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("$FreeBSD$");
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_scan);
144 #if !defined(NO_SWAPPING)
145 /* the kernel process "vm_daemon"*/
146 static void vm_daemon(void);
147 static struct proc *vmproc;
149 static struct kproc_desc vm_kp = {
154 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
158 int vm_pageout_deficit; /* Estimated number of pages deficit */
159 u_int vm_pageout_wakeup_thresh;
160 static int vm_pageout_oom_seq = 12;
161 bool vm_pageout_wanted; /* Event on which pageout daemon sleeps */
162 bool vm_pages_needed; /* Are threads waiting for free pages? */
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 #if !defined(NO_SWAPPING)
241 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
242 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
243 static void vm_req_vmdaemon(int req);
245 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
248 * Initialize a dummy page for marking the caller's place in the specified
249 * paging queue. In principle, this function only needs to set the flag
250 * PG_MARKER. Nonetheless, it write busies and initializes the hold count
251 * to one as safety precautions.
254 vm_pageout_init_marker(vm_page_t marker, u_short queue)
257 bzero(marker, sizeof(*marker));
258 marker->flags = PG_MARKER;
259 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
260 marker->queue = queue;
261 marker->hold_count = 1;
265 * vm_pageout_fallback_object_lock:
267 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
268 * known to have failed and page queue must be either PQ_ACTIVE or
269 * PQ_INACTIVE. To avoid lock order violation, unlock the page queue
270 * while locking the vm object. Use marker page to detect page queue
271 * changes and maintain notion of next page on page queue. Return
272 * TRUE if no changes were detected, FALSE otherwise. vm object is
275 * This function depends on both the lock portion of struct vm_object
276 * and normal struct vm_page being type stable.
279 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
281 struct vm_page marker;
282 struct vm_pagequeue *pq;
288 vm_pageout_init_marker(&marker, queue);
289 pq = vm_page_pagequeue(m);
292 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
293 vm_pagequeue_unlock(pq);
295 VM_OBJECT_WLOCK(object);
297 vm_pagequeue_lock(pq);
300 * The page's object might have changed, and/or the page might
301 * have moved from its original position in the queue. If the
302 * page's object has changed, then the caller should abandon
303 * processing the page because the wrong object lock was
304 * acquired. Use the marker's plinks.q, not the page's, to
305 * determine if the page has been moved. The state of the
306 * page's plinks.q can be indeterminate; whereas, the marker's
307 * plinks.q must be valid.
309 *next = TAILQ_NEXT(&marker, plinks.q);
310 unchanged = m->object == object &&
311 m == TAILQ_PREV(&marker, pglist, plinks.q);
312 KASSERT(!unchanged || m->queue == queue,
313 ("page %p queue %d %d", m, queue, m->queue));
314 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
319 * Lock the page while holding the page queue lock. Use marker page
320 * to detect page queue changes and maintain notion of next page on
321 * page queue. Return TRUE if no changes were detected, FALSE
322 * otherwise. The page is locked on return. The page queue lock might
323 * be dropped and reacquired.
325 * This function depends on normal struct vm_page being type stable.
328 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
330 struct vm_page marker;
331 struct vm_pagequeue *pq;
335 vm_page_lock_assert(m, MA_NOTOWNED);
336 if (vm_page_trylock(m))
340 vm_pageout_init_marker(&marker, queue);
341 pq = vm_page_pagequeue(m);
343 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
344 vm_pagequeue_unlock(pq);
346 vm_pagequeue_lock(pq);
348 /* Page queue might have changed. */
349 *next = TAILQ_NEXT(&marker, plinks.q);
350 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
351 KASSERT(!unchanged || m->queue == queue,
352 ("page %p queue %d %d", m, queue, m->queue));
353 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
358 * Scan for pages at adjacent offsets within the given page's object that are
359 * eligible for laundering, form a cluster of these pages and the given page,
360 * and launder that cluster.
363 vm_pageout_cluster(vm_page_t m)
366 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
368 int ib, is, page_base, pageout_count;
370 vm_page_assert_locked(m);
372 VM_OBJECT_ASSERT_WLOCKED(object);
376 * We can't clean the page if it is busy or held.
378 vm_page_assert_unbusied(m);
379 KASSERT(m->hold_count == 0, ("page %p is held", m));
382 mc[vm_pageout_page_count] = pb = ps = m;
384 page_base = vm_pageout_page_count;
389 * We can cluster only if the page is not clean, busy, or held, and
390 * the page is inactive.
392 * During heavy mmap/modification loads the pageout
393 * daemon can really fragment the underlying file
394 * due to flushing pages out of order and not trying to
395 * align the clusters (which leaves sporadic out-of-order
396 * holes). To solve this problem we do the reverse scan
397 * first and attempt to align our cluster, then do a
398 * forward scan if room remains.
401 while (ib != 0 && pageout_count < vm_pageout_page_count) {
406 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
410 vm_page_test_dirty(p);
416 if (p->queue != PQ_INACTIVE ||
417 p->hold_count != 0) { /* may be undergoing I/O */
423 mc[--page_base] = pb = p;
428 * We are at an alignment boundary. Stop here, and switch
429 * directions. Do not clear ib.
431 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
434 while (pageout_count < vm_pageout_page_count &&
435 pindex + is < object->size) {
436 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
438 vm_page_test_dirty(p);
442 if (p->queue != PQ_INACTIVE ||
443 p->hold_count != 0) { /* may be undergoing I/O */
448 mc[page_base + pageout_count] = ps = p;
454 * If we exhausted our forward scan, continue with the reverse scan
455 * when possible, even past an alignment boundary. This catches
456 * boundary conditions.
458 if (ib != 0 && pageout_count < vm_pageout_page_count)
461 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
466 * vm_pageout_flush() - launder the given pages
468 * The given pages are laundered. Note that we setup for the start of
469 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
470 * reference count all in here rather then in the parent. If we want
471 * the parent to do more sophisticated things we may have to change
474 * Returned runlen is the count of pages between mreq and first
475 * page after mreq with status VM_PAGER_AGAIN.
476 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
477 * for any page in runlen set.
480 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
483 vm_object_t object = mc[0]->object;
484 int pageout_status[count];
488 VM_OBJECT_ASSERT_WLOCKED(object);
491 * Initiate I/O. Bump the vm_page_t->busy counter and
492 * mark the pages read-only.
494 * We do not have to fixup the clean/dirty bits here... we can
495 * allow the pager to do it after the I/O completes.
497 * NOTE! mc[i]->dirty may be partial or fragmented due to an
498 * edge case with file fragments.
500 for (i = 0; i < count; i++) {
501 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
502 ("vm_pageout_flush: partially invalid page %p index %d/%d",
504 vm_page_sbusy(mc[i]);
505 pmap_remove_write(mc[i]);
507 vm_object_pip_add(object, count);
509 vm_pager_put_pages(object, mc, count, flags, pageout_status);
511 runlen = count - mreq;
514 for (i = 0; i < count; i++) {
515 vm_page_t mt = mc[i];
517 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
518 !pmap_page_is_write_mapped(mt),
519 ("vm_pageout_flush: page %p is not write protected", mt));
520 switch (pageout_status[i]) {
527 * Page outside of range of object. Right now we
528 * essentially lose the changes by pretending it
536 * If page couldn't be paged out, then reactivate the
537 * page so it doesn't clog the inactive list. (We
538 * will try paging out it again later).
541 vm_page_activate(mt);
543 if (eio != NULL && i >= mreq && i - mreq < runlen)
547 if (i >= mreq && i - mreq < runlen)
553 * If the operation is still going, leave the page busy to
554 * block all other accesses. Also, leave the paging in
555 * progress indicator set so that we don't attempt an object
558 if (pageout_status[i] != VM_PAGER_PEND) {
559 vm_object_pip_wakeup(object);
565 return (numpagedout);
568 #if !defined(NO_SWAPPING)
570 * vm_pageout_object_deactivate_pages
572 * Deactivate enough pages to satisfy the inactive target
575 * The object and map must be locked.
578 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
581 vm_object_t backing_object, object;
583 int act_delta, remove_mode;
585 VM_OBJECT_ASSERT_LOCKED(first_object);
586 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
588 for (object = first_object;; object = backing_object) {
589 if (pmap_resident_count(pmap) <= desired)
591 VM_OBJECT_ASSERT_LOCKED(object);
592 if ((object->flags & OBJ_UNMANAGED) != 0 ||
593 object->paging_in_progress != 0)
597 if (object->shadow_count > 1)
600 * Scan the object's entire memory queue.
602 TAILQ_FOREACH(p, &object->memq, listq) {
603 if (pmap_resident_count(pmap) <= desired)
605 if (vm_page_busied(p))
607 PCPU_INC(cnt.v_pdpages);
609 if (p->wire_count != 0 || p->hold_count != 0 ||
610 !pmap_page_exists_quick(pmap, p)) {
614 act_delta = pmap_ts_referenced(p);
615 if ((p->aflags & PGA_REFERENCED) != 0) {
618 vm_page_aflag_clear(p, PGA_REFERENCED);
620 if (p->queue != PQ_ACTIVE && act_delta != 0) {
622 p->act_count += act_delta;
623 } else if (p->queue == PQ_ACTIVE) {
624 if (act_delta == 0) {
625 p->act_count -= min(p->act_count,
627 if (!remove_mode && p->act_count == 0) {
629 vm_page_deactivate(p);
634 if (p->act_count < ACT_MAX -
636 p->act_count += ACT_ADVANCE;
639 } else if (p->queue == PQ_INACTIVE)
643 if ((backing_object = object->backing_object) == NULL)
645 VM_OBJECT_RLOCK(backing_object);
646 if (object != first_object)
647 VM_OBJECT_RUNLOCK(object);
650 if (object != first_object)
651 VM_OBJECT_RUNLOCK(object);
655 * deactivate some number of pages in a map, try to do it fairly, but
656 * that is really hard to do.
659 vm_pageout_map_deactivate_pages(map, desired)
664 vm_object_t obj, bigobj;
667 if (!vm_map_trylock(map))
674 * first, search out the biggest object, and try to free pages from
677 tmpe = map->header.next;
678 while (tmpe != &map->header) {
679 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
680 obj = tmpe->object.vm_object;
681 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
682 if (obj->shadow_count <= 1 &&
684 bigobj->resident_page_count < obj->resident_page_count)) {
686 VM_OBJECT_RUNLOCK(bigobj);
689 VM_OBJECT_RUNLOCK(obj);
692 if (tmpe->wired_count > 0)
693 nothingwired = FALSE;
697 if (bigobj != NULL) {
698 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
699 VM_OBJECT_RUNLOCK(bigobj);
702 * Next, hunt around for other pages to deactivate. We actually
703 * do this search sort of wrong -- .text first is not the best idea.
705 tmpe = map->header.next;
706 while (tmpe != &map->header) {
707 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
709 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
710 obj = tmpe->object.vm_object;
712 VM_OBJECT_RLOCK(obj);
713 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
714 VM_OBJECT_RUNLOCK(obj);
721 * Remove all mappings if a process is swapped out, this will free page
724 if (desired == 0 && nothingwired) {
725 pmap_remove(vm_map_pmap(map), vm_map_min(map),
731 #endif /* !defined(NO_SWAPPING) */
734 * Attempt to acquire all of the necessary locks to launder a page and
735 * then call through the clustering layer to PUTPAGES. Wait a short
736 * time for a vnode lock.
738 * Requires the page and object lock on entry, releases both before return.
739 * Returns 0 on success and an errno otherwise.
742 vm_pageout_clean(vm_page_t m)
750 vm_page_assert_locked(m);
752 VM_OBJECT_ASSERT_WLOCKED(object);
758 * The object is already known NOT to be dead. It
759 * is possible for the vget() to block the whole
760 * pageout daemon, but the new low-memory handling
761 * code should prevent it.
763 * We can't wait forever for the vnode lock, we might
764 * deadlock due to a vn_read() getting stuck in
765 * vm_wait while holding this vnode. We skip the
766 * vnode if we can't get it in a reasonable amount
769 if (object->type == OBJT_VNODE) {
772 if (vp->v_type == VREG &&
773 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
779 ("vp %p with NULL v_mount", vp));
780 vm_object_reference_locked(object);
782 VM_OBJECT_WUNLOCK(object);
783 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
784 LK_SHARED : LK_EXCLUSIVE;
785 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
790 VM_OBJECT_WLOCK(object);
793 * While the object and page were unlocked, the page
795 * (1) moved to a different queue,
796 * (2) reallocated to a different object,
797 * (3) reallocated to a different offset, or
800 if (m->queue != PQ_INACTIVE || m->object != object ||
801 m->pindex != pindex || m->dirty == 0) {
808 * The page may have been busied or held while the object
809 * and page locks were released.
811 if (vm_page_busied(m) || m->hold_count != 0) {
819 * If a page is dirty, then it is either being washed
820 * (but not yet cleaned) or it is still in the
821 * laundry. If it is still in the laundry, then we
822 * start the cleaning operation.
824 if (vm_pageout_cluster(m) == 0)
828 VM_OBJECT_WUNLOCK(object);
831 vm_page_lock_assert(m, MA_NOTOWNED);
835 vm_object_deallocate(object);
836 vn_finished_write(mp);
843 * vm_pageout_scan does the dirty work for the pageout daemon.
845 * pass 0 - Update active LRU/deactivate pages
846 * pass 1 - Free inactive pages
847 * pass 2 - Launder dirty pages
850 vm_pageout_scan(struct vm_domain *vmd, int pass)
853 struct vm_pagequeue *pq;
856 int act_delta, addl_page_shortage, deficit, error, maxlaunder, maxscan;
857 int page_shortage, scan_tick, scanned, starting_page_shortage;
859 boolean_t pageout_ok, queue_locked;
862 * If we need to reclaim memory ask kernel caches to return
863 * some. We rate limit to avoid thrashing.
865 if (vmd == &vm_dom[0] && pass > 0 &&
866 (time_uptime - lowmem_uptime) >= lowmem_period) {
868 * Decrease registered cache sizes.
870 SDT_PROBE0(vm, , , vm__lowmem_scan);
871 EVENTHANDLER_INVOKE(vm_lowmem, 0);
873 * We do this explicitly after the caches have been
877 lowmem_uptime = time_uptime;
881 * The addl_page_shortage is the number of temporarily
882 * stuck pages in the inactive queue. In other words, the
883 * number of pages from the inactive count that should be
884 * discounted in setting the target for the active queue scan.
886 addl_page_shortage = 0;
889 * Calculate the number of pages that we want to free.
892 deficit = atomic_readandclear_int(&vm_pageout_deficit);
893 page_shortage = vm_paging_target() + deficit;
895 page_shortage = deficit = 0;
896 starting_page_shortage = page_shortage;
899 * maxlaunder limits the number of dirty pages we flush per scan.
900 * For most systems a smaller value (16 or 32) is more robust under
901 * extreme memory and disk pressure because any unnecessary writes
902 * to disk can result in extreme performance degredation. However,
903 * systems with excessive dirty pages (especially when MAP_NOSYNC is
904 * used) will die horribly with limited laundering. If the pageout
905 * daemon cannot clean enough pages in the first pass, we let it go
906 * all out in succeeding passes.
908 if ((maxlaunder = vm_max_launder) <= 1)
916 * Start scanning the inactive queue for pages that we can free. The
917 * scan will stop when we reach the target or we have scanned the
918 * entire queue. (Note that m->act_count is not used to make
919 * decisions for the inactive queue, only for the active queue.)
921 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
922 maxscan = pq->pq_cnt;
923 vm_pagequeue_lock(pq);
925 for (m = TAILQ_FIRST(&pq->pq_pl);
926 m != NULL && maxscan-- > 0 && page_shortage > 0;
928 vm_pagequeue_assert_locked(pq);
929 KASSERT(queue_locked, ("unlocked inactive queue"));
930 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
932 PCPU_INC(cnt.v_pdpages);
933 next = TAILQ_NEXT(m, plinks.q);
938 if (m->flags & PG_MARKER)
941 KASSERT((m->flags & PG_FICTITIOUS) == 0,
942 ("Fictitious page %p cannot be in inactive queue", m));
943 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
944 ("Unmanaged page %p cannot be in inactive queue", m));
947 * The page or object lock acquisitions fail if the
948 * page was removed from the queue or moved to a
949 * different position within the queue. In either
950 * case, addl_page_shortage should not be incremented.
952 if (!vm_pageout_page_lock(m, &next))
954 else if (m->hold_count != 0) {
956 * Held pages are essentially stuck in the
957 * queue. So, they ought to be discounted
958 * from the inactive count. See the
959 * calculation of the page_shortage for the
960 * loop over the active queue below.
962 addl_page_shortage++;
966 if (!VM_OBJECT_TRYWLOCK(object)) {
967 if (!vm_pageout_fallback_object_lock(m, &next))
969 else if (m->hold_count != 0) {
970 addl_page_shortage++;
974 if (vm_page_busied(m)) {
976 * Don't mess with busy pages. Leave them at
977 * the front of the queue. Most likely, they
978 * are being paged out and will leave the
979 * queue shortly after the scan finishes. So,
980 * they ought to be discounted from the
983 addl_page_shortage++;
985 VM_OBJECT_WUNLOCK(object);
990 KASSERT(m->hold_count == 0, ("Held page %p", m));
993 * We unlock the inactive page queue, invalidating the
994 * 'next' pointer. Use our marker to remember our
997 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
998 vm_pagequeue_unlock(pq);
999 queue_locked = FALSE;
1002 * Invalid pages can be easily freed. They cannot be
1003 * mapped, vm_page_free() asserts this.
1009 * If the page has been referenced and the object is not dead,
1010 * reactivate or requeue the page depending on whether the
1013 if ((m->aflags & PGA_REFERENCED) != 0) {
1014 vm_page_aflag_clear(m, PGA_REFERENCED);
1018 if (object->ref_count != 0) {
1019 act_delta += pmap_ts_referenced(m);
1021 KASSERT(!pmap_page_is_mapped(m),
1022 ("vm_pageout_scan: page %p is mapped", m));
1024 if (act_delta != 0) {
1025 if (object->ref_count != 0) {
1026 vm_page_activate(m);
1029 * Increase the activation count if the page
1030 * was referenced while in the inactive queue.
1031 * This makes it less likely that the page will
1032 * be returned prematurely to the inactive
1035 m->act_count += act_delta + ACT_ADVANCE;
1037 } else if ((object->flags & OBJ_DEAD) == 0)
1042 * If the page appears to be clean at the machine-independent
1043 * layer, then remove all of its mappings from the pmap in
1044 * anticipation of freeing it. If, however, any of the page's
1045 * mappings allow write access, then the page may still be
1046 * modified until the last of those mappings are removed.
1048 if (object->ref_count != 0) {
1049 vm_page_test_dirty(m);
1054 if (m->dirty == 0) {
1056 * Clean pages can be freed.
1060 PCPU_INC(cnt.v_dfree);
1062 } else if ((object->flags & OBJ_DEAD) != 0) {
1064 * Leave dirty pages from dead objects at the front of
1065 * the queue. They are being paged out and freed by
1066 * the thread that destroyed the object. They will
1067 * leave the queue shortly after the scan finishes, so
1068 * they should be discounted from the inactive count.
1070 addl_page_shortage++;
1071 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1073 * Dirty pages need to be paged out, but flushing
1074 * a page is extremely expensive versus freeing
1075 * a clean page. Rather then artificially limiting
1076 * the number of pages we can flush, we instead give
1077 * dirty pages extra priority on the inactive queue
1078 * by forcing them to be cycled through the queue
1079 * twice before being flushed, after which the
1080 * (now clean) page will cycle through once more
1081 * before being freed. This significantly extends
1082 * the thrash point for a heavily loaded machine.
1084 m->flags |= PG_WINATCFLS;
1086 vm_pagequeue_lock(pq);
1087 queue_locked = TRUE;
1088 vm_page_requeue_locked(m);
1089 } else if (maxlaunder > 0) {
1091 * We always want to try to flush some dirty pages if
1092 * we encounter them, to keep the system stable.
1093 * Normally this number is small, but under extreme
1094 * pressure where there are insufficient clean pages
1095 * on the inactive queue, we may have to go all out.
1098 if (object->type != OBJT_SWAP &&
1099 object->type != OBJT_DEFAULT)
1101 else if (disable_swap_pageouts)
1103 else if (defer_swap_pageouts)
1104 pageout_ok = vm_page_count_min();
1109 error = vm_pageout_clean(m);
1111 * Decrement page_shortage on success to account for
1112 * the (future) cleaned page. Otherwise we could wind
1113 * up laundering or cleaning too many pages.
1118 } else if (error == EDEADLK) {
1119 pageout_lock_miss++;
1121 } else if (error == EBUSY) {
1122 addl_page_shortage++;
1124 vm_page_lock_assert(m, MA_NOTOWNED);
1129 VM_OBJECT_WUNLOCK(object);
1131 if (!queue_locked) {
1132 vm_pagequeue_lock(pq);
1133 queue_locked = TRUE;
1135 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1136 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1138 vm_pagequeue_unlock(pq);
1140 #if !defined(NO_SWAPPING)
1142 * Wakeup the swapout daemon if we didn't free the targeted number of
1145 if (vm_swap_enabled && page_shortage > 0)
1146 vm_req_vmdaemon(VM_SWAP_NORMAL);
1150 * Wakeup the sync daemon if we skipped a vnode in a writeable object
1151 * and we didn't free enough pages.
1153 if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target -
1155 (void)speedup_syncer();
1158 * If the inactive queue scan fails repeatedly to meet its
1159 * target, kill the largest process.
1161 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1164 * Compute the number of pages we want to try to move from the
1165 * active queue to the inactive queue.
1167 page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count +
1168 vm_paging_target() + deficit + addl_page_shortage;
1170 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1171 vm_pagequeue_lock(pq);
1172 maxscan = pq->pq_cnt;
1175 * If we're just idle polling attempt to visit every
1176 * active page within 'update_period' seconds.
1179 if (vm_pageout_update_period != 0) {
1180 min_scan = pq->pq_cnt;
1181 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1182 min_scan /= hz * vm_pageout_update_period;
1185 if (min_scan > 0 || (page_shortage > 0 && maxscan > 0))
1186 vmd->vmd_last_active_scan = scan_tick;
1189 * Scan the active queue for pages that can be deactivated. Update
1190 * the per-page activity counter and use it to identify deactivation
1193 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1194 min_scan || (page_shortage > 0 && scanned < maxscan)); m = next,
1197 KASSERT(m->queue == PQ_ACTIVE,
1198 ("vm_pageout_scan: page %p isn't active", m));
1200 next = TAILQ_NEXT(m, plinks.q);
1201 if ((m->flags & PG_MARKER) != 0)
1203 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1204 ("Fictitious page %p cannot be in active queue", m));
1205 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1206 ("Unmanaged page %p cannot be in active queue", m));
1207 if (!vm_pageout_page_lock(m, &next)) {
1213 * The count for pagedaemon pages is done after checking the
1214 * page for eligibility...
1216 PCPU_INC(cnt.v_pdpages);
1219 * Check to see "how much" the page has been used.
1221 if ((m->aflags & PGA_REFERENCED) != 0) {
1222 vm_page_aflag_clear(m, PGA_REFERENCED);
1228 * Unlocked object ref count check. Two races are possible.
1229 * 1) The ref was transitioning to zero and we saw non-zero,
1230 * the pmap bits will be checked unnecessarily.
1231 * 2) The ref was transitioning to one and we saw zero.
1232 * The page lock prevents a new reference to this page so
1233 * we need not check the reference bits.
1235 if (m->object->ref_count != 0)
1236 act_delta += pmap_ts_referenced(m);
1239 * Advance or decay the act_count based on recent usage.
1241 if (act_delta != 0) {
1242 m->act_count += ACT_ADVANCE + act_delta;
1243 if (m->act_count > ACT_MAX)
1244 m->act_count = ACT_MAX;
1246 m->act_count -= min(m->act_count, ACT_DECLINE);
1249 * Move this page to the tail of the active or inactive
1250 * queue depending on usage.
1252 if (m->act_count == 0) {
1253 /* Dequeue to avoid later lock recursion. */
1254 vm_page_dequeue_locked(m);
1255 vm_page_deactivate(m);
1258 vm_page_requeue_locked(m);
1261 vm_pagequeue_unlock(pq);
1262 #if !defined(NO_SWAPPING)
1264 * Idle process swapout -- run once per second when we are reclaiming
1267 if (vm_swap_idle_enabled && pass > 0) {
1269 if (time_second != lsec) {
1270 vm_req_vmdaemon(VM_SWAP_IDLE);
1277 static int vm_pageout_oom_vote;
1280 * The pagedaemon threads randlomly select one to perform the
1281 * OOM. Trying to kill processes before all pagedaemons
1282 * failed to reach free target is premature.
1285 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1286 int starting_page_shortage)
1290 if (starting_page_shortage <= 0 || starting_page_shortage !=
1292 vmd->vmd_oom_seq = 0;
1295 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1297 vmd->vmd_oom = FALSE;
1298 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1304 * Do not follow the call sequence until OOM condition is
1307 vmd->vmd_oom_seq = 0;
1312 vmd->vmd_oom = TRUE;
1313 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1314 if (old_vote != vm_ndomains - 1)
1318 * The current pagedaemon thread is the last in the quorum to
1319 * start OOM. Initiate the selection and signaling of the
1322 vm_pageout_oom(VM_OOM_MEM);
1325 * After one round of OOM terror, recall our vote. On the
1326 * next pass, current pagedaemon would vote again if the low
1327 * memory condition is still there, due to vmd_oom being
1330 vmd->vmd_oom = FALSE;
1331 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1335 * The OOM killer is the page daemon's action of last resort when
1336 * memory allocation requests have been stalled for a prolonged period
1337 * of time because it cannot reclaim memory. This function computes
1338 * the approximate number of physical pages that could be reclaimed if
1339 * the specified address space is destroyed.
1341 * Private, anonymous memory owned by the address space is the
1342 * principal resource that we expect to recover after an OOM kill.
1343 * Since the physical pages mapped by the address space's COW entries
1344 * are typically shared pages, they are unlikely to be released and so
1345 * they are not counted.
1347 * To get to the point where the page daemon runs the OOM killer, its
1348 * efforts to write-back vnode-backed pages may have stalled. This
1349 * could be caused by a memory allocation deadlock in the write path
1350 * that might be resolved by an OOM kill. Therefore, physical pages
1351 * belonging to vnode-backed objects are counted, because they might
1352 * be freed without being written out first if the address space holds
1353 * the last reference to an unlinked vnode.
1355 * Similarly, physical pages belonging to OBJT_PHYS objects are
1356 * counted because the address space might hold the last reference to
1360 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1363 vm_map_entry_t entry;
1367 map = &vmspace->vm_map;
1368 KASSERT(!map->system_map, ("system map"));
1369 sx_assert(&map->lock, SA_LOCKED);
1371 for (entry = map->header.next; entry != &map->header;
1372 entry = entry->next) {
1373 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1375 obj = entry->object.vm_object;
1378 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1379 obj->ref_count != 1)
1381 switch (obj->type) {
1386 res += obj->resident_page_count;
1394 vm_pageout_oom(int shortage)
1396 struct proc *p, *bigproc;
1397 vm_offset_t size, bigsize;
1402 * We keep the process bigproc locked once we find it to keep anyone
1403 * from messing with it; however, there is a possibility of
1404 * deadlock if process B is bigproc and one of it's child processes
1405 * attempts to propagate a signal to B while we are waiting for A's
1406 * lock while walking this list. To avoid this, we don't block on
1407 * the process lock but just skip a process if it is already locked.
1411 sx_slock(&allproc_lock);
1412 FOREACH_PROC_IN_SYSTEM(p) {
1418 * If this is a system, protected or killed process, skip it.
1420 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1421 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1422 p->p_pid == 1 || P_KILLED(p) ||
1423 (p->p_pid < 48 && swap_pager_avail != 0)) {
1428 * If the process is in a non-running type state,
1429 * don't touch it. Check all the threads individually.
1432 FOREACH_THREAD_IN_PROC(p, td) {
1434 if (!TD_ON_RUNQ(td) &&
1435 !TD_IS_RUNNING(td) &&
1436 !TD_IS_SLEEPING(td) &&
1437 !TD_IS_SUSPENDED(td) &&
1438 !TD_IS_SWAPPED(td)) {
1450 * get the process size
1452 vm = vmspace_acquire_ref(p);
1459 sx_sunlock(&allproc_lock);
1460 if (!vm_map_trylock_read(&vm->vm_map)) {
1462 sx_slock(&allproc_lock);
1466 size = vmspace_swap_count(vm);
1467 if (shortage == VM_OOM_MEM)
1468 size += vm_pageout_oom_pagecount(vm);
1469 vm_map_unlock_read(&vm->vm_map);
1471 sx_slock(&allproc_lock);
1474 * If this process is bigger than the biggest one,
1477 if (size > bigsize) {
1478 if (bigproc != NULL)
1486 sx_sunlock(&allproc_lock);
1487 if (bigproc != NULL) {
1488 if (vm_panic_on_oom != 0)
1489 panic("out of swap space");
1491 killproc(bigproc, "out of swap space");
1492 sched_nice(bigproc, PRIO_MIN);
1494 PROC_UNLOCK(bigproc);
1495 wakeup(&vm_cnt.v_free_count);
1500 vm_pageout_worker(void *arg)
1502 struct vm_domain *domain;
1505 domidx = (uintptr_t)arg;
1506 domain = &vm_dom[domidx];
1509 * XXXKIB It could be useful to bind pageout daemon threads to
1510 * the cores belonging to the domain, from which vm_page_array
1514 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1515 domain->vmd_last_active_scan = ticks;
1516 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1517 vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
1518 TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1519 &domain->vmd_inacthead, plinks.q);
1522 * The pageout daemon worker is never done, so loop forever.
1525 mtx_lock(&vm_page_queue_free_mtx);
1528 * Generally, after a level >= 1 scan, if there are enough
1529 * free pages to wakeup the waiters, then they are already
1530 * awake. A call to vm_page_free() during the scan awakened
1531 * them. However, in the following case, this wakeup serves
1532 * to bound the amount of time that a thread might wait.
1533 * Suppose a thread's call to vm_page_alloc() fails, but
1534 * before that thread calls VM_WAIT, enough pages are freed by
1535 * other threads to alleviate the free page shortage. The
1536 * thread will, nonetheless, wait until another page is freed
1537 * or this wakeup is performed.
1539 if (vm_pages_needed && !vm_page_count_min()) {
1540 vm_pages_needed = false;
1541 wakeup(&vm_cnt.v_free_count);
1545 * Do not clear vm_pageout_wanted until we reach our target.
1546 * Otherwise, we may be awakened over and over again, wasting
1549 if (vm_pageout_wanted && !vm_paging_needed())
1550 vm_pageout_wanted = false;
1553 * Might the page daemon receive a wakeup call?
1555 if (vm_pageout_wanted) {
1557 * No. Either vm_pageout_wanted was set by another
1558 * thread during the previous scan, which must have
1559 * been a level 0 scan, or vm_pageout_wanted was
1560 * already set and the scan failed to free enough
1561 * pages. If we haven't yet performed a level >= 2
1562 * scan (unlimited dirty cleaning), then upgrade the
1563 * level and scan again now. Otherwise, sleep a bit
1564 * and try again later.
1566 mtx_unlock(&vm_page_queue_free_mtx);
1567 if (domain->vmd_pass > 1)
1568 pause("psleep", hz / 2);
1572 * Yes. Sleep until pages need to be reclaimed or
1573 * have their reference stats updated.
1575 if (mtx_sleep(&vm_pageout_wanted,
1576 &vm_page_queue_free_mtx, PDROP | PVM, "psleep",
1578 PCPU_INC(cnt.v_pdwakeups);
1579 domain->vmd_pass = 1;
1581 domain->vmd_pass = 0;
1584 vm_pageout_scan(domain, domain->vmd_pass);
1589 * vm_pageout_init initialises basic pageout daemon settings.
1592 vm_pageout_init(void)
1595 * Initialize some paging parameters.
1597 vm_cnt.v_interrupt_free_min = 2;
1598 if (vm_cnt.v_page_count < 2000)
1599 vm_pageout_page_count = 8;
1602 * v_free_reserved needs to include enough for the largest
1603 * swap pager structures plus enough for any pv_entry structs
1606 if (vm_cnt.v_page_count > 1024)
1607 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1609 vm_cnt.v_free_min = 4;
1610 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1611 vm_cnt.v_interrupt_free_min;
1612 vm_cnt.v_free_reserved = vm_pageout_page_count +
1613 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1614 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1615 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1616 vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1617 vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1618 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1619 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1620 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1623 * Set the default wakeup threshold to be 10% above the minimum
1624 * page limit. This keeps the steady state out of shortfall.
1626 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1629 * Set interval in seconds for active scan. We want to visit each
1630 * page at least once every ten minutes. This is to prevent worst
1631 * case paging behaviors with stale active LRU.
1633 if (vm_pageout_update_period == 0)
1634 vm_pageout_update_period = 600;
1636 /* XXX does not really belong here */
1637 if (vm_page_max_wired == 0)
1638 vm_page_max_wired = vm_cnt.v_free_count / 3;
1642 * vm_pageout is the high level pageout daemon.
1648 #ifdef VM_NUMA_ALLOC
1652 swap_pager_swap_init();
1653 #ifdef VM_NUMA_ALLOC
1654 for (i = 1; i < vm_ndomains; i++) {
1655 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1656 curproc, NULL, 0, 0, "dom%d", i);
1658 panic("starting pageout for domain %d, error %d\n",
1663 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1666 panic("starting uma_reclaim helper, error %d\n", error);
1667 vm_pageout_worker((void *)(uintptr_t)0);
1671 * Unless the free page queue lock is held by the caller, this function
1672 * should be regarded as advisory. Specifically, the caller should
1673 * not msleep() on &vm_cnt.v_free_count following this function unless
1674 * the free page queue lock is held until the msleep() is performed.
1677 pagedaemon_wakeup(void)
1680 if (!vm_pageout_wanted && curthread->td_proc != pageproc) {
1681 vm_pageout_wanted = true;
1682 wakeup(&vm_pageout_wanted);
1686 #if !defined(NO_SWAPPING)
1688 vm_req_vmdaemon(int req)
1690 static int lastrun = 0;
1692 mtx_lock(&vm_daemon_mtx);
1693 vm_pageout_req_swapout |= req;
1694 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1695 wakeup(&vm_daemon_needed);
1698 mtx_unlock(&vm_daemon_mtx);
1704 struct rlimit rsslim;
1708 int breakout, swapout_flags, tryagain, attempts;
1710 uint64_t rsize, ravailable;
1714 mtx_lock(&vm_daemon_mtx);
1715 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
1717 racct_enable ? hz : 0
1722 swapout_flags = vm_pageout_req_swapout;
1723 vm_pageout_req_swapout = 0;
1724 mtx_unlock(&vm_daemon_mtx);
1726 swapout_procs(swapout_flags);
1729 * scan the processes for exceeding their rlimits or if
1730 * process is swapped out -- deactivate pages
1736 sx_slock(&allproc_lock);
1737 FOREACH_PROC_IN_SYSTEM(p) {
1738 vm_pindex_t limit, size;
1741 * if this is a system process or if we have already
1742 * looked at this process, skip it.
1745 if (p->p_state != PRS_NORMAL ||
1746 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1751 * if the process is in a non-running type state,
1755 FOREACH_THREAD_IN_PROC(p, td) {
1757 if (!TD_ON_RUNQ(td) &&
1758 !TD_IS_RUNNING(td) &&
1759 !TD_IS_SLEEPING(td) &&
1760 !TD_IS_SUSPENDED(td)) {
1774 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
1776 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1779 * let processes that are swapped out really be
1780 * swapped out set the limit to nothing (will force a
1783 if ((p->p_flag & P_INMEM) == 0)
1784 limit = 0; /* XXX */
1785 vm = vmspace_acquire_ref(p);
1792 sx_sunlock(&allproc_lock);
1794 size = vmspace_resident_count(vm);
1795 if (size >= limit) {
1796 vm_pageout_map_deactivate_pages(
1797 &vm->vm_map, limit);
1801 rsize = IDX_TO_OFF(size);
1803 racct_set(p, RACCT_RSS, rsize);
1804 ravailable = racct_get_available(p, RACCT_RSS);
1806 if (rsize > ravailable) {
1808 * Don't be overly aggressive; this
1809 * might be an innocent process,
1810 * and the limit could've been exceeded
1811 * by some memory hog. Don't try
1812 * to deactivate more than 1/4th
1813 * of process' resident set size.
1815 if (attempts <= 8) {
1816 if (ravailable < rsize -
1818 ravailable = rsize -
1822 vm_pageout_map_deactivate_pages(
1824 OFF_TO_IDX(ravailable));
1825 /* Update RSS usage after paging out. */
1826 size = vmspace_resident_count(vm);
1827 rsize = IDX_TO_OFF(size);
1829 racct_set(p, RACCT_RSS, rsize);
1831 if (rsize > ravailable)
1837 sx_slock(&allproc_lock);
1840 sx_sunlock(&allproc_lock);
1841 if (tryagain != 0 && attempts <= 10)
1845 #endif /* !defined(NO_SWAPPING) */