2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * This code is derived from software contributed to Berkeley by
12 * The Mach Operating System project at Carnegie-Mellon University.
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
50 * Permission to use, copy, modify and distribute this software and
51 * its documentation is hereby granted, provided that both the copyright
52 * notice and this permission notice appear in all copies of the
53 * software, derivative works or modified versions, and any portions
54 * thereof, and that both notices appear in supporting documentation.
56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
60 * Carnegie Mellon requests users of this software to return to
62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
63 * School of Computer Science
64 * Carnegie Mellon University
65 * Pittsburgh PA 15213-3890
67 * any improvements or extensions that they make and grant Carnegie the
68 * rights to redistribute these changes.
72 * The proverbial page-out daemon.
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
84 #include <sys/mutex.h>
86 #include <sys/kthread.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
95 #include <sys/sysctl.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
108 #include <machine/mutex.h>
111 * System initialization
114 /* the kernel process "vm_pageout"*/
115 static void vm_pageout(void);
116 static int vm_pageout_clean(vm_page_t);
117 static void vm_pageout_pmap_collect(void);
118 static void vm_pageout_scan(int pass);
120 struct proc *pageproc;
122 static struct kproc_desc page_kp = {
127 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
129 #if !defined(NO_SWAPPING)
130 /* the kernel process "vm_daemon"*/
131 static void vm_daemon(void);
132 static struct proc *vmproc;
134 static struct kproc_desc vm_kp = {
139 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
143 int vm_pages_needed; /* Event on which pageout daemon sleeps */
144 int vm_pageout_deficit; /* Estimated number of pages deficit */
145 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
147 #if !defined(NO_SWAPPING)
148 static int vm_pageout_req_swapout; /* XXX */
149 static int vm_daemon_needed;
150 static struct mtx vm_daemon_mtx;
151 /* Allow for use by vm_pageout before vm_daemon is initialized. */
152 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
154 static int vm_max_launder = 32;
155 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
156 static int vm_pageout_full_stats_interval = 0;
157 static int vm_pageout_algorithm=0;
158 static int defer_swap_pageouts=0;
159 static int disable_swap_pageouts=0;
161 #if defined(NO_SWAPPING)
162 static int vm_swap_enabled=0;
163 static int vm_swap_idle_enabled=0;
165 static int vm_swap_enabled=1;
166 static int vm_swap_idle_enabled=0;
169 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
170 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
172 SYSCTL_INT(_vm, OID_AUTO, max_launder,
173 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
176 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
178 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
179 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
181 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
182 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
184 #if defined(NO_SWAPPING)
185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
186 CTLFLAG_RD, &vm_swap_enabled, 0, "");
187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
188 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
190 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
191 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
192 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
193 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
196 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
197 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
199 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
200 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
202 static int pageout_lock_miss;
203 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
204 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
206 #define VM_PAGEOUT_PAGE_COUNT 16
207 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
209 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
211 #if !defined(NO_SWAPPING)
212 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
213 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
214 static void vm_req_vmdaemon(int req);
216 static void vm_pageout_page_stats(void);
219 * vm_pageout_fallback_object_lock:
221 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
222 * known to have failed and page queue must be either PQ_ACTIVE or
223 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
224 * while locking the vm object. Use marker page to detect page queue
225 * changes and maintain notion of next page on page queue. Return
226 * TRUE if no changes were detected, FALSE otherwise. vm object is
229 * This function depends on both the lock portion of struct vm_object
230 * and normal struct vm_page being type stable.
233 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
235 struct vm_page marker;
241 * Initialize our marker
243 bzero(&marker, sizeof(marker));
244 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
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 * Don't mess with the page if it's busy, held, or special
301 if ((m->hold_count != 0) ||
302 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
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 (((p->queue - p->pc) == PQ_CACHE) ||
346 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 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 (((p->queue - p->pc) == PQ_CACHE) ||
376 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 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_page_protect(mc[i], VM_PROT_READ);
441 vm_page_unlock_queues();
442 vm_object_pip_add(object, count);
444 vm_pager_put_pages(object, mc, count,
445 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
448 vm_page_lock_queues();
449 for (i = 0; i < count; i++) {
450 vm_page_t mt = mc[i];
452 KASSERT((mt->flags & PG_WRITEABLE) == 0,
453 ("vm_pageout_flush: page %p is not write protected", mt));
454 switch (pageout_status[i]) {
461 * Page outside of range of object. Right now we
462 * essentially lose the changes by pretending it
465 pmap_clear_modify(mt);
471 * If page couldn't be paged out, then reactivate the
472 * page so it doesn't clog the inactive list. (We
473 * will try paging out it again later).
475 vm_page_activate(mt);
482 * If the operation is still going, leave the page busy to
483 * block all other accesses. Also, leave the paging in
484 * progress indicator set so that we don't attempt an object
487 if (pageout_status[i] != VM_PAGER_PEND) {
488 vm_object_pip_wakeup(object);
489 vm_page_io_finish(mt);
490 if (vm_page_count_severe())
491 vm_page_try_to_cache(mt);
497 #if !defined(NO_SWAPPING)
499 * vm_pageout_object_deactivate_pages
501 * deactivate enough pages to satisfy the inactive target
502 * requirements or if vm_page_proc_limit is set, then
503 * deactivate all of the pages in the object and its
506 * The object and map must be locked.
509 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
511 vm_object_t first_object;
514 vm_object_t backing_object, object;
516 int actcount, rcount, remove_mode;
518 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
519 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
521 for (object = first_object;; object = backing_object) {
522 if (pmap_resident_count(pmap) <= desired)
524 if (object->paging_in_progress)
528 if (object->shadow_count > 1)
531 * scan the objects entire memory queue
533 rcount = object->resident_page_count;
534 p = TAILQ_FIRST(&object->memq);
535 vm_page_lock_queues();
536 while (p && (rcount-- > 0)) {
537 if (pmap_resident_count(pmap) <= desired) {
538 vm_page_unlock_queues();
541 next = TAILQ_NEXT(p, listq);
543 if (p->wire_count != 0 ||
544 p->hold_count != 0 ||
546 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
547 !pmap_page_exists_quick(pmap, p)) {
551 actcount = pmap_ts_referenced(p);
553 vm_page_flag_set(p, PG_REFERENCED);
554 } else if (p->flags & PG_REFERENCED) {
557 if ((p->queue != PQ_ACTIVE) &&
558 (p->flags & PG_REFERENCED)) {
560 p->act_count += actcount;
561 vm_page_flag_clear(p, PG_REFERENCED);
562 } else if (p->queue == PQ_ACTIVE) {
563 if ((p->flags & PG_REFERENCED) == 0) {
564 p->act_count -= min(p->act_count, ACT_DECLINE);
565 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
567 vm_page_deactivate(p);
573 vm_page_flag_clear(p, PG_REFERENCED);
574 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
575 p->act_count += ACT_ADVANCE;
578 } else if (p->queue == PQ_INACTIVE) {
583 vm_page_unlock_queues();
584 if ((backing_object = object->backing_object) == NULL)
586 VM_OBJECT_LOCK(backing_object);
587 if (object != first_object)
588 VM_OBJECT_UNLOCK(object);
591 if (object != first_object)
592 VM_OBJECT_UNLOCK(object);
596 * deactivate some number of pages in a map, try to do it fairly, but
597 * that is really hard to do.
600 vm_pageout_map_deactivate_pages(map, desired)
605 vm_object_t obj, bigobj;
608 if (!vm_map_trylock(map))
615 * first, search out the biggest object, and try to free pages from
618 tmpe = map->header.next;
619 while (tmpe != &map->header) {
620 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
621 obj = tmpe->object.vm_object;
622 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
623 if (obj->shadow_count <= 1 &&
625 bigobj->resident_page_count < obj->resident_page_count)) {
627 VM_OBJECT_UNLOCK(bigobj);
630 VM_OBJECT_UNLOCK(obj);
633 if (tmpe->wired_count > 0)
634 nothingwired = FALSE;
638 if (bigobj != NULL) {
639 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
640 VM_OBJECT_UNLOCK(bigobj);
643 * Next, hunt around for other pages to deactivate. We actually
644 * do this search sort of wrong -- .text first is not the best idea.
646 tmpe = map->header.next;
647 while (tmpe != &map->header) {
648 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
650 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
651 obj = tmpe->object.vm_object;
654 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
655 VM_OBJECT_UNLOCK(obj);
662 * Remove all mappings if a process is swapped out, this will free page
665 if (desired == 0 && nothingwired) {
666 pmap_remove(vm_map_pmap(map), vm_map_min(map),
671 #endif /* !defined(NO_SWAPPING) */
674 * This routine is very drastic, but can save the system
678 vm_pageout_pmap_collect(void)
682 static int warningdone;
684 if (pmap_pagedaemon_waken == 0)
686 if (warningdone < 5) {
687 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
690 vm_page_lock_queues();
691 for (i = 0; i < vm_page_array_size; i++) {
692 m = &vm_page_array[i];
693 if (m->wire_count || m->hold_count || m->busy ||
694 (m->flags & (PG_BUSY | PG_UNMANAGED)))
698 vm_page_unlock_queues();
699 pmap_pagedaemon_waken = 0;
703 * vm_pageout_scan does the dirty work for the pageout daemon.
706 vm_pageout_scan(int pass)
709 struct vm_page marker;
710 int page_shortage, maxscan, pcount;
711 int addl_page_shortage, addl_page_shortage_init;
712 struct proc *p, *bigproc;
714 vm_offset_t size, bigsize;
716 int actcount, cache_cur, cache_first_failure;
717 static int cache_last_free;
718 int vnodes_skipped = 0;
722 * Decrease registered cache sizes.
724 EVENTHANDLER_INVOKE(vm_lowmem, 0);
726 * We do this explicitly after the caches have been drained above.
730 * Do whatever cleanup that the pmap code can.
732 vm_pageout_pmap_collect();
734 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
737 * Calculate the number of pages we want to either free or move
740 page_shortage = vm_paging_target() + addl_page_shortage_init;
743 * Initialize our marker
745 bzero(&marker, sizeof(marker));
746 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
747 marker.queue = PQ_INACTIVE;
748 marker.wire_count = 1;
751 * Start scanning the inactive queue for pages we can move to the
752 * cache or free. The scan will stop when the target is reached or
753 * we have scanned the entire inactive queue. Note that m->act_count
754 * is not used to form decisions for the inactive queue, only for the
757 * maxlaunder limits the number of dirty pages we flush per scan.
758 * For most systems a smaller value (16 or 32) is more robust under
759 * extreme memory and disk pressure because any unnecessary writes
760 * to disk can result in extreme performance degredation. However,
761 * systems with excessive dirty pages (especially when MAP_NOSYNC is
762 * used) will die horribly with limited laundering. If the pageout
763 * daemon cannot clean enough pages in the first pass, we let it go
764 * all out in succeeding passes.
766 if ((maxlaunder = vm_max_launder) <= 1)
770 vm_page_lock_queues();
772 addl_page_shortage = addl_page_shortage_init;
773 maxscan = cnt.v_inactive_count;
775 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
776 m != NULL && maxscan-- > 0 && page_shortage > 0;
781 if (m->queue != PQ_INACTIVE) {
785 next = TAILQ_NEXT(m, pageq);
791 if (m->flags & PG_MARKER)
795 * A held page may be undergoing I/O, so skip it.
799 addl_page_shortage++;
803 * Don't mess with busy pages, keep in the front of the
804 * queue, most likely are being paged out.
806 if (!VM_OBJECT_TRYLOCK(object) &&
807 (!vm_pageout_fallback_object_lock(m, &next) ||
808 m->hold_count != 0)) {
809 VM_OBJECT_UNLOCK(object);
810 addl_page_shortage++;
813 if (m->busy || (m->flags & PG_BUSY)) {
814 VM_OBJECT_UNLOCK(object);
815 addl_page_shortage++;
820 * If the object is not being used, we ignore previous
823 if (object->ref_count == 0) {
824 vm_page_flag_clear(m, PG_REFERENCED);
825 pmap_clear_reference(m);
828 * Otherwise, if the page has been referenced while in the
829 * inactive queue, we bump the "activation count" upwards,
830 * making it less likely that the page will be added back to
831 * the inactive queue prematurely again. Here we check the
832 * page tables (or emulated bits, if any), given the upper
833 * level VM system not knowing anything about existing
836 } else if (((m->flags & PG_REFERENCED) == 0) &&
837 (actcount = pmap_ts_referenced(m))) {
839 VM_OBJECT_UNLOCK(object);
840 m->act_count += (actcount + ACT_ADVANCE);
845 * If the upper level VM system knows about any page
846 * references, we activate the page. We also set the
847 * "activation count" higher than normal so that we will less
848 * likely place pages back onto the inactive queue again.
850 if ((m->flags & PG_REFERENCED) != 0) {
851 vm_page_flag_clear(m, PG_REFERENCED);
852 actcount = pmap_ts_referenced(m);
854 VM_OBJECT_UNLOCK(object);
855 m->act_count += (actcount + ACT_ADVANCE + 1);
860 * If the upper level VM system doesn't know anything about
861 * the page being dirty, we have to check for it again. As
862 * far as the VM code knows, any partially dirty pages are
865 if (m->dirty == 0 && !pmap_is_modified(m)) {
867 * Avoid a race condition: Unless write access is
868 * removed from the page, another processor could
869 * modify it before all access is removed by the call
870 * to vm_page_cache() below. If vm_page_cache() finds
871 * that the page has been modified when it removes all
872 * access, it panics because it cannot cache dirty
873 * pages. In principle, we could eliminate just write
874 * access here rather than all access. In the expected
875 * case, when there are no last instant modifications
876 * to the page, removing all access will be cheaper
879 if ((m->flags & PG_WRITEABLE) != 0)
887 * Invalid pages can be easily freed
893 } else if (m->dirty == 0) {
895 * Clean pages can be placed onto the cache queue.
896 * This effectively frees them.
900 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
902 * Dirty pages need to be paged out, but flushing
903 * a page is extremely expensive verses freeing
904 * a clean page. Rather then artificially limiting
905 * the number of pages we can flush, we instead give
906 * dirty pages extra priority on the inactive queue
907 * by forcing them to be cycled through the queue
908 * twice before being flushed, after which the
909 * (now clean) page will cycle through once more
910 * before being freed. This significantly extends
911 * the thrash point for a heavily loaded machine.
913 vm_page_flag_set(m, PG_WINATCFLS);
915 } else if (maxlaunder > 0) {
917 * We always want to try to flush some dirty pages if
918 * we encounter them, to keep the system stable.
919 * Normally this number is small, but under extreme
920 * pressure where there are insufficient clean pages
921 * on the inactive queue, we may have to go all out.
923 int swap_pageouts_ok, vfslocked = 0;
924 struct vnode *vp = NULL;
925 struct mount *mp = NULL;
927 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
928 swap_pageouts_ok = 1;
930 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
931 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
932 vm_page_count_min());
937 * We don't bother paging objects that are "dead".
938 * Those objects are in a "rundown" state.
940 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
941 VM_OBJECT_UNLOCK(object);
947 * Following operations may unlock
948 * vm_page_queue_mtx, invalidating the 'next'
949 * pointer. To prevent an inordinate number
950 * of restarts we use our marker to remember
954 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
957 * The object is already known NOT to be dead. It
958 * is possible for the vget() to block the whole
959 * pageout daemon, but the new low-memory handling
960 * code should prevent it.
962 * The previous code skipped locked vnodes and, worse,
963 * reordered pages in the queue. This results in
964 * completely non-deterministic operation and, on a
965 * busy system, can lead to extremely non-optimal
966 * pageouts. For example, it can cause clean pages
967 * to be freed and dirty pages to be moved to the end
968 * of the queue. Since dirty pages are also moved to
969 * the end of the queue once-cleaned, this gives
970 * way too large a weighting to defering the freeing
973 * We can't wait forever for the vnode lock, we might
974 * deadlock due to a vn_read() getting stuck in
975 * vm_wait while holding this vnode. We skip the
976 * vnode if we can't get it in a reasonable amount
979 if (object->type == OBJT_VNODE) {
981 if (vp->v_type == VREG &&
982 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
984 ("vm_pageout_scan: mp != NULL"));
986 if (object->flags & OBJ_MIGHTBEDIRTY)
988 goto unlock_and_continue;
990 vm_page_unlock_queues();
991 vm_object_reference_locked(object);
992 VM_OBJECT_UNLOCK(object);
993 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
994 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
996 VM_OBJECT_LOCK(object);
997 vm_page_lock_queues();
999 if (object->flags & OBJ_MIGHTBEDIRTY)
1002 goto unlock_and_continue;
1004 VM_OBJECT_LOCK(object);
1005 vm_page_lock_queues();
1007 * The page might have been moved to another
1008 * queue during potential blocking in vget()
1009 * above. The page might have been freed and
1010 * reused for another vnode.
1012 if (m->queue != PQ_INACTIVE ||
1013 m->object != object ||
1014 TAILQ_NEXT(m, pageq) != &marker) {
1015 if (object->flags & OBJ_MIGHTBEDIRTY)
1017 goto unlock_and_continue;
1021 * The page may have been busied during the
1022 * blocking in vget(). We don't move the
1023 * page back onto the end of the queue so that
1024 * statistics are more correct if we don't.
1026 if (m->busy || (m->flags & PG_BUSY)) {
1027 goto unlock_and_continue;
1031 * If the page has become held it might
1032 * be undergoing I/O, so skip it
1034 if (m->hold_count) {
1035 vm_pageq_requeue(m);
1036 if (object->flags & OBJ_MIGHTBEDIRTY)
1038 goto unlock_and_continue;
1043 * If a page is dirty, then it is either being washed
1044 * (but not yet cleaned) or it is still in the
1045 * laundry. If it is still in the laundry, then we
1046 * start the cleaning operation.
1048 * decrement page_shortage on success to account for
1049 * the (future) cleaned page. Otherwise we could wind
1050 * up laundering or cleaning too many pages.
1052 if (vm_pageout_clean(m) != 0) {
1056 unlock_and_continue:
1057 VM_OBJECT_UNLOCK(object);
1059 vm_page_unlock_queues();
1062 VFS_UNLOCK_GIANT(vfslocked);
1063 vm_object_deallocate(object);
1064 vn_finished_write(mp);
1065 vm_page_lock_queues();
1067 next = TAILQ_NEXT(&marker, pageq);
1068 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1072 VM_OBJECT_UNLOCK(object);
1076 * Compute the number of pages we want to try to move from the
1077 * active queue to the inactive queue.
1079 page_shortage = vm_paging_target() +
1080 cnt.v_inactive_target - cnt.v_inactive_count;
1081 page_shortage += addl_page_shortage;
1084 * Scan the active queue for things we can deactivate. We nominally
1085 * track the per-page activity counter and use it to locate
1086 * deactivation candidates.
1088 pcount = cnt.v_active_count;
1089 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1091 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1093 KASSERT(m->queue == PQ_ACTIVE,
1094 ("vm_pageout_scan: page %p isn't active", m));
1096 next = TAILQ_NEXT(m, pageq);
1098 if ((m->flags & PG_MARKER) != 0) {
1102 if (!VM_OBJECT_TRYLOCK(object) &&
1103 !vm_pageout_fallback_object_lock(m, &next)) {
1104 VM_OBJECT_UNLOCK(object);
1110 * Don't deactivate pages that are busy.
1112 if ((m->busy != 0) ||
1113 (m->flags & PG_BUSY) ||
1114 (m->hold_count != 0)) {
1115 VM_OBJECT_UNLOCK(object);
1116 vm_pageq_requeue(m);
1122 * The count for pagedaemon pages is done after checking the
1123 * page for eligibility...
1128 * Check to see "how much" the page has been used.
1131 if (object->ref_count != 0) {
1132 if (m->flags & PG_REFERENCED) {
1135 actcount += pmap_ts_referenced(m);
1137 m->act_count += ACT_ADVANCE + actcount;
1138 if (m->act_count > ACT_MAX)
1139 m->act_count = ACT_MAX;
1144 * Since we have "tested" this bit, we need to clear it now.
1146 vm_page_flag_clear(m, PG_REFERENCED);
1149 * Only if an object is currently being used, do we use the
1150 * page activation count stats.
1152 if (actcount && (object->ref_count != 0)) {
1153 vm_pageq_requeue(m);
1155 m->act_count -= min(m->act_count, ACT_DECLINE);
1156 if (vm_pageout_algorithm ||
1157 object->ref_count == 0 ||
1158 m->act_count == 0) {
1160 if (object->ref_count == 0) {
1165 vm_page_deactivate(m);
1167 vm_page_deactivate(m);
1170 vm_pageq_requeue(m);
1173 VM_OBJECT_UNLOCK(object);
1178 * We try to maintain some *really* free pages, this allows interrupt
1179 * code to be guaranteed space. Since both cache and free queues
1180 * are considered basically 'free', moving pages from cache to free
1181 * does not effect other calculations.
1183 cache_cur = cache_last_free;
1184 cache_first_failure = -1;
1185 while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur =
1186 (cache_cur + PQ_PRIME2) & PQ_L2_MASK) != cache_first_failure) {
1187 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl,
1189 KASSERT(m->dirty == 0,
1190 ("Found dirty cache page %p", m));
1191 KASSERT(!pmap_page_is_mapped(m),
1192 ("Found mapped cache page %p", m));
1193 KASSERT((m->flags & PG_UNMANAGED) == 0,
1194 ("Found unmanaged cache page %p", m));
1195 KASSERT(m->wire_count == 0,
1196 ("Found wired cache page %p", m));
1197 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1199 KASSERT((m->flags & PG_BUSY) == 0 &&
1200 m->busy == 0, ("Found busy cache page %p",
1203 VM_OBJECT_UNLOCK(object);
1205 cache_last_free = cache_cur;
1206 cache_first_failure = -1;
1210 if (m == NULL && cache_first_failure == -1)
1211 cache_first_failure = cache_cur;
1213 vm_page_unlock_queues();
1214 #if !defined(NO_SWAPPING)
1216 * Idle process swapout -- run once per second.
1218 if (vm_swap_idle_enabled) {
1220 if (time_second != lsec) {
1221 vm_req_vmdaemon(VM_SWAP_IDLE);
1228 * If we didn't get enough free pages, and we have skipped a vnode
1229 * in a writeable object, wakeup the sync daemon. And kick swapout
1230 * if we did not get enough free pages.
1232 if (vm_paging_target() > 0) {
1233 if (vnodes_skipped && vm_page_count_min())
1234 (void) speedup_syncer();
1235 #if !defined(NO_SWAPPING)
1236 if (vm_swap_enabled && vm_page_count_target())
1237 vm_req_vmdaemon(VM_SWAP_NORMAL);
1242 * If we are critically low on one of RAM or swap and low on
1243 * the other, kill the largest process. However, we avoid
1244 * doing this on the first pass in order to give ourselves a
1245 * chance to flush out dirty vnode-backed pages and to allow
1246 * active pages to be moved to the inactive queue and reclaimed.
1248 * We keep the process bigproc locked once we find it to keep anyone
1249 * from messing with it; however, there is a possibility of
1250 * deadlock if process B is bigproc and one of it's child processes
1251 * attempts to propagate a signal to B while we are waiting for A's
1252 * lock while walking this list. To avoid this, we don't block on
1253 * the process lock but just skip a process if it is already locked.
1256 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1257 (swap_pager_full && vm_paging_target() > 0))) {
1260 sx_slock(&allproc_lock);
1261 FOREACH_PROC_IN_SYSTEM(p) {
1264 if (PROC_TRYLOCK(p) == 0)
1267 * If this is a system or protected process, skip it.
1269 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1270 (p->p_flag & P_PROTECTED) ||
1271 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1276 * If the process is in a non-running type state,
1277 * don't touch it. Check all the threads individually.
1279 mtx_lock_spin(&sched_lock);
1281 FOREACH_THREAD_IN_PROC(p, td) {
1282 if (!TD_ON_RUNQ(td) &&
1283 !TD_IS_RUNNING(td) &&
1284 !TD_IS_SLEEPING(td)) {
1290 mtx_unlock_spin(&sched_lock);
1294 mtx_unlock_spin(&sched_lock);
1296 * get the process size
1298 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1302 size = vmspace_swap_count(p->p_vmspace);
1303 vm_map_unlock_read(&p->p_vmspace->vm_map);
1304 size += vmspace_resident_count(p->p_vmspace);
1306 * if the this process is bigger than the biggest one
1309 if (size > bigsize) {
1310 if (bigproc != NULL)
1311 PROC_UNLOCK(bigproc);
1317 sx_sunlock(&allproc_lock);
1318 if (bigproc != NULL) {
1319 killproc(bigproc, "out of swap space");
1320 mtx_lock_spin(&sched_lock);
1321 sched_nice(bigproc, PRIO_MIN);
1322 mtx_unlock_spin(&sched_lock);
1323 PROC_UNLOCK(bigproc);
1324 wakeup(&cnt.v_free_count);
1330 * This routine tries to maintain the pseudo LRU active queue,
1331 * so that during long periods of time where there is no paging,
1332 * that some statistic accumulation still occurs. This code
1333 * helps the situation where paging just starts to occur.
1336 vm_pageout_page_stats()
1340 int pcount,tpcount; /* Number of pages to check */
1341 static int fullintervalcount = 0;
1344 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1346 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1347 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1349 if (page_shortage <= 0)
1352 pcount = cnt.v_active_count;
1353 fullintervalcount += vm_pageout_stats_interval;
1354 if (fullintervalcount < vm_pageout_full_stats_interval) {
1355 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1356 if (pcount > tpcount)
1359 fullintervalcount = 0;
1362 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1363 while ((m != NULL) && (pcount-- > 0)) {
1366 KASSERT(m->queue == PQ_ACTIVE,
1367 ("vm_pageout_page_stats: page %p isn't active", m));
1369 next = TAILQ_NEXT(m, pageq);
1372 if ((m->flags & PG_MARKER) != 0) {
1376 if (!VM_OBJECT_TRYLOCK(object) &&
1377 !vm_pageout_fallback_object_lock(m, &next)) {
1378 VM_OBJECT_UNLOCK(object);
1384 * Don't deactivate pages that are busy.
1386 if ((m->busy != 0) ||
1387 (m->flags & PG_BUSY) ||
1388 (m->hold_count != 0)) {
1389 VM_OBJECT_UNLOCK(object);
1390 vm_pageq_requeue(m);
1396 if (m->flags & PG_REFERENCED) {
1397 vm_page_flag_clear(m, PG_REFERENCED);
1401 actcount += pmap_ts_referenced(m);
1403 m->act_count += ACT_ADVANCE + actcount;
1404 if (m->act_count > ACT_MAX)
1405 m->act_count = ACT_MAX;
1406 vm_pageq_requeue(m);
1408 if (m->act_count == 0) {
1410 * We turn off page access, so that we have
1411 * more accurate RSS stats. We don't do this
1412 * in the normal page deactivation when the
1413 * system is loaded VM wise, because the
1414 * cost of the large number of page protect
1415 * operations would be higher than the value
1416 * of doing the operation.
1419 vm_page_deactivate(m);
1421 m->act_count -= min(m->act_count, ACT_DECLINE);
1422 vm_pageq_requeue(m);
1425 VM_OBJECT_UNLOCK(object);
1431 * vm_pageout is the high level pageout daemon.
1439 * Initialize some paging parameters.
1441 cnt.v_interrupt_free_min = 2;
1442 if (cnt.v_page_count < 2000)
1443 vm_pageout_page_count = 8;
1446 * v_free_reserved needs to include enough for the largest
1447 * swap pager structures plus enough for any pv_entry structs
1450 if (cnt.v_page_count > 1024)
1451 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1454 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1455 cnt.v_interrupt_free_min;
1456 cnt.v_free_reserved = vm_pageout_page_count +
1457 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_L2_SIZE;
1458 cnt.v_free_severe = cnt.v_free_min / 2;
1459 cnt.v_free_min += cnt.v_free_reserved;
1460 cnt.v_free_severe += cnt.v_free_reserved;
1463 * v_free_target and v_cache_min control pageout hysteresis. Note
1464 * that these are more a measure of the VM cache queue hysteresis
1465 * then the VM free queue. Specifically, v_free_target is the
1466 * high water mark (free+cache pages).
1468 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1469 * low water mark, while v_free_min is the stop. v_cache_min must
1470 * be big enough to handle memory needs while the pageout daemon
1471 * is signalled and run to free more pages.
1473 if (cnt.v_free_count > 6144)
1474 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1476 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1478 if (cnt.v_free_count > 2048) {
1479 cnt.v_cache_min = cnt.v_free_target;
1480 cnt.v_cache_max = 2 * cnt.v_cache_min;
1481 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1483 cnt.v_cache_min = 0;
1484 cnt.v_cache_max = 0;
1485 cnt.v_inactive_target = cnt.v_free_count / 4;
1487 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1488 cnt.v_inactive_target = cnt.v_free_count / 3;
1490 /* XXX does not really belong here */
1491 if (vm_page_max_wired == 0)
1492 vm_page_max_wired = cnt.v_free_count / 3;
1494 if (vm_pageout_stats_max == 0)
1495 vm_pageout_stats_max = cnt.v_free_target;
1498 * Set interval in seconds for stats scan.
1500 if (vm_pageout_stats_interval == 0)
1501 vm_pageout_stats_interval = 5;
1502 if (vm_pageout_full_stats_interval == 0)
1503 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1505 swap_pager_swap_init();
1508 * The pageout daemon is never done, so loop forever.
1511 vm_page_lock_queues();
1513 * If we have enough free memory, wakeup waiters. Do
1514 * not clear vm_pages_needed until we reach our target,
1515 * otherwise we may be woken up over and over again and
1516 * waste a lot of cpu.
1518 if (vm_pages_needed && !vm_page_count_min()) {
1519 if (!vm_paging_needed())
1520 vm_pages_needed = 0;
1521 wakeup(&cnt.v_free_count);
1523 if (vm_pages_needed) {
1525 * Still not done, take a second pass without waiting
1526 * (unlimited dirty cleaning), otherwise sleep a bit
1531 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1535 * Good enough, sleep & handle stats. Prime the pass
1542 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1543 "psleep", vm_pageout_stats_interval * hz);
1544 if (error && !vm_pages_needed) {
1546 vm_pageout_page_stats();
1547 vm_page_unlock_queues();
1551 if (vm_pages_needed)
1553 vm_page_unlock_queues();
1554 vm_pageout_scan(pass);
1559 * Unless the page queue lock is held by the caller, this function
1560 * should be regarded as advisory. Specifically, the caller should
1561 * not msleep() on &cnt.v_free_count following this function unless
1562 * the page queue lock is held until the msleep() is performed.
1568 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1569 vm_pages_needed = 1;
1570 wakeup(&vm_pages_needed);
1574 #if !defined(NO_SWAPPING)
1576 vm_req_vmdaemon(int req)
1578 static int lastrun = 0;
1580 mtx_lock(&vm_daemon_mtx);
1581 vm_pageout_req_swapout |= req;
1582 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1583 wakeup(&vm_daemon_needed);
1586 mtx_unlock(&vm_daemon_mtx);
1592 struct rlimit rsslim;
1595 int breakout, swapout_flags;
1598 mtx_lock(&vm_daemon_mtx);
1599 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1600 swapout_flags = vm_pageout_req_swapout;
1601 vm_pageout_req_swapout = 0;
1602 mtx_unlock(&vm_daemon_mtx);
1604 swapout_procs(swapout_flags);
1607 * scan the processes for exceeding their rlimits or if
1608 * process is swapped out -- deactivate pages
1610 sx_slock(&allproc_lock);
1611 LIST_FOREACH(p, &allproc, p_list) {
1612 vm_pindex_t limit, size;
1615 * if this is a system process or if we have already
1616 * looked at this process, skip it.
1619 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1624 * if the process is in a non-running type state,
1627 mtx_lock_spin(&sched_lock);
1629 FOREACH_THREAD_IN_PROC(p, td) {
1630 if (!TD_ON_RUNQ(td) &&
1631 !TD_IS_RUNNING(td) &&
1632 !TD_IS_SLEEPING(td)) {
1637 mtx_unlock_spin(&sched_lock);
1645 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1647 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1650 * let processes that are swapped out really be
1651 * swapped out set the limit to nothing (will force a
1654 if ((p->p_sflag & PS_INMEM) == 0)
1655 limit = 0; /* XXX */
1658 size = vmspace_resident_count(p->p_vmspace);
1659 if (limit >= 0 && size >= limit) {
1660 vm_pageout_map_deactivate_pages(
1661 &p->p_vmspace->vm_map, limit);
1664 sx_sunlock(&allproc_lock);
1667 #endif /* !defined(NO_SWAPPING) */