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_scan(int pass);
119 struct proc *pageproc;
121 static struct kproc_desc page_kp = {
126 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
128 #if !defined(NO_SWAPPING)
129 /* the kernel process "vm_daemon"*/
130 static void vm_daemon(void);
131 static struct proc *vmproc;
133 static struct kproc_desc vm_kp = {
138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
142 int vm_pages_needed; /* Event on which pageout daemon sleeps */
143 int vm_pageout_deficit; /* Estimated number of pages deficit */
144 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
146 #if !defined(NO_SWAPPING)
147 static int vm_pageout_req_swapout; /* XXX */
148 static int vm_daemon_needed;
149 static struct mtx vm_daemon_mtx;
150 /* Allow for use by vm_pageout before vm_daemon is initialized. */
151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
153 static int vm_max_launder = 32;
154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
155 static int vm_pageout_full_stats_interval = 0;
156 static int vm_pageout_algorithm=0;
157 static int defer_swap_pageouts=0;
158 static int disable_swap_pageouts=0;
160 #if defined(NO_SWAPPING)
161 static int vm_swap_enabled=0;
162 static int vm_swap_idle_enabled=0;
164 static int vm_swap_enabled=1;
165 static int vm_swap_idle_enabled=0;
168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
169 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
171 SYSCTL_INT(_vm, OID_AUTO, max_launder,
172 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
175 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
178 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
181 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
183 #if defined(NO_SWAPPING)
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 CTLFLAG_RD, &vm_swap_enabled, 0, "");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
190 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
192 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
196 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
199 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
201 static int pageout_lock_miss;
202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
203 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
205 #define VM_PAGEOUT_PAGE_COUNT 16
206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
208 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
210 #if !defined(NO_SWAPPING)
211 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
212 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
213 static void vm_req_vmdaemon(int req);
215 static void vm_pageout_page_stats(void);
218 * vm_pageout_fallback_object_lock:
220 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
221 * known to have failed and page queue must be either PQ_ACTIVE or
222 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
223 * while locking the vm object. Use marker page to detect page queue
224 * changes and maintain notion of next page on page queue. Return
225 * TRUE if no changes were detected, FALSE otherwise. vm object is
228 * This function depends on both the lock portion of struct vm_object
229 * and normal struct vm_page being type stable.
232 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
234 struct vm_page marker;
240 * Initialize our marker
242 bzero(&marker, sizeof(marker));
243 marker.flags = PG_FICTITIOUS | PG_MARKER;
244 marker.oflags = VPO_BUSY;
245 marker.queue = m->queue;
246 marker.wire_count = 1;
251 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
253 vm_page_unlock_queues();
254 VM_OBJECT_LOCK(object);
255 vm_page_lock_queues();
257 /* Page queue might have changed. */
258 *next = TAILQ_NEXT(&marker, pageq);
259 unchanged = (m->queue == queue &&
260 m->object == object &&
261 &marker == TAILQ_NEXT(m, pageq));
262 TAILQ_REMOVE(&vm_page_queues[queue].pl,
270 * Clean the page and remove it from the laundry.
272 * We set the busy bit to cause potential page faults on this page to
273 * block. Note the careful timing, however, the busy bit isn't set till
274 * late and we cannot do anything that will mess with the page.
281 vm_page_t mc[2*vm_pageout_page_count];
283 int ib, is, page_base;
284 vm_pindex_t pindex = m->pindex;
286 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
287 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
290 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
291 * with the new swapper, but we could have serious problems paging
292 * out other object types if there is insufficient memory.
294 * Unfortunately, checking free memory here is far too late, so the
295 * check has been moved up a procedural level.
299 * Can't clean the page if it's busy or held.
301 if ((m->hold_count != 0) ||
302 ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
306 mc[vm_pageout_page_count] = m;
308 page_base = vm_pageout_page_count;
313 * Scan object for clusterable pages.
315 * We can cluster ONLY if: ->> the page is NOT
316 * clean, wired, busy, held, or mapped into a
317 * buffer, and one of the following:
318 * 1) The page is inactive, or a seldom used
321 * 2) we force the issue.
323 * During heavy mmap/modification loads the pageout
324 * daemon can really fragment the underlying file
325 * due to flushing pages out of order and not trying
326 * align the clusters (which leave sporatic out-of-order
327 * holes). To solve this problem we do the reverse scan
328 * first and attempt to align our cluster, then do a
329 * forward scan if room remains.
333 while (ib && pageout_count < vm_pageout_page_count) {
341 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
345 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
346 (p->oflags & VPO_BUSY) || p->busy) {
350 vm_page_test_dirty(p);
351 if ((p->dirty & p->valid) == 0 ||
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 (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
376 (p->oflags & VPO_BUSY) || p->busy) {
379 vm_page_test_dirty(p);
380 if ((p->dirty & p->valid) == 0 ||
381 p->queue != PQ_INACTIVE ||
382 p->wire_count != 0 || /* may be held by buf cache */
383 p->hold_count != 0) { /* may be undergoing I/O */
386 mc[page_base + pageout_count] = p;
392 * If we exhausted our forward scan, continue with the reverse scan
393 * when possible, even past a page boundry. This catches boundry
396 if (ib && pageout_count < vm_pageout_page_count)
400 * we allow reads during pageouts...
402 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
406 * vm_pageout_flush() - launder the given pages
408 * The given pages are laundered. Note that we setup for the start of
409 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
410 * reference count all in here rather then in the parent. If we want
411 * the parent to do more sophisticated things we may have to change
415 vm_pageout_flush(vm_page_t *mc, int count, int flags)
417 vm_object_t object = mc[0]->object;
418 int pageout_status[count];
422 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
423 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
425 * Initiate I/O. Bump the vm_page_t->busy counter and
426 * mark the pages read-only.
428 * We do not have to fixup the clean/dirty bits here... we can
429 * allow the pager to do it after the I/O completes.
431 * NOTE! mc[i]->dirty may be partial or fragmented due to an
432 * edge case with file fragments.
434 for (i = 0; i < count; i++) {
435 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
436 ("vm_pageout_flush: partially invalid page %p index %d/%d",
438 vm_page_io_start(mc[i]);
439 pmap_remove_write(mc[i]);
441 vm_page_unlock_queues();
442 vm_object_pip_add(object, count);
444 vm_pager_put_pages(object, mc, count, flags, pageout_status);
446 vm_page_lock_queues();
447 for (i = 0; i < count; i++) {
448 vm_page_t mt = mc[i];
450 KASSERT((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
463 pmap_clear_modify(mt);
469 * If page couldn't be paged out, then reactivate the
470 * page so it doesn't clog the inactive list. (We
471 * will try paging out it again later).
473 vm_page_activate(mt);
480 * If the operation is still going, leave the page busy to
481 * block all other accesses. Also, leave the paging in
482 * progress indicator set so that we don't attempt an object
485 if (pageout_status[i] != VM_PAGER_PEND) {
486 vm_object_pip_wakeup(object);
487 vm_page_io_finish(mt);
488 if (vm_page_count_severe())
489 vm_page_try_to_cache(mt);
495 #if !defined(NO_SWAPPING)
497 * vm_pageout_object_deactivate_pages
499 * deactivate enough pages to satisfy the inactive target
500 * requirements or if vm_page_proc_limit is set, then
501 * deactivate all of the pages in the object and its
504 * The object and map must be locked.
507 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
509 vm_object_t first_object;
512 vm_object_t backing_object, object;
514 int actcount, rcount, remove_mode;
516 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
517 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
519 for (object = first_object;; object = backing_object) {
520 if (pmap_resident_count(pmap) <= desired)
522 if (object->paging_in_progress)
526 if (object->shadow_count > 1)
529 * scan the objects entire memory queue
531 rcount = object->resident_page_count;
532 p = TAILQ_FIRST(&object->memq);
533 vm_page_lock_queues();
534 while (p && (rcount-- > 0)) {
535 if (pmap_resident_count(pmap) <= desired) {
536 vm_page_unlock_queues();
539 next = TAILQ_NEXT(p, listq);
541 if (p->wire_count != 0 ||
542 p->hold_count != 0 ||
544 (p->oflags & VPO_BUSY) ||
545 (p->flags & PG_UNMANAGED) ||
546 !pmap_page_exists_quick(pmap, p)) {
550 actcount = pmap_ts_referenced(p);
552 vm_page_flag_set(p, PG_REFERENCED);
553 } else if (p->flags & PG_REFERENCED) {
556 if ((p->queue != PQ_ACTIVE) &&
557 (p->flags & PG_REFERENCED)) {
559 p->act_count += actcount;
560 vm_page_flag_clear(p, PG_REFERENCED);
561 } else if (p->queue == PQ_ACTIVE) {
562 if ((p->flags & PG_REFERENCED) == 0) {
563 p->act_count -= min(p->act_count, ACT_DECLINE);
564 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
566 vm_page_deactivate(p);
572 vm_page_flag_clear(p, PG_REFERENCED);
573 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
574 p->act_count += ACT_ADVANCE;
577 } else if (p->queue == PQ_INACTIVE) {
582 vm_page_unlock_queues();
583 if ((backing_object = object->backing_object) == NULL)
585 VM_OBJECT_LOCK(backing_object);
586 if (object != first_object)
587 VM_OBJECT_UNLOCK(object);
590 if (object != first_object)
591 VM_OBJECT_UNLOCK(object);
595 * deactivate some number of pages in a map, try to do it fairly, but
596 * that is really hard to do.
599 vm_pageout_map_deactivate_pages(map, desired)
604 vm_object_t obj, bigobj;
607 if (!vm_map_trylock(map))
614 * first, search out the biggest object, and try to free pages from
617 tmpe = map->header.next;
618 while (tmpe != &map->header) {
619 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
620 obj = tmpe->object.vm_object;
621 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
622 if (obj->shadow_count <= 1 &&
624 bigobj->resident_page_count < obj->resident_page_count)) {
626 VM_OBJECT_UNLOCK(bigobj);
629 VM_OBJECT_UNLOCK(obj);
632 if (tmpe->wired_count > 0)
633 nothingwired = FALSE;
637 if (bigobj != NULL) {
638 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
639 VM_OBJECT_UNLOCK(bigobj);
642 * Next, hunt around for other pages to deactivate. We actually
643 * do this search sort of wrong -- .text first is not the best idea.
645 tmpe = map->header.next;
646 while (tmpe != &map->header) {
647 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
649 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
650 obj = tmpe->object.vm_object;
653 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
654 VM_OBJECT_UNLOCK(obj);
661 * Remove all mappings if a process is swapped out, this will free page
664 if (desired == 0 && nothingwired) {
665 pmap_remove(vm_map_pmap(map), vm_map_min(map),
670 #endif /* !defined(NO_SWAPPING) */
673 * vm_pageout_scan does the dirty work for the pageout daemon.
676 vm_pageout_scan(int pass)
679 struct vm_page marker;
680 int page_shortage, maxscan, pcount;
681 int addl_page_shortage, addl_page_shortage_init;
682 struct proc *p, *bigproc;
684 vm_offset_t size, bigsize;
687 int vnodes_skipped = 0;
691 * Decrease registered cache sizes.
693 EVENTHANDLER_INVOKE(vm_lowmem, 0);
695 * We do this explicitly after the caches have been drained above.
699 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
702 * Calculate the number of pages we want to either free or move
705 page_shortage = vm_paging_target() + addl_page_shortage_init;
708 * Initialize our marker
710 bzero(&marker, sizeof(marker));
711 marker.flags = PG_FICTITIOUS | PG_MARKER;
712 marker.oflags = VPO_BUSY;
713 marker.queue = PQ_INACTIVE;
714 marker.wire_count = 1;
717 * Start scanning the inactive queue for pages we can move to the
718 * cache or free. The scan will stop when the target is reached or
719 * we have scanned the entire inactive queue. Note that m->act_count
720 * is not used to form decisions for the inactive queue, only for the
723 * maxlaunder limits the number of dirty pages we flush per scan.
724 * For most systems a smaller value (16 or 32) is more robust under
725 * extreme memory and disk pressure because any unnecessary writes
726 * to disk can result in extreme performance degredation. However,
727 * systems with excessive dirty pages (especially when MAP_NOSYNC is
728 * used) will die horribly with limited laundering. If the pageout
729 * daemon cannot clean enough pages in the first pass, we let it go
730 * all out in succeeding passes.
732 if ((maxlaunder = vm_max_launder) <= 1)
736 vm_page_lock_queues();
738 addl_page_shortage = addl_page_shortage_init;
739 maxscan = cnt.v_inactive_count;
741 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
742 m != NULL && maxscan-- > 0 && page_shortage > 0;
747 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
751 next = TAILQ_NEXT(m, pageq);
757 if (m->flags & PG_MARKER)
761 * A held page may be undergoing I/O, so skip it.
765 addl_page_shortage++;
769 * Don't mess with busy pages, keep in the front of the
770 * queue, most likely are being paged out.
772 if (!VM_OBJECT_TRYLOCK(object) &&
773 (!vm_pageout_fallback_object_lock(m, &next) ||
774 m->hold_count != 0)) {
775 VM_OBJECT_UNLOCK(object);
776 addl_page_shortage++;
779 if (m->busy || (m->oflags & VPO_BUSY)) {
780 VM_OBJECT_UNLOCK(object);
781 addl_page_shortage++;
786 * If the object is not being used, we ignore previous
789 if (object->ref_count == 0) {
790 vm_page_flag_clear(m, PG_REFERENCED);
791 pmap_clear_reference(m);
794 * Otherwise, if the page has been referenced while in the
795 * inactive queue, we bump the "activation count" upwards,
796 * making it less likely that the page will be added back to
797 * the inactive queue prematurely again. Here we check the
798 * page tables (or emulated bits, if any), given the upper
799 * level VM system not knowing anything about existing
802 } else if (((m->flags & PG_REFERENCED) == 0) &&
803 (actcount = pmap_ts_referenced(m))) {
805 VM_OBJECT_UNLOCK(object);
806 m->act_count += (actcount + ACT_ADVANCE);
811 * If the upper level VM system knows about any page
812 * references, we activate the page. We also set the
813 * "activation count" higher than normal so that we will less
814 * likely place pages back onto the inactive queue again.
816 if ((m->flags & PG_REFERENCED) != 0) {
817 vm_page_flag_clear(m, PG_REFERENCED);
818 actcount = pmap_ts_referenced(m);
820 VM_OBJECT_UNLOCK(object);
821 m->act_count += (actcount + ACT_ADVANCE + 1);
826 * If the upper level VM system doesn't know anything about
827 * the page being dirty, we have to check for it again. As
828 * far as the VM code knows, any partially dirty pages are
831 if (m->dirty == 0 && !pmap_is_modified(m)) {
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 ((m->flags & PG_WRITEABLE) != 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) {
949 ("vm_pageout_scan: mp != NULL"));
951 if (object->flags & OBJ_MIGHTBEDIRTY)
953 goto unlock_and_continue;
955 vm_page_unlock_queues();
956 vm_object_reference_locked(object);
957 VM_OBJECT_UNLOCK(object);
958 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
959 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
961 VM_OBJECT_LOCK(object);
962 vm_page_lock_queues();
964 if (object->flags & OBJ_MIGHTBEDIRTY)
967 goto unlock_and_continue;
969 VM_OBJECT_LOCK(object);
970 vm_page_lock_queues();
972 * The page might have been moved to another
973 * queue during potential blocking in vget()
974 * above. The page might have been freed and
975 * reused for another vnode.
977 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
978 m->object != object ||
979 TAILQ_NEXT(m, pageq) != &marker) {
980 if (object->flags & OBJ_MIGHTBEDIRTY)
982 goto unlock_and_continue;
986 * The page may have been busied during the
987 * blocking in vget(). We don't move the
988 * page back onto the end of the queue so that
989 * statistics are more correct if we don't.
991 if (m->busy || (m->oflags & VPO_BUSY)) {
992 goto unlock_and_continue;
996 * If the page has become held it might
997 * be undergoing I/O, so skip it
1000 vm_pageq_requeue(m);
1001 if (object->flags & OBJ_MIGHTBEDIRTY)
1003 goto unlock_and_continue;
1008 * If a page is dirty, then it is either being washed
1009 * (but not yet cleaned) or it is still in the
1010 * laundry. If it is still in the laundry, then we
1011 * start the cleaning operation.
1013 * decrement page_shortage on success to account for
1014 * the (future) cleaned page. Otherwise we could wind
1015 * up laundering or cleaning too many pages.
1017 if (vm_pageout_clean(m) != 0) {
1021 unlock_and_continue:
1022 VM_OBJECT_UNLOCK(object);
1024 vm_page_unlock_queues();
1027 VFS_UNLOCK_GIANT(vfslocked);
1028 vm_object_deallocate(object);
1029 vn_finished_write(mp);
1030 vm_page_lock_queues();
1032 next = TAILQ_NEXT(&marker, pageq);
1033 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1037 VM_OBJECT_UNLOCK(object);
1041 * Compute the number of pages we want to try to move from the
1042 * active queue to the inactive queue.
1044 page_shortage = vm_paging_target() +
1045 cnt.v_inactive_target - cnt.v_inactive_count;
1046 page_shortage += addl_page_shortage;
1049 * Scan the active queue for things we can deactivate. We nominally
1050 * track the per-page activity counter and use it to locate
1051 * deactivation candidates.
1053 pcount = cnt.v_active_count;
1054 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1056 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1058 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1059 ("vm_pageout_scan: page %p isn't active", m));
1061 next = TAILQ_NEXT(m, pageq);
1063 if ((m->flags & PG_MARKER) != 0) {
1067 if (!VM_OBJECT_TRYLOCK(object) &&
1068 !vm_pageout_fallback_object_lock(m, &next)) {
1069 VM_OBJECT_UNLOCK(object);
1075 * Don't deactivate pages that are busy.
1077 if ((m->busy != 0) ||
1078 (m->oflags & VPO_BUSY) ||
1079 (m->hold_count != 0)) {
1080 VM_OBJECT_UNLOCK(object);
1081 vm_pageq_requeue(m);
1087 * The count for pagedaemon pages is done after checking the
1088 * page for eligibility...
1093 * Check to see "how much" the page has been used.
1096 if (object->ref_count != 0) {
1097 if (m->flags & PG_REFERENCED) {
1100 actcount += pmap_ts_referenced(m);
1102 m->act_count += ACT_ADVANCE + actcount;
1103 if (m->act_count > ACT_MAX)
1104 m->act_count = ACT_MAX;
1109 * Since we have "tested" this bit, we need to clear it now.
1111 vm_page_flag_clear(m, PG_REFERENCED);
1114 * Only if an object is currently being used, do we use the
1115 * page activation count stats.
1117 if (actcount && (object->ref_count != 0)) {
1118 vm_pageq_requeue(m);
1120 m->act_count -= min(m->act_count, ACT_DECLINE);
1121 if (vm_pageout_algorithm ||
1122 object->ref_count == 0 ||
1123 m->act_count == 0) {
1125 if (object->ref_count == 0) {
1130 vm_page_deactivate(m);
1132 vm_page_deactivate(m);
1135 vm_pageq_requeue(m);
1138 VM_OBJECT_UNLOCK(object);
1143 * We try to maintain some *really* free pages, this allows interrupt
1144 * code to be guaranteed space. Since both cache and free queues
1145 * are considered basically 'free', moving pages from cache to free
1146 * does not effect other calculations.
1148 while (cnt.v_free_count < cnt.v_free_reserved) {
1149 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE].pl, pageq) {
1150 KASSERT(m->dirty == 0,
1151 ("Found dirty cache page %p", m));
1152 KASSERT(!pmap_page_is_mapped(m),
1153 ("Found mapped cache page %p", m));
1154 KASSERT((m->flags & PG_UNMANAGED) == 0,
1155 ("Found unmanaged cache page %p", m));
1156 KASSERT(m->wire_count == 0,
1157 ("Found wired cache page %p", m));
1158 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1160 KASSERT((m->oflags & VPO_BUSY) == 0 &&
1161 m->busy == 0, ("Found busy cache page %p",
1164 VM_OBJECT_UNLOCK(object);
1172 vm_page_unlock_queues();
1173 #if !defined(NO_SWAPPING)
1175 * Idle process swapout -- run once per second.
1177 if (vm_swap_idle_enabled) {
1179 if (time_second != lsec) {
1180 vm_req_vmdaemon(VM_SWAP_IDLE);
1187 * If we didn't get enough free pages, and we have skipped a vnode
1188 * in a writeable object, wakeup the sync daemon. And kick swapout
1189 * if we did not get enough free pages.
1191 if (vm_paging_target() > 0) {
1192 if (vnodes_skipped && vm_page_count_min())
1193 (void) speedup_syncer();
1194 #if !defined(NO_SWAPPING)
1195 if (vm_swap_enabled && vm_page_count_target())
1196 vm_req_vmdaemon(VM_SWAP_NORMAL);
1201 * If we are critically low on one of RAM or swap and low on
1202 * the other, kill the largest process. However, we avoid
1203 * doing this on the first pass in order to give ourselves a
1204 * chance to flush out dirty vnode-backed pages and to allow
1205 * active pages to be moved to the inactive queue and reclaimed.
1207 * We keep the process bigproc locked once we find it to keep anyone
1208 * from messing with it; however, there is a possibility of
1209 * deadlock if process B is bigproc and one of it's child processes
1210 * attempts to propagate a signal to B while we are waiting for A's
1211 * lock while walking this list. To avoid this, we don't block on
1212 * the process lock but just skip a process if it is already locked.
1215 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1216 (swap_pager_full && vm_paging_target() > 0))) {
1219 sx_slock(&allproc_lock);
1220 FOREACH_PROC_IN_SYSTEM(p) {
1223 if (PROC_TRYLOCK(p) == 0)
1226 * If this is a system or protected process, skip it.
1228 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1229 (p->p_flag & P_PROTECTED) ||
1230 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1235 * If the process is in a non-running type state,
1236 * don't touch it. Check all the threads individually.
1240 FOREACH_THREAD_IN_PROC(p, td) {
1242 if (!TD_ON_RUNQ(td) &&
1243 !TD_IS_RUNNING(td) &&
1244 !TD_IS_SLEEPING(td)) {
1257 * get the process size
1259 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1263 size = vmspace_swap_count(p->p_vmspace);
1264 vm_map_unlock_read(&p->p_vmspace->vm_map);
1265 size += vmspace_resident_count(p->p_vmspace);
1267 * if the this process is bigger than the biggest one
1270 if (size > bigsize) {
1271 if (bigproc != NULL)
1272 PROC_UNLOCK(bigproc);
1278 sx_sunlock(&allproc_lock);
1279 if (bigproc != NULL) {
1280 killproc(bigproc, "out of swap space");
1281 PROC_SLOCK(bigproc);
1282 sched_nice(bigproc, PRIO_MIN);
1283 PROC_SUNLOCK(bigproc);
1284 PROC_UNLOCK(bigproc);
1285 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)
1541 vm_req_vmdaemon(int req)
1543 static int lastrun = 0;
1545 mtx_lock(&vm_daemon_mtx);
1546 vm_pageout_req_swapout |= req;
1547 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1548 wakeup(&vm_daemon_needed);
1551 mtx_unlock(&vm_daemon_mtx);
1557 struct rlimit rsslim;
1560 int breakout, swapout_flags;
1563 mtx_lock(&vm_daemon_mtx);
1564 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1565 swapout_flags = vm_pageout_req_swapout;
1566 vm_pageout_req_swapout = 0;
1567 mtx_unlock(&vm_daemon_mtx);
1569 swapout_procs(swapout_flags);
1572 * scan the processes for exceeding their rlimits or if
1573 * process is swapped out -- deactivate pages
1575 sx_slock(&allproc_lock);
1576 FOREACH_PROC_IN_SYSTEM(p) {
1577 vm_pindex_t limit, size;
1580 * if this is a system process or if we have already
1581 * looked at this process, skip it.
1584 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1589 * if the process is in a non-running type state,
1594 FOREACH_THREAD_IN_PROC(p, td) {
1596 if (!TD_ON_RUNQ(td) &&
1597 !TD_IS_RUNNING(td) &&
1598 !TD_IS_SLEEPING(td)) {
1613 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1615 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1618 * let processes that are swapped out really be
1619 * swapped out set the limit to nothing (will force a
1622 if ((p->p_sflag & PS_INMEM) == 0)
1623 limit = 0; /* XXX */
1626 size = vmspace_resident_count(p->p_vmspace);
1627 if (limit >= 0 && size >= limit) {
1628 vm_pageout_map_deactivate_pages(
1629 &p->p_vmspace->vm_map, limit);
1632 sx_sunlock(&allproc_lock);
1635 #endif /* !defined(NO_SWAPPING) */