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 "opt_kdtrace.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
83 #include <sys/eventhandler.h>
85 #include <sys/mutex.h>
87 #include <sys/kthread.h>
89 #include <sys/mount.h>
90 #include <sys/racct.h>
91 #include <sys/resourcevar.h>
92 #include <sys/sched.h>
94 #include <sys/signalvar.h>
97 #include <sys/vnode.h>
98 #include <sys/vmmeter.h>
99 #include <sys/rwlock.h>
101 #include <sys/sysctl.h>
104 #include <vm/vm_param.h>
105 #include <vm/vm_object.h>
106 #include <vm/vm_page.h>
107 #include <vm/vm_map.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_pager.h>
110 #include <vm/vm_phys.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
116 * System initialization
119 /* the kernel process "vm_pageout"*/
120 static void vm_pageout(void);
121 static void vm_pageout_init(void);
122 static int vm_pageout_clean(vm_page_t);
123 static void vm_pageout_scan(struct vm_domain *vmd, int pass);
124 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass);
126 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
129 struct proc *pageproc;
131 static struct kproc_desc page_kp = {
136 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
139 SDT_PROVIDER_DEFINE(vm);
140 SDT_PROBE_DEFINE(vm, , , vm__lowmem_cache);
141 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
143 #if !defined(NO_SWAPPING)
144 /* the kernel process "vm_daemon"*/
145 static void vm_daemon(void);
146 static struct proc *vmproc;
148 static struct kproc_desc vm_kp = {
153 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
157 int vm_pages_needed; /* Event on which pageout daemon sleeps */
158 int vm_pageout_deficit; /* Estimated number of pages deficit */
159 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
160 int vm_pageout_wakeup_thresh;
162 #if !defined(NO_SWAPPING)
163 static int vm_pageout_req_swapout; /* XXX */
164 static int vm_daemon_needed;
165 static struct mtx vm_daemon_mtx;
166 /* Allow for use by vm_pageout before vm_daemon is initialized. */
167 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
169 static int vm_max_launder = 32;
170 static int vm_pageout_update_period;
171 static int defer_swap_pageouts;
172 static int disable_swap_pageouts;
173 static int lowmem_period = 10;
174 static time_t lowmem_uptime;
176 #if defined(NO_SWAPPING)
177 static int vm_swap_enabled = 0;
178 static int vm_swap_idle_enabled = 0;
180 static int vm_swap_enabled = 1;
181 static int vm_swap_idle_enabled = 0;
184 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
185 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
186 "free page threshold for waking up the pageout daemon");
188 SYSCTL_INT(_vm, OID_AUTO, max_launder,
189 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
191 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
192 CTLFLAG_RW, &vm_pageout_update_period, 0,
193 "Maximum active LRU update period");
195 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
196 "Low memory callback period");
198 #if defined(NO_SWAPPING)
199 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
200 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
201 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
202 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
204 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
205 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
206 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
207 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
210 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
211 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
213 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
214 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
216 static int pageout_lock_miss;
217 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
218 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
220 #define VM_PAGEOUT_PAGE_COUNT 16
221 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
223 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
224 SYSCTL_INT(_vm, OID_AUTO, max_wired,
225 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
227 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
228 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t,
230 #if !defined(NO_SWAPPING)
231 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
232 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
233 static void vm_req_vmdaemon(int req);
235 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
238 * Initialize a dummy page for marking the caller's place in the specified
239 * paging queue. In principle, this function only needs to set the flag
240 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count
241 * to one as safety precautions.
244 vm_pageout_init_marker(vm_page_t marker, u_short queue)
247 bzero(marker, sizeof(*marker));
248 marker->flags = PG_MARKER;
249 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
250 marker->queue = queue;
251 marker->hold_count = 1;
255 * vm_pageout_fallback_object_lock:
257 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
258 * known to have failed and page queue must be either PQ_ACTIVE or
259 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
260 * while locking the vm object. Use marker page to detect page queue
261 * changes and maintain notion of next page on page queue. Return
262 * TRUE if no changes were detected, FALSE otherwise. vm object is
265 * This function depends on both the lock portion of struct vm_object
266 * and normal struct vm_page being type stable.
269 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
271 struct vm_page marker;
272 struct vm_pagequeue *pq;
278 vm_pageout_init_marker(&marker, queue);
279 pq = vm_page_pagequeue(m);
282 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
283 vm_pagequeue_unlock(pq);
285 VM_OBJECT_WLOCK(object);
287 vm_pagequeue_lock(pq);
290 * The page's object might have changed, and/or the page might
291 * have moved from its original position in the queue. If the
292 * page's object has changed, then the caller should abandon
293 * processing the page because the wrong object lock was
294 * acquired. Use the marker's plinks.q, not the page's, to
295 * determine if the page has been moved. The state of the
296 * page's plinks.q can be indeterminate; whereas, the marker's
297 * plinks.q must be valid.
299 *next = TAILQ_NEXT(&marker, plinks.q);
300 unchanged = m->object == object &&
301 m == TAILQ_PREV(&marker, pglist, plinks.q);
302 KASSERT(!unchanged || m->queue == queue,
303 ("page %p queue %d %d", m, queue, m->queue));
304 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
309 * Lock the page while holding the page queue lock. Use marker page
310 * to detect page queue changes and maintain notion of next page on
311 * page queue. Return TRUE if no changes were detected, FALSE
312 * otherwise. The page is locked on return. The page queue lock might
313 * be dropped and reacquired.
315 * This function depends on normal struct vm_page being type stable.
318 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
320 struct vm_page marker;
321 struct vm_pagequeue *pq;
325 vm_page_lock_assert(m, MA_NOTOWNED);
326 if (vm_page_trylock(m))
330 vm_pageout_init_marker(&marker, queue);
331 pq = vm_page_pagequeue(m);
333 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
334 vm_pagequeue_unlock(pq);
336 vm_pagequeue_lock(pq);
338 /* Page queue might have changed. */
339 *next = TAILQ_NEXT(&marker, plinks.q);
340 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
341 KASSERT(!unchanged || m->queue == queue,
342 ("page %p queue %d %d", m, queue, m->queue));
343 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
350 * Clean the page and remove it from the laundry.
352 * We set the busy bit to cause potential page faults on this page to
353 * block. Note the careful timing, however, the busy bit isn't set till
354 * late and we cannot do anything that will mess with the page.
357 vm_pageout_clean(vm_page_t m)
360 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
362 int ib, is, page_base;
363 vm_pindex_t pindex = m->pindex;
365 vm_page_lock_assert(m, MA_OWNED);
367 VM_OBJECT_ASSERT_WLOCKED(object);
370 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
371 * with the new swapper, but we could have serious problems paging
372 * out other object types if there is insufficient memory.
374 * Unfortunately, checking free memory here is far too late, so the
375 * check has been moved up a procedural level.
379 * Can't clean the page if it's busy or held.
381 vm_page_assert_unbusied(m);
382 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
385 mc[vm_pageout_page_count] = pb = ps = m;
387 page_base = vm_pageout_page_count;
392 * Scan object for clusterable pages.
394 * We can cluster ONLY if: ->> the page is NOT
395 * clean, wired, busy, held, or mapped into a
396 * buffer, and one of the following:
397 * 1) The page is inactive, or a seldom used
400 * 2) we force the issue.
402 * During heavy mmap/modification loads the pageout
403 * daemon can really fragment the underlying file
404 * due to flushing pages out of order and not trying
405 * align the clusters (which leave sporatic out-of-order
406 * holes). To solve this problem we do the reverse scan
407 * first and attempt to align our cluster, then do a
408 * forward scan if room remains.
411 while (ib && pageout_count < vm_pageout_page_count) {
419 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
423 vm_page_test_dirty(p);
429 if (p->queue != PQ_INACTIVE ||
430 p->hold_count != 0) { /* may be undergoing I/O */
436 mc[--page_base] = pb = p;
440 * alignment boundry, stop here and switch directions. Do
443 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
447 while (pageout_count < vm_pageout_page_count &&
448 pindex + is < object->size) {
451 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
453 vm_page_test_dirty(p);
457 if (p->queue != PQ_INACTIVE ||
458 p->hold_count != 0) { /* may be undergoing I/O */
463 mc[page_base + pageout_count] = ps = p;
469 * If we exhausted our forward scan, continue with the reverse scan
470 * when possible, even past a page boundry. This catches boundry
473 if (ib && pageout_count < vm_pageout_page_count)
477 * we allow reads during pageouts...
479 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
484 * vm_pageout_flush() - launder the given pages
486 * The given pages are laundered. Note that we setup for the start of
487 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
488 * reference count all in here rather then in the parent. If we want
489 * the parent to do more sophisticated things we may have to change
492 * Returned runlen is the count of pages between mreq and first
493 * page after mreq with status VM_PAGER_AGAIN.
494 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
495 * for any page in runlen set.
498 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
501 vm_object_t object = mc[0]->object;
502 int pageout_status[count];
506 VM_OBJECT_ASSERT_WLOCKED(object);
509 * Initiate I/O. Bump the vm_page_t->busy counter and
510 * mark the pages read-only.
512 * We do not have to fixup the clean/dirty bits here... we can
513 * allow the pager to do it after the I/O completes.
515 * NOTE! mc[i]->dirty may be partial or fragmented due to an
516 * edge case with file fragments.
518 for (i = 0; i < count; i++) {
519 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
520 ("vm_pageout_flush: partially invalid page %p index %d/%d",
522 vm_page_sbusy(mc[i]);
523 pmap_remove_write(mc[i]);
525 vm_object_pip_add(object, count);
527 vm_pager_put_pages(object, mc, count, flags, pageout_status);
529 runlen = count - mreq;
532 for (i = 0; i < count; i++) {
533 vm_page_t mt = mc[i];
535 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
536 !pmap_page_is_write_mapped(mt),
537 ("vm_pageout_flush: page %p is not write protected", mt));
538 switch (pageout_status[i]) {
545 * Page outside of range of object. Right now we
546 * essentially lose the changes by pretending it
554 * If page couldn't be paged out, then reactivate the
555 * page so it doesn't clog the inactive list. (We
556 * will try paging out it again later).
559 vm_page_activate(mt);
561 if (eio != NULL && i >= mreq && i - mreq < runlen)
565 if (i >= mreq && i - mreq < runlen)
571 * If the operation is still going, leave the page busy to
572 * block all other accesses. Also, leave the paging in
573 * progress indicator set so that we don't attempt an object
576 if (pageout_status[i] != VM_PAGER_PEND) {
577 vm_object_pip_wakeup(object);
579 if (vm_page_count_severe()) {
581 vm_page_try_to_cache(mt);
588 return (numpagedout);
592 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low,
599 vm_page_t m, m_tmp, next;
602 vm_pagequeue_lock(pq);
603 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) {
604 if ((m->flags & PG_MARKER) != 0)
606 pa = VM_PAGE_TO_PHYS(m);
607 if (pa < low || pa + PAGE_SIZE > high)
609 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
614 if ((!VM_OBJECT_TRYWLOCK(object) &&
615 (!vm_pageout_fallback_object_lock(m, &next) ||
616 m->hold_count != 0)) || vm_page_busied(m)) {
618 VM_OBJECT_WUNLOCK(object);
621 vm_page_test_dirty(m);
622 if (m->dirty == 0 && object->ref_count != 0)
626 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) {
627 VM_OBJECT_WUNLOCK(object);
630 if (object->type == OBJT_VNODE) {
631 vm_pagequeue_unlock(pq);
633 vm_object_reference_locked(object);
634 VM_OBJECT_WUNLOCK(object);
635 (void)vn_start_write(vp, &mp, V_WAIT);
636 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
637 LK_SHARED : LK_EXCLUSIVE;
638 vn_lock(vp, lockmode | LK_RETRY);
639 VM_OBJECT_WLOCK(object);
640 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
641 VM_OBJECT_WUNLOCK(object);
643 vm_object_deallocate(object);
644 vn_finished_write(mp);
646 } else if (object->type == OBJT_SWAP ||
647 object->type == OBJT_DEFAULT) {
648 vm_pagequeue_unlock(pq);
650 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC,
652 VM_OBJECT_WUNLOCK(object);
657 * Dequeue here to prevent lock recursion in
660 vm_page_dequeue_locked(m);
664 VM_OBJECT_WUNLOCK(object);
666 vm_pagequeue_unlock(pq);
671 * Increase the number of cached pages. The specified value, "tries",
672 * determines which categories of pages are cached:
674 * 0: All clean, inactive pages within the specified physical address range
675 * are cached. Will not sleep.
676 * 1: The vm_lowmem handlers are called. All inactive pages within
677 * the specified physical address range are cached. May sleep.
678 * 2: The vm_lowmem handlers are called. All inactive and active pages
679 * within the specified physical address range are cached. May sleep.
682 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
684 int actl, actmax, inactl, inactmax, dom, initial_dom;
685 static int start_dom = 0;
689 * Decrease registered cache sizes. The vm_lowmem handlers
690 * may acquire locks and/or sleep, so they can only be invoked
691 * when "tries" is greater than zero.
693 SDT_PROBE0(vm, , , vm__lowmem_cache);
694 EVENTHANDLER_INVOKE(vm_lowmem, 0);
697 * We do this explicitly after the caches have been drained
704 * Make the next scan start on the next domain.
706 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains;
709 inactmax = cnt.v_inactive_count;
711 actmax = tries < 2 ? 0 : cnt.v_active_count;
715 * Scan domains in round-robin order, first inactive queues,
716 * then active. Since domain usually owns large physically
717 * contiguous chunk of memory, it makes sense to completely
718 * exhaust one domain before switching to next, while growing
719 * the pool of contiguous physical pages.
721 * Do not even start launder a domain which cannot contain
722 * the specified address range, as indicated by segments
723 * constituting the domain.
726 if (inactl < inactmax) {
727 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
729 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE],
734 if (++dom == vm_ndomains)
736 if (dom != initial_dom)
740 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
742 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE],
747 if (++dom == vm_ndomains)
749 if (dom != initial_dom)
754 #if !defined(NO_SWAPPING)
756 * vm_pageout_object_deactivate_pages
758 * Deactivate enough pages to satisfy the inactive target
761 * The object and map must be locked.
764 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
767 vm_object_t backing_object, object;
769 int act_delta, remove_mode;
771 VM_OBJECT_ASSERT_LOCKED(first_object);
772 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
774 for (object = first_object;; object = backing_object) {
775 if (pmap_resident_count(pmap) <= desired)
777 VM_OBJECT_ASSERT_LOCKED(object);
778 if ((object->flags & OBJ_UNMANAGED) != 0 ||
779 object->paging_in_progress != 0)
783 if (object->shadow_count > 1)
786 * Scan the object's entire memory queue.
788 TAILQ_FOREACH(p, &object->memq, listq) {
789 if (pmap_resident_count(pmap) <= desired)
791 if (vm_page_busied(p))
793 PCPU_INC(cnt.v_pdpages);
795 if (p->wire_count != 0 || p->hold_count != 0 ||
796 !pmap_page_exists_quick(pmap, p)) {
800 act_delta = pmap_ts_referenced(p);
801 if ((p->aflags & PGA_REFERENCED) != 0) {
804 vm_page_aflag_clear(p, PGA_REFERENCED);
806 if (p->queue != PQ_ACTIVE && act_delta != 0) {
808 p->act_count += act_delta;
809 } else if (p->queue == PQ_ACTIVE) {
810 if (act_delta == 0) {
811 p->act_count -= min(p->act_count,
813 if (!remove_mode && p->act_count == 0) {
815 vm_page_deactivate(p);
820 if (p->act_count < ACT_MAX -
822 p->act_count += ACT_ADVANCE;
825 } else if (p->queue == PQ_INACTIVE)
829 if ((backing_object = object->backing_object) == NULL)
831 VM_OBJECT_RLOCK(backing_object);
832 if (object != first_object)
833 VM_OBJECT_RUNLOCK(object);
836 if (object != first_object)
837 VM_OBJECT_RUNLOCK(object);
841 * deactivate some number of pages in a map, try to do it fairly, but
842 * that is really hard to do.
845 vm_pageout_map_deactivate_pages(map, desired)
850 vm_object_t obj, bigobj;
853 if (!vm_map_trylock(map))
860 * first, search out the biggest object, and try to free pages from
863 tmpe = map->header.next;
864 while (tmpe != &map->header) {
865 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
866 obj = tmpe->object.vm_object;
867 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
868 if (obj->shadow_count <= 1 &&
870 bigobj->resident_page_count < obj->resident_page_count)) {
872 VM_OBJECT_RUNLOCK(bigobj);
875 VM_OBJECT_RUNLOCK(obj);
878 if (tmpe->wired_count > 0)
879 nothingwired = FALSE;
883 if (bigobj != NULL) {
884 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
885 VM_OBJECT_RUNLOCK(bigobj);
888 * Next, hunt around for other pages to deactivate. We actually
889 * do this search sort of wrong -- .text first is not the best idea.
891 tmpe = map->header.next;
892 while (tmpe != &map->header) {
893 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
895 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
896 obj = tmpe->object.vm_object;
898 VM_OBJECT_RLOCK(obj);
899 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
900 VM_OBJECT_RUNLOCK(obj);
908 * Remove all non-wired, managed mappings if a process is swapped out.
909 * This will free page table pages.
912 pmap_remove_pages(map->pmap);
915 * Remove all mappings if a process is swapped out, this will free page
918 if (desired == 0 && nothingwired) {
919 pmap_remove(vm_map_pmap(map), vm_map_min(map),
926 #endif /* !defined(NO_SWAPPING) */
929 * vm_pageout_scan does the dirty work for the pageout daemon.
931 * pass 0 - Update active LRU/deactivate pages
932 * pass 1 - Move inactive to cache or free
933 * pass 2 - Launder dirty pages
936 vm_pageout_scan(struct vm_domain *vmd, int pass)
939 struct vm_pagequeue *pq;
942 int act_delta, addl_page_shortage, deficit, maxscan, page_shortage;
943 int vnodes_skipped = 0;
944 int maxlaunder, scan_tick, scanned;
946 boolean_t queues_locked;
949 * If we need to reclaim memory ask kernel caches to return
950 * some. We rate limit to avoid thrashing.
952 if (vmd == &vm_dom[0] && pass > 0 &&
953 (time_uptime - lowmem_uptime) >= lowmem_period) {
955 * Decrease registered cache sizes.
957 SDT_PROBE0(vm, , , vm__lowmem_scan);
958 EVENTHANDLER_INVOKE(vm_lowmem, 0);
960 * We do this explicitly after the caches have been
964 lowmem_uptime = time_uptime;
968 * The addl_page_shortage is the number of temporarily
969 * stuck pages in the inactive queue. In other words, the
970 * number of pages from the inactive count that should be
971 * discounted in setting the target for the active queue scan.
973 addl_page_shortage = 0;
976 * Calculate the number of pages we want to either free or move
980 deficit = atomic_readandclear_int(&vm_pageout_deficit);
981 page_shortage = vm_paging_target() + deficit;
983 page_shortage = deficit = 0;
986 * maxlaunder limits the number of dirty pages we flush per scan.
987 * For most systems a smaller value (16 or 32) is more robust under
988 * extreme memory and disk pressure because any unnecessary writes
989 * to disk can result in extreme performance degredation. However,
990 * systems with excessive dirty pages (especially when MAP_NOSYNC is
991 * used) will die horribly with limited laundering. If the pageout
992 * daemon cannot clean enough pages in the first pass, we let it go
993 * all out in succeeding passes.
995 if ((maxlaunder = vm_max_launder) <= 1)
1001 * Start scanning the inactive queue for pages we can move to the
1002 * cache or free. The scan will stop when the target is reached or
1003 * we have scanned the entire inactive queue. Note that m->act_count
1004 * is not used to form decisions for the inactive queue, only for the
1007 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1008 maxscan = pq->pq_cnt;
1009 vm_pagequeue_lock(pq);
1010 queues_locked = TRUE;
1011 for (m = TAILQ_FIRST(&pq->pq_pl);
1012 m != NULL && maxscan-- > 0 && page_shortage > 0;
1014 vm_pagequeue_assert_locked(pq);
1015 KASSERT(queues_locked, ("unlocked queues"));
1016 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
1018 PCPU_INC(cnt.v_pdpages);
1019 next = TAILQ_NEXT(m, plinks.q);
1024 if (m->flags & PG_MARKER)
1027 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1028 ("Fictitious page %p cannot be in inactive queue", m));
1029 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1030 ("Unmanaged page %p cannot be in inactive queue", m));
1033 * The page or object lock acquisitions fail if the
1034 * page was removed from the queue or moved to a
1035 * different position within the queue. In either
1036 * case, addl_page_shortage should not be incremented.
1038 if (!vm_pageout_page_lock(m, &next)) {
1043 if (!VM_OBJECT_TRYWLOCK(object) &&
1044 !vm_pageout_fallback_object_lock(m, &next)) {
1046 VM_OBJECT_WUNLOCK(object);
1051 * Don't mess with busy pages, keep them at at the
1052 * front of the queue, most likely they are being
1053 * paged out. Increment addl_page_shortage for busy
1054 * pages, because they may leave the inactive queue
1055 * shortly after page scan is finished.
1057 if (vm_page_busied(m)) {
1059 VM_OBJECT_WUNLOCK(object);
1060 addl_page_shortage++;
1065 * We unlock the inactive page queue, invalidating the
1066 * 'next' pointer. Use our marker to remember our
1069 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1070 vm_pagequeue_unlock(pq);
1071 queues_locked = FALSE;
1074 * We bump the activation count if the page has been
1075 * referenced while in the inactive queue. This makes
1076 * it less likely that the page will be added back to the
1077 * inactive queue prematurely again. Here we check the
1078 * page tables (or emulated bits, if any), given the upper
1079 * level VM system not knowing anything about existing
1083 if ((m->aflags & PGA_REFERENCED) != 0) {
1084 vm_page_aflag_clear(m, PGA_REFERENCED);
1087 if (object->ref_count != 0) {
1088 act_delta += pmap_ts_referenced(m);
1090 KASSERT(!pmap_page_is_mapped(m),
1091 ("vm_pageout_scan: page %p is mapped", m));
1095 * If the upper level VM system knows about any page
1096 * references, we reactivate the page or requeue it.
1098 if (act_delta != 0) {
1099 if (object->ref_count) {
1100 vm_page_activate(m);
1101 m->act_count += act_delta + ACT_ADVANCE;
1103 vm_pagequeue_lock(pq);
1104 queues_locked = TRUE;
1105 vm_page_requeue_locked(m);
1107 VM_OBJECT_WUNLOCK(object);
1112 if (m->hold_count != 0) {
1114 VM_OBJECT_WUNLOCK(object);
1117 * Held pages are essentially stuck in the
1118 * queue. So, they ought to be discounted
1119 * from the inactive count. See the
1120 * calculation of the page_shortage for the
1121 * loop over the active queue below.
1123 addl_page_shortage++;
1128 * If the page appears to be clean at the machine-independent
1129 * layer, then remove all of its mappings from the pmap in
1130 * anticipation of placing it onto the cache queue. If,
1131 * however, any of the page's mappings allow write access,
1132 * then the page may still be modified until the last of those
1133 * mappings are removed.
1135 if (object->ref_count != 0) {
1136 vm_page_test_dirty(m);
1141 if (m->valid == 0) {
1143 * Invalid pages can be easily freed
1146 PCPU_INC(cnt.v_dfree);
1148 } else if (m->dirty == 0) {
1150 * Clean pages can be placed onto the cache queue.
1151 * This effectively frees them.
1155 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1157 * Dirty pages need to be paged out, but flushing
1158 * a page is extremely expensive verses freeing
1159 * a clean page. Rather then artificially limiting
1160 * the number of pages we can flush, we instead give
1161 * dirty pages extra priority on the inactive queue
1162 * by forcing them to be cycled through the queue
1163 * twice before being flushed, after which the
1164 * (now clean) page will cycle through once more
1165 * before being freed. This significantly extends
1166 * the thrash point for a heavily loaded machine.
1168 m->flags |= PG_WINATCFLS;
1169 vm_pagequeue_lock(pq);
1170 queues_locked = TRUE;
1171 vm_page_requeue_locked(m);
1172 } else if (maxlaunder > 0) {
1174 * We always want to try to flush some dirty pages if
1175 * we encounter them, to keep the system stable.
1176 * Normally this number is small, but under extreme
1177 * pressure where there are insufficient clean pages
1178 * on the inactive queue, we may have to go all out.
1180 int swap_pageouts_ok;
1181 struct vnode *vp = NULL;
1182 struct mount *mp = NULL;
1184 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
1185 swap_pageouts_ok = 1;
1187 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
1188 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
1189 vm_page_count_min());
1194 * We don't bother paging objects that are "dead".
1195 * Those objects are in a "rundown" state.
1197 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
1198 vm_pagequeue_lock(pq);
1200 VM_OBJECT_WUNLOCK(object);
1201 queues_locked = TRUE;
1202 vm_page_requeue_locked(m);
1207 * The object is already known NOT to be dead. It
1208 * is possible for the vget() to block the whole
1209 * pageout daemon, but the new low-memory handling
1210 * code should prevent it.
1212 * The previous code skipped locked vnodes and, worse,
1213 * reordered pages in the queue. This results in
1214 * completely non-deterministic operation and, on a
1215 * busy system, can lead to extremely non-optimal
1216 * pageouts. For example, it can cause clean pages
1217 * to be freed and dirty pages to be moved to the end
1218 * of the queue. Since dirty pages are also moved to
1219 * the end of the queue once-cleaned, this gives
1220 * way too large a weighting to defering the freeing
1223 * We can't wait forever for the vnode lock, we might
1224 * deadlock due to a vn_read() getting stuck in
1225 * vm_wait while holding this vnode. We skip the
1226 * vnode if we can't get it in a reasonable amount
1229 if (object->type == OBJT_VNODE) {
1231 vp = object->handle;
1232 if (vp->v_type == VREG &&
1233 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1235 ++pageout_lock_miss;
1236 if (object->flags & OBJ_MIGHTBEDIRTY)
1238 goto unlock_and_continue;
1241 ("vp %p with NULL v_mount", vp));
1242 vm_object_reference_locked(object);
1243 VM_OBJECT_WUNLOCK(object);
1244 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
1245 LK_SHARED : LK_EXCLUSIVE;
1246 if (vget(vp, lockmode | LK_TIMELOCK,
1248 VM_OBJECT_WLOCK(object);
1249 ++pageout_lock_miss;
1250 if (object->flags & OBJ_MIGHTBEDIRTY)
1253 goto unlock_and_continue;
1255 VM_OBJECT_WLOCK(object);
1257 vm_pagequeue_lock(pq);
1258 queues_locked = TRUE;
1260 * The page might have been moved to another
1261 * queue during potential blocking in vget()
1262 * above. The page might have been freed and
1263 * reused for another vnode.
1265 if (m->queue != PQ_INACTIVE ||
1266 m->object != object ||
1267 TAILQ_NEXT(m, plinks.q) != &vmd->vmd_marker) {
1269 if (object->flags & OBJ_MIGHTBEDIRTY)
1271 goto unlock_and_continue;
1275 * The page may have been busied during the
1276 * blocking in vget(). We don't move the
1277 * page back onto the end of the queue so that
1278 * statistics are more correct if we don't.
1280 if (vm_page_busied(m)) {
1282 addl_page_shortage++;
1283 goto unlock_and_continue;
1287 * If the page has become held it might
1288 * be undergoing I/O, so skip it
1290 if (m->hold_count != 0) {
1292 addl_page_shortage++;
1293 if (object->flags & OBJ_MIGHTBEDIRTY)
1295 goto unlock_and_continue;
1297 vm_pagequeue_unlock(pq);
1298 queues_locked = FALSE;
1302 * If a page is dirty, then it is either being washed
1303 * (but not yet cleaned) or it is still in the
1304 * laundry. If it is still in the laundry, then we
1305 * start the cleaning operation.
1307 * decrement page_shortage on success to account for
1308 * the (future) cleaned page. Otherwise we could wind
1309 * up laundering or cleaning too many pages.
1311 if (vm_pageout_clean(m) != 0) {
1315 unlock_and_continue:
1316 vm_page_lock_assert(m, MA_NOTOWNED);
1317 VM_OBJECT_WUNLOCK(object);
1319 if (queues_locked) {
1320 vm_pagequeue_unlock(pq);
1321 queues_locked = FALSE;
1325 vm_object_deallocate(object);
1326 vn_finished_write(mp);
1328 vm_page_lock_assert(m, MA_NOTOWNED);
1332 VM_OBJECT_WUNLOCK(object);
1334 if (!queues_locked) {
1335 vm_pagequeue_lock(pq);
1336 queues_locked = TRUE;
1338 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1339 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1341 vm_pagequeue_unlock(pq);
1343 #if !defined(NO_SWAPPING)
1345 * Wakeup the swapout daemon if we didn't cache or free the targeted
1348 if (vm_swap_enabled && page_shortage > 0)
1349 vm_req_vmdaemon(VM_SWAP_NORMAL);
1353 * Wakeup the sync daemon if we skipped a vnode in a writeable object
1354 * and we didn't cache or free enough pages.
1356 if (vnodes_skipped > 0 && page_shortage > cnt.v_free_target -
1358 (void)speedup_syncer();
1361 * Compute the number of pages we want to try to move from the
1362 * active queue to the inactive queue.
1364 page_shortage = cnt.v_inactive_target - cnt.v_inactive_count +
1365 vm_paging_target() + deficit + addl_page_shortage;
1367 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1368 vm_pagequeue_lock(pq);
1369 maxscan = pq->pq_cnt;
1372 * If we're just idle polling attempt to visit every
1373 * active page within 'update_period' seconds.
1376 if (vm_pageout_update_period != 0) {
1377 min_scan = pq->pq_cnt;
1378 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1379 min_scan /= hz * vm_pageout_update_period;
1382 if (min_scan > 0 || (page_shortage > 0 && maxscan > 0))
1383 vmd->vmd_last_active_scan = scan_tick;
1386 * Scan the active queue for pages that can be deactivated. Update
1387 * the per-page activity counter and use it to identify deactivation
1390 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1391 min_scan || (page_shortage > 0 && scanned < maxscan)); m = next,
1394 KASSERT(m->queue == PQ_ACTIVE,
1395 ("vm_pageout_scan: page %p isn't active", m));
1397 next = TAILQ_NEXT(m, plinks.q);
1398 if ((m->flags & PG_MARKER) != 0)
1400 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1401 ("Fictitious page %p cannot be in active queue", m));
1402 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1403 ("Unmanaged page %p cannot be in active queue", m));
1404 if (!vm_pageout_page_lock(m, &next)) {
1410 * The count for pagedaemon pages is done after checking the
1411 * page for eligibility...
1413 PCPU_INC(cnt.v_pdpages);
1416 * Check to see "how much" the page has been used.
1419 if (m->aflags & PGA_REFERENCED) {
1420 vm_page_aflag_clear(m, PGA_REFERENCED);
1424 * Unlocked object ref count check. Two races are possible.
1425 * 1) The ref was transitioning to zero and we saw non-zero,
1426 * the pmap bits will be checked unnecessarily.
1427 * 2) The ref was transitioning to one and we saw zero.
1428 * The page lock prevents a new reference to this page so
1429 * we need not check the reference bits.
1431 if (m->object->ref_count != 0)
1432 act_delta += pmap_ts_referenced(m);
1435 * Advance or decay the act_count based on recent usage.
1438 m->act_count += ACT_ADVANCE + act_delta;
1439 if (m->act_count > ACT_MAX)
1440 m->act_count = ACT_MAX;
1442 m->act_count -= min(m->act_count, ACT_DECLINE);
1443 act_delta = m->act_count;
1447 * Move this page to the tail of the active or inactive
1448 * queue depending on usage.
1450 if (act_delta == 0) {
1451 /* Dequeue to avoid later lock recursion. */
1452 vm_page_dequeue_locked(m);
1453 vm_page_deactivate(m);
1456 vm_page_requeue_locked(m);
1459 vm_pagequeue_unlock(pq);
1460 #if !defined(NO_SWAPPING)
1462 * Idle process swapout -- run once per second.
1464 if (vm_swap_idle_enabled) {
1466 if (time_second != lsec) {
1467 vm_req_vmdaemon(VM_SWAP_IDLE);
1474 * If we are critically low on one of RAM or swap and low on
1475 * the other, kill the largest process. However, we avoid
1476 * doing this on the first pass in order to give ourselves a
1477 * chance to flush out dirty vnode-backed pages and to allow
1478 * active pages to be moved to the inactive queue and reclaimed.
1480 vm_pageout_mightbe_oom(vmd, pass);
1483 static int vm_pageout_oom_vote;
1486 * The pagedaemon threads randlomly select one to perform the
1487 * OOM. Trying to kill processes before all pagedaemons
1488 * failed to reach free target is premature.
1491 vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass)
1495 if (pass <= 1 || !((swap_pager_avail < 64 && vm_page_count_min()) ||
1496 (swap_pager_full && vm_paging_target() > 0))) {
1498 vmd->vmd_oom = FALSE;
1499 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1507 vmd->vmd_oom = TRUE;
1508 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1509 if (old_vote != vm_ndomains - 1)
1513 * The current pagedaemon thread is the last in the quorum to
1514 * start OOM. Initiate the selection and signaling of the
1517 vm_pageout_oom(VM_OOM_MEM);
1520 * After one round of OOM terror, recall our vote. On the
1521 * next pass, current pagedaemon would vote again if the low
1522 * memory condition is still there, due to vmd_oom being
1525 vmd->vmd_oom = FALSE;
1526 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1530 vm_pageout_oom(int shortage)
1532 struct proc *p, *bigproc;
1533 vm_offset_t size, bigsize;
1538 * We keep the process bigproc locked once we find it to keep anyone
1539 * from messing with it; however, there is a possibility of
1540 * deadlock if process B is bigproc and one of it's child processes
1541 * attempts to propagate a signal to B while we are waiting for A's
1542 * lock while walking this list. To avoid this, we don't block on
1543 * the process lock but just skip a process if it is already locked.
1547 sx_slock(&allproc_lock);
1548 FOREACH_PROC_IN_SYSTEM(p) {
1554 * If this is a system, protected or killed process, skip it.
1556 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1557 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1558 p->p_pid == 1 || P_KILLED(p) ||
1559 (p->p_pid < 48 && swap_pager_avail != 0)) {
1564 * If the process is in a non-running type state,
1565 * don't touch it. Check all the threads individually.
1568 FOREACH_THREAD_IN_PROC(p, td) {
1570 if (!TD_ON_RUNQ(td) &&
1571 !TD_IS_RUNNING(td) &&
1572 !TD_IS_SLEEPING(td) &&
1573 !TD_IS_SUSPENDED(td) &&
1574 !TD_IS_SWAPPED(td)) {
1586 * get the process size
1588 vm = vmspace_acquire_ref(p);
1594 if (!vm_map_trylock_read(&vm->vm_map)) {
1601 size = vmspace_swap_count(vm);
1602 vm_map_unlock_read(&vm->vm_map);
1603 if (shortage == VM_OOM_MEM)
1604 size += vmspace_resident_count(vm);
1607 * if the this process is bigger than the biggest one
1610 if (size > bigsize) {
1611 if (bigproc != NULL)
1619 sx_sunlock(&allproc_lock);
1620 if (bigproc != NULL) {
1622 killproc(bigproc, "out of swap space");
1623 sched_nice(bigproc, PRIO_MIN);
1625 PROC_UNLOCK(bigproc);
1626 wakeup(&cnt.v_free_count);
1631 vm_pageout_worker(void *arg)
1633 struct vm_domain *domain;
1636 domidx = (uintptr_t)arg;
1637 domain = &vm_dom[domidx];
1640 * XXXKIB It could be useful to bind pageout daemon threads to
1641 * the cores belonging to the domain, from which vm_page_array
1645 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1646 domain->vmd_last_active_scan = ticks;
1647 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1650 * The pageout daemon worker is never done, so loop forever.
1654 * If we have enough free memory, wakeup waiters. Do
1655 * not clear vm_pages_needed until we reach our target,
1656 * otherwise we may be woken up over and over again and
1657 * waste a lot of cpu.
1659 mtx_lock(&vm_page_queue_free_mtx);
1660 if (vm_pages_needed && !vm_page_count_min()) {
1661 if (!vm_paging_needed())
1662 vm_pages_needed = 0;
1663 wakeup(&cnt.v_free_count);
1665 if (vm_pages_needed) {
1667 * We're still not done. Either vm_pages_needed was
1668 * set by another thread during the previous scan
1669 * (typically, this happens during a level 0 scan) or
1670 * vm_pages_needed was already set and the scan failed
1671 * to free enough pages. If we haven't yet performed
1672 * a level >= 2 scan (unlimited dirty cleaning), then
1673 * upgrade the level and scan again now. Otherwise,
1674 * sleep a bit and try again later. While sleeping,
1675 * vm_pages_needed can be cleared.
1677 if (domain->vmd_pass > 1)
1678 msleep(&vm_pages_needed,
1679 &vm_page_queue_free_mtx, PVM, "psleep",
1683 * Good enough, sleep until required to refresh
1686 msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1689 if (vm_pages_needed) {
1693 domain->vmd_pass = 0;
1694 mtx_unlock(&vm_page_queue_free_mtx);
1695 vm_pageout_scan(domain, domain->vmd_pass);
1700 * vm_pageout_init initialises basic pageout daemon settings.
1703 vm_pageout_init(void)
1706 * Initialize some paging parameters.
1708 cnt.v_interrupt_free_min = 2;
1709 if (cnt.v_page_count < 2000)
1710 vm_pageout_page_count = 8;
1713 * v_free_reserved needs to include enough for the largest
1714 * swap pager structures plus enough for any pv_entry structs
1717 if (cnt.v_page_count > 1024)
1718 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1721 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1722 cnt.v_interrupt_free_min;
1723 cnt.v_free_reserved = vm_pageout_page_count +
1724 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1725 cnt.v_free_severe = cnt.v_free_min / 2;
1726 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1727 cnt.v_free_min += cnt.v_free_reserved;
1728 cnt.v_free_severe += cnt.v_free_reserved;
1729 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1730 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1731 cnt.v_inactive_target = cnt.v_free_count / 3;
1734 * Set the default wakeup threshold to be 10% above the minimum
1735 * page limit. This keeps the steady state out of shortfall.
1737 vm_pageout_wakeup_thresh = (cnt.v_free_min / 10) * 11;
1740 * Set interval in seconds for active scan. We want to visit each
1741 * page at least once every ten minutes. This is to prevent worst
1742 * case paging behaviors with stale active LRU.
1744 if (vm_pageout_update_period == 0)
1745 vm_pageout_update_period = 600;
1747 /* XXX does not really belong here */
1748 if (vm_page_max_wired == 0)
1749 vm_page_max_wired = cnt.v_free_count / 3;
1753 * vm_pageout is the high level pageout daemon.
1763 swap_pager_swap_init();
1765 for (i = 1; i < vm_ndomains; i++) {
1766 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1767 curproc, NULL, 0, 0, "dom%d", i);
1769 panic("starting pageout for domain %d, error %d\n",
1774 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1777 panic("starting uma_reclaim helper, error %d\n", error);
1778 vm_pageout_worker((void *)(uintptr_t)0);
1782 * Unless the free page queue lock is held by the caller, this function
1783 * should be regarded as advisory. Specifically, the caller should
1784 * not msleep() on &cnt.v_free_count following this function unless
1785 * the free page queue lock is held until the msleep() is performed.
1788 pagedaemon_wakeup(void)
1791 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1792 vm_pages_needed = 1;
1793 wakeup(&vm_pages_needed);
1797 #if !defined(NO_SWAPPING)
1799 vm_req_vmdaemon(int req)
1801 static int lastrun = 0;
1803 mtx_lock(&vm_daemon_mtx);
1804 vm_pageout_req_swapout |= req;
1805 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1806 wakeup(&vm_daemon_needed);
1809 mtx_unlock(&vm_daemon_mtx);
1815 struct rlimit rsslim;
1819 int breakout, swapout_flags, tryagain, attempts;
1821 uint64_t rsize, ravailable;
1825 mtx_lock(&vm_daemon_mtx);
1826 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
1828 racct_enable ? hz : 0
1833 swapout_flags = vm_pageout_req_swapout;
1834 vm_pageout_req_swapout = 0;
1835 mtx_unlock(&vm_daemon_mtx);
1837 swapout_procs(swapout_flags);
1840 * scan the processes for exceeding their rlimits or if
1841 * process is swapped out -- deactivate pages
1847 sx_slock(&allproc_lock);
1848 FOREACH_PROC_IN_SYSTEM(p) {
1849 vm_pindex_t limit, size;
1852 * if this is a system process or if we have already
1853 * looked at this process, skip it.
1856 if (p->p_state != PRS_NORMAL ||
1857 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1862 * if the process is in a non-running type state,
1866 FOREACH_THREAD_IN_PROC(p, td) {
1868 if (!TD_ON_RUNQ(td) &&
1869 !TD_IS_RUNNING(td) &&
1870 !TD_IS_SLEEPING(td) &&
1871 !TD_IS_SUSPENDED(td)) {
1885 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1887 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1890 * let processes that are swapped out really be
1891 * swapped out set the limit to nothing (will force a
1894 if ((p->p_flag & P_INMEM) == 0)
1895 limit = 0; /* XXX */
1896 vm = vmspace_acquire_ref(p);
1901 size = vmspace_resident_count(vm);
1902 if (size >= limit) {
1903 vm_pageout_map_deactivate_pages(
1904 &vm->vm_map, limit);
1908 rsize = IDX_TO_OFF(size);
1910 racct_set(p, RACCT_RSS, rsize);
1911 ravailable = racct_get_available(p, RACCT_RSS);
1913 if (rsize > ravailable) {
1915 * Don't be overly aggressive; this
1916 * might be an innocent process,
1917 * and the limit could've been exceeded
1918 * by some memory hog. Don't try
1919 * to deactivate more than 1/4th
1920 * of process' resident set size.
1922 if (attempts <= 8) {
1923 if (ravailable < rsize -
1925 ravailable = rsize -
1929 vm_pageout_map_deactivate_pages(
1931 OFF_TO_IDX(ravailable));
1932 /* Update RSS usage after paging out. */
1933 size = vmspace_resident_count(vm);
1934 rsize = IDX_TO_OFF(size);
1936 racct_set(p, RACCT_RSS, rsize);
1938 if (rsize > ravailable)
1945 sx_sunlock(&allproc_lock);
1946 if (tryagain != 0 && attempts <= 10)
1950 #endif /* !defined(NO_SWAPPING) */