vm_page_flag_clear(p, PG_REFERENCED);
574 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
575 p->act_count += ACT_ADVANCE;
578 } else if (p->queue == PQ_INACTIVE) {
583 vm_page_unlock_queues();
584 if ((backing_object = object->backing_object) == NULL)
586 VM_OBJECT_LOCK(backing_object);
587 if (object != first_object)
588 VM_OBJECT_UNLOCK(object);
591 if (object != first_object)
592 VM_OBJECT_UNLOCK(object);
596 * deactivate some number of pages in a map, try to do it fairly, but
597 * that is really hard to do.
600 vm_pageout_map_deactivate_pages(map, desired)
605 vm_object_t obj, bigobj;
608 if (!vm_map_trylock(map))
615 * first, search out the biggest object, and try to free pages from
618 tmpe = map->header.next;
619 while (tmpe != &map->header) {
620 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
621 obj = tmpe->object.vm_object;
622 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
623 if (obj->shadow_count <= 1 &&
625 bigobj->resident_page_count < obj->resident_page_count)) {
627 VM_OBJECT_UNLOCK(bigobj);
630 VM_OBJECT_UNLOCK(obj);
633 if (tmpe->wired_count > 0)
634 nothingwired = FALSE;
638 if (bigobj != NULL) {
639 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
640 VM_OBJECT_UNLOCK(bigobj);
643 * Next, hunt around for other pages to deactivate. We actually
644 * do this search sort of wrong -- .text first is not the best idea.
646 tmpe = map->header.next;
647 while (tmpe != &map->header) {
648 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
650 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
651 obj = tmpe->object.vm_object;
654 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
655 VM_OBJECT_UNLOCK(obj);
662 * Remove all mappings if a process is swapped out, this will free page
665 if (desired == 0 && nothingwired) {
666 pmap_remove(vm_map_pmap(map), vm_map_min(map),
671 #endif /* !defined(NO_SWAPPING) */
674 * vm_pageout_scan does the dirty work for the pageout daemon.
677 vm_pageout_scan(int pass)
680 struct vm_page marker;
681 int page_shortage, maxscan, pcount;
682 int addl_page_shortage, addl_page_shortage_init;
685 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_FICTITIOUS | PG_MARKER;
710 marker.oflags = VPO_BUSY;
711 marker.queue = PQ_INACTIVE;
712 marker.wire_count = 1;
715 * Start scanning the inactive queue for pages we can move to the
716 * cache or free. The scan will stop when the target is reached or
717 * we have scanned the entire inactive queue. Note that m->act_count
718 * is not used to form decisions for the inactive queue, only for the
721 * maxlaunder limits the number of dirty pages we flush per scan.
722 * For most systems a smaller value (16 or 32) is more robust under
723 * extreme memory and disk pressure because any unnecessary writes
724 * to disk can result in extreme performance degredation. However,
725 * systems with excessive dirty pages (especially when MAP_NOSYNC is
726 * used) will die horribly with limited laundering. If the pageout
727 * daemon cannot clean enough pages in the first pass, we let it go
728 * all out in succeeding passes.
730 if ((maxlaunder = vm_max_launder) <= 1)
734 vm_page_lock_queues();
736 addl_page_shortage = addl_page_shortage_init;
737 maxscan = cnt.v_inactive_count;
739 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
740 m != NULL && maxscan-- > 0 && page_shortage > 0;
745 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
749 next = TAILQ_NEXT(m, pageq);
755 if (m->flags & PG_MARKER)
759 * A held page may be undergoing I/O, so skip it.
763 addl_page_shortage++;
767 * Don't mess with busy pages, keep in the front of the
768 * queue, most likely are being paged out.
770 if (!VM_OBJECT_TRYLOCK(object) &&
771 (!vm_pageout_fallback_object_lock(m, &next) ||
772 m->hold_count != 0)) {
773 VM_OBJECT_UNLOCK(object);
774 addl_page_shortage++;
777 if (m->busy || (m->oflags & VPO_BUSY)) {
778 VM_OBJECT_UNLOCK(object);
779 addl_page_shortage++;
784 * If the object is not being used, we ignore previous
787 if (object->ref_count == 0) {
788 vm_page_flag_clear(m, PG_REFERENCED);
789 KASSERT(!pmap_page_is_mapped(m),
790 ("vm_pageout_scan: page %p is mapped", m));
793 * Otherwise, if the page has been referenced while in the
794 * inactive queue, we bump the "activation count" upwards,
795 * making it less likely that the page will be added back to
796 * the inactive queue prematurely again. Here we check the
797 * page tables (or emulated bits, if any), given the upper
798 * level VM system not knowing anything about existing
801 } else if (((m->flags & PG_REFERENCED) == 0) &&
802 (actcount = pmap_ts_referenced(m))) {
804 VM_OBJECT_UNLOCK(object);
805 m->act_count += (actcount + ACT_ADVANCE);
810 * If the upper level VM system knows about any page
811 * references, we activate the page. We also set the
812 * "activation count" higher than normal so that we will less
813 * likely place pages back onto the inactive queue again.
815 if ((m->flags & PG_REFERENCED) != 0) {
816 vm_page_flag_clear(m, PG_REFERENCED);
817 actcount = pmap_ts_referenced(m);
819 VM_OBJECT_UNLOCK(object);
820 m->act_count += (actcount + ACT_ADVANCE + 1);
825 * If the upper level VM system does not believe that the page
826 * is fully dirty, but it is mapped for write access, then we
827 * consult the pmap to see if the page's dirty status should
830 if (m->dirty != VM_PAGE_BITS_ALL &&
831 (m->flags & PG_WRITEABLE) != 0) {
833 * Avoid a race condition: Unless write access is
834 * removed from the page, another processor could
835 * modify it before all access is removed by the call
836 * to vm_page_cache() below. If vm_page_cache() finds
837 * that the page has been modified when it removes all
838 * access, it panics because it cannot cache dirty
839 * pages. In principle, we could eliminate just write
840 * access here rather than all access. In the expected
841 * case, when there are no last instant modifications
842 * to the page, removing all access will be cheaper
845 if (pmap_is_modified(m))
847 else if (m->dirty == 0)
853 * Invalid pages can be easily freed
858 } else if (m->dirty == 0) {
860 * Clean pages can be placed onto the cache queue.
861 * This effectively frees them.
865 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
867 * Dirty pages need to be paged out, but flushing
868 * a page is extremely expensive verses freeing
869 * a clean page. Rather then artificially limiting
870 * the number of pages we can flush, we instead give
871 * dirty pages extra priority on the inactive queue
872 * by forcing them to be cycled through the queue
873 * twice before being flushed, after which the
874 * (now clean) page will cycle through once more
875 * before being freed. This significantly extends
876 * the thrash point for a heavily loaded machine.
878 vm_page_flag_set(m, PG_WINATCFLS);
880 } else if (maxlaunder > 0) {
882 * We always want to try to flush some dirty pages if
883 * we encounter them, to keep the system stable.
884 * Normally this number is small, but under extreme
885 * pressure where there are insufficient clean pages
886 * on the inactive queue, we may have to go all out.
888 int swap_pageouts_ok, vfslocked = 0;
889 struct vnode *vp = NULL;
890 struct mount *mp = NULL;
892 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
893 swap_pageouts_ok = 1;
895 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
896 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
897 vm_page_count_min());
902 * We don't bother paging objects that are "dead".
903 * Those objects are in a "rundown" state.
905 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
906 VM_OBJECT_UNLOCK(object);
912 * Following operations may unlock
913 * vm_page_queue_mtx, invalidating the 'next'
914 * pointer. To prevent an inordinate number
915 * of restarts we use our marker to remember
919 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
922 * The object is already known NOT to be dead. It
923 * is possible for the vget() to block the whole
924 * pageout daemon, but the new low-memory handling
925 * code should prevent it.
927 * The previous code skipped locked vnodes and, worse,
928 * reordered pages in the queue. This results in
929 * completely non-deterministic operation and, on a
930 * busy system, can lead to extremely non-optimal
931 * pageouts. For example, it can cause clean pages
932 * to be freed and dirty pages to be moved to the end
933 * of the queue. Since dirty pages are also moved to
934 * the end of the queue once-cleaned, this gives
935 * way too large a weighting to defering the freeing
938 * We can't wait forever for the vnode lock, we might
939 * deadlock due to a vn_read() getting stuck in
940 * vm_wait while holding this vnode. We skip the
941 * vnode if we can't get it in a reasonable amount
944 if (object->type == OBJT_VNODE) {
946 if (vp->v_type == VREG &&
947 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
950 if (object->flags & OBJ_MIGHTBEDIRTY)
952 goto unlock_and_continue;
955 ("vp %p with NULL v_mount", vp));
956 vm_page_unlock_queues();
957 vm_object_reference_locked(object);
958 VM_OBJECT_UNLOCK(object);
959 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
960 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
962 VM_OBJECT_LOCK(object);
963 vm_page_lock_queues();
965 if (object->flags & OBJ_MIGHTBEDIRTY)
968 goto unlock_and_continue;
970 VM_OBJECT_LOCK(object);
971 vm_page_lock_queues();
973 * The page might have been moved to another
974 * queue during potential blocking in vget()
975 * above. The page might have been freed and
976 * reused for another vnode.
978 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
979 m->object != object ||
980 TAILQ_NEXT(m, pageq) != &marker) {
981 if (object->flags & OBJ_MIGHTBEDIRTY)
983 goto unlock_and_continue;
987 * The page may have been busied during the
988 * blocking in vget(). We don't move the
989 * page back onto the end of the queue so that
990 * statistics are more correct if we don't.
992 if (m->busy || (m->oflags & VPO_BUSY)) {
993 goto unlock_and_continue;
997 * If the page has become held it might
998 * be undergoing I/O, so skip it
1000 if (m->hold_count) {
1002 if (object->flags & OBJ_MIGHTBEDIRTY)
1004 goto unlock_and_continue;
1009 * If a page is dirty, then it is either being washed
1010 * (but not yet cleaned) or it is still in the
1011 * laundry. If it is still in the laundry, then we
1012 * start the cleaning operation.
1014 * decrement page_shortage on success to account for
1015 * the (future) cleaned page. Otherwise we could wind
1016 * up laundering or cleaning too many pages.
1018 if (vm_pageout_clean(m) != 0) {
1022 unlock_and_continue:
1023 VM_OBJECT_UNLOCK(object);
1025 vm_page_unlock_queues();
1028 VFS_UNLOCK_GIANT(vfslocked);
1029 vm_object_deallocate(object);
1030 vn_finished_write(mp);
1031 vm_page_lock_queues();
1033 next = TAILQ_NEXT(&marker, pageq);
1034 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1038 VM_OBJECT_UNLOCK(object);
1042 * Compute the number of pages we want to try to move from the
1043 * active queue to the inactive queue.
1045 page_shortage = vm_paging_target() +
1046 cnt.v_inactive_target - cnt.v_inactive_count;
1047 page_shortage += addl_page_shortage;
1050 * Scan the active queue for things we can deactivate. We nominally
1051 * track the per-page activity counter and use it to locate
1052 * deactivation candidates.
1054 pcount = cnt.v_active_count;
1055 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1057 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1059 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1060 ("vm_pageout_scan: page %p isn't active", m));
1062 next = TAILQ_NEXT(m, pageq);
1064 if ((m->flags & PG_MARKER) != 0) {
1068 if (!VM_OBJECT_TRYLOCK(object) &&
1069 !vm_pageout_fallback_object_lock(m, &next)) {
1070 VM_OBJECT_UNLOCK(object);
1076 * Don't deactivate pages that are busy.
1078 if ((m->busy != 0) ||
1079 (m->oflags & VPO_BUSY) ||
1080 (m->hold_count != 0)) {
1081 VM_OBJECT_UNLOCK(object);
1088 * The count for pagedaemon pages is done after checking the
1089 * page for eligibility...
1094 * Check to see "how much" the page has been used.
1097 if (object->ref_count != 0) {
1098 if (m->flags & PG_REFERENCED) {
1101 actcount += pmap_ts_referenced(m);
1103 m->act_count += ACT_ADVANCE + actcount;
1104 if (m->act_count > ACT_MAX)
1105 m->act_count = ACT_MAX;
1110 * Since we have "tested" this bit, we need to clear it now.
1112 vm_page_flag_clear(m, PG_REFERENCED);
1115 * Only if an object is currently being used, do we use the
1116 * page activation count stats.
1118 if (actcount && (object->ref_count != 0)) {
1121 m->act_count -= min(m->act_count, ACT_DECLINE);
1122 if (vm_pageout_algorithm ||
1123 object->ref_count == 0 ||
1124 m->act_count == 0) {
1126 if (object->ref_count == 0) {
1131 vm_page_deactivate(m);
1133 vm_page_deactivate(m);
1139 VM_OBJECT_UNLOCK(object);
1142 vm_page_unlock_queues();
1143 #if !defined(NO_SWAPPING)
1145 * Idle process swapout -- run once per second.
1147 if (vm_swap_idle_enabled) {
1149 if (time_second != lsec) {
1150 vm_req_vmdaemon(VM_SWAP_IDLE);
1157 * If we didn't get enough free pages, and we have skipped a vnode
1158 * in a writeable object, wakeup the sync daemon. And kick swapout
1159 * if we did not get enough free pages.
1161 if (vm_paging_target() > 0) {
1162 if (vnodes_skipped && vm_page_count_min())
1163 (void) speedup_syncer();
1164 #if !defined(NO_SWAPPING)
1165 if (vm_swap_enabled && vm_page_count_target())
1166 vm_req_vmdaemon(VM_SWAP_NORMAL);
1171 * If we are critically low on one of RAM or swap and low on
1172 * the other, kill the largest process. However, we avoid
1173 * doing this on the first pass in order to give ourselves a
1174 * chance to flush out dirty vnode-backed pages and to allow
1175 * active pages to be moved to the inactive queue and reclaimed.
1178 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1179 (swap_pager_full && vm_paging_target() > 0)))
1180 vm_pageout_oom(VM_OOM_MEM);
1185 vm_pageout_oom(int shortage)
1187 struct proc *p, *bigproc;
1188 vm_offset_t size, bigsize;
1193 * We keep the process bigproc locked once we find it to keep anyone
1194 * from messing with it; however, there is a possibility of
1195 * deadlock if process B is bigproc and one of it's child processes
1196 * attempts to propagate a signal to B while we are waiting for A's
1197 * lock while walking this list. To avoid this, we don't block on
1198 * the process lock but just skip a process if it is already locked.
1202 sx_slock(&allproc_lock);
1203 FOREACH_PROC_IN_SYSTEM(p) {
1206 if (PROC_TRYLOCK(p) == 0)
1209 * If this is a system, protected or killed process, skip it.
1211 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1212 (p->p_pid == 1) || P_KILLED(p) ||
1213 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1218 * If the process is in a non-running type state,
1219 * don't touch it. Check all the threads individually.
1222 FOREACH_THREAD_IN_PROC(p, td) {
1224 if (!TD_ON_RUNQ(td) &&
1225 !TD_IS_RUNNING(td) &&
1226 !TD_IS_SLEEPING(td)) {
1238 * get the process size
1240 vm = vmspace_acquire_ref(p);
1245 if (!vm_map_trylock_read(&vm->vm_map)) {
1250 size = vmspace_swap_count(vm);
1251 vm_map_unlock_read(&vm->vm_map);
1252 if (shortage == VM_OOM_MEM)
1253 size += vmspace_resident_count(vm);
1256 * if the this process is bigger than the biggest one
1259 if (size > bigsize) {
1260 if (bigproc != NULL)
1261 PROC_UNLOCK(bigproc);
1267 sx_sunlock(&allproc_lock);
1268 if (bigproc != NULL) {
1269 killproc(bigproc, "out of swap space");
1270 sched_nice(bigproc, PRIO_MIN);
1271 PROC_UNLOCK(bigproc);
1272 wakeup(&cnt.v_free_count);
1277 * This routine tries to maintain the pseudo LRU active queue,
1278 * so that during long periods of time where there is no paging,
1279 * that some statistic accumulation still occurs. This code
1280 * helps the situation where paging just starts to occur.
1283 vm_pageout_page_stats()
1287 int pcount,tpcount; /* Number of pages to check */
1288 static int fullintervalcount = 0;
1291 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1293 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1294 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1296 if (page_shortage <= 0)
1299 pcount = cnt.v_active_count;
1300 fullintervalcount += vm_pageout_stats_interval;
1301 if (fullintervalcount < vm_pageout_full_stats_interval) {
1302 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1304 if (pcount > tpcount)
1307 fullintervalcount = 0;
1310 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1311 while ((m != NULL) && (pcount-- > 0)) {
1314 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1315 ("vm_pageout_page_stats: page %p isn't active", m));
1317 next = TAILQ_NEXT(m, pageq);
1320 if ((m->flags & PG_MARKER) != 0) {
1324 if (!VM_OBJECT_TRYLOCK(object) &&
1325 !vm_pageout_fallback_object_lock(m, &next)) {
1326 VM_OBJECT_UNLOCK(object);
1332 * Don't deactivate pages that are busy.
1334 if ((m->busy != 0) ||
1335 (m->oflags & VPO_BUSY) ||
1336 (m->hold_count != 0)) {
1337 VM_OBJECT_UNLOCK(object);
1344 if (m->flags & PG_REFERENCED) {
1345 vm_page_flag_clear(m, PG_REFERENCED);
1349 actcount += pmap_ts_referenced(m);
1351 m->act_count += ACT_ADVANCE + actcount;
1352 if (m->act_count > ACT_MAX)
1353 m->act_count = ACT_MAX;
1356 if (m->act_count == 0) {
1358 * We turn off page access, so that we have
1359 * more accurate RSS stats. We don't do this
1360 * in the normal page deactivation when the
1361 * system is loaded VM wise, because the
1362 * cost of the large number of page protect
1363 * operations would be higher than the value
1364 * of doing the operation.
1367 vm_page_deactivate(m);
1369 m->act_count -= min(m->act_count, ACT_DECLINE);
1373 VM_OBJECT_UNLOCK(object);
1379 * vm_pageout is the high level pageout daemon.
1387 * Initialize some paging parameters.
1389 cnt.v_interrupt_free_min = 2;
1390 if (cnt.v_page_count < 2000)
1391 vm_pageout_page_count = 8;
1394 * v_free_reserved needs to include enough for the largest
1395 * swap pager structures plus enough for any pv_entry structs
1398 if (cnt.v_page_count > 1024)
1399 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1402 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1403 cnt.v_interrupt_free_min;
1404 cnt.v_free_reserved = vm_pageout_page_count +
1405 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1406 cnt.v_free_severe = cnt.v_free_min / 2;
1407 cnt.v_free_min += cnt.v_free_reserved;
1408 cnt.v_free_severe += cnt.v_free_reserved;
1411 * v_free_target and v_cache_min control pageout hysteresis. Note
1412 * that these are more a measure of the VM cache queue hysteresis
1413 * then the VM free queue. Specifically, v_free_target is the
1414 * high water mark (free+cache pages).
1416 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1417 * low water mark, while v_free_min is the stop. v_cache_min must
1418 * be big enough to handle memory needs while the pageout daemon
1419 * is signalled and run to free more pages.
1421 if (cnt.v_free_count > 6144)
1422 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1424 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1426 if (cnt.v_free_count > 2048) {
1427 cnt.v_cache_min = cnt.v_free_target;
1428 cnt.v_cache_max = 2 * cnt.v_cache_min;
1429 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1431 cnt.v_cache_min = 0;
1432 cnt.v_cache_max = 0;
1433 cnt.v_inactive_target = cnt.v_free_count / 4;
1435 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1436 cnt.v_inactive_target = cnt.v_free_count / 3;
1438 /* XXX does not really belong here */
1439 if (vm_page_max_wired == 0)
1440 vm_page_max_wired = cnt.v_free_count / 3;
1442 if (vm_pageout_stats_max == 0)
1443 vm_pageout_stats_max = cnt.v_free_target;
1446 * Set interval in seconds for stats scan.
1448 if (vm_pageout_stats_interval == 0)
1449 vm_pageout_stats_interval = 5;
1450 if (vm_pageout_full_stats_interval == 0)
1451 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1453 swap_pager_swap_init();
1456 * The pageout daemon is never done, so loop forever.
1460 * If we have enough free memory, wakeup waiters. Do
1461 * not clear vm_pages_needed until we reach our target,
1462 * otherwise we may be woken up over and over again and
1463 * waste a lot of cpu.
1465 mtx_lock(&vm_page_queue_free_mtx);
1466 if (vm_pages_needed && !vm_page_count_min()) {
1467 if (!vm_paging_needed())
1468 vm_pages_needed = 0;
1469 wakeup(&cnt.v_free_count);
1471 if (vm_pages_needed) {
1473 * Still not done, take a second pass without waiting
1474 * (unlimited dirty cleaning), otherwise sleep a bit
1479 msleep(&vm_pages_needed,
1480 &vm_page_queue_free_mtx, PVM, "psleep",
1484 * Good enough, sleep & handle stats. Prime the pass
1491 error = msleep(&vm_pages_needed,
1492 &vm_page_queue_free_mtx, PVM, "psleep",
1493 vm_pageout_stats_interval * hz);
1494 if (error && !vm_pages_needed) {
1495 mtx_unlock(&vm_page_queue_free_mtx);
1497 vm_page_lock_queues();
1498 vm_pageout_page_stats();
1499 vm_page_unlock_queues();
1503 if (vm_pages_needed)
1505 mtx_unlock(&vm_page_queue_free_mtx);
1506 vm_pageout_scan(pass);
1511 * Unless the free 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 free 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)
1528 vm_req_vmdaemon(int req)
1530 static int lastrun = 0;
1532 mtx_lock(&vm_daemon_mtx);
1533 vm_pageout_req_swapout |= req;
1534 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1535 wakeup(&vm_daemon_needed);
1538 mtx_unlock(&vm_daemon_mtx);
1544 struct rlimit rsslim;
1548 int breakout, swapout_flags;
1551 mtx_lock(&vm_daemon_mtx);
1552 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1553 swapout_flags = vm_pageout_req_swapout;
1554 vm_pageout_req_swapout = 0;
1555 mtx_unlock(&vm_daemon_mtx);
1557 swapout_procs(swapout_flags);
1560 * scan the processes for exceeding their rlimits or if
1561 * process is swapped out -- deactivate pages
1563 sx_slock(&allproc_lock);
1564 FOREACH_PROC_IN_SYSTEM(p) {
1565 vm_pindex_t limit, size;
1568 * if this is a system process or if we have already
1569 * looked at this process, skip it.
1572 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1577 * if the process is in a non-running type state,
1581 FOREACH_THREAD_IN_PROC(p, td) {
1583 if (!TD_ON_RUNQ(td) &&
1584 !TD_IS_RUNNING(td) &&
1585 !TD_IS_SLEEPING(td)) {
1599 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1601 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1604 * let processes that are swapped out really be
1605 * swapped out set the limit to nothing (will force a
1608 if ((p->p_flag & P_INMEM) == 0)
1609 limit = 0; /* XXX */
1610 vm = vmspace_acquire_ref(p);
1615 size = vmspace_resident_count(vm);
1616 if (limit >= 0 && size >= limit) {
1617 vm_pageout_map_deactivate_pages(
1618 &vm->vm_map, limit);
1622 sx_sunlock(&allproc_lock);
1625 #endif /* !defined(NO_SWAPPING) */