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 m);
123 static int vm_pageout_cluster(vm_page_t m);
124 static void vm_pageout_scan(struct vm_domain *vmd, int pass);
125 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
126 int starting_page_shortage);
128 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
131 struct proc *pageproc;
133 static struct kproc_desc page_kp = {
138 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
141 SDT_PROVIDER_DEFINE(vm);
142 SDT_PROBE_DEFINE(vm, , , vm__lowmem_cache);
143 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
145 #if !defined(NO_SWAPPING)
146 /* the kernel process "vm_daemon"*/
147 static void vm_daemon(void);
148 static struct proc *vmproc;
150 static struct kproc_desc vm_kp = {
155 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
159 int vm_pages_needed; /* Event on which pageout daemon sleeps */
160 int vm_pageout_deficit; /* Estimated number of pages deficit */
161 int vm_pageout_wakeup_thresh;
162 static int vm_pageout_oom_seq = 12;
164 #if !defined(NO_SWAPPING)
165 static int vm_pageout_req_swapout; /* XXX */
166 static int vm_daemon_needed;
167 static struct mtx vm_daemon_mtx;
168 /* Allow for use by vm_pageout before vm_daemon is initialized. */
169 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
171 static int vm_max_launder = 32;
172 static int vm_pageout_update_period;
173 static int defer_swap_pageouts;
174 static int disable_swap_pageouts;
175 static int lowmem_period = 10;
176 static time_t lowmem_uptime;
178 #if defined(NO_SWAPPING)
179 static int vm_swap_enabled = 0;
180 static int vm_swap_idle_enabled = 0;
182 static int vm_swap_enabled = 1;
183 static int vm_swap_idle_enabled = 0;
186 static int vm_panic_on_oom = 0;
188 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
189 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
190 "panic on out of memory instead of killing the largest process");
192 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
193 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
194 "free page threshold for waking up the pageout daemon");
196 SYSCTL_INT(_vm, OID_AUTO, max_launder,
197 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
199 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
200 CTLFLAG_RW, &vm_pageout_update_period, 0,
201 "Maximum active LRU update period");
203 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
204 "Low memory callback period");
206 #if defined(NO_SWAPPING)
207 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
208 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
209 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
210 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
212 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
213 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
214 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
215 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
218 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
219 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
221 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
222 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
224 static int pageout_lock_miss;
225 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
226 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
228 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
229 CTLFLAG_RW, &vm_pageout_oom_seq, 0,
230 "back-to-back calls to oom detector to start OOM");
232 #define VM_PAGEOUT_PAGE_COUNT 16
233 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
235 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
236 SYSCTL_INT(_vm, OID_AUTO, max_wired,
237 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
239 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
240 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t,
242 #if !defined(NO_SWAPPING)
243 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
244 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
245 static void vm_req_vmdaemon(int req);
247 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
250 * Initialize a dummy page for marking the caller's place in the specified
251 * paging queue. In principle, this function only needs to set the flag
252 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count
253 * to one as safety precautions.
256 vm_pageout_init_marker(vm_page_t marker, u_short queue)
259 bzero(marker, sizeof(*marker));
260 marker->flags = PG_MARKER;
261 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
262 marker->queue = queue;
263 marker->hold_count = 1;
267 * vm_pageout_fallback_object_lock:
269 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
270 * known to have failed and page queue must be either PQ_ACTIVE or
271 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
272 * while locking the vm object. Use marker page to detect page queue
273 * changes and maintain notion of next page on page queue. Return
274 * TRUE if no changes were detected, FALSE otherwise. vm object is
277 * This function depends on both the lock portion of struct vm_object
278 * and normal struct vm_page being type stable.
281 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
283 struct vm_page marker;
284 struct vm_pagequeue *pq;
290 vm_pageout_init_marker(&marker, queue);
291 pq = vm_page_pagequeue(m);
294 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
295 vm_pagequeue_unlock(pq);
297 VM_OBJECT_WLOCK(object);
299 vm_pagequeue_lock(pq);
302 * The page's object might have changed, and/or the page might
303 * have moved from its original position in the queue. If the
304 * page's object has changed, then the caller should abandon
305 * processing the page because the wrong object lock was
306 * acquired. Use the marker's plinks.q, not the page's, to
307 * determine if the page has been moved. The state of the
308 * page's plinks.q can be indeterminate; whereas, the marker's
309 * plinks.q must be valid.
311 *next = TAILQ_NEXT(&marker, plinks.q);
312 unchanged = m->object == object &&
313 m == TAILQ_PREV(&marker, pglist, plinks.q);
314 KASSERT(!unchanged || m->queue == queue,
315 ("page %p queue %d %d", m, queue, m->queue));
316 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
321 * Lock the page while holding the page queue lock. Use marker page
322 * to detect page queue changes and maintain notion of next page on
323 * page queue. Return TRUE if no changes were detected, FALSE
324 * otherwise. The page is locked on return. The page queue lock might
325 * be dropped and reacquired.
327 * This function depends on normal struct vm_page being type stable.
330 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
332 struct vm_page marker;
333 struct vm_pagequeue *pq;
337 vm_page_lock_assert(m, MA_NOTOWNED);
338 if (vm_page_trylock(m))
342 vm_pageout_init_marker(&marker, queue);
343 pq = vm_page_pagequeue(m);
345 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
346 vm_pagequeue_unlock(pq);
348 vm_pagequeue_lock(pq);
350 /* Page queue might have changed. */
351 *next = TAILQ_NEXT(&marker, plinks.q);
352 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
353 KASSERT(!unchanged || m->queue == queue,
354 ("page %p queue %d %d", m, queue, m->queue));
355 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
362 * Clean the page and remove it from the laundry.
364 * We set the busy bit to cause potential page faults on this page to
365 * block. Note the careful timing, however, the busy bit isn't set till
366 * late and we cannot do anything that will mess with the page.
369 vm_pageout_cluster(vm_page_t m)
372 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
374 int ib, is, page_base;
375 vm_pindex_t pindex = m->pindex;
377 vm_page_lock_assert(m, MA_OWNED);
379 VM_OBJECT_ASSERT_WLOCKED(object);
382 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
383 * with the new swapper, but we could have serious problems paging
384 * out other object types if there is insufficient memory.
386 * Unfortunately, checking free memory here is far too late, so the
387 * check has been moved up a procedural level.
391 * Can't clean the page if it's busy or held.
393 vm_page_assert_unbusied(m);
394 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
397 mc[vm_pageout_page_count] = pb = ps = m;
399 page_base = vm_pageout_page_count;
404 * Scan object for clusterable pages.
406 * We can cluster ONLY if: ->> the page is NOT
407 * clean, wired, busy, held, or mapped into a
408 * buffer, and one of the following:
409 * 1) The page is inactive, or a seldom used
412 * 2) we force the issue.
414 * During heavy mmap/modification loads the pageout
415 * daemon can really fragment the underlying file
416 * due to flushing pages out of order and not trying
417 * align the clusters (which leave sporatic out-of-order
418 * holes). To solve this problem we do the reverse scan
419 * first and attempt to align our cluster, then do a
420 * forward scan if room remains.
423 while (ib && pageout_count < vm_pageout_page_count) {
431 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
435 vm_page_test_dirty(p);
441 if (p->queue != PQ_INACTIVE ||
442 p->hold_count != 0) { /* may be undergoing I/O */
448 mc[--page_base] = pb = p;
452 * alignment boundry, stop here and switch directions. Do
455 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
459 while (pageout_count < vm_pageout_page_count &&
460 pindex + is < object->size) {
463 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
465 vm_page_test_dirty(p);
469 if (p->queue != PQ_INACTIVE ||
470 p->hold_count != 0) { /* may be undergoing I/O */
475 mc[page_base + pageout_count] = ps = p;
481 * If we exhausted our forward scan, continue with the reverse scan
482 * when possible, even past a page boundry. This catches boundry
485 if (ib && pageout_count < vm_pageout_page_count)
489 * we allow reads during pageouts...
491 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
496 * vm_pageout_flush() - launder the given pages
498 * The given pages are laundered. Note that we setup for the start of
499 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
500 * reference count all in here rather then in the parent. If we want
501 * the parent to do more sophisticated things we may have to change
504 * Returned runlen is the count of pages between mreq and first
505 * page after mreq with status VM_PAGER_AGAIN.
506 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
507 * for any page in runlen set.
510 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
513 vm_object_t object = mc[0]->object;
514 int pageout_status[count];
518 VM_OBJECT_ASSERT_WLOCKED(object);
521 * Initiate I/O. Bump the vm_page_t->busy counter and
522 * mark the pages read-only.
524 * We do not have to fixup the clean/dirty bits here... we can
525 * allow the pager to do it after the I/O completes.
527 * NOTE! mc[i]->dirty may be partial or fragmented due to an
528 * edge case with file fragments.
530 for (i = 0; i < count; i++) {
531 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
532 ("vm_pageout_flush: partially invalid page %p index %d/%d",
534 vm_page_sbusy(mc[i]);
535 pmap_remove_write(mc[i]);
537 vm_object_pip_add(object, count);
539 vm_pager_put_pages(object, mc, count, flags, pageout_status);
541 runlen = count - mreq;
544 for (i = 0; i < count; i++) {
545 vm_page_t mt = mc[i];
547 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
548 !pmap_page_is_write_mapped(mt),
549 ("vm_pageout_flush: page %p is not write protected", mt));
550 switch (pageout_status[i]) {
557 * Page outside of range of object. Right now we
558 * essentially lose the changes by pretending it
566 * If page couldn't be paged out, then reactivate the
567 * page so it doesn't clog the inactive list. (We
568 * will try paging out it again later).
571 vm_page_activate(mt);
573 if (eio != NULL && i >= mreq && i - mreq < runlen)
577 if (i >= mreq && i - mreq < runlen)
583 * If the operation is still going, leave the page busy to
584 * block all other accesses. Also, leave the paging in
585 * progress indicator set so that we don't attempt an object
588 if (pageout_status[i] != VM_PAGER_PEND) {
589 vm_object_pip_wakeup(object);
595 return (numpagedout);
599 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low,
606 vm_page_t m, m_tmp, next;
609 vm_pagequeue_lock(pq);
610 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) {
611 if ((m->flags & PG_MARKER) != 0)
613 pa = VM_PAGE_TO_PHYS(m);
614 if (pa < low || pa + PAGE_SIZE > high)
616 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
621 if ((!VM_OBJECT_TRYWLOCK(object) &&
622 (!vm_pageout_fallback_object_lock(m, &next) ||
623 m->hold_count != 0)) || vm_page_busied(m)) {
625 VM_OBJECT_WUNLOCK(object);
628 vm_page_test_dirty(m);
629 if (m->dirty == 0 && object->ref_count != 0)
633 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) {
634 VM_OBJECT_WUNLOCK(object);
637 if (object->type == OBJT_VNODE) {
638 vm_pagequeue_unlock(pq);
640 vm_object_reference_locked(object);
641 VM_OBJECT_WUNLOCK(object);
642 (void)vn_start_write(vp, &mp, V_WAIT);
643 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
644 LK_SHARED : LK_EXCLUSIVE;
645 vn_lock(vp, lockmode | LK_RETRY);
646 VM_OBJECT_WLOCK(object);
647 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
648 VM_OBJECT_WUNLOCK(object);
650 vm_object_deallocate(object);
651 vn_finished_write(mp);
653 } else if (object->type == OBJT_SWAP ||
654 object->type == OBJT_DEFAULT) {
655 vm_pagequeue_unlock(pq);
657 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC,
659 VM_OBJECT_WUNLOCK(object);
664 * Dequeue here to prevent lock recursion in
667 vm_page_dequeue_locked(m);
671 VM_OBJECT_WUNLOCK(object);
673 vm_pagequeue_unlock(pq);
678 * Increase the number of cached pages. The specified value, "tries",
679 * determines which categories of pages are cached:
681 * 0: All clean, inactive pages within the specified physical address range
682 * are cached. Will not sleep.
683 * 1: The vm_lowmem handlers are called. All inactive pages within
684 * the specified physical address range are cached. May sleep.
685 * 2: The vm_lowmem handlers are called. All inactive and active pages
686 * within the specified physical address range are cached. May sleep.
689 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
691 int actl, actmax, inactl, inactmax, dom, initial_dom;
692 static int start_dom = 0;
696 * Decrease registered cache sizes. The vm_lowmem handlers
697 * may acquire locks and/or sleep, so they can only be invoked
698 * when "tries" is greater than zero.
700 SDT_PROBE0(vm, , , vm__lowmem_cache);
701 EVENTHANDLER_INVOKE(vm_lowmem, 0);
704 * We do this explicitly after the caches have been drained
711 * Make the next scan start on the next domain.
713 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains;
716 inactmax = vm_cnt.v_inactive_count;
718 actmax = tries < 2 ? 0 : vm_cnt.v_active_count;
722 * Scan domains in round-robin order, first inactive queues,
723 * then active. Since domain usually owns large physically
724 * contiguous chunk of memory, it makes sense to completely
725 * exhaust one domain before switching to next, while growing
726 * the pool of contiguous physical pages.
728 * Do not even start launder a domain which cannot contain
729 * the specified address range, as indicated by segments
730 * constituting the domain.
733 if (inactl < inactmax) {
734 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
736 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE],
741 if (++dom == vm_ndomains)
743 if (dom != initial_dom)
747 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
749 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE],
754 if (++dom == vm_ndomains)
756 if (dom != initial_dom)
761 #if !defined(NO_SWAPPING)
763 * vm_pageout_object_deactivate_pages
765 * Deactivate enough pages to satisfy the inactive target
768 * The object and map must be locked.
771 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
774 vm_object_t backing_object, object;
776 int act_delta, remove_mode;
778 VM_OBJECT_ASSERT_LOCKED(first_object);
779 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
781 for (object = first_object;; object = backing_object) {
782 if (pmap_resident_count(pmap) <= desired)
784 VM_OBJECT_ASSERT_LOCKED(object);
785 if ((object->flags & OBJ_UNMANAGED) != 0 ||
786 object->paging_in_progress != 0)
790 if (object->shadow_count > 1)
793 * Scan the object's entire memory queue.
795 TAILQ_FOREACH(p, &object->memq, listq) {
796 if (pmap_resident_count(pmap) <= desired)
798 if (vm_page_busied(p))
800 PCPU_INC(cnt.v_pdpages);
802 if (p->wire_count != 0 || p->hold_count != 0 ||
803 !pmap_page_exists_quick(pmap, p)) {
807 act_delta = pmap_ts_referenced(p);
808 if ((p->aflags & PGA_REFERENCED) != 0) {
811 vm_page_aflag_clear(p, PGA_REFERENCED);
813 if (p->queue != PQ_ACTIVE && act_delta != 0) {
815 p->act_count += act_delta;
816 } else if (p->queue == PQ_ACTIVE) {
817 if (act_delta == 0) {
818 p->act_count -= min(p->act_count,
820 if (!remove_mode && p->act_count == 0) {
822 vm_page_deactivate(p);
827 if (p->act_count < ACT_MAX -
829 p->act_count += ACT_ADVANCE;
832 } else if (p->queue == PQ_INACTIVE)
836 if ((backing_object = object->backing_object) == NULL)
838 VM_OBJECT_RLOCK(backing_object);
839 if (object != first_object)
840 VM_OBJECT_RUNLOCK(object);
843 if (object != first_object)
844 VM_OBJECT_RUNLOCK(object);
848 * deactivate some number of pages in a map, try to do it fairly, but
849 * that is really hard to do.
852 vm_pageout_map_deactivate_pages(map, desired)
857 vm_object_t obj, bigobj;
860 if (!vm_map_trylock(map))
867 * first, search out the biggest object, and try to free pages from
870 tmpe = map->header.next;
871 while (tmpe != &map->header) {
872 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
873 obj = tmpe->object.vm_object;
874 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
875 if (obj->shadow_count <= 1 &&
877 bigobj->resident_page_count < obj->resident_page_count)) {
879 VM_OBJECT_RUNLOCK(bigobj);
882 VM_OBJECT_RUNLOCK(obj);
885 if (tmpe->wired_count > 0)
886 nothingwired = FALSE;
890 if (bigobj != NULL) {
891 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
892 VM_OBJECT_RUNLOCK(bigobj);
895 * Next, hunt around for other pages to deactivate. We actually
896 * do this search sort of wrong -- .text first is not the best idea.
898 tmpe = map->header.next;
899 while (tmpe != &map->header) {
900 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
902 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
903 obj = tmpe->object.vm_object;
905 VM_OBJECT_RLOCK(obj);
906 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
907 VM_OBJECT_RUNLOCK(obj);
914 * Remove all mappings if a process is swapped out, this will free page
917 if (desired == 0 && nothingwired) {
918 pmap_remove(vm_map_pmap(map), vm_map_min(map),
924 #endif /* !defined(NO_SWAPPING) */
927 * Attempt to acquire all of the necessary locks to launder a page and
928 * then call through the clustering layer to PUTPAGES. Wait a short
929 * time for a vnode lock.
931 * Requires the page and object lock on entry, releases both before return.
932 * Returns 0 on success and an errno otherwise.
935 vm_pageout_clean(vm_page_t m)
943 vm_page_assert_locked(m);
945 VM_OBJECT_ASSERT_WLOCKED(object);
951 * The object is already known NOT to be dead. It
952 * is possible for the vget() to block the whole
953 * pageout daemon, but the new low-memory handling
954 * code should prevent it.
956 * We can't wait forever for the vnode lock, we might
957 * deadlock due to a vn_read() getting stuck in
958 * vm_wait while holding this vnode. We skip the
959 * vnode if we can't get it in a reasonable amount
962 if (object->type == OBJT_VNODE) {
965 if (vp->v_type == VREG &&
966 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
972 ("vp %p with NULL v_mount", vp));
973 vm_object_reference_locked(object);
975 VM_OBJECT_WUNLOCK(object);
976 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
977 LK_SHARED : LK_EXCLUSIVE;
978 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
983 VM_OBJECT_WLOCK(object);
986 * While the object and page were unlocked, the page
988 * (1) moved to a different queue,
989 * (2) reallocated to a different object,
990 * (3) reallocated to a different offset, or
993 if (m->queue != PQ_INACTIVE || m->object != object ||
994 m->pindex != pindex || m->dirty == 0) {
1001 * The page may have been busied or held while the object
1002 * and page locks were released.
1004 if (vm_page_busied(m) || m->hold_count != 0) {
1012 * If a page is dirty, then it is either being washed
1013 * (but not yet cleaned) or it is still in the
1014 * laundry. If it is still in the laundry, then we
1015 * start the cleaning operation.
1017 if (vm_pageout_cluster(m) == 0)
1021 VM_OBJECT_WUNLOCK(object);
1024 vm_page_lock_assert(m, MA_NOTOWNED);
1028 vm_object_deallocate(object);
1029 vn_finished_write(mp);
1036 * vm_pageout_scan does the dirty work for the pageout daemon.
1038 * pass 0 - Update active LRU/deactivate pages
1039 * pass 1 - Move inactive to cache or free
1040 * pass 2 - Launder dirty pages
1043 vm_pageout_scan(struct vm_domain *vmd, int pass)
1046 struct vm_pagequeue *pq;
1049 int act_delta, addl_page_shortage, deficit, error, maxlaunder, maxscan;
1050 int page_shortage, scan_tick, scanned, starting_page_shortage;
1052 boolean_t pageout_ok, queues_locked;
1055 * If we need to reclaim memory ask kernel caches to return
1056 * some. We rate limit to avoid thrashing.
1058 if (vmd == &vm_dom[0] && pass > 0 &&
1059 (time_uptime - lowmem_uptime) >= lowmem_period) {
1061 * Decrease registered cache sizes.
1063 SDT_PROBE0(vm, , , vm__lowmem_scan);
1064 EVENTHANDLER_INVOKE(vm_lowmem, 0);
1066 * We do this explicitly after the caches have been
1070 lowmem_uptime = time_uptime;
1074 * The addl_page_shortage is the number of temporarily
1075 * stuck pages in the inactive queue. In other words, the
1076 * number of pages from the inactive count that should be
1077 * discounted in setting the target for the active queue scan.
1079 addl_page_shortage = 0;
1082 * Calculate the number of pages we want to either free or move
1086 deficit = atomic_readandclear_int(&vm_pageout_deficit);
1087 page_shortage = vm_paging_target() + deficit;
1089 page_shortage = deficit = 0;
1090 starting_page_shortage = page_shortage;
1093 * maxlaunder limits the number of dirty pages we flush per scan.
1094 * For most systems a smaller value (16 or 32) is more robust under
1095 * extreme memory and disk pressure because any unnecessary writes
1096 * to disk can result in extreme performance degredation. However,
1097 * systems with excessive dirty pages (especially when MAP_NOSYNC is
1098 * used) will die horribly with limited laundering. If the pageout
1099 * daemon cannot clean enough pages in the first pass, we let it go
1100 * all out in succeeding passes.
1102 if ((maxlaunder = vm_max_launder) <= 1)
1110 * Start scanning the inactive queue for pages we can move to the
1111 * cache or free. The scan will stop when the target is reached or
1112 * we have scanned the entire inactive queue. Note that m->act_count
1113 * is not used to form decisions for the inactive queue, only for the
1116 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1117 maxscan = pq->pq_cnt;
1118 vm_pagequeue_lock(pq);
1119 queues_locked = TRUE;
1120 for (m = TAILQ_FIRST(&pq->pq_pl);
1121 m != NULL && maxscan-- > 0 && page_shortage > 0;
1123 vm_pagequeue_assert_locked(pq);
1124 KASSERT(queues_locked, ("unlocked queues"));
1125 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
1127 PCPU_INC(cnt.v_pdpages);
1128 next = TAILQ_NEXT(m, plinks.q);
1133 if (m->flags & PG_MARKER)
1136 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1137 ("Fictitious page %p cannot be in inactive queue", m));
1138 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1139 ("Unmanaged page %p cannot be in inactive queue", m));
1142 * The page or object lock acquisitions fail if the
1143 * page was removed from the queue or moved to a
1144 * different position within the queue. In either
1145 * case, addl_page_shortage should not be incremented.
1147 if (!vm_pageout_page_lock(m, &next))
1149 else if (m->hold_count != 0) {
1151 * Held pages are essentially stuck in the
1152 * queue. So, they ought to be discounted
1153 * from the inactive count. See the
1154 * calculation of the page_shortage for the
1155 * loop over the active queue below.
1157 addl_page_shortage++;
1161 if (!VM_OBJECT_TRYWLOCK(object)) {
1162 if (!vm_pageout_fallback_object_lock(m, &next))
1164 else if (m->hold_count != 0) {
1165 addl_page_shortage++;
1169 if (vm_page_busied(m)) {
1171 * Don't mess with busy pages. Leave them at
1172 * the front of the queue. Most likely, they
1173 * are being paged out and will leave the
1174 * queue shortly after the scan finishes. So,
1175 * they ought to be discounted from the
1178 addl_page_shortage++;
1180 VM_OBJECT_WUNLOCK(object);
1185 KASSERT(m->hold_count == 0, ("Held page %p", m));
1188 * We unlock the inactive page queue, invalidating the
1189 * 'next' pointer. Use our marker to remember our
1192 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1193 vm_pagequeue_unlock(pq);
1194 queues_locked = FALSE;
1197 * Invalid pages can be easily freed. They cannot be
1198 * mapped, vm_page_free() asserts this.
1204 * If the page has been referenced and the object is not dead,
1205 * reactivate or requeue the page depending on whether the
1208 if ((m->aflags & PGA_REFERENCED) != 0) {
1209 vm_page_aflag_clear(m, PGA_REFERENCED);
1213 if (object->ref_count != 0) {
1214 act_delta += pmap_ts_referenced(m);
1216 KASSERT(!pmap_page_is_mapped(m),
1217 ("vm_pageout_scan: page %p is mapped", m));
1219 if (act_delta != 0) {
1220 if (object->ref_count != 0) {
1221 vm_page_activate(m);
1224 * Increase the activation count if the page
1225 * was referenced while in the inactive queue.
1226 * This makes it less likely that the page will
1227 * be returned prematurely to the inactive
1230 m->act_count += act_delta + ACT_ADVANCE;
1232 } else if ((object->flags & OBJ_DEAD) == 0)
1237 * If the page appears to be clean at the machine-independent
1238 * layer, then remove all of its mappings from the pmap in
1239 * anticipation of placing it onto the cache queue. If,
1240 * however, any of the page's mappings allow write access,
1241 * then the page may still be modified until the last of those
1242 * mappings are removed.
1244 if (object->ref_count != 0) {
1245 vm_page_test_dirty(m);
1250 if (m->dirty == 0) {
1252 * Clean pages can be freed.
1256 PCPU_INC(cnt.v_dfree);
1258 } else if ((object->flags & OBJ_DEAD) != 0) {
1260 * Leave dirty pages from dead objects at the front of
1261 * the queue. They are being paged out and freed by
1262 * the thread that destroyed the object. They will
1263 * leave the queue shortly after the scan finishes, so
1264 * they should be discounted from the inactive count.
1266 addl_page_shortage++;
1267 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1269 * Dirty pages need to be paged out, but flushing
1270 * a page is extremely expensive versus freeing
1271 * a clean page. Rather then artificially limiting
1272 * the number of pages we can flush, we instead give
1273 * dirty pages extra priority on the inactive queue
1274 * by forcing them to be cycled through the queue
1275 * twice before being flushed, after which the
1276 * (now clean) page will cycle through once more
1277 * before being freed. This significantly extends
1278 * the thrash point for a heavily loaded machine.
1280 m->flags |= PG_WINATCFLS;
1282 vm_pagequeue_lock(pq);
1283 queues_locked = TRUE;
1284 vm_page_requeue_locked(m);
1285 } else if (maxlaunder > 0) {
1287 * We always want to try to flush some dirty pages if
1288 * we encounter them, to keep the system stable.
1289 * Normally this number is small, but under extreme
1290 * pressure where there are insufficient clean pages
1291 * on the inactive queue, we may have to go all out.
1294 if (object->type != OBJT_SWAP &&
1295 object->type != OBJT_DEFAULT)
1297 else if (disable_swap_pageouts)
1299 else if (defer_swap_pageouts)
1300 pageout_ok = vm_page_count_min();
1305 error = vm_pageout_clean(m);
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.
1314 } else if (error == EDEADLK) {
1315 pageout_lock_miss++;
1317 } else if (error == EBUSY) {
1318 addl_page_shortage++;
1320 vm_page_lock_assert(m, MA_NOTOWNED);
1325 VM_OBJECT_WUNLOCK(object);
1327 if (!queues_locked) {
1328 vm_pagequeue_lock(pq);
1329 queues_locked = TRUE;
1331 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1332 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1334 vm_pagequeue_unlock(pq);
1336 #if !defined(NO_SWAPPING)
1338 * Wakeup the swapout daemon if we didn't cache or free the targeted
1341 if (vm_swap_enabled && page_shortage > 0)
1342 vm_req_vmdaemon(VM_SWAP_NORMAL);
1346 * Wakeup the sync daemon if we skipped a vnode in a writeable object
1347 * and we didn't cache or free enough pages.
1349 if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target -
1351 (void)speedup_syncer();
1354 * If the inactive queue scan fails repeatedly to meet its
1355 * target, kill the largest process.
1357 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1360 * Compute the number of pages we want to try to move from the
1361 * active queue to the inactive queue.
1363 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count +
1364 vm_paging_target() + deficit + addl_page_shortage;
1366 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1367 vm_pagequeue_lock(pq);
1368 maxscan = pq->pq_cnt;
1371 * If we're just idle polling attempt to visit every
1372 * active page within 'update_period' seconds.
1375 if (vm_pageout_update_period != 0) {
1376 min_scan = pq->pq_cnt;
1377 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1378 min_scan /= hz * vm_pageout_update_period;
1381 if (min_scan > 0 || (page_shortage > 0 && maxscan > 0))
1382 vmd->vmd_last_active_scan = scan_tick;
1385 * Scan the active queue for pages that can be deactivated. Update
1386 * the per-page activity counter and use it to identify deactivation
1389 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1390 min_scan || (page_shortage > 0 && scanned < maxscan)); m = next,
1393 KASSERT(m->queue == PQ_ACTIVE,
1394 ("vm_pageout_scan: page %p isn't active", m));
1396 next = TAILQ_NEXT(m, plinks.q);
1397 if ((m->flags & PG_MARKER) != 0)
1399 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1400 ("Fictitious page %p cannot be in active queue", m));
1401 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1402 ("Unmanaged page %p cannot be in active queue", m));
1403 if (!vm_pageout_page_lock(m, &next)) {
1409 * The count for pagedaemon pages is done after checking the
1410 * page for eligibility...
1412 PCPU_INC(cnt.v_pdpages);
1415 * Check to see "how much" the page has been used.
1417 if ((m->aflags & PGA_REFERENCED) != 0) {
1418 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.
1437 if (act_delta != 0) {
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);
1445 * Move this page to the tail of the active or inactive
1446 * queue depending on usage.
1448 if (m->act_count == 0) {
1449 /* Dequeue to avoid later lock recursion. */
1450 vm_page_dequeue_locked(m);
1451 vm_page_deactivate(m);
1454 vm_page_requeue_locked(m);
1457 vm_pagequeue_unlock(pq);
1458 #if !defined(NO_SWAPPING)
1460 * Idle process swapout -- run once per second.
1462 if (vm_swap_idle_enabled) {
1464 if (time_second != lsec) {
1465 vm_req_vmdaemon(VM_SWAP_IDLE);
1472 static int vm_pageout_oom_vote;
1475 * The pagedaemon threads randlomly select one to perform the
1476 * OOM. Trying to kill processes before all pagedaemons
1477 * failed to reach free target is premature.
1480 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1481 int starting_page_shortage)
1485 if (starting_page_shortage <= 0 || starting_page_shortage !=
1487 vmd->vmd_oom_seq = 0;
1490 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1492 vmd->vmd_oom = FALSE;
1493 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1499 * Do not follow the call sequence until OOM condition is
1502 vmd->vmd_oom_seq = 0;
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 * The OOM killer is the page daemon's action of last resort when
1531 * memory allocation requests have been stalled for a prolonged period
1532 * of time because it cannot reclaim memory. This function computes
1533 * the approximate number of physical pages that could be reclaimed if
1534 * the specified address space is destroyed.
1536 * Private, anonymous memory owned by the address space is the
1537 * principal resource that we expect to recover after an OOM kill.
1538 * Since the physical pages mapped by the address space's COW entries
1539 * are typically shared pages, they are unlikely to be released and so
1540 * they are not counted.
1542 * To get to the point where the page daemon runs the OOM killer, its
1543 * efforts to write-back vnode-backed pages may have stalled. This
1544 * could be caused by a memory allocation deadlock in the write path
1545 * that might be resolved by an OOM kill. Therefore, physical pages
1546 * belonging to vnode-backed objects are counted, because they might
1547 * be freed without being written out first if the address space holds
1548 * the last reference to an unlinked vnode.
1550 * Similarly, physical pages belonging to OBJT_PHYS objects are
1551 * counted because the address space might hold the last reference to
1555 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1558 vm_map_entry_t entry;
1562 map = &vmspace->vm_map;
1563 KASSERT(!map->system_map, ("system map"));
1564 sx_assert(&map->lock, SA_LOCKED);
1566 for (entry = map->header.next; entry != &map->header;
1567 entry = entry->next) {
1568 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1570 obj = entry->object.vm_object;
1573 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1574 obj->ref_count != 1)
1576 switch (obj->type) {
1581 res += obj->resident_page_count;
1589 vm_pageout_oom(int shortage)
1591 struct proc *p, *bigproc;
1592 vm_offset_t size, bigsize;
1597 * We keep the process bigproc locked once we find it to keep anyone
1598 * from messing with it; however, there is a possibility of
1599 * deadlock if process B is bigproc and one of it's child processes
1600 * attempts to propagate a signal to B while we are waiting for A's
1601 * lock while walking this list. To avoid this, we don't block on
1602 * the process lock but just skip a process if it is already locked.
1606 sx_slock(&allproc_lock);
1607 FOREACH_PROC_IN_SYSTEM(p) {
1613 * If this is a system, protected or killed process, skip it.
1615 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1616 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1617 p->p_pid == 1 || P_KILLED(p) ||
1618 (p->p_pid < 48 && swap_pager_avail != 0)) {
1623 * If the process is in a non-running type state,
1624 * don't touch it. Check all the threads individually.
1627 FOREACH_THREAD_IN_PROC(p, td) {
1629 if (!TD_ON_RUNQ(td) &&
1630 !TD_IS_RUNNING(td) &&
1631 !TD_IS_SLEEPING(td) &&
1632 !TD_IS_SUSPENDED(td) &&
1633 !TD_IS_SWAPPED(td)) {
1645 * get the process size
1647 vm = vmspace_acquire_ref(p);
1653 if (!vm_map_trylock_read(&vm->vm_map)) {
1660 size = vmspace_swap_count(vm);
1661 if (shortage == VM_OOM_MEM)
1662 size += vm_pageout_oom_pagecount(vm);
1663 vm_map_unlock_read(&vm->vm_map);
1667 * If this process is bigger than the biggest one,
1670 if (size > bigsize) {
1671 if (bigproc != NULL)
1679 sx_sunlock(&allproc_lock);
1680 if (bigproc != NULL) {
1681 if (vm_panic_on_oom != 0)
1682 panic("out of swap space");
1684 killproc(bigproc, "out of swap space");
1685 sched_nice(bigproc, PRIO_MIN);
1687 PROC_UNLOCK(bigproc);
1688 wakeup(&vm_cnt.v_free_count);
1693 vm_pageout_worker(void *arg)
1695 struct vm_domain *domain;
1698 domidx = (uintptr_t)arg;
1699 domain = &vm_dom[domidx];
1702 * XXXKIB It could be useful to bind pageout daemon threads to
1703 * the cores belonging to the domain, from which vm_page_array
1707 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1708 domain->vmd_last_active_scan = ticks;
1709 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1710 vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
1711 TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1712 &domain->vmd_inacthead, plinks.q);
1715 * The pageout daemon worker is never done, so loop forever.
1719 * If we have enough free memory, wakeup waiters. Do
1720 * not clear vm_pages_needed until we reach our target,
1721 * otherwise we may be woken up over and over again and
1722 * waste a lot of cpu.
1724 mtx_lock(&vm_page_queue_free_mtx);
1725 if (vm_pages_needed && !vm_page_count_min()) {
1726 if (!vm_paging_needed())
1727 vm_pages_needed = 0;
1728 wakeup(&vm_cnt.v_free_count);
1730 if (vm_pages_needed) {
1732 * We're still not done. Either vm_pages_needed was
1733 * set by another thread during the previous scan
1734 * (typically, this happens during a level 0 scan) or
1735 * vm_pages_needed was already set and the scan failed
1736 * to free enough pages. If we haven't yet performed
1737 * a level >= 2 scan (unlimited dirty cleaning), then
1738 * upgrade the level and scan again now. Otherwise,
1739 * sleep a bit and try again later. While sleeping,
1740 * vm_pages_needed can be cleared.
1742 if (domain->vmd_pass > 1)
1743 msleep(&vm_pages_needed,
1744 &vm_page_queue_free_mtx, PVM, "psleep",
1748 * Good enough, sleep until required to refresh
1751 msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1754 if (vm_pages_needed) {
1755 vm_cnt.v_pdwakeups++;
1758 domain->vmd_pass = 0;
1759 mtx_unlock(&vm_page_queue_free_mtx);
1760 vm_pageout_scan(domain, domain->vmd_pass);
1765 * vm_pageout_init initialises basic pageout daemon settings.
1768 vm_pageout_init(void)
1771 * Initialize some paging parameters.
1773 vm_cnt.v_interrupt_free_min = 2;
1774 if (vm_cnt.v_page_count < 2000)
1775 vm_pageout_page_count = 8;
1778 * v_free_reserved needs to include enough for the largest
1779 * swap pager structures plus enough for any pv_entry structs
1782 if (vm_cnt.v_page_count > 1024)
1783 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1785 vm_cnt.v_free_min = 4;
1786 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1787 vm_cnt.v_interrupt_free_min;
1788 vm_cnt.v_free_reserved = vm_pageout_page_count +
1789 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1790 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1791 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1792 vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1793 vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1794 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1795 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1796 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1799 * Set the default wakeup threshold to be 10% above the minimum
1800 * page limit. This keeps the steady state out of shortfall.
1802 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1805 * Set interval in seconds for active scan. We want to visit each
1806 * page at least once every ten minutes. This is to prevent worst
1807 * case paging behaviors with stale active LRU.
1809 if (vm_pageout_update_period == 0)
1810 vm_pageout_update_period = 600;
1812 /* XXX does not really belong here */
1813 if (vm_page_max_wired == 0)
1814 vm_page_max_wired = vm_cnt.v_free_count / 3;
1818 * vm_pageout is the high level pageout daemon.
1828 swap_pager_swap_init();
1830 for (i = 1; i < vm_ndomains; i++) {
1831 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1832 curproc, NULL, 0, 0, "dom%d", i);
1834 panic("starting pageout for domain %d, error %d\n",
1839 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1842 panic("starting uma_reclaim helper, error %d\n", error);
1843 vm_pageout_worker((void *)(uintptr_t)0);
1847 * Unless the free page queue lock is held by the caller, this function
1848 * should be regarded as advisory. Specifically, the caller should
1849 * not msleep() on &vm_cnt.v_free_count following this function unless
1850 * the free page queue lock is held until the msleep() is performed.
1853 pagedaemon_wakeup(void)
1856 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1857 vm_pages_needed = 1;
1858 wakeup(&vm_pages_needed);
1862 #if !defined(NO_SWAPPING)
1864 vm_req_vmdaemon(int req)
1866 static int lastrun = 0;
1868 mtx_lock(&vm_daemon_mtx);
1869 vm_pageout_req_swapout |= req;
1870 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1871 wakeup(&vm_daemon_needed);
1874 mtx_unlock(&vm_daemon_mtx);
1880 struct rlimit rsslim;
1884 int breakout, swapout_flags, tryagain, attempts;
1886 uint64_t rsize, ravailable;
1890 mtx_lock(&vm_daemon_mtx);
1891 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
1893 racct_enable ? hz : 0
1898 swapout_flags = vm_pageout_req_swapout;
1899 vm_pageout_req_swapout = 0;
1900 mtx_unlock(&vm_daemon_mtx);
1902 swapout_procs(swapout_flags);
1905 * scan the processes for exceeding their rlimits or if
1906 * process is swapped out -- deactivate pages
1912 sx_slock(&allproc_lock);
1913 FOREACH_PROC_IN_SYSTEM(p) {
1914 vm_pindex_t limit, size;
1917 * if this is a system process or if we have already
1918 * looked at this process, skip it.
1921 if (p->p_state != PRS_NORMAL ||
1922 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1927 * if the process is in a non-running type state,
1931 FOREACH_THREAD_IN_PROC(p, td) {
1933 if (!TD_ON_RUNQ(td) &&
1934 !TD_IS_RUNNING(td) &&
1935 !TD_IS_SLEEPING(td) &&
1936 !TD_IS_SUSPENDED(td)) {
1950 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
1952 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1955 * let processes that are swapped out really be
1956 * swapped out set the limit to nothing (will force a
1959 if ((p->p_flag & P_INMEM) == 0)
1960 limit = 0; /* XXX */
1961 vm = vmspace_acquire_ref(p);
1966 size = vmspace_resident_count(vm);
1967 if (size >= limit) {
1968 vm_pageout_map_deactivate_pages(
1969 &vm->vm_map, limit);
1973 rsize = IDX_TO_OFF(size);
1975 racct_set(p, RACCT_RSS, rsize);
1976 ravailable = racct_get_available(p, RACCT_RSS);
1978 if (rsize > ravailable) {
1980 * Don't be overly aggressive; this
1981 * might be an innocent process,
1982 * and the limit could've been exceeded
1983 * by some memory hog. Don't try
1984 * to deactivate more than 1/4th
1985 * of process' resident set size.
1987 if (attempts <= 8) {
1988 if (ravailable < rsize -
1990 ravailable = rsize -
1994 vm_pageout_map_deactivate_pages(
1996 OFF_TO_IDX(ravailable));
1997 /* Update RSS usage after paging out. */
1998 size = vmspace_resident_count(vm);
1999 rsize = IDX_TO_OFF(size);
2001 racct_set(p, RACCT_RSS, rsize);
2003 if (rsize > ravailable)
2010 sx_sunlock(&allproc_lock);
2011 if (tryagain != 0 && attempts <= 10)
2015 #endif /* !defined(NO_SWAPPING) */