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);
329 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
332 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
333 * with the new swapper, but we could have serious problems paging
334 * out other object types if there is insufficient memory.
336 * Unfortunately, checking free memory here is far too late, so the
337 * check has been moved up a procedural level.
341 * Can't clean the page if it's busy or held.
343 KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
344 ("vm_pageout_clean: page %p is busy", m));
345 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
347 mc[vm_pageout_page_count] = pb = ps = m;
349 page_base = vm_pageout_page_count;
354 * Scan object for clusterable pages.
356 * We can cluster ONLY if: ->> the page is NOT
357 * clean, wired, busy, held, or mapped into a
358 * buffer, and one of the following:
359 * 1) The page is inactive, or a seldom used
362 * 2) we force the issue.
364 * During heavy mmap/modification loads the pageout
365 * daemon can really fragment the underlying file
366 * due to flushing pages out of order and not trying
367 * align the clusters (which leave sporatic out-of-order
368 * holes). To solve this problem we do the reverse scan
369 * first and attempt to align our cluster, then do a
370 * forward scan if room remains.
374 while (ib && pageout_count < vm_pageout_page_count) {
382 if ((p = vm_page_prev(pb)) == NULL ||
383 (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
388 vm_page_test_dirty(p);
390 p->queue != PQ_INACTIVE ||
391 p->hold_count != 0) { /* may be undergoing I/O */
397 mc[--page_base] = pb = p;
401 * alignment boundry, stop here and switch directions. Do
404 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
408 while (pageout_count < vm_pageout_page_count &&
409 pindex + is < object->size) {
412 if ((p = vm_page_next(ps)) == NULL ||
413 (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
416 vm_page_test_dirty(p);
418 p->queue != PQ_INACTIVE ||
419 p->hold_count != 0) { /* may be undergoing I/O */
424 mc[page_base + pageout_count] = ps = p;
430 * If we exhausted our forward scan, continue with the reverse scan
431 * when possible, even past a page boundry. This catches boundry
434 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));
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
454 vm_pageout_flush(vm_page_t *mc, int count, int flags)
456 vm_object_t object = mc[0]->object;
457 int pageout_status[count];
461 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
462 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
465 * Initiate I/O. Bump the vm_page_t->busy counter and
466 * mark the pages read-only.
468 * We do not have to fixup the clean/dirty bits here... we can
469 * allow the pager to do it after the I/O completes.
471 * NOTE! mc[i]->dirty may be partial or fragmented due to an
472 * edge case with file fragments.
474 for (i = 0; i < count; i++) {
475 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
476 ("vm_pageout_flush: partially invalid page %p index %d/%d",
478 vm_page_io_start(mc[i]);
479 pmap_remove_write(mc[i]);
481 vm_object_pip_add(object, count);
483 vm_pager_put_pages(object, mc, count, flags, pageout_status);
485 for (i = 0; i < count; i++) {
486 vm_page_t mt = mc[i];
488 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
489 (mt->flags & PG_WRITEABLE) == 0,
490 ("vm_pageout_flush: page %p is not write protected", mt));
491 switch (pageout_status[i]) {
498 * Page outside of range of object. Right now we
499 * essentially lose the changes by pretending it
507 * If page couldn't be paged out, then reactivate the
508 * page so it doesn't clog the inactive list. (We
509 * will try paging out it again later).
512 vm_page_activate(mt);
520 * If the operation is still going, leave the page busy to
521 * block all other accesses. Also, leave the paging in
522 * progress indicator set so that we don't attempt an object
525 if (pageout_status[i] != VM_PAGER_PEND) {
526 vm_object_pip_wakeup(object);
527 vm_page_io_finish(mt);
528 if (vm_page_count_severe()) {
530 vm_page_try_to_cache(mt);
535 return (numpagedout);
538 #if !defined(NO_SWAPPING)
540 * vm_pageout_object_deactivate_pages
542 * Deactivate enough pages to satisfy the inactive target
545 * The object and map must be locked.
548 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
551 vm_object_t backing_object, object;
553 int actcount, remove_mode;
555 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
556 if (first_object->type == OBJT_DEVICE ||
557 first_object->type == OBJT_SG)
559 for (object = first_object;; object = backing_object) {
560 if (pmap_resident_count(pmap) <= desired)
562 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
563 if (object->type == OBJT_PHYS || object->paging_in_progress)
567 if (object->shadow_count > 1)
570 * Scan the object's entire memory queue.
572 TAILQ_FOREACH(p, &object->memq, listq) {
573 if (pmap_resident_count(pmap) <= desired)
575 if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
577 PCPU_INC(cnt.v_pdpages);
579 if (p->wire_count != 0 || p->hold_count != 0 ||
580 !pmap_page_exists_quick(pmap, p)) {
584 actcount = pmap_ts_referenced(p);
585 if ((p->flags & PG_REFERENCED) != 0) {
588 vm_page_lock_queues();
589 vm_page_flag_clear(p, PG_REFERENCED);
590 vm_page_unlock_queues();
592 if (p->queue != PQ_ACTIVE && actcount != 0) {
594 p->act_count += actcount;
595 } else if (p->queue == PQ_ACTIVE) {
597 p->act_count -= min(p->act_count,
600 (vm_pageout_algorithm ||
601 p->act_count == 0)) {
603 vm_page_deactivate(p);
605 vm_page_lock_queues();
607 vm_page_unlock_queues();
611 if (p->act_count < ACT_MAX -
613 p->act_count += ACT_ADVANCE;
614 vm_page_lock_queues();
616 vm_page_unlock_queues();
618 } else if (p->queue == PQ_INACTIVE)
622 if ((backing_object = object->backing_object) == NULL)
624 VM_OBJECT_LOCK(backing_object);
625 if (object != first_object)
626 VM_OBJECT_UNLOCK(object);
629 if (object != first_object)
630 VM_OBJECT_UNLOCK(object);
634 * deactivate some number of pages in a map, try to do it fairly, but
635 * that is really hard to do.
638 vm_pageout_map_deactivate_pages(map, desired)
643 vm_object_t obj, bigobj;
646 if (!vm_map_trylock(map))
653 * first, search out the biggest object, and try to free pages from
656 tmpe = map->header.next;
657 while (tmpe != &map->header) {
658 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
659 obj = tmpe->object.vm_object;
660 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
661 if (obj->shadow_count <= 1 &&
663 bigobj->resident_page_count < obj->resident_page_count)) {
665 VM_OBJECT_UNLOCK(bigobj);
668 VM_OBJECT_UNLOCK(obj);
671 if (tmpe->wired_count > 0)
672 nothingwired = FALSE;
676 if (bigobj != NULL) {
677 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
678 VM_OBJECT_UNLOCK(bigobj);
681 * Next, hunt around for other pages to deactivate. We actually
682 * do this search sort of wrong -- .text first is not the best idea.
684 tmpe = map->header.next;
685 while (tmpe != &map->header) {
686 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
688 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
689 obj = tmpe->object.vm_object;
692 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
693 VM_OBJECT_UNLOCK(obj);
700 * Remove all mappings if a process is swapped out, this will free page
703 if (desired == 0 && nothingwired) {
704 tmpe = map->header.next;
705 while (tmpe != &map->header) {
706 pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end);
712 #endif /* !defined(NO_SWAPPING) */
715 * vm_pageout_scan does the dirty work for the pageout daemon.
718 vm_pageout_scan(int pass)
721 struct vm_page marker;
722 int page_shortage, maxscan, pcount;
723 int addl_page_shortage, addl_page_shortage_init;
726 int vnodes_skipped = 0;
730 * Decrease registered cache sizes.
732 EVENTHANDLER_INVOKE(vm_lowmem, 0);
734 * We do this explicitly after the caches have been drained above.
738 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
741 * Calculate the number of pages we want to either free or move
744 page_shortage = vm_paging_target() + addl_page_shortage_init;
746 vm_pageout_init_marker(&marker, PQ_INACTIVE);
749 * Start scanning the inactive queue for pages we can move to the
750 * cache or free. The scan will stop when the target is reached or
751 * we have scanned the entire inactive queue. Note that m->act_count
752 * is not used to form decisions for the inactive queue, only for the
755 * maxlaunder limits the number of dirty pages we flush per scan.
756 * For most systems a smaller value (16 or 32) is more robust under
757 * extreme memory and disk pressure because any unnecessary writes
758 * to disk can result in extreme performance degredation. However,
759 * systems with excessive dirty pages (especially when MAP_NOSYNC is
760 * used) will die horribly with limited laundering. If the pageout
761 * daemon cannot clean enough pages in the first pass, we let it go
762 * all out in succeeding passes.
764 if ((maxlaunder = vm_max_launder) <= 1)
768 vm_page_lock_queues();
770 addl_page_shortage = addl_page_shortage_init;
771 maxscan = cnt.v_inactive_count;
773 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
774 m != NULL && maxscan-- > 0 && page_shortage > 0;
779 if (m->queue != PQ_INACTIVE)
782 next = TAILQ_NEXT(m, pageq);
787 if (m->flags & PG_MARKER)
793 if (!vm_pageout_page_lock(m, &next)) {
795 addl_page_shortage++;
800 * A held page may be undergoing I/O, so skip it.
805 addl_page_shortage++;
810 * Don't mess with busy pages, keep in the front of the
811 * queue, most likely are being paged out.
814 if (!VM_OBJECT_TRYLOCK(object) &&
815 (!vm_pageout_fallback_object_lock(m, &next) ||
816 m->hold_count != 0)) {
817 VM_OBJECT_UNLOCK(object);
819 addl_page_shortage++;
822 if (m->busy || (m->oflags & VPO_BUSY)) {
824 VM_OBJECT_UNLOCK(object);
825 addl_page_shortage++;
830 * If the object is not being used, we ignore previous
833 if (object->ref_count == 0) {
834 vm_page_flag_clear(m, PG_REFERENCED);
835 KASSERT(!pmap_page_is_mapped(m),
836 ("vm_pageout_scan: page %p is mapped", m));
839 * Otherwise, if the page has been referenced while in the
840 * inactive queue, we bump the "activation count" upwards,
841 * making it less likely that the page will be added back to
842 * the inactive queue prematurely again. Here we check the
843 * page tables (or emulated bits, if any), given the upper
844 * level VM system not knowing anything about existing
847 } else if (((m->flags & PG_REFERENCED) == 0) &&
848 (actcount = pmap_ts_referenced(m))) {
850 VM_OBJECT_UNLOCK(object);
851 m->act_count += (actcount + ACT_ADVANCE);
857 * If the upper level VM system knows about any page
858 * references, we activate the page. We also set the
859 * "activation count" higher than normal so that we will less
860 * likely place pages back onto the inactive queue again.
862 if ((m->flags & PG_REFERENCED) != 0) {
863 vm_page_flag_clear(m, PG_REFERENCED);
864 actcount = pmap_ts_referenced(m);
866 VM_OBJECT_UNLOCK(object);
867 m->act_count += (actcount + ACT_ADVANCE + 1);
873 * If the upper level VM system does not believe that the page
874 * is fully dirty, but it is mapped for write access, then we
875 * consult the pmap to see if the page's dirty status should
878 if (m->dirty != VM_PAGE_BITS_ALL &&
879 (m->flags & PG_WRITEABLE) != 0) {
881 * Avoid a race condition: Unless write access is
882 * removed from the page, another processor could
883 * modify it before all access is removed by the call
884 * to vm_page_cache() below. If vm_page_cache() finds
885 * that the page has been modified when it removes all
886 * access, it panics because it cannot cache dirty
887 * pages. In principle, we could eliminate just write
888 * access here rather than all access. In the expected
889 * case, when there are no last instant modifications
890 * to the page, removing all access will be cheaper
893 if (pmap_is_modified(m))
895 else if (m->dirty == 0)
901 * Invalid pages can be easily freed
906 } else if (m->dirty == 0) {
908 * Clean pages can be placed onto the cache queue.
909 * This effectively frees them.
913 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
915 * Dirty pages need to be paged out, but flushing
916 * a page is extremely expensive verses freeing
917 * a clean page. Rather then artificially limiting
918 * the number of pages we can flush, we instead give
919 * dirty pages extra priority on the inactive queue
920 * by forcing them to be cycled through the queue
921 * twice before being flushed, after which the
922 * (now clean) page will cycle through once more
923 * before being freed. This significantly extends
924 * the thrash point for a heavily loaded machine.
926 vm_page_flag_set(m, PG_WINATCFLS);
928 } else if (maxlaunder > 0) {
930 * We always want to try to flush some dirty pages if
931 * we encounter them, to keep the system stable.
932 * Normally this number is small, but under extreme
933 * pressure where there are insufficient clean pages
934 * on the inactive queue, we may have to go all out.
936 int swap_pageouts_ok, vfslocked = 0;
937 struct vnode *vp = NULL;
938 struct mount *mp = NULL;
940 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
941 swap_pageouts_ok = 1;
943 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
944 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
945 vm_page_count_min());
950 * We don't bother paging objects that are "dead".
951 * Those objects are in a "rundown" state.
953 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
955 VM_OBJECT_UNLOCK(object);
961 * Following operations may unlock
962 * vm_page_queue_mtx, invalidating the 'next'
963 * pointer. To prevent an inordinate number
964 * of restarts we use our marker to remember
968 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
971 * The object is already known NOT to be dead. It
972 * is possible for the vget() to block the whole
973 * pageout daemon, but the new low-memory handling
974 * code should prevent it.
976 * The previous code skipped locked vnodes and, worse,
977 * reordered pages in the queue. This results in
978 * completely non-deterministic operation and, on a
979 * busy system, can lead to extremely non-optimal
980 * pageouts. For example, it can cause clean pages
981 * to be freed and dirty pages to be moved to the end
982 * of the queue. Since dirty pages are also moved to
983 * the end of the queue once-cleaned, this gives
984 * way too large a weighting to defering the freeing
987 * We can't wait forever for the vnode lock, we might
988 * deadlock due to a vn_read() getting stuck in
989 * vm_wait while holding this vnode. We skip the
990 * vnode if we can't get it in a reasonable amount
993 if (object->type == OBJT_VNODE) {
994 vm_page_unlock_queues();
997 if (vp->v_type == VREG &&
998 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1000 ++pageout_lock_miss;
1001 if (object->flags & OBJ_MIGHTBEDIRTY)
1003 vm_page_lock_queues();
1004 goto unlock_and_continue;
1007 ("vp %p with NULL v_mount", vp));
1008 vm_object_reference_locked(object);
1009 VM_OBJECT_UNLOCK(object);
1010 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1011 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1013 VM_OBJECT_LOCK(object);
1014 vm_page_lock_queues();
1015 ++pageout_lock_miss;
1016 if (object->flags & OBJ_MIGHTBEDIRTY)
1019 goto unlock_and_continue;
1021 VM_OBJECT_LOCK(object);
1023 vm_page_lock_queues();
1025 * The page might have been moved to another
1026 * queue during potential blocking in vget()
1027 * above. The page might have been freed and
1028 * reused for another vnode.
1030 if (m->queue != PQ_INACTIVE ||
1031 m->object != object ||
1032 TAILQ_NEXT(m, pageq) != &marker) {
1034 if (object->flags & OBJ_MIGHTBEDIRTY)
1036 goto unlock_and_continue;
1040 * The page may have been busied during the
1041 * blocking in vget(). We don't move the
1042 * page back onto the end of the queue so that
1043 * statistics are more correct if we don't.
1045 if (m->busy || (m->oflags & VPO_BUSY)) {
1047 goto unlock_and_continue;
1051 * If the page has become held it might
1052 * be undergoing I/O, so skip it
1054 if (m->hold_count) {
1057 if (object->flags & OBJ_MIGHTBEDIRTY)
1059 goto unlock_and_continue;
1064 * If a page is dirty, then it is either being washed
1065 * (but not yet cleaned) or it is still in the
1066 * laundry. If it is still in the laundry, then we
1067 * start the cleaning operation.
1069 * decrement page_shortage on success to account for
1070 * the (future) cleaned page. Otherwise we could wind
1071 * up laundering or cleaning too many pages.
1073 vm_page_unlock_queues();
1074 if (vm_pageout_clean(m) != 0) {
1078 vm_page_lock_queues();
1079 unlock_and_continue:
1080 vm_page_lock_assert(m, MA_NOTOWNED);
1081 VM_OBJECT_UNLOCK(object);
1083 vm_page_unlock_queues();
1086 VFS_UNLOCK_GIANT(vfslocked);
1087 vm_object_deallocate(object);
1088 vn_finished_write(mp);
1089 vm_page_lock_queues();
1091 next = TAILQ_NEXT(&marker, pageq);
1092 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1094 vm_page_lock_assert(m, MA_NOTOWNED);
1098 VM_OBJECT_UNLOCK(object);
1102 * Compute the number of pages we want to try to move from the
1103 * active queue to the inactive queue.
1105 page_shortage = vm_paging_target() +
1106 cnt.v_inactive_target - cnt.v_inactive_count;
1107 page_shortage += addl_page_shortage;
1110 * Scan the active queue for things we can deactivate. We nominally
1111 * track the per-page activity counter and use it to locate
1112 * deactivation candidates.
1114 pcount = cnt.v_active_count;
1115 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1116 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1118 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1120 KASSERT(m->queue == PQ_ACTIVE,
1121 ("vm_pageout_scan: page %p isn't active", m));
1123 next = TAILQ_NEXT(m, pageq);
1124 if ((m->flags & PG_MARKER) != 0) {
1128 if (!vm_pageout_page_lock(m, &next)) {
1134 if (!VM_OBJECT_TRYLOCK(object) &&
1135 !vm_pageout_fallback_object_lock(m, &next)) {
1136 VM_OBJECT_UNLOCK(object);
1143 * Don't deactivate pages that are busy.
1145 if ((m->busy != 0) ||
1146 (m->oflags & VPO_BUSY) ||
1147 (m->hold_count != 0)) {
1149 VM_OBJECT_UNLOCK(object);
1156 * The count for pagedaemon pages is done after checking the
1157 * page for eligibility...
1162 * Check to see "how much" the page has been used.
1165 if (object->ref_count != 0) {
1166 if (m->flags & PG_REFERENCED) {
1169 actcount += pmap_ts_referenced(m);
1171 m->act_count += ACT_ADVANCE + actcount;
1172 if (m->act_count > ACT_MAX)
1173 m->act_count = ACT_MAX;
1178 * Since we have "tested" this bit, we need to clear it now.
1180 vm_page_flag_clear(m, PG_REFERENCED);
1183 * Only if an object is currently being used, do we use the
1184 * page activation count stats.
1186 if (actcount && (object->ref_count != 0)) {
1189 m->act_count -= min(m->act_count, ACT_DECLINE);
1190 if (vm_pageout_algorithm ||
1191 object->ref_count == 0 ||
1192 m->act_count == 0) {
1194 if (object->ref_count == 0) {
1195 KASSERT(!pmap_page_is_mapped(m),
1196 ("vm_pageout_scan: page %p is mapped", m));
1200 vm_page_deactivate(m);
1202 vm_page_deactivate(m);
1209 VM_OBJECT_UNLOCK(object);
1212 vm_page_unlock_queues();
1213 #if !defined(NO_SWAPPING)
1215 * Idle process swapout -- run once per second.
1217 if (vm_swap_idle_enabled) {
1219 if (time_second != lsec) {
1220 vm_req_vmdaemon(VM_SWAP_IDLE);
1227 * If we didn't get enough free pages, and we have skipped a vnode
1228 * in a writeable object, wakeup the sync daemon. And kick swapout
1229 * if we did not get enough free pages.
1231 if (vm_paging_target() > 0) {
1232 if (vnodes_skipped && vm_page_count_min())
1233 (void) speedup_syncer();
1234 #if !defined(NO_SWAPPING)
1235 if (vm_swap_enabled && vm_page_count_target())
1236 vm_req_vmdaemon(VM_SWAP_NORMAL);
1241 * If we are critically low on one of RAM or swap and low on
1242 * the other, kill the largest process. However, we avoid
1243 * doing this on the first pass in order to give ourselves a
1244 * chance to flush out dirty vnode-backed pages and to allow
1245 * active pages to be moved to the inactive queue and reclaimed.
1248 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1249 (swap_pager_full && vm_paging_target() > 0)))
1250 vm_pageout_oom(VM_OOM_MEM);
1255 vm_pageout_oom(int shortage)
1257 struct proc *p, *bigproc;
1258 vm_offset_t size, bigsize;
1263 * We keep the process bigproc locked once we find it to keep anyone
1264 * from messing with it; however, there is a possibility of
1265 * deadlock if process B is bigproc and one of it's child processes
1266 * attempts to propagate a signal to B while we are waiting for A's
1267 * lock while walking this list. To avoid this, we don't block on
1268 * the process lock but just skip a process if it is already locked.
1272 sx_slock(&allproc_lock);
1273 FOREACH_PROC_IN_SYSTEM(p) {
1276 if (PROC_TRYLOCK(p) == 0)
1279 * If this is a system, protected or killed process, skip it.
1281 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1282 (p->p_pid == 1) || P_KILLED(p) ||
1283 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1288 * If the process is in a non-running type state,
1289 * don't touch it. Check all the threads individually.
1292 FOREACH_THREAD_IN_PROC(p, td) {
1294 if (!TD_ON_RUNQ(td) &&
1295 !TD_IS_RUNNING(td) &&
1296 !TD_IS_SLEEPING(td)) {
1308 * get the process size
1310 vm = vmspace_acquire_ref(p);
1315 if (!vm_map_trylock_read(&vm->vm_map)) {
1320 size = vmspace_swap_count(vm);
1321 vm_map_unlock_read(&vm->vm_map);
1322 if (shortage == VM_OOM_MEM)
1323 size += vmspace_resident_count(vm);
1326 * if the this process is bigger than the biggest one
1329 if (size > bigsize) {
1330 if (bigproc != NULL)
1331 PROC_UNLOCK(bigproc);
1337 sx_sunlock(&allproc_lock);
1338 if (bigproc != NULL) {
1339 killproc(bigproc, "out of swap space");
1340 sched_nice(bigproc, PRIO_MIN);
1341 PROC_UNLOCK(bigproc);
1342 wakeup(&cnt.v_free_count);
1347 * This routine tries to maintain the pseudo LRU active queue,
1348 * so that during long periods of time where there is no paging,
1349 * that some statistic accumulation still occurs. This code
1350 * helps the situation where paging just starts to occur.
1353 vm_pageout_page_stats()
1357 int pcount,tpcount; /* Number of pages to check */
1358 static int fullintervalcount = 0;
1362 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1363 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1365 if (page_shortage <= 0)
1368 vm_page_lock_queues();
1369 pcount = cnt.v_active_count;
1370 fullintervalcount += vm_pageout_stats_interval;
1371 if (fullintervalcount < vm_pageout_full_stats_interval) {
1372 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1374 if (pcount > tpcount)
1377 fullintervalcount = 0;
1380 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1381 while ((m != NULL) && (pcount-- > 0)) {
1384 KASSERT(m->queue == PQ_ACTIVE,
1385 ("vm_pageout_page_stats: page %p isn't active", m));
1387 next = TAILQ_NEXT(m, pageq);
1388 if ((m->flags & PG_MARKER) != 0) {
1392 vm_page_lock_assert(m, MA_NOTOWNED);
1393 if (!vm_pageout_page_lock(m, &next)) {
1399 if (!VM_OBJECT_TRYLOCK(object) &&
1400 !vm_pageout_fallback_object_lock(m, &next)) {
1401 VM_OBJECT_UNLOCK(object);
1408 * Don't deactivate pages that are busy.
1410 if ((m->busy != 0) ||
1411 (m->oflags & VPO_BUSY) ||
1412 (m->hold_count != 0)) {
1414 VM_OBJECT_UNLOCK(object);
1421 if (m->flags & PG_REFERENCED) {
1422 vm_page_flag_clear(m, PG_REFERENCED);
1426 actcount += pmap_ts_referenced(m);
1428 m->act_count += ACT_ADVANCE + actcount;
1429 if (m->act_count > ACT_MAX)
1430 m->act_count = ACT_MAX;
1433 if (m->act_count == 0) {
1435 * We turn off page access, so that we have
1436 * more accurate RSS stats. We don't do this
1437 * in the normal page deactivation when the
1438 * system is loaded VM wise, because the
1439 * cost of the large number of page protect
1440 * operations would be higher than the value
1441 * of doing the operation.
1444 vm_page_deactivate(m);
1446 m->act_count -= min(m->act_count, ACT_DECLINE);
1451 VM_OBJECT_UNLOCK(object);
1454 vm_page_unlock_queues();
1458 * vm_pageout is the high level pageout daemon.
1466 * Initialize some paging parameters.
1468 cnt.v_interrupt_free_min = 2;
1469 if (cnt.v_page_count < 2000)
1470 vm_pageout_page_count = 8;
1473 * v_free_reserved needs to include enough for the largest
1474 * swap pager structures plus enough for any pv_entry structs
1477 if (cnt.v_page_count > 1024)
1478 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1481 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1482 cnt.v_interrupt_free_min;
1483 cnt.v_free_reserved = vm_pageout_page_count +
1484 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1485 cnt.v_free_severe = cnt.v_free_min / 2;
1486 cnt.v_free_min += cnt.v_free_reserved;
1487 cnt.v_free_severe += cnt.v_free_reserved;
1490 * v_free_target and v_cache_min control pageout hysteresis. Note
1491 * that these are more a measure of the VM cache queue hysteresis
1492 * then the VM free queue. Specifically, v_free_target is the
1493 * high water mark (free+cache pages).
1495 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1496 * low water mark, while v_free_min is the stop. v_cache_min must
1497 * be big enough to handle memory needs while the pageout daemon
1498 * is signalled and run to free more pages.
1500 if (cnt.v_free_count > 6144)
1501 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1503 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1505 if (cnt.v_free_count > 2048) {
1506 cnt.v_cache_min = cnt.v_free_target;
1507 cnt.v_cache_max = 2 * cnt.v_cache_min;
1508 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1510 cnt.v_cache_min = 0;
1511 cnt.v_cache_max = 0;
1512 cnt.v_inactive_target = cnt.v_free_count / 4;
1514 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1515 cnt.v_inactive_target = cnt.v_free_count / 3;
1517 /* XXX does not really belong here */
1518 if (vm_page_max_wired == 0)
1519 vm_page_max_wired = cnt.v_free_count / 3;
1521 if (vm_pageout_stats_max == 0)
1522 vm_pageout_stats_max = cnt.v_free_target;
1525 * Set interval in seconds for stats scan.
1527 if (vm_pageout_stats_interval == 0)
1528 vm_pageout_stats_interval = 5;
1529 if (vm_pageout_full_stats_interval == 0)
1530 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1532 swap_pager_swap_init();
1535 * The pageout daemon is never done, so loop forever.
1539 * If we have enough free memory, wakeup waiters. Do
1540 * not clear vm_pages_needed until we reach our target,
1541 * otherwise we may be woken up over and over again and
1542 * waste a lot of cpu.
1544 mtx_lock(&vm_page_queue_free_mtx);
1545 if (vm_pages_needed && !vm_page_count_min()) {
1546 if (!vm_paging_needed())
1547 vm_pages_needed = 0;
1548 wakeup(&cnt.v_free_count);
1550 if (vm_pages_needed) {
1552 * Still not done, take a second pass without waiting
1553 * (unlimited dirty cleaning), otherwise sleep a bit
1558 msleep(&vm_pages_needed,
1559 &vm_page_queue_free_mtx, PVM, "psleep",
1563 * Good enough, sleep & handle stats. Prime the pass
1570 error = msleep(&vm_pages_needed,
1571 &vm_page_queue_free_mtx, PVM, "psleep",
1572 vm_pageout_stats_interval * hz);
1573 if (error && !vm_pages_needed) {
1574 mtx_unlock(&vm_page_queue_free_mtx);
1576 vm_pageout_page_stats();
1580 if (vm_pages_needed)
1582 mtx_unlock(&vm_page_queue_free_mtx);
1583 vm_pageout_scan(pass);
1588 * Unless the free page queue lock is held by the caller, this function
1589 * should be regarded as advisory. Specifically, the caller should
1590 * not msleep() on &cnt.v_free_count following this function unless
1591 * the free page queue lock is held until the msleep() is performed.
1597 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1598 vm_pages_needed = 1;
1599 wakeup(&vm_pages_needed);
1603 #if !defined(NO_SWAPPING)
1605 vm_req_vmdaemon(int req)
1607 static int lastrun = 0;
1609 mtx_lock(&vm_daemon_mtx);
1610 vm_pageout_req_swapout |= req;
1611 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1612 wakeup(&vm_daemon_needed);
1615 mtx_unlock(&vm_daemon_mtx);
1621 struct rlimit rsslim;
1625 int breakout, swapout_flags;
1628 mtx_lock(&vm_daemon_mtx);
1629 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1630 swapout_flags = vm_pageout_req_swapout;
1631 vm_pageout_req_swapout = 0;
1632 mtx_unlock(&vm_daemon_mtx);
1634 swapout_procs(swapout_flags);
1637 * scan the processes for exceeding their rlimits or if
1638 * process is swapped out -- deactivate pages
1640 sx_slock(&allproc_lock);
1641 FOREACH_PROC_IN_SYSTEM(p) {
1642 vm_pindex_t limit, size;
1645 * if this is a system process or if we have already
1646 * looked at this process, skip it.
1649 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1654 * if the process is in a non-running type state,
1658 FOREACH_THREAD_IN_PROC(p, td) {
1660 if (!TD_ON_RUNQ(td) &&
1661 !TD_IS_RUNNING(td) &&
1662 !TD_IS_SLEEPING(td)) {
1676 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1678 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1681 * let processes that are swapped out really be
1682 * swapped out set the limit to nothing (will force a
1685 if ((p->p_flag & P_INMEM) == 0)
1686 limit = 0; /* XXX */
1687 vm = vmspace_acquire_ref(p);
1692 size = vmspace_resident_count(vm);
1693 if (limit >= 0 && size >= limit) {
1694 vm_pageout_map_deactivate_pages(
1695 &vm->vm_map, limit);
1699 sx_sunlock(&allproc_lock);
1702 #endif /* !defined(NO_SWAPPING) */