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/racct.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sched.h>
92 #include <sys/signalvar.h>
94 #include <sys/vnode.h>
95 #include <sys/vmmeter.h>
96 #include <sys/rwlock.h>
98 #include <sys/sysctl.h>
101 #include <vm/vm_param.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_map.h>
105 #include <vm/vm_pageout.h>
106 #include <vm/vm_pager.h>
107 #include <vm/vm_phys.h>
108 #include <vm/swap_pager.h>
109 #include <vm/vm_extern.h>
113 * System initialization
116 /* the kernel process "vm_pageout"*/
117 static void vm_pageout(void);
118 static int vm_pageout_clean(vm_page_t);
119 static void vm_pageout_scan(struct vm_domain *vmd, int pass);
120 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass);
122 struct proc *pageproc;
124 static struct kproc_desc page_kp = {
129 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
132 #if !defined(NO_SWAPPING)
133 /* the kernel process "vm_daemon"*/
134 static void vm_daemon(void);
135 static struct proc *vmproc;
137 static struct kproc_desc vm_kp = {
142 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
146 int vm_pages_needed; /* Event on which pageout daemon sleeps */
147 int vm_pageout_deficit; /* Estimated number of pages deficit */
148 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
149 int vm_pageout_wakeup_thresh;
151 #if !defined(NO_SWAPPING)
152 static int vm_pageout_req_swapout; /* XXX */
153 static int vm_daemon_needed;
154 static struct mtx vm_daemon_mtx;
155 /* Allow for use by vm_pageout before vm_daemon is initialized. */
156 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
158 static int vm_max_launder = 32;
159 static int vm_pageout_update_period;
160 static int defer_swap_pageouts;
161 static int disable_swap_pageouts;
162 static int lowmem_period = 10;
163 static int lowmem_ticks;
165 #if defined(NO_SWAPPING)
166 static int vm_swap_enabled = 0;
167 static int vm_swap_idle_enabled = 0;
169 static int vm_swap_enabled = 1;
170 static int vm_swap_idle_enabled = 0;
173 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
174 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
175 "free page threshold for waking up the pageout daemon");
177 SYSCTL_INT(_vm, OID_AUTO, max_launder,
178 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
180 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
181 CTLFLAG_RW, &vm_pageout_update_period, 0,
182 "Maximum active LRU update period");
184 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
185 "Low memory callback period");
187 #if defined(NO_SWAPPING)
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
193 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
194 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
195 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
196 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
199 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
200 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
202 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
203 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
205 static int pageout_lock_miss;
206 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
207 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
209 #define VM_PAGEOUT_PAGE_COUNT 16
210 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
212 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
213 SYSCTL_INT(_vm, OID_AUTO, max_wired,
214 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
216 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
217 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t,
219 #if !defined(NO_SWAPPING)
220 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
221 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
222 static void vm_req_vmdaemon(int req);
224 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
227 * Initialize a dummy page for marking the caller's place in the specified
228 * paging queue. In principle, this function only needs to set the flag
229 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count
230 * to one as safety precautions.
233 vm_pageout_init_marker(vm_page_t marker, u_short queue)
236 bzero(marker, sizeof(*marker));
237 marker->flags = PG_MARKER;
238 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
239 marker->queue = queue;
240 marker->hold_count = 1;
244 * vm_pageout_fallback_object_lock:
246 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
247 * known to have failed and page queue must be either PQ_ACTIVE or
248 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
249 * while locking the vm object. Use marker page to detect page queue
250 * changes and maintain notion of next page on page queue. Return
251 * TRUE if no changes were detected, FALSE otherwise. vm object is
254 * This function depends on both the lock portion of struct vm_object
255 * and normal struct vm_page being type stable.
258 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
260 struct vm_page marker;
261 struct vm_pagequeue *pq;
267 vm_pageout_init_marker(&marker, queue);
268 pq = vm_page_pagequeue(m);
271 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
272 vm_pagequeue_unlock(pq);
274 VM_OBJECT_WLOCK(object);
276 vm_pagequeue_lock(pq);
278 /* Page queue might have changed. */
279 *next = TAILQ_NEXT(&marker, plinks.q);
280 unchanged = (m->queue == queue &&
281 m->object == object &&
282 &marker == TAILQ_NEXT(m, plinks.q));
283 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
288 * Lock the page while holding the page queue lock. Use marker page
289 * to detect page queue changes and maintain notion of next page on
290 * page queue. Return TRUE if no changes were detected, FALSE
291 * otherwise. The page is locked on return. The page queue lock might
292 * be dropped and reacquired.
294 * This function depends on normal struct vm_page being type stable.
297 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
299 struct vm_page marker;
300 struct vm_pagequeue *pq;
304 vm_page_lock_assert(m, MA_NOTOWNED);
305 if (vm_page_trylock(m))
309 vm_pageout_init_marker(&marker, queue);
310 pq = vm_page_pagequeue(m);
312 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
313 vm_pagequeue_unlock(pq);
315 vm_pagequeue_lock(pq);
317 /* Page queue might have changed. */
318 *next = TAILQ_NEXT(&marker, plinks.q);
319 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, plinks.q));
320 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
327 * Clean the page and remove it from the laundry.
329 * We set the busy bit to cause potential page faults on this page to
330 * block. Note the careful timing, however, the busy bit isn't set till
331 * late and we cannot do anything that will mess with the page.
334 vm_pageout_clean(vm_page_t m)
337 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
339 int ib, is, page_base;
340 vm_pindex_t pindex = m->pindex;
342 vm_page_lock_assert(m, MA_OWNED);
344 VM_OBJECT_ASSERT_WLOCKED(object);
347 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
348 * with the new swapper, but we could have serious problems paging
349 * out other object types if there is insufficient memory.
351 * Unfortunately, checking free memory here is far too late, so the
352 * check has been moved up a procedural level.
356 * Can't clean the page if it's busy or held.
358 vm_page_assert_unbusied(m);
359 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
362 mc[vm_pageout_page_count] = pb = ps = m;
364 page_base = vm_pageout_page_count;
369 * Scan object for clusterable pages.
371 * We can cluster ONLY if: ->> the page is NOT
372 * clean, wired, busy, held, or mapped into a
373 * buffer, and one of the following:
374 * 1) The page is inactive, or a seldom used
377 * 2) we force the issue.
379 * During heavy mmap/modification loads the pageout
380 * daemon can really fragment the underlying file
381 * due to flushing pages out of order and not trying
382 * align the clusters (which leave sporatic out-of-order
383 * holes). To solve this problem we do the reverse scan
384 * first and attempt to align our cluster, then do a
385 * forward scan if room remains.
388 while (ib && pageout_count < vm_pageout_page_count) {
396 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
401 vm_page_test_dirty(p);
403 p->queue != PQ_INACTIVE ||
404 p->hold_count != 0) { /* may be undergoing I/O */
410 mc[--page_base] = pb = p;
414 * alignment boundry, stop here and switch directions. Do
417 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
421 while (pageout_count < vm_pageout_page_count &&
422 pindex + is < object->size) {
425 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
428 vm_page_test_dirty(p);
430 p->queue != PQ_INACTIVE ||
431 p->hold_count != 0) { /* may be undergoing I/O */
436 mc[page_base + pageout_count] = ps = p;
442 * If we exhausted our forward scan, continue with the reverse scan
443 * when possible, even past a page boundry. This catches boundry
446 if (ib && pageout_count < vm_pageout_page_count)
450 * we allow reads during pageouts...
452 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
457 * vm_pageout_flush() - launder the given pages
459 * The given pages are laundered. Note that we setup for the start of
460 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
461 * reference count all in here rather then in the parent. If we want
462 * the parent to do more sophisticated things we may have to change
465 * Returned runlen is the count of pages between mreq and first
466 * page after mreq with status VM_PAGER_AGAIN.
467 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
468 * for any page in runlen set.
471 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
474 vm_object_t object = mc[0]->object;
475 int pageout_status[count];
479 VM_OBJECT_ASSERT_WLOCKED(object);
482 * Initiate I/O. Bump the vm_page_t->busy counter and
483 * mark the pages read-only.
485 * We do not have to fixup the clean/dirty bits here... we can
486 * allow the pager to do it after the I/O completes.
488 * NOTE! mc[i]->dirty may be partial or fragmented due to an
489 * edge case with file fragments.
491 for (i = 0; i < count; i++) {
492 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
493 ("vm_pageout_flush: partially invalid page %p index %d/%d",
495 vm_page_sbusy(mc[i]);
496 pmap_remove_write(mc[i]);
498 vm_object_pip_add(object, count);
500 vm_pager_put_pages(object, mc, count, flags, pageout_status);
502 runlen = count - mreq;
505 for (i = 0; i < count; i++) {
506 vm_page_t mt = mc[i];
508 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
509 !pmap_page_is_write_mapped(mt),
510 ("vm_pageout_flush: page %p is not write protected", mt));
511 switch (pageout_status[i]) {
518 * Page outside of range of object. Right now we
519 * essentially lose the changes by pretending it
527 * If page couldn't be paged out, then reactivate the
528 * page so it doesn't clog the inactive list. (We
529 * will try paging out it again later).
532 vm_page_activate(mt);
534 if (eio != NULL && i >= mreq && i - mreq < runlen)
538 if (i >= mreq && i - mreq < runlen)
544 * If the operation is still going, leave the page busy to
545 * block all other accesses. Also, leave the paging in
546 * progress indicator set so that we don't attempt an object
549 if (pageout_status[i] != VM_PAGER_PEND) {
550 vm_object_pip_wakeup(object);
552 if (vm_page_count_severe()) {
554 vm_page_try_to_cache(mt);
561 return (numpagedout);
565 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low,
572 vm_page_t m, m_tmp, next;
575 vm_pagequeue_lock(pq);
576 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) {
577 if ((m->flags & PG_MARKER) != 0)
579 pa = VM_PAGE_TO_PHYS(m);
580 if (pa < low || pa + PAGE_SIZE > high)
582 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
587 if ((!VM_OBJECT_TRYWLOCK(object) &&
588 (!vm_pageout_fallback_object_lock(m, &next) ||
589 m->hold_count != 0)) || vm_page_busied(m)) {
591 VM_OBJECT_WUNLOCK(object);
594 vm_page_test_dirty(m);
595 if (m->dirty == 0 && object->ref_count != 0)
599 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) {
600 VM_OBJECT_WUNLOCK(object);
603 if (object->type == OBJT_VNODE) {
604 vm_pagequeue_unlock(pq);
606 vm_object_reference_locked(object);
607 VM_OBJECT_WUNLOCK(object);
608 (void)vn_start_write(vp, &mp, V_WAIT);
609 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
610 LK_SHARED : LK_EXCLUSIVE;
611 vn_lock(vp, lockmode | LK_RETRY);
612 VM_OBJECT_WLOCK(object);
613 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
614 VM_OBJECT_WUNLOCK(object);
616 vm_object_deallocate(object);
617 vn_finished_write(mp);
619 } else if (object->type == OBJT_SWAP ||
620 object->type == OBJT_DEFAULT) {
621 vm_pagequeue_unlock(pq);
623 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC,
625 VM_OBJECT_WUNLOCK(object);
630 * Dequeue here to prevent lock recursion in
633 vm_page_dequeue_locked(m);
637 VM_OBJECT_WUNLOCK(object);
639 vm_pagequeue_unlock(pq);
644 * Increase the number of cached pages. The specified value, "tries",
645 * determines which categories of pages are cached:
647 * 0: All clean, inactive pages within the specified physical address range
648 * are cached. Will not sleep.
649 * 1: The vm_lowmem handlers are called. All inactive pages within
650 * the specified physical address range are cached. May sleep.
651 * 2: The vm_lowmem handlers are called. All inactive and active pages
652 * within the specified physical address range are cached. May sleep.
655 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
657 int actl, actmax, inactl, inactmax, dom, initial_dom;
658 static int start_dom = 0;
662 * Decrease registered cache sizes. The vm_lowmem handlers
663 * may acquire locks and/or sleep, so they can only be invoked
664 * when "tries" is greater than zero.
666 EVENTHANDLER_INVOKE(vm_lowmem, 0);
669 * We do this explicitly after the caches have been drained
676 * Make the next scan start on the next domain.
678 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains;
681 inactmax = vm_cnt.v_inactive_count;
683 actmax = tries < 2 ? 0 : vm_cnt.v_active_count;
687 * Scan domains in round-robin order, first inactive queues,
688 * then active. Since domain usually owns large physically
689 * contiguous chunk of memory, it makes sense to completely
690 * exhaust one domain before switching to next, while growing
691 * the pool of contiguous physical pages.
693 * Do not even start launder a domain which cannot contain
694 * the specified address range, as indicated by segments
695 * constituting the domain.
698 if (inactl < inactmax) {
699 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
701 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE],
706 if (++dom == vm_ndomains)
708 if (dom != initial_dom)
712 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
714 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE],
719 if (++dom == vm_ndomains)
721 if (dom != initial_dom)
726 #if !defined(NO_SWAPPING)
728 * vm_pageout_object_deactivate_pages
730 * Deactivate enough pages to satisfy the inactive target
733 * The object and map must be locked.
736 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
739 vm_object_t backing_object, object;
741 int act_delta, remove_mode;
743 VM_OBJECT_ASSERT_LOCKED(first_object);
744 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
746 for (object = first_object;; object = backing_object) {
747 if (pmap_resident_count(pmap) <= desired)
749 VM_OBJECT_ASSERT_LOCKED(object);
750 if ((object->flags & OBJ_UNMANAGED) != 0 ||
751 object->paging_in_progress != 0)
755 if (object->shadow_count > 1)
758 * Scan the object's entire memory queue.
760 TAILQ_FOREACH(p, &object->memq, listq) {
761 if (pmap_resident_count(pmap) <= desired)
763 if (vm_page_busied(p))
765 PCPU_INC(cnt.v_pdpages);
767 if (p->wire_count != 0 || p->hold_count != 0 ||
768 !pmap_page_exists_quick(pmap, p)) {
772 act_delta = pmap_ts_referenced(p);
773 if ((p->aflags & PGA_REFERENCED) != 0) {
776 vm_page_aflag_clear(p, PGA_REFERENCED);
778 if (p->queue != PQ_ACTIVE && act_delta != 0) {
780 p->act_count += act_delta;
781 } else if (p->queue == PQ_ACTIVE) {
782 if (act_delta == 0) {
783 p->act_count -= min(p->act_count,
785 if (!remove_mode && p->act_count == 0) {
787 vm_page_deactivate(p);
792 if (p->act_count < ACT_MAX -
794 p->act_count += ACT_ADVANCE;
797 } else if (p->queue == PQ_INACTIVE)
801 if ((backing_object = object->backing_object) == NULL)
803 VM_OBJECT_RLOCK(backing_object);
804 if (object != first_object)
805 VM_OBJECT_RUNLOCK(object);
808 if (object != first_object)
809 VM_OBJECT_RUNLOCK(object);
813 * deactivate some number of pages in a map, try to do it fairly, but
814 * that is really hard to do.
817 vm_pageout_map_deactivate_pages(map, desired)
822 vm_object_t obj, bigobj;
825 if (!vm_map_trylock(map))
832 * first, search out the biggest object, and try to free pages from
835 tmpe = map->header.next;
836 while (tmpe != &map->header) {
837 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
838 obj = tmpe->object.vm_object;
839 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
840 if (obj->shadow_count <= 1 &&
842 bigobj->resident_page_count < obj->resident_page_count)) {
844 VM_OBJECT_RUNLOCK(bigobj);
847 VM_OBJECT_RUNLOCK(obj);
850 if (tmpe->wired_count > 0)
851 nothingwired = FALSE;
855 if (bigobj != NULL) {
856 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
857 VM_OBJECT_RUNLOCK(bigobj);
860 * Next, hunt around for other pages to deactivate. We actually
861 * do this search sort of wrong -- .text first is not the best idea.
863 tmpe = map->header.next;
864 while (tmpe != &map->header) {
865 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
867 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
868 obj = tmpe->object.vm_object;
870 VM_OBJECT_RLOCK(obj);
871 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
872 VM_OBJECT_RUNLOCK(obj);
880 * Remove all non-wired, managed mappings if a process is swapped out.
881 * This will free page table pages.
884 pmap_remove_pages(map->pmap);
887 * Remove all mappings if a process is swapped out, this will free page
890 if (desired == 0 && nothingwired) {
891 pmap_remove(vm_map_pmap(map), vm_map_min(map),
898 #endif /* !defined(NO_SWAPPING) */
901 * vm_pageout_scan does the dirty work for the pageout daemon.
903 * pass 0 - Update active LRU/deactivate pages
904 * pass 1 - Move inactive to cache or free
905 * pass 2 - Launder dirty pages
908 vm_pageout_scan(struct vm_domain *vmd, int pass)
911 struct vm_pagequeue *pq;
913 int act_delta, addl_page_shortage, deficit, maxscan, page_shortage;
914 int vnodes_skipped = 0;
917 boolean_t queues_locked;
920 * If we need to reclaim memory ask kernel caches to return
921 * some. We rate limit to avoid thrashing.
923 if (vmd == &vm_dom[0] && pass > 0 &&
924 lowmem_ticks + (lowmem_period * hz) < ticks) {
926 * Decrease registered cache sizes.
928 EVENTHANDLER_INVOKE(vm_lowmem, 0);
930 * We do this explicitly after the caches have been
934 lowmem_ticks = ticks;
938 * The addl_page_shortage is the number of temporarily
939 * stuck pages in the inactive queue. In other words, the
940 * number of pages from the inactive count that should be
941 * discounted in setting the target for the active queue scan.
943 addl_page_shortage = 0;
946 * Calculate the number of pages we want to either free or move
950 deficit = atomic_readandclear_int(&vm_pageout_deficit);
951 page_shortage = vm_paging_target() + deficit;
953 page_shortage = deficit = 0;
956 * maxlaunder limits the number of dirty pages we flush per scan.
957 * For most systems a smaller value (16 or 32) is more robust under
958 * extreme memory and disk pressure because any unnecessary writes
959 * to disk can result in extreme performance degredation. However,
960 * systems with excessive dirty pages (especially when MAP_NOSYNC is
961 * used) will die horribly with limited laundering. If the pageout
962 * daemon cannot clean enough pages in the first pass, we let it go
963 * all out in succeeding passes.
965 if ((maxlaunder = vm_max_launder) <= 1)
971 * Start scanning the inactive queue for pages we can move to the
972 * cache or free. The scan will stop when the target is reached or
973 * we have scanned the entire inactive queue. Note that m->act_count
974 * is not used to form decisions for the inactive queue, only for the
977 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
978 maxscan = pq->pq_cnt;
979 vm_pagequeue_lock(pq);
980 queues_locked = TRUE;
981 for (m = TAILQ_FIRST(&pq->pq_pl);
982 m != NULL && maxscan-- > 0 && page_shortage > 0;
984 vm_pagequeue_assert_locked(pq);
985 KASSERT(queues_locked, ("unlocked queues"));
986 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
988 PCPU_INC(cnt.v_pdpages);
989 next = TAILQ_NEXT(m, plinks.q);
994 if (m->flags & PG_MARKER)
997 KASSERT((m->flags & PG_FICTITIOUS) == 0,
998 ("Fictitious page %p cannot be in inactive queue", m));
999 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1000 ("Unmanaged page %p cannot be in inactive queue", m));
1003 * The page or object lock acquisitions fail if the
1004 * page was removed from the queue or moved to a
1005 * different position within the queue. In either
1006 * case, addl_page_shortage should not be incremented.
1008 if (!vm_pageout_page_lock(m, &next)) {
1013 if (!VM_OBJECT_TRYWLOCK(object) &&
1014 !vm_pageout_fallback_object_lock(m, &next)) {
1016 VM_OBJECT_WUNLOCK(object);
1021 * Don't mess with busy pages, keep them at at the
1022 * front of the queue, most likely they are being
1023 * paged out. Increment addl_page_shortage for busy
1024 * pages, because they may leave the inactive queue
1025 * shortly after page scan is finished.
1027 if (vm_page_busied(m)) {
1029 VM_OBJECT_WUNLOCK(object);
1030 addl_page_shortage++;
1035 * We unlock the inactive page queue, invalidating the
1036 * 'next' pointer. Use our marker to remember our
1039 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1040 vm_pagequeue_unlock(pq);
1041 queues_locked = FALSE;
1044 * We bump the activation count if the page has been
1045 * referenced while in the inactive queue. This makes
1046 * it less likely that the page will be added back to the
1047 * inactive queue prematurely again. Here we check the
1048 * page tables (or emulated bits, if any), given the upper
1049 * level VM system not knowing anything about existing
1052 if ((m->aflags & PGA_REFERENCED) != 0) {
1053 vm_page_aflag_clear(m, PGA_REFERENCED);
1057 if (object->ref_count != 0) {
1058 act_delta += pmap_ts_referenced(m);
1060 KASSERT(!pmap_page_is_mapped(m),
1061 ("vm_pageout_scan: page %p is mapped", m));
1065 * If the upper level VM system knows about any page
1066 * references, we reactivate the page or requeue it.
1068 if (act_delta != 0) {
1069 if (object->ref_count != 0) {
1070 vm_page_activate(m);
1071 m->act_count += act_delta + ACT_ADVANCE;
1073 vm_pagequeue_lock(pq);
1074 queues_locked = TRUE;
1075 vm_page_requeue_locked(m);
1077 VM_OBJECT_WUNLOCK(object);
1082 if (m->hold_count != 0) {
1084 VM_OBJECT_WUNLOCK(object);
1087 * Held pages are essentially stuck in the
1088 * queue. So, they ought to be discounted
1089 * from the inactive count. See the
1090 * calculation of the page_shortage for the
1091 * loop over the active queue below.
1093 addl_page_shortage++;
1098 * If the page appears to be clean at the machine-independent
1099 * layer, then remove all of its mappings from the pmap in
1100 * anticipation of placing it onto the cache queue. If,
1101 * however, any of the page's mappings allow write access,
1102 * then the page may still be modified until the last of those
1103 * mappings are removed.
1105 vm_page_test_dirty(m);
1106 if (m->dirty == 0 && object->ref_count != 0)
1109 if (m->valid == 0) {
1111 * Invalid pages can be easily freed
1114 PCPU_INC(cnt.v_dfree);
1116 } else if (m->dirty == 0) {
1118 * Clean pages can be placed onto the cache queue.
1119 * This effectively frees them.
1123 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1125 * Dirty pages need to be paged out, but flushing
1126 * a page is extremely expensive versus freeing
1127 * a clean page. Rather then artificially limiting
1128 * the number of pages we can flush, we instead give
1129 * dirty pages extra priority on the inactive queue
1130 * by forcing them to be cycled through the queue
1131 * twice before being flushed, after which the
1132 * (now clean) page will cycle through once more
1133 * before being freed. This significantly extends
1134 * the thrash point for a heavily loaded machine.
1136 m->flags |= PG_WINATCFLS;
1137 vm_pagequeue_lock(pq);
1138 queues_locked = TRUE;
1139 vm_page_requeue_locked(m);
1140 } else if (maxlaunder > 0) {
1142 * We always want to try to flush some dirty pages if
1143 * we encounter them, to keep the system stable.
1144 * Normally this number is small, but under extreme
1145 * pressure where there are insufficient clean pages
1146 * on the inactive queue, we may have to go all out.
1148 int swap_pageouts_ok;
1149 struct vnode *vp = NULL;
1150 struct mount *mp = NULL;
1152 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
1153 swap_pageouts_ok = 1;
1155 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
1156 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
1157 vm_page_count_min());
1162 * We don't bother paging objects that are "dead".
1163 * Those objects are in a "rundown" state.
1165 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
1166 vm_pagequeue_lock(pq);
1168 VM_OBJECT_WUNLOCK(object);
1169 queues_locked = TRUE;
1170 vm_page_requeue_locked(m);
1175 * The object is already known NOT to be dead. It
1176 * is possible for the vget() to block the whole
1177 * pageout daemon, but the new low-memory handling
1178 * code should prevent it.
1180 * The previous code skipped locked vnodes and, worse,
1181 * reordered pages in the queue. This results in
1182 * completely non-deterministic operation and, on a
1183 * busy system, can lead to extremely non-optimal
1184 * pageouts. For example, it can cause clean pages
1185 * to be freed and dirty pages to be moved to the end
1186 * of the queue. Since dirty pages are also moved to
1187 * the end of the queue once-cleaned, this gives
1188 * way too large a weighting to deferring the freeing
1191 * We can't wait forever for the vnode lock, we might
1192 * deadlock due to a vn_read() getting stuck in
1193 * vm_wait while holding this vnode. We skip the
1194 * vnode if we can't get it in a reasonable amount
1197 if (object->type == OBJT_VNODE) {
1199 vp = object->handle;
1200 if (vp->v_type == VREG &&
1201 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1203 ++pageout_lock_miss;
1204 if (object->flags & OBJ_MIGHTBEDIRTY)
1206 goto unlock_and_continue;
1209 ("vp %p with NULL v_mount", vp));
1210 vm_object_reference_locked(object);
1211 VM_OBJECT_WUNLOCK(object);
1212 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
1213 LK_SHARED : LK_EXCLUSIVE;
1214 if (vget(vp, lockmode | LK_TIMELOCK,
1216 VM_OBJECT_WLOCK(object);
1217 ++pageout_lock_miss;
1218 if (object->flags & OBJ_MIGHTBEDIRTY)
1221 goto unlock_and_continue;
1223 VM_OBJECT_WLOCK(object);
1225 vm_pagequeue_lock(pq);
1226 queues_locked = TRUE;
1228 * The page might have been moved to another
1229 * queue during potential blocking in vget()
1230 * above. The page might have been freed and
1231 * reused for another vnode.
1233 if (m->queue != PQ_INACTIVE ||
1234 m->object != object ||
1235 TAILQ_NEXT(m, plinks.q) != &vmd->vmd_marker) {
1237 if (object->flags & OBJ_MIGHTBEDIRTY)
1239 goto unlock_and_continue;
1243 * The page may have been busied during the
1244 * blocking in vget(). We don't move the
1245 * page back onto the end of the queue so that
1246 * statistics are more correct if we don't.
1248 if (vm_page_busied(m)) {
1250 addl_page_shortage++;
1251 goto unlock_and_continue;
1255 * If the page has become held it might
1256 * be undergoing I/O, so skip it
1258 if (m->hold_count != 0) {
1260 addl_page_shortage++;
1261 if (object->flags & OBJ_MIGHTBEDIRTY)
1263 goto unlock_and_continue;
1265 vm_pagequeue_unlock(pq);
1266 queues_locked = FALSE;
1270 * If a page is dirty, then it is either being washed
1271 * (but not yet cleaned) or it is still in the
1272 * laundry. If it is still in the laundry, then we
1273 * start the cleaning operation.
1275 * decrement page_shortage on success to account for
1276 * the (future) cleaned page. Otherwise we could wind
1277 * up laundering or cleaning too many pages.
1279 if (vm_pageout_clean(m) != 0) {
1283 unlock_and_continue:
1284 vm_page_lock_assert(m, MA_NOTOWNED);
1285 VM_OBJECT_WUNLOCK(object);
1287 if (queues_locked) {
1288 vm_pagequeue_unlock(pq);
1289 queues_locked = FALSE;
1293 vm_object_deallocate(object);
1294 vn_finished_write(mp);
1296 vm_page_lock_assert(m, MA_NOTOWNED);
1300 VM_OBJECT_WUNLOCK(object);
1302 if (!queues_locked) {
1303 vm_pagequeue_lock(pq);
1304 queues_locked = TRUE;
1306 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1307 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1309 vm_pagequeue_unlock(pq);
1312 * Compute the number of pages we want to try to move from the
1313 * active queue to the inactive queue.
1315 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count +
1316 vm_paging_target() + deficit + addl_page_shortage;
1318 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1319 vm_pagequeue_lock(pq);
1320 maxscan = pq->pq_cnt;
1323 * If we're just idle polling attempt to visit every
1324 * active page within 'update_period' seconds.
1326 if (pass == 0 && vm_pageout_update_period != 0) {
1327 maxscan /= vm_pageout_update_period;
1328 page_shortage = maxscan;
1332 * Scan the active queue for things we can deactivate. We nominally
1333 * track the per-page activity counter and use it to locate
1334 * deactivation candidates.
1336 m = TAILQ_FIRST(&pq->pq_pl);
1337 while (m != NULL && maxscan-- > 0 && page_shortage > 0) {
1339 KASSERT(m->queue == PQ_ACTIVE,
1340 ("vm_pageout_scan: page %p isn't active", m));
1342 next = TAILQ_NEXT(m, plinks.q);
1343 if ((m->flags & PG_MARKER) != 0) {
1347 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1348 ("Fictitious page %p cannot be in active queue", m));
1349 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1350 ("Unmanaged page %p cannot be in active queue", m));
1351 if (!vm_pageout_page_lock(m, &next)) {
1358 * The count for pagedaemon pages is done after checking the
1359 * page for eligibility...
1361 PCPU_INC(cnt.v_pdpages);
1364 * Check to see "how much" the page has been used.
1366 if ((m->aflags & PGA_REFERENCED) != 0) {
1367 vm_page_aflag_clear(m, PGA_REFERENCED);
1373 * Unlocked object ref count check. Two races are possible.
1374 * 1) The ref was transitioning to zero and we saw non-zero,
1375 * the pmap bits will be checked unnecessarily.
1376 * 2) The ref was transitioning to one and we saw zero.
1377 * The page lock prevents a new reference to this page so
1378 * we need not check the reference bits.
1380 if (m->object->ref_count != 0)
1381 act_delta += pmap_ts_referenced(m);
1384 * Advance or decay the act_count based on recent usage.
1386 if (act_delta != 0) {
1387 m->act_count += ACT_ADVANCE + act_delta;
1388 if (m->act_count > ACT_MAX)
1389 m->act_count = ACT_MAX;
1391 m->act_count -= min(m->act_count, ACT_DECLINE);
1394 * Move this page to the tail of the active or inactive
1395 * queue depending on usage.
1397 if (m->act_count == 0) {
1398 /* Dequeue to avoid later lock recursion. */
1399 vm_page_dequeue_locked(m);
1400 vm_page_deactivate(m);
1403 vm_page_requeue_locked(m);
1407 vm_pagequeue_unlock(pq);
1408 #if !defined(NO_SWAPPING)
1410 * Idle process swapout -- run once per second.
1412 if (vm_swap_idle_enabled) {
1414 if (time_second != lsec) {
1415 vm_req_vmdaemon(VM_SWAP_IDLE);
1422 * If we didn't get enough free pages, and we have skipped a vnode
1423 * in a writeable object, wakeup the sync daemon. And kick swapout
1424 * if we did not get enough free pages.
1426 if (vm_paging_target() > 0) {
1427 if (vnodes_skipped && vm_page_count_min())
1428 (void) speedup_syncer();
1429 #if !defined(NO_SWAPPING)
1430 if (vm_swap_enabled && vm_page_count_target())
1431 vm_req_vmdaemon(VM_SWAP_NORMAL);
1436 * If we are critically low on one of RAM or swap and low on
1437 * the other, kill the largest process. However, we avoid
1438 * doing this on the first pass in order to give ourselves a
1439 * chance to flush out dirty vnode-backed pages and to allow
1440 * active pages to be moved to the inactive queue and reclaimed.
1442 vm_pageout_mightbe_oom(vmd, pass);
1445 static int vm_pageout_oom_vote;
1448 * The pagedaemon threads randlomly select one to perform the
1449 * OOM. Trying to kill processes before all pagedaemons
1450 * failed to reach free target is premature.
1453 vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass)
1457 if (pass <= 1 || !((swap_pager_avail < 64 && vm_page_count_min()) ||
1458 (swap_pager_full && vm_paging_target() > 0))) {
1460 vmd->vmd_oom = FALSE;
1461 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1469 vmd->vmd_oom = TRUE;
1470 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1471 if (old_vote != vm_ndomains - 1)
1475 * The current pagedaemon thread is the last in the quorum to
1476 * start OOM. Initiate the selection and signaling of the
1479 vm_pageout_oom(VM_OOM_MEM);
1482 * After one round of OOM terror, recall our vote. On the
1483 * next pass, current pagedaemon would vote again if the low
1484 * memory condition is still there, due to vmd_oom being
1487 vmd->vmd_oom = FALSE;
1488 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1492 vm_pageout_oom(int shortage)
1494 struct proc *p, *bigproc;
1495 vm_offset_t size, bigsize;
1500 * We keep the process bigproc locked once we find it to keep anyone
1501 * from messing with it; however, there is a possibility of
1502 * deadlock if process B is bigproc and one of it's child processes
1503 * attempts to propagate a signal to B while we are waiting for A's
1504 * lock while walking this list. To avoid this, we don't block on
1505 * the process lock but just skip a process if it is already locked.
1509 sx_slock(&allproc_lock);
1510 FOREACH_PROC_IN_SYSTEM(p) {
1513 if (PROC_TRYLOCK(p) == 0)
1516 * If this is a system, protected or killed process, skip it.
1518 if (p->p_state != PRS_NORMAL ||
1519 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1520 (p->p_pid == 1) || P_KILLED(p) ||
1521 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1526 * If the process is in a non-running type state,
1527 * don't touch it. Check all the threads individually.
1530 FOREACH_THREAD_IN_PROC(p, td) {
1532 if (!TD_ON_RUNQ(td) &&
1533 !TD_IS_RUNNING(td) &&
1534 !TD_IS_SLEEPING(td) &&
1535 !TD_IS_SUSPENDED(td)) {
1547 * get the process size
1549 vm = vmspace_acquire_ref(p);
1554 if (!vm_map_trylock_read(&vm->vm_map)) {
1559 size = vmspace_swap_count(vm);
1560 vm_map_unlock_read(&vm->vm_map);
1561 if (shortage == VM_OOM_MEM)
1562 size += vmspace_resident_count(vm);
1565 * if the this process is bigger than the biggest one
1568 if (size > bigsize) {
1569 if (bigproc != NULL)
1570 PROC_UNLOCK(bigproc);
1576 sx_sunlock(&allproc_lock);
1577 if (bigproc != NULL) {
1578 killproc(bigproc, "out of swap space");
1579 sched_nice(bigproc, PRIO_MIN);
1580 PROC_UNLOCK(bigproc);
1581 wakeup(&vm_cnt.v_free_count);
1586 vm_pageout_worker(void *arg)
1588 struct vm_domain *domain;
1591 domidx = (uintptr_t)arg;
1592 domain = &vm_dom[domidx];
1595 * XXXKIB It could be useful to bind pageout daemon threads to
1596 * the cores belonging to the domain, from which vm_page_array
1600 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1601 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1604 * The pageout daemon worker is never done, so loop forever.
1608 * If we have enough free memory, wakeup waiters. Do
1609 * not clear vm_pages_needed until we reach our target,
1610 * otherwise we may be woken up over and over again and
1611 * waste a lot of cpu.
1613 mtx_lock(&vm_page_queue_free_mtx);
1614 if (vm_pages_needed && !vm_page_count_min()) {
1615 if (!vm_paging_needed())
1616 vm_pages_needed = 0;
1617 wakeup(&vm_cnt.v_free_count);
1619 if (vm_pages_needed) {
1621 * Still not done, take a second pass without waiting
1622 * (unlimited dirty cleaning), otherwise sleep a bit
1625 if (domain->vmd_pass > 1)
1626 msleep(&vm_pages_needed,
1627 &vm_page_queue_free_mtx, PVM, "psleep",
1631 * Good enough, sleep until required to refresh
1634 domain->vmd_pass = 0;
1635 msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1639 if (vm_pages_needed) {
1640 vm_cnt.v_pdwakeups++;
1643 mtx_unlock(&vm_page_queue_free_mtx);
1644 vm_pageout_scan(domain, domain->vmd_pass);
1649 * vm_pageout is the high level pageout daemon.
1659 * Initialize some paging parameters.
1661 vm_cnt.v_interrupt_free_min = 2;
1662 if (vm_cnt.v_page_count < 2000)
1663 vm_pageout_page_count = 8;
1666 * v_free_reserved needs to include enough for the largest
1667 * swap pager structures plus enough for any pv_entry structs
1670 if (vm_cnt.v_page_count > 1024)
1671 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1673 vm_cnt.v_free_min = 4;
1674 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1675 vm_cnt.v_interrupt_free_min;
1676 vm_cnt.v_free_reserved = vm_pageout_page_count +
1677 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1678 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1679 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1680 vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1681 vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1682 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1683 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1684 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1687 * Set the default wakeup threshold to be 10% above the minimum
1688 * page limit. This keeps the steady state out of shortfall.
1690 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1693 * Set interval in seconds for active scan. We want to visit each
1694 * page at least once every ten minutes. This is to prevent worst
1695 * case paging behaviors with stale active LRU.
1697 if (vm_pageout_update_period == 0)
1698 vm_pageout_update_period = 600;
1700 /* XXX does not really belong here */
1701 if (vm_page_max_wired == 0)
1702 vm_page_max_wired = vm_cnt.v_free_count / 3;
1704 swap_pager_swap_init();
1706 for (i = 1; i < vm_ndomains; i++) {
1707 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1708 curproc, NULL, 0, 0, "dom%d", i);
1710 panic("starting pageout for domain %d, error %d\n",
1715 vm_pageout_worker((void *)(uintptr_t)0);
1719 * Unless the free page queue lock is held by the caller, this function
1720 * should be regarded as advisory. Specifically, the caller should
1721 * not msleep() on &vm_cnt.v_free_count following this function unless
1722 * the free page queue lock is held until the msleep() is performed.
1725 pagedaemon_wakeup(void)
1728 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1729 vm_pages_needed = 1;
1730 wakeup(&vm_pages_needed);
1734 #if !defined(NO_SWAPPING)
1736 vm_req_vmdaemon(int req)
1738 static int lastrun = 0;
1740 mtx_lock(&vm_daemon_mtx);
1741 vm_pageout_req_swapout |= req;
1742 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1743 wakeup(&vm_daemon_needed);
1746 mtx_unlock(&vm_daemon_mtx);
1752 struct rlimit rsslim;
1756 int breakout, swapout_flags, tryagain, attempts;
1758 uint64_t rsize, ravailable;
1762 mtx_lock(&vm_daemon_mtx);
1764 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz);
1766 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1768 swapout_flags = vm_pageout_req_swapout;
1769 vm_pageout_req_swapout = 0;
1770 mtx_unlock(&vm_daemon_mtx);
1772 swapout_procs(swapout_flags);
1775 * scan the processes for exceeding their rlimits or if
1776 * process is swapped out -- deactivate pages
1782 sx_slock(&allproc_lock);
1783 FOREACH_PROC_IN_SYSTEM(p) {
1784 vm_pindex_t limit, size;
1787 * if this is a system process or if we have already
1788 * looked at this process, skip it.
1791 if (p->p_state != PRS_NORMAL ||
1792 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1797 * if the process is in a non-running type state,
1801 FOREACH_THREAD_IN_PROC(p, td) {
1803 if (!TD_ON_RUNQ(td) &&
1804 !TD_IS_RUNNING(td) &&
1805 !TD_IS_SLEEPING(td) &&
1806 !TD_IS_SUSPENDED(td)) {
1820 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1822 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1825 * let processes that are swapped out really be
1826 * swapped out set the limit to nothing (will force a
1829 if ((p->p_flag & P_INMEM) == 0)
1830 limit = 0; /* XXX */
1831 vm = vmspace_acquire_ref(p);
1836 size = vmspace_resident_count(vm);
1837 if (size >= limit) {
1838 vm_pageout_map_deactivate_pages(
1839 &vm->vm_map, limit);
1842 rsize = IDX_TO_OFF(size);
1844 racct_set(p, RACCT_RSS, rsize);
1845 ravailable = racct_get_available(p, RACCT_RSS);
1847 if (rsize > ravailable) {
1849 * Don't be overly aggressive; this might be
1850 * an innocent process, and the limit could've
1851 * been exceeded by some memory hog. Don't
1852 * try to deactivate more than 1/4th of process'
1853 * resident set size.
1855 if (attempts <= 8) {
1856 if (ravailable < rsize - (rsize / 4))
1857 ravailable = rsize - (rsize / 4);
1859 vm_pageout_map_deactivate_pages(
1860 &vm->vm_map, OFF_TO_IDX(ravailable));
1861 /* Update RSS usage after paging out. */
1862 size = vmspace_resident_count(vm);
1863 rsize = IDX_TO_OFF(size);
1865 racct_set(p, RACCT_RSS, rsize);
1867 if (rsize > ravailable)
1873 sx_sunlock(&allproc_lock);
1874 if (tryagain != 0 && attempts <= 10)
1878 #endif /* !defined(NO_SWAPPING) */