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_init_marker(vm_page_t marker, u_short queue)
222 bzero(marker, sizeof(*marker));
223 marker->flags = PG_FICTITIOUS | PG_MARKER;
224 marker->oflags = VPO_BUSY;
225 marker->queue = queue;
226 marker->wire_count = 1;
230 * vm_pageout_fallback_object_lock:
232 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
233 * known to have failed and page queue must be either PQ_ACTIVE or
234 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
235 * while locking the vm object. Use marker page to detect page queue
236 * changes and maintain notion of next page on page queue. Return
237 * TRUE if no changes were detected, FALSE otherwise. vm object is
240 * This function depends on both the lock portion of struct vm_object
241 * and normal struct vm_page being type stable.
244 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
246 struct vm_page marker;
252 vm_pageout_init_marker(&marker, queue);
255 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
257 vm_page_unlock_queues();
259 VM_OBJECT_LOCK(object);
261 vm_page_lock_queues();
263 /* Page queue might have changed. */
264 *next = TAILQ_NEXT(&marker, pageq);
265 unchanged = (m->queue == queue &&
266 m->object == object &&
267 &marker == TAILQ_NEXT(m, pageq));
268 TAILQ_REMOVE(&vm_page_queues[queue].pl,
274 * Lock the page while holding the page queue lock. Use marker page
275 * to detect page queue changes and maintain notion of next page on
276 * page queue. Return TRUE if no changes were detected, FALSE
277 * otherwise. The page is locked on return. The page queue lock might
278 * be dropped and reacquired.
280 * This function depends on normal struct vm_page being type stable.
283 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
285 struct vm_page marker;
289 vm_page_lock_assert(m, MA_NOTOWNED);
290 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
292 if (vm_page_trylock(m))
296 vm_pageout_init_marker(&marker, queue);
298 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq);
299 vm_page_unlock_queues();
301 vm_page_lock_queues();
303 /* Page queue might have changed. */
304 *next = TAILQ_NEXT(&marker, pageq);
305 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
306 TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq);
313 * Clean the page and remove it from the laundry.
315 * We set the busy bit to cause potential page faults on this page to
316 * block. Note the careful timing, however, the busy bit isn't set till
317 * late and we cannot do anything that will mess with the page.
320 vm_pageout_clean(vm_page_t m)
323 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
325 int ib, is, page_base;
326 vm_pindex_t pindex = m->pindex;
328 vm_page_lock_assert(m, MA_OWNED);
330 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
333 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
334 * with the new swapper, but we could have serious problems paging
335 * out other object types if there is insufficient memory.
337 * Unfortunately, checking free memory here is far too late, so the
338 * check has been moved up a procedural level.
342 * Can't clean the page if it's busy or held.
344 KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
345 ("vm_pageout_clean: page %p is busy", m));
346 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
349 mc[vm_pageout_page_count] = pb = ps = m;
351 page_base = vm_pageout_page_count;
356 * Scan object for clusterable pages.
358 * We can cluster ONLY if: ->> the page is NOT
359 * clean, wired, busy, held, or mapped into a
360 * buffer, and one of the following:
361 * 1) The page is inactive, or a seldom used
364 * 2) we force the issue.
366 * During heavy mmap/modification loads the pageout
367 * daemon can really fragment the underlying file
368 * due to flushing pages out of order and not trying
369 * align the clusters (which leave sporatic out-of-order
370 * holes). To solve this problem we do the reverse scan
371 * first and attempt to align our cluster, then do a
372 * forward scan if room remains.
375 while (ib && pageout_count < vm_pageout_page_count) {
383 if ((p = vm_page_prev(pb)) == NULL ||
384 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
389 vm_page_test_dirty(p);
391 p->queue != PQ_INACTIVE ||
392 p->hold_count != 0) { /* may be undergoing I/O */
398 mc[--page_base] = pb = p;
402 * alignment boundry, stop here and switch directions. Do
405 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
409 while (pageout_count < vm_pageout_page_count &&
410 pindex + is < object->size) {
413 if ((p = vm_page_next(ps)) == NULL ||
414 (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
417 vm_page_test_dirty(p);
419 p->queue != PQ_INACTIVE ||
420 p->hold_count != 0) { /* may be undergoing I/O */
425 mc[page_base + pageout_count] = ps = p;
431 * If we exhausted our forward scan, continue with the reverse scan
432 * when possible, even past a page boundry. This catches boundry
435 if (ib && pageout_count < vm_pageout_page_count)
439 * we allow reads during pageouts...
441 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL));
445 * vm_pageout_flush() - launder the given pages
447 * The given pages are laundered. Note that we setup for the start of
448 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
449 * reference count all in here rather then in the parent. If we want
450 * the parent to do more sophisticated things we may have to change
453 * Returned runlen is the count of pages between mreq and first
454 * page after mreq with status VM_PAGER_AGAIN.
457 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen)
459 vm_object_t object = mc[0]->object;
460 int pageout_status[count];
464 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
465 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
468 * Initiate I/O. Bump the vm_page_t->busy counter and
469 * mark the pages read-only.
471 * We do not have to fixup the clean/dirty bits here... we can
472 * allow the pager to do it after the I/O completes.
474 * NOTE! mc[i]->dirty may be partial or fragmented due to an
475 * edge case with file fragments.
477 for (i = 0; i < count; i++) {
478 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
479 ("vm_pageout_flush: partially invalid page %p index %d/%d",
481 vm_page_io_start(mc[i]);
482 pmap_remove_write(mc[i]);
484 vm_object_pip_add(object, count);
486 vm_pager_put_pages(object, mc, count, flags, pageout_status);
488 runlen = count - mreq;
489 for (i = 0; i < count; i++) {
490 vm_page_t mt = mc[i];
492 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
493 (mt->flags & PG_WRITEABLE) == 0,
494 ("vm_pageout_flush: page %p is not write protected", mt));
495 switch (pageout_status[i]) {
502 * Page outside of range of object. Right now we
503 * essentially lose the changes by pretending it
511 * If page couldn't be paged out, then reactivate the
512 * page so it doesn't clog the inactive list. (We
513 * will try paging out it again later).
516 vm_page_activate(mt);
520 if (i >= mreq && i - mreq < runlen)
526 * If the operation is still going, leave the page busy to
527 * block all other accesses. Also, leave the paging in
528 * progress indicator set so that we don't attempt an object
531 if (pageout_status[i] != VM_PAGER_PEND) {
532 vm_object_pip_wakeup(object);
533 vm_page_io_finish(mt);
534 if (vm_page_count_severe()) {
536 vm_page_try_to_cache(mt);
543 return (numpagedout);
546 #if !defined(NO_SWAPPING)
548 * vm_pageout_object_deactivate_pages
550 * Deactivate enough pages to satisfy the inactive target
553 * The object and map must be locked.
556 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
559 vm_object_t backing_object, object;
561 int actcount, remove_mode;
563 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
564 if (first_object->type == OBJT_DEVICE ||
565 first_object->type == OBJT_SG)
567 for (object = first_object;; object = backing_object) {
568 if (pmap_resident_count(pmap) <= desired)
570 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
571 if (object->type == OBJT_PHYS || object->paging_in_progress)
575 if (object->shadow_count > 1)
578 * Scan the object's entire memory queue.
580 TAILQ_FOREACH(p, &object->memq, listq) {
581 if (pmap_resident_count(pmap) <= desired)
583 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
585 PCPU_INC(cnt.v_pdpages);
587 if (p->wire_count != 0 || p->hold_count != 0 ||
588 !pmap_page_exists_quick(pmap, p)) {
592 actcount = pmap_ts_referenced(p);
593 if ((p->flags & PG_REFERENCED) != 0) {
596 vm_page_lock_queues();
597 vm_page_flag_clear(p, PG_REFERENCED);
598 vm_page_unlock_queues();
600 if (p->queue != PQ_ACTIVE && actcount != 0) {
602 p->act_count += actcount;
603 } else if (p->queue == PQ_ACTIVE) {
605 p->act_count -= min(p->act_count,
608 (vm_pageout_algorithm ||
609 p->act_count == 0)) {
611 vm_page_deactivate(p);
613 vm_page_lock_queues();
615 vm_page_unlock_queues();
619 if (p->act_count < ACT_MAX -
621 p->act_count += ACT_ADVANCE;
622 vm_page_lock_queues();
624 vm_page_unlock_queues();
626 } else if (p->queue == PQ_INACTIVE)
630 if ((backing_object = object->backing_object) == NULL)
632 VM_OBJECT_LOCK(backing_object);
633 if (object != first_object)
634 VM_OBJECT_UNLOCK(object);
637 if (object != first_object)
638 VM_OBJECT_UNLOCK(object);
642 * deactivate some number of pages in a map, try to do it fairly, but
643 * that is really hard to do.
646 vm_pageout_map_deactivate_pages(map, desired)
651 vm_object_t obj, bigobj;
654 if (!vm_map_trylock(map))
661 * first, search out the biggest object, and try to free pages from
664 tmpe = map->header.next;
665 while (tmpe != &map->header) {
666 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
667 obj = tmpe->object.vm_object;
668 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
669 if (obj->shadow_count <= 1 &&
671 bigobj->resident_page_count < obj->resident_page_count)) {
673 VM_OBJECT_UNLOCK(bigobj);
676 VM_OBJECT_UNLOCK(obj);
679 if (tmpe->wired_count > 0)
680 nothingwired = FALSE;
684 if (bigobj != NULL) {
685 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
686 VM_OBJECT_UNLOCK(bigobj);
689 * Next, hunt around for other pages to deactivate. We actually
690 * do this search sort of wrong -- .text first is not the best idea.
692 tmpe = map->header.next;
693 while (tmpe != &map->header) {
694 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
696 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
697 obj = tmpe->object.vm_object;
700 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
701 VM_OBJECT_UNLOCK(obj);
708 * Remove all mappings if a process is swapped out, this will free page
711 if (desired == 0 && nothingwired) {
712 tmpe = map->header.next;
713 while (tmpe != &map->header) {
714 pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end);
720 #endif /* !defined(NO_SWAPPING) */
723 * vm_pageout_scan does the dirty work for the pageout daemon.
726 vm_pageout_scan(int pass)
729 struct vm_page marker;
730 int page_shortage, maxscan, pcount;
731 int addl_page_shortage, addl_page_shortage_init;
734 int vnodes_skipped = 0;
738 * Decrease registered cache sizes.
740 EVENTHANDLER_INVOKE(vm_lowmem, 0);
742 * We do this explicitly after the caches have been drained above.
746 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
749 * Calculate the number of pages we want to either free or move
752 page_shortage = vm_paging_target() + addl_page_shortage_init;
754 vm_pageout_init_marker(&marker, PQ_INACTIVE);
757 * Start scanning the inactive queue for pages we can move to the
758 * cache or free. The scan will stop when the target is reached or
759 * we have scanned the entire inactive queue. Note that m->act_count
760 * is not used to form decisions for the inactive queue, only for the
763 * maxlaunder limits the number of dirty pages we flush per scan.
764 * For most systems a smaller value (16 or 32) is more robust under
765 * extreme memory and disk pressure because any unnecessary writes
766 * to disk can result in extreme performance degredation. However,
767 * systems with excessive dirty pages (especially when MAP_NOSYNC is
768 * used) will die horribly with limited laundering. If the pageout
769 * daemon cannot clean enough pages in the first pass, we let it go
770 * all out in succeeding passes.
772 if ((maxlaunder = vm_max_launder) <= 1)
776 vm_page_lock_queues();
778 addl_page_shortage = addl_page_shortage_init;
779 maxscan = cnt.v_inactive_count;
781 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
782 m != NULL && maxscan-- > 0 && page_shortage > 0;
787 if (m->queue != PQ_INACTIVE)
790 next = TAILQ_NEXT(m, pageq);
795 if (m->flags & PG_MARKER)
801 if (!vm_pageout_page_lock(m, &next)) {
803 addl_page_shortage++;
808 * A held page may be undergoing I/O, so skip it.
813 addl_page_shortage++;
818 * Don't mess with busy pages, keep in the front of the
819 * queue, most likely are being paged out.
822 if (!VM_OBJECT_TRYLOCK(object) &&
823 (!vm_pageout_fallback_object_lock(m, &next) ||
824 m->hold_count != 0)) {
825 VM_OBJECT_UNLOCK(object);
827 addl_page_shortage++;
830 if (m->busy || (m->oflags & VPO_BUSY)) {
832 VM_OBJECT_UNLOCK(object);
833 addl_page_shortage++;
838 * If the object is not being used, we ignore previous
841 if (object->ref_count == 0) {
842 vm_page_flag_clear(m, PG_REFERENCED);
843 KASSERT(!pmap_page_is_mapped(m),
844 ("vm_pageout_scan: page %p is mapped", m));
847 * Otherwise, if the page has been referenced while in the
848 * inactive queue, we bump the "activation count" upwards,
849 * making it less likely that the page will be added back to
850 * the inactive queue prematurely again. Here we check the
851 * page tables (or emulated bits, if any), given the upper
852 * level VM system not knowing anything about existing
855 } else if (((m->flags & PG_REFERENCED) == 0) &&
856 (actcount = pmap_ts_referenced(m))) {
859 m->act_count += actcount + ACT_ADVANCE;
860 VM_OBJECT_UNLOCK(object);
865 * If the upper level VM system knows about any page
866 * references, we activate the page. We also set the
867 * "activation count" higher than normal so that we will less
868 * likely place pages back onto the inactive queue again.
870 if ((m->flags & PG_REFERENCED) != 0) {
871 vm_page_flag_clear(m, PG_REFERENCED);
872 actcount = pmap_ts_referenced(m);
875 m->act_count += actcount + ACT_ADVANCE + 1;
876 VM_OBJECT_UNLOCK(object);
881 * If the upper level VM system does not believe that the page
882 * is fully dirty, but it is mapped for write access, then we
883 * consult the pmap to see if the page's dirty status should
886 if (m->dirty != VM_PAGE_BITS_ALL &&
887 (m->flags & PG_WRITEABLE) != 0) {
889 * Avoid a race condition: Unless write access is
890 * removed from the page, another processor could
891 * modify it before all access is removed by the call
892 * to vm_page_cache() below. If vm_page_cache() finds
893 * that the page has been modified when it removes all
894 * access, it panics because it cannot cache dirty
895 * pages. In principle, we could eliminate just write
896 * access here rather than all access. In the expected
897 * case, when there are no last instant modifications
898 * to the page, removing all access will be cheaper
901 if (pmap_is_modified(m))
903 else if (m->dirty == 0)
909 * Invalid pages can be easily freed
914 } else if (m->dirty == 0) {
916 * Clean pages can be placed onto the cache queue.
917 * This effectively frees them.
921 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
923 * Dirty pages need to be paged out, but flushing
924 * a page is extremely expensive verses freeing
925 * a clean page. Rather then artificially limiting
926 * the number of pages we can flush, we instead give
927 * dirty pages extra priority on the inactive queue
928 * by forcing them to be cycled through the queue
929 * twice before being flushed, after which the
930 * (now clean) page will cycle through once more
931 * before being freed. This significantly extends
932 * the thrash point for a heavily loaded machine.
934 vm_page_flag_set(m, PG_WINATCFLS);
936 } else if (maxlaunder > 0) {
938 * We always want to try to flush some dirty pages if
939 * we encounter them, to keep the system stable.
940 * Normally this number is small, but under extreme
941 * pressure where there are insufficient clean pages
942 * on the inactive queue, we may have to go all out.
944 int swap_pageouts_ok, vfslocked = 0;
945 struct vnode *vp = NULL;
946 struct mount *mp = NULL;
948 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
949 swap_pageouts_ok = 1;
951 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
952 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
953 vm_page_count_min());
958 * We don't bother paging objects that are "dead".
959 * Those objects are in a "rundown" state.
961 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
963 VM_OBJECT_UNLOCK(object);
969 * Following operations may unlock
970 * vm_page_queue_mtx, invalidating the 'next'
971 * pointer. To prevent an inordinate number
972 * of restarts we use our marker to remember
976 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
979 * The object is already known NOT to be dead. It
980 * is possible for the vget() to block the whole
981 * pageout daemon, but the new low-memory handling
982 * code should prevent it.
984 * The previous code skipped locked vnodes and, worse,
985 * reordered pages in the queue. This results in
986 * completely non-deterministic operation and, on a
987 * busy system, can lead to extremely non-optimal
988 * pageouts. For example, it can cause clean pages
989 * to be freed and dirty pages to be moved to the end
990 * of the queue. Since dirty pages are also moved to
991 * the end of the queue once-cleaned, this gives
992 * way too large a weighting to defering the freeing
995 * We can't wait forever for the vnode lock, we might
996 * deadlock due to a vn_read() getting stuck in
997 * vm_wait while holding this vnode. We skip the
998 * vnode if we can't get it in a reasonable amount
1001 if (object->type == OBJT_VNODE) {
1002 vm_page_unlock_queues();
1004 vp = object->handle;
1005 if (vp->v_type == VREG &&
1006 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1008 ++pageout_lock_miss;
1009 if (object->flags & OBJ_MIGHTBEDIRTY)
1011 vm_page_lock_queues();
1012 goto unlock_and_continue;
1015 ("vp %p with NULL v_mount", vp));
1016 vm_object_reference_locked(object);
1017 VM_OBJECT_UNLOCK(object);
1018 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1019 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1021 VM_OBJECT_LOCK(object);
1022 vm_page_lock_queues();
1023 ++pageout_lock_miss;
1024 if (object->flags & OBJ_MIGHTBEDIRTY)
1027 goto unlock_and_continue;
1029 VM_OBJECT_LOCK(object);
1031 vm_page_lock_queues();
1033 * The page might have been moved to another
1034 * queue during potential blocking in vget()
1035 * above. The page might have been freed and
1036 * reused for another vnode.
1038 if (m->queue != PQ_INACTIVE ||
1039 m->object != object ||
1040 TAILQ_NEXT(m, pageq) != &marker) {
1042 if (object->flags & OBJ_MIGHTBEDIRTY)
1044 goto unlock_and_continue;
1048 * The page may have been busied during the
1049 * blocking in vget(). We don't move the
1050 * page back onto the end of the queue so that
1051 * statistics are more correct if we don't.
1053 if (m->busy || (m->oflags & VPO_BUSY)) {
1055 goto unlock_and_continue;
1059 * If the page has become held it might
1060 * be undergoing I/O, so skip it
1062 if (m->hold_count) {
1065 if (object->flags & OBJ_MIGHTBEDIRTY)
1067 goto unlock_and_continue;
1072 * If a page is dirty, then it is either being washed
1073 * (but not yet cleaned) or it is still in the
1074 * laundry. If it is still in the laundry, then we
1075 * start the cleaning operation.
1077 * decrement page_shortage on success to account for
1078 * the (future) cleaned page. Otherwise we could wind
1079 * up laundering or cleaning too many pages.
1081 vm_page_unlock_queues();
1082 if (vm_pageout_clean(m) != 0) {
1086 vm_page_lock_queues();
1087 unlock_and_continue:
1088 vm_page_lock_assert(m, MA_NOTOWNED);
1089 VM_OBJECT_UNLOCK(object);
1091 vm_page_unlock_queues();
1094 VFS_UNLOCK_GIANT(vfslocked);
1095 vm_object_deallocate(object);
1096 vn_finished_write(mp);
1097 vm_page_lock_queues();
1099 next = TAILQ_NEXT(&marker, pageq);
1100 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1102 vm_page_lock_assert(m, MA_NOTOWNED);
1106 VM_OBJECT_UNLOCK(object);
1110 * Compute the number of pages we want to try to move from the
1111 * active queue to the inactive queue.
1113 page_shortage = vm_paging_target() +
1114 cnt.v_inactive_target - cnt.v_inactive_count;
1115 page_shortage += addl_page_shortage;
1118 * Scan the active queue for things we can deactivate. We nominally
1119 * track the per-page activity counter and use it to locate
1120 * deactivation candidates.
1122 pcount = cnt.v_active_count;
1123 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1124 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1126 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1128 KASSERT(m->queue == PQ_ACTIVE,
1129 ("vm_pageout_scan: page %p isn't active", m));
1131 next = TAILQ_NEXT(m, pageq);
1132 if ((m->flags & PG_MARKER) != 0) {
1136 if (!vm_pageout_page_lock(m, &next)) {
1142 if (!VM_OBJECT_TRYLOCK(object) &&
1143 !vm_pageout_fallback_object_lock(m, &next)) {
1144 VM_OBJECT_UNLOCK(object);
1151 * Don't deactivate pages that are busy.
1153 if ((m->busy != 0) ||
1154 (m->oflags & VPO_BUSY) ||
1155 (m->hold_count != 0)) {
1157 VM_OBJECT_UNLOCK(object);
1164 * The count for pagedaemon pages is done after checking the
1165 * page for eligibility...
1170 * Check to see "how much" the page has been used.
1173 if (object->ref_count != 0) {
1174 if (m->flags & PG_REFERENCED) {
1177 actcount += pmap_ts_referenced(m);
1179 m->act_count += ACT_ADVANCE + actcount;
1180 if (m->act_count > ACT_MAX)
1181 m->act_count = ACT_MAX;
1186 * Since we have "tested" this bit, we need to clear it now.
1188 vm_page_flag_clear(m, PG_REFERENCED);
1191 * Only if an object is currently being used, do we use the
1192 * page activation count stats.
1194 if (actcount && (object->ref_count != 0)) {
1197 m->act_count -= min(m->act_count, ACT_DECLINE);
1198 if (vm_pageout_algorithm ||
1199 object->ref_count == 0 ||
1200 m->act_count == 0) {
1202 if (object->ref_count == 0) {
1203 KASSERT(!pmap_page_is_mapped(m),
1204 ("vm_pageout_scan: page %p is mapped", m));
1208 vm_page_deactivate(m);
1210 vm_page_deactivate(m);
1217 VM_OBJECT_UNLOCK(object);
1220 vm_page_unlock_queues();
1221 #if !defined(NO_SWAPPING)
1223 * Idle process swapout -- run once per second.
1225 if (vm_swap_idle_enabled) {
1227 if (time_second != lsec) {
1228 vm_req_vmdaemon(VM_SWAP_IDLE);
1235 * If we didn't get enough free pages, and we have skipped a vnode
1236 * in a writeable object, wakeup the sync daemon. And kick swapout
1237 * if we did not get enough free pages.
1239 if (vm_paging_target() > 0) {
1240 if (vnodes_skipped && vm_page_count_min())
1241 (void) speedup_syncer();
1242 #if !defined(NO_SWAPPING)
1243 if (vm_swap_enabled && vm_page_count_target())
1244 vm_req_vmdaemon(VM_SWAP_NORMAL);
1249 * If we are critically low on one of RAM or swap and low on
1250 * the other, kill the largest process. However, we avoid
1251 * doing this on the first pass in order to give ourselves a
1252 * chance to flush out dirty vnode-backed pages and to allow
1253 * active pages to be moved to the inactive queue and reclaimed.
1256 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1257 (swap_pager_full && vm_paging_target() > 0)))
1258 vm_pageout_oom(VM_OOM_MEM);
1263 vm_pageout_oom(int shortage)
1265 struct proc *p, *bigproc;
1266 vm_offset_t size, bigsize;
1271 * We keep the process bigproc locked once we find it to keep anyone
1272 * from messing with it; however, there is a possibility of
1273 * deadlock if process B is bigproc and one of it's child processes
1274 * attempts to propagate a signal to B while we are waiting for A's
1275 * lock while walking this list. To avoid this, we don't block on
1276 * the process lock but just skip a process if it is already locked.
1280 sx_slock(&allproc_lock);
1281 FOREACH_PROC_IN_SYSTEM(p) {
1284 if (PROC_TRYLOCK(p) == 0)
1287 * If this is a system, protected or killed process, skip it.
1289 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1290 (p->p_pid == 1) || P_KILLED(p) ||
1291 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1296 * If the process is in a non-running type state,
1297 * don't touch it. Check all the threads individually.
1300 FOREACH_THREAD_IN_PROC(p, td) {
1302 if (!TD_ON_RUNQ(td) &&
1303 !TD_IS_RUNNING(td) &&
1304 !TD_IS_SLEEPING(td)) {
1316 * get the process size
1318 vm = vmspace_acquire_ref(p);
1323 if (!vm_map_trylock_read(&vm->vm_map)) {
1328 size = vmspace_swap_count(vm);
1329 vm_map_unlock_read(&vm->vm_map);
1330 if (shortage == VM_OOM_MEM)
1331 size += vmspace_resident_count(vm);
1334 * if the this process is bigger than the biggest one
1337 if (size > bigsize) {
1338 if (bigproc != NULL)
1339 PROC_UNLOCK(bigproc);
1345 sx_sunlock(&allproc_lock);
1346 if (bigproc != NULL) {
1347 killproc(bigproc, "out of swap space");
1348 sched_nice(bigproc, PRIO_MIN);
1349 PROC_UNLOCK(bigproc);
1350 wakeup(&cnt.v_free_count);
1355 * This routine tries to maintain the pseudo LRU active queue,
1356 * so that during long periods of time where there is no paging,
1357 * that some statistic accumulation still occurs. This code
1358 * helps the situation where paging just starts to occur.
1361 vm_pageout_page_stats()
1365 int pcount,tpcount; /* Number of pages to check */
1366 static int fullintervalcount = 0;
1370 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1371 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1373 if (page_shortage <= 0)
1376 vm_page_lock_queues();
1377 pcount = cnt.v_active_count;
1378 fullintervalcount += vm_pageout_stats_interval;
1379 if (fullintervalcount < vm_pageout_full_stats_interval) {
1380 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1382 if (pcount > tpcount)
1385 fullintervalcount = 0;
1388 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1389 while ((m != NULL) && (pcount-- > 0)) {
1392 KASSERT(m->queue == PQ_ACTIVE,
1393 ("vm_pageout_page_stats: page %p isn't active", m));
1395 next = TAILQ_NEXT(m, pageq);
1396 if ((m->flags & PG_MARKER) != 0) {
1400 vm_page_lock_assert(m, MA_NOTOWNED);
1401 if (!vm_pageout_page_lock(m, &next)) {
1407 if (!VM_OBJECT_TRYLOCK(object) &&
1408 !vm_pageout_fallback_object_lock(m, &next)) {
1409 VM_OBJECT_UNLOCK(object);
1416 * Don't deactivate pages that are busy.
1418 if ((m->busy != 0) ||
1419 (m->oflags & VPO_BUSY) ||
1420 (m->hold_count != 0)) {
1422 VM_OBJECT_UNLOCK(object);
1429 if (m->flags & PG_REFERENCED) {
1430 vm_page_flag_clear(m, PG_REFERENCED);
1434 actcount += pmap_ts_referenced(m);
1436 m->act_count += ACT_ADVANCE + actcount;
1437 if (m->act_count > ACT_MAX)
1438 m->act_count = ACT_MAX;
1441 if (m->act_count == 0) {
1443 * We turn off page access, so that we have
1444 * more accurate RSS stats. We don't do this
1445 * in the normal page deactivation when the
1446 * system is loaded VM wise, because the
1447 * cost of the large number of page protect
1448 * operations would be higher than the value
1449 * of doing the operation.
1452 vm_page_deactivate(m);
1454 m->act_count -= min(m->act_count, ACT_DECLINE);
1459 VM_OBJECT_UNLOCK(object);
1462 vm_page_unlock_queues();
1466 * vm_pageout is the high level pageout daemon.
1474 * Initialize some paging parameters.
1476 cnt.v_interrupt_free_min = 2;
1477 if (cnt.v_page_count < 2000)
1478 vm_pageout_page_count = 8;
1481 * v_free_reserved needs to include enough for the largest
1482 * swap pager structures plus enough for any pv_entry structs
1485 if (cnt.v_page_count > 1024)
1486 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1489 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1490 cnt.v_interrupt_free_min;
1491 cnt.v_free_reserved = vm_pageout_page_count +
1492 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1493 cnt.v_free_severe = cnt.v_free_min / 2;
1494 cnt.v_free_min += cnt.v_free_reserved;
1495 cnt.v_free_severe += cnt.v_free_reserved;
1498 * v_free_target and v_cache_min control pageout hysteresis. Note
1499 * that these are more a measure of the VM cache queue hysteresis
1500 * then the VM free queue. Specifically, v_free_target is the
1501 * high water mark (free+cache pages).
1503 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1504 * low water mark, while v_free_min is the stop. v_cache_min must
1505 * be big enough to handle memory needs while the pageout daemon
1506 * is signalled and run to free more pages.
1508 if (cnt.v_free_count > 6144)
1509 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1511 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1513 if (cnt.v_free_count > 2048) {
1514 cnt.v_cache_min = cnt.v_free_target;
1515 cnt.v_cache_max = 2 * cnt.v_cache_min;
1516 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1518 cnt.v_cache_min = 0;
1519 cnt.v_cache_max = 0;
1520 cnt.v_inactive_target = cnt.v_free_count / 4;
1522 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1523 cnt.v_inactive_target = cnt.v_free_count / 3;
1525 /* XXX does not really belong here */
1526 if (vm_page_max_wired == 0)
1527 vm_page_max_wired = cnt.v_free_count / 3;
1529 if (vm_pageout_stats_max == 0)
1530 vm_pageout_stats_max = cnt.v_free_target;
1533 * Set interval in seconds for stats scan.
1535 if (vm_pageout_stats_interval == 0)
1536 vm_pageout_stats_interval = 5;
1537 if (vm_pageout_full_stats_interval == 0)
1538 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1540 swap_pager_swap_init();
1543 * The pageout daemon is never done, so loop forever.
1547 * If we have enough free memory, wakeup waiters. Do
1548 * not clear vm_pages_needed until we reach our target,
1549 * otherwise we may be woken up over and over again and
1550 * waste a lot of cpu.
1552 mtx_lock(&vm_page_queue_free_mtx);
1553 if (vm_pages_needed && !vm_page_count_min()) {
1554 if (!vm_paging_needed())
1555 vm_pages_needed = 0;
1556 wakeup(&cnt.v_free_count);
1558 if (vm_pages_needed) {
1560 * Still not done, take a second pass without waiting
1561 * (unlimited dirty cleaning), otherwise sleep a bit
1566 msleep(&vm_pages_needed,
1567 &vm_page_queue_free_mtx, PVM, "psleep",
1571 * Good enough, sleep & handle stats. Prime the pass
1578 error = msleep(&vm_pages_needed,
1579 &vm_page_queue_free_mtx, PVM, "psleep",
1580 vm_pageout_stats_interval * hz);
1581 if (error && !vm_pages_needed) {
1582 mtx_unlock(&vm_page_queue_free_mtx);
1584 vm_pageout_page_stats();
1588 if (vm_pages_needed)
1590 mtx_unlock(&vm_page_queue_free_mtx);
1591 vm_pageout_scan(pass);
1596 * Unless the free page queue lock is held by the caller, this function
1597 * should be regarded as advisory. Specifically, the caller should
1598 * not msleep() on &cnt.v_free_count following this function unless
1599 * the free page queue lock is held until the msleep() is performed.
1605 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1606 vm_pages_needed = 1;
1607 wakeup(&vm_pages_needed);
1611 #if !defined(NO_SWAPPING)
1613 vm_req_vmdaemon(int req)
1615 static int lastrun = 0;
1617 mtx_lock(&vm_daemon_mtx);
1618 vm_pageout_req_swapout |= req;
1619 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1620 wakeup(&vm_daemon_needed);
1623 mtx_unlock(&vm_daemon_mtx);
1629 struct rlimit rsslim;
1633 int breakout, swapout_flags;
1636 mtx_lock(&vm_daemon_mtx);
1637 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1638 swapout_flags = vm_pageout_req_swapout;
1639 vm_pageout_req_swapout = 0;
1640 mtx_unlock(&vm_daemon_mtx);
1642 swapout_procs(swapout_flags);
1645 * scan the processes for exceeding their rlimits or if
1646 * process is swapped out -- deactivate pages
1648 sx_slock(&allproc_lock);
1649 FOREACH_PROC_IN_SYSTEM(p) {
1650 vm_pindex_t limit, size;
1653 * if this is a system process or if we have already
1654 * looked at this process, skip it.
1657 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1662 * if the process is in a non-running type state,
1666 FOREACH_THREAD_IN_PROC(p, td) {
1668 if (!TD_ON_RUNQ(td) &&
1669 !TD_IS_RUNNING(td) &&
1670 !TD_IS_SLEEPING(td)) {
1684 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1686 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1689 * let processes that are swapped out really be
1690 * swapped out set the limit to nothing (will force a
1693 if ((p->p_flag & P_INMEM) == 0)
1694 limit = 0; /* XXX */
1695 vm = vmspace_acquire_ref(p);
1700 size = vmspace_resident_count(vm);
1701 if (limit >= 0 && size >= limit) {
1702 vm_pageout_map_deactivate_pages(
1703 &vm->vm_map, limit);
1707 sx_sunlock(&allproc_lock);
1710 #endif /* !defined(NO_SWAPPING) */