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, int *numpagedout);
123 static int vm_pageout_cluster(vm_page_t m);
124 static bool 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 /* Pagedaemon activity rates, in subdivisions of one second. */
145 #define VM_LAUNDER_RATE 10
146 #define VM_INACT_SCAN_RATE 2
148 int vm_pageout_deficit; /* Estimated number of pages deficit */
149 u_int vm_pageout_wakeup_thresh;
150 static int vm_pageout_oom_seq = 12;
151 bool vm_pageout_wanted; /* Event on which pageout daemon sleeps */
152 bool vm_pages_needed; /* Are threads waiting for free pages? */
154 /* Pending request for dirty page laundering. */
157 VM_LAUNDRY_BACKGROUND,
159 } vm_laundry_request = VM_LAUNDRY_IDLE;
161 static int vm_pageout_update_period;
162 static int disable_swap_pageouts;
163 static int lowmem_period = 10;
164 static time_t lowmem_uptime;
165 static int swapdev_enabled;
167 static int vm_panic_on_oom = 0;
169 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
170 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
171 "panic on out of memory instead of killing the largest process");
173 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
174 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
175 "free page threshold for waking up the pageout daemon");
177 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
178 CTLFLAG_RW, &vm_pageout_update_period, 0,
179 "Maximum active LRU update period");
181 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
182 "Low memory callback period");
184 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
185 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
187 static int pageout_lock_miss;
188 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
189 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
191 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
192 CTLFLAG_RW, &vm_pageout_oom_seq, 0,
193 "back-to-back calls to oom detector to start OOM");
195 static int act_scan_laundry_weight = 3;
196 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RW,
197 &act_scan_laundry_weight, 0,
198 "weight given to clean vs. dirty pages in active queue scans");
200 static u_int vm_background_launder_target;
201 SYSCTL_UINT(_vm, OID_AUTO, background_launder_target, CTLFLAG_RW,
202 &vm_background_launder_target, 0,
203 "background laundering target, in pages");
205 static u_int vm_background_launder_rate = 4096;
206 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RW,
207 &vm_background_launder_rate, 0,
208 "background laundering rate, in kilobytes per second");
210 static u_int vm_background_launder_max = 20 * 1024;
211 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RW,
212 &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
214 int vm_pageout_page_count = 32;
216 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
217 SYSCTL_INT(_vm, OID_AUTO, max_wired,
218 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
220 static u_int isqrt(u_int num);
221 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
222 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
224 static void vm_pageout_laundry_worker(void *arg);
225 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
228 * Initialize a dummy page for marking the caller's place in the specified
229 * paging queue. In principle, this function only needs to set the flag
230 * PG_MARKER. Nonetheless, it write busies and initializes the hold count
231 * to one as safety precautions.
234 vm_pageout_init_marker(vm_page_t marker, u_short queue)
237 bzero(marker, sizeof(*marker));
238 marker->flags = PG_MARKER;
239 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
240 marker->queue = queue;
241 marker->hold_count = 1;
245 * vm_pageout_fallback_object_lock:
247 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
248 * known to have failed and page queue must be either PQ_ACTIVE or
249 * PQ_INACTIVE. To avoid lock order violation, unlock the page queue
250 * while locking the vm object. Use marker page to detect page queue
251 * changes and maintain notion of next page on page queue. Return
252 * TRUE if no changes were detected, FALSE otherwise. vm object is
255 * This function depends on both the lock portion of struct vm_object
256 * and normal struct vm_page being type stable.
259 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
261 struct vm_page marker;
262 struct vm_pagequeue *pq;
268 vm_pageout_init_marker(&marker, queue);
269 pq = vm_page_pagequeue(m);
272 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
273 vm_pagequeue_unlock(pq);
275 VM_OBJECT_WLOCK(object);
277 vm_pagequeue_lock(pq);
280 * The page's object might have changed, and/or the page might
281 * have moved from its original position in the queue. If the
282 * page's object has changed, then the caller should abandon
283 * processing the page because the wrong object lock was
284 * acquired. Use the marker's plinks.q, not the page's, to
285 * determine if the page has been moved. The state of the
286 * page's plinks.q can be indeterminate; whereas, the marker's
287 * plinks.q must be valid.
289 *next = TAILQ_NEXT(&marker, plinks.q);
290 unchanged = m->object == object &&
291 m == TAILQ_PREV(&marker, pglist, plinks.q);
292 KASSERT(!unchanged || m->queue == queue,
293 ("page %p queue %d %d", m, queue, m->queue));
294 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
299 * Lock the page while holding the page queue lock. Use marker page
300 * to detect page queue changes and maintain notion of next page on
301 * page queue. Return TRUE if no changes were detected, FALSE
302 * otherwise. The page is locked on return. The page queue lock might
303 * be dropped and reacquired.
305 * This function depends on normal struct vm_page being type stable.
308 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
310 struct vm_page marker;
311 struct vm_pagequeue *pq;
315 vm_page_lock_assert(m, MA_NOTOWNED);
316 if (vm_page_trylock(m))
320 vm_pageout_init_marker(&marker, queue);
321 pq = vm_page_pagequeue(m);
323 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
324 vm_pagequeue_unlock(pq);
326 vm_pagequeue_lock(pq);
328 /* Page queue might have changed. */
329 *next = TAILQ_NEXT(&marker, plinks.q);
330 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
331 KASSERT(!unchanged || m->queue == queue,
332 ("page %p queue %d %d", m, queue, m->queue));
333 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
338 * Scan for pages at adjacent offsets within the given page's object that are
339 * eligible for laundering, form a cluster of these pages and the given page,
340 * and launder that cluster.
343 vm_pageout_cluster(vm_page_t m)
346 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
348 int ib, is, page_base, pageout_count;
350 vm_page_assert_locked(m);
352 VM_OBJECT_ASSERT_WLOCKED(object);
356 * We can't clean the page if it is busy or held.
358 vm_page_assert_unbusied(m);
359 KASSERT(m->hold_count == 0, ("page %p is held", m));
361 pmap_remove_write(m);
364 mc[vm_pageout_page_count] = pb = ps = m;
366 page_base = vm_pageout_page_count;
371 * We can cluster only if the page is not clean, busy, or held, and
372 * the page is in the laundry queue.
374 * During heavy mmap/modification loads the pageout
375 * daemon can really fragment the underlying file
376 * due to flushing pages out of order and not trying to
377 * align the clusters (which leaves sporadic out-of-order
378 * holes). To solve this problem we do the reverse scan
379 * first and attempt to align our cluster, then do a
380 * forward scan if room remains.
383 while (ib != 0 && pageout_count < vm_pageout_page_count) {
388 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
392 vm_page_test_dirty(p);
398 if (!vm_page_in_laundry(p) ||
399 p->hold_count != 0) { /* may be undergoing I/O */
404 pmap_remove_write(p);
406 mc[--page_base] = pb = p;
411 * We are at an alignment boundary. Stop here, and switch
412 * directions. Do not clear ib.
414 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
417 while (pageout_count < vm_pageout_page_count &&
418 pindex + is < object->size) {
419 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
421 vm_page_test_dirty(p);
425 if (!vm_page_in_laundry(p) ||
426 p->hold_count != 0) { /* may be undergoing I/O */
430 pmap_remove_write(p);
432 mc[page_base + pageout_count] = ps = p;
438 * If we exhausted our forward scan, continue with the reverse scan
439 * when possible, even past an alignment boundary. This catches
440 * boundary conditions.
442 if (ib != 0 && pageout_count < vm_pageout_page_count)
445 return (vm_pageout_flush(&mc[page_base], pageout_count,
446 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
450 * vm_pageout_flush() - launder the given pages
452 * The given pages are laundered. Note that we setup for the start of
453 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
454 * reference count all in here rather then in the parent. If we want
455 * the parent to do more sophisticated things we may have to change
458 * Returned runlen is the count of pages between mreq and first
459 * page after mreq with status VM_PAGER_AGAIN.
460 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
461 * for any page in runlen set.
464 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
467 vm_object_t object = mc[0]->object;
468 int pageout_status[count];
472 VM_OBJECT_ASSERT_WLOCKED(object);
475 * Initiate I/O. Mark the pages busy and verify that they're valid
478 * We do not have to fixup the clean/dirty bits here... we can
479 * allow the pager to do it after the I/O completes.
481 * NOTE! mc[i]->dirty may be partial or fragmented due to an
482 * edge case with file fragments.
484 for (i = 0; i < count; i++) {
485 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
486 ("vm_pageout_flush: partially invalid page %p index %d/%d",
488 KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
489 ("vm_pageout_flush: writeable page %p", mc[i]));
490 vm_page_sbusy(mc[i]);
492 vm_object_pip_add(object, count);
494 vm_pager_put_pages(object, mc, count, flags, pageout_status);
496 runlen = count - mreq;
499 for (i = 0; i < count; i++) {
500 vm_page_t mt = mc[i];
502 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
503 !pmap_page_is_write_mapped(mt),
504 ("vm_pageout_flush: page %p is not write protected", mt));
505 switch (pageout_status[i]) {
508 if (vm_page_in_laundry(mt))
509 vm_page_deactivate_noreuse(mt);
517 * The page is outside the object's range. We pretend
518 * that the page out worked and clean the page, so the
519 * changes will be lost if the page is reclaimed by
524 if (vm_page_in_laundry(mt))
525 vm_page_deactivate_noreuse(mt);
531 * If the page couldn't be paged out to swap because the
532 * pager wasn't able to find space, place the page in
533 * the PQ_UNSWAPPABLE holding queue. This is an
534 * optimization that prevents the page daemon from
535 * wasting CPU cycles on pages that cannot be reclaimed
536 * becase no swap device is configured.
538 * Otherwise, reactivate the page so that it doesn't
539 * clog the laundry and inactive queues. (We will try
540 * paging it out again later.)
543 if (object->type == OBJT_SWAP &&
544 pageout_status[i] == VM_PAGER_FAIL) {
545 vm_page_unswappable(mt);
548 vm_page_activate(mt);
550 if (eio != NULL && i >= mreq && i - mreq < runlen)
554 if (i >= mreq && i - mreq < runlen)
560 * If the operation is still going, leave the page busy to
561 * block all other accesses. Also, leave the paging in
562 * progress indicator set so that we don't attempt an object
565 if (pageout_status[i] != VM_PAGER_PEND) {
566 vm_object_pip_wakeup(object);
572 return (numpagedout);
576 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
579 atomic_store_rel_int(&swapdev_enabled, 1);
583 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
586 if (swap_pager_nswapdev() == 1)
587 atomic_store_rel_int(&swapdev_enabled, 0);
591 * Attempt to acquire all of the necessary locks to launder a page and
592 * then call through the clustering layer to PUTPAGES. Wait a short
593 * time for a vnode lock.
595 * Requires the page and object lock on entry, releases both before return.
596 * Returns 0 on success and an errno otherwise.
599 vm_pageout_clean(vm_page_t m, int *numpagedout)
607 vm_page_assert_locked(m);
609 VM_OBJECT_ASSERT_WLOCKED(object);
615 * The object is already known NOT to be dead. It
616 * is possible for the vget() to block the whole
617 * pageout daemon, but the new low-memory handling
618 * code should prevent it.
620 * We can't wait forever for the vnode lock, we might
621 * deadlock due to a vn_read() getting stuck in
622 * vm_wait while holding this vnode. We skip the
623 * vnode if we can't get it in a reasonable amount
626 if (object->type == OBJT_VNODE) {
629 if (vp->v_type == VREG &&
630 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
636 ("vp %p with NULL v_mount", vp));
637 vm_object_reference_locked(object);
639 VM_OBJECT_WUNLOCK(object);
640 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
641 LK_SHARED : LK_EXCLUSIVE;
642 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
647 VM_OBJECT_WLOCK(object);
650 * While the object and page were unlocked, the page
652 * (1) moved to a different queue,
653 * (2) reallocated to a different object,
654 * (3) reallocated to a different offset, or
657 if (!vm_page_in_laundry(m) || m->object != object ||
658 m->pindex != pindex || m->dirty == 0) {
665 * The page may have been busied or held while the object
666 * and page locks were released.
668 if (vm_page_busied(m) || m->hold_count != 0) {
676 * If a page is dirty, then it is either being washed
677 * (but not yet cleaned) or it is still in the
678 * laundry. If it is still in the laundry, then we
679 * start the cleaning operation.
681 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
685 VM_OBJECT_WUNLOCK(object);
688 vm_page_lock_assert(m, MA_NOTOWNED);
692 vm_object_deallocate(object);
693 vn_finished_write(mp);
700 * Attempt to launder the specified number of pages.
702 * Returns the number of pages successfully laundered.
705 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
707 struct vm_pagequeue *pq;
710 int act_delta, error, maxscan, numpagedout, starting_target;
712 bool pageout_ok, queue_locked;
714 starting_target = launder;
718 * Scan the laundry queues for pages eligible to be laundered. We stop
719 * once the target number of dirty pages have been laundered, or once
720 * we've reached the end of the queue. A single iteration of this loop
721 * may cause more than one page to be laundered because of clustering.
723 * maxscan ensures that we don't re-examine requeued pages. Any
724 * additional pages written as part of a cluster are subtracted from
725 * maxscan since they must be taken from the laundry queue.
727 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
728 * swap devices are configured.
730 if (atomic_load_acq_int(&swapdev_enabled))
731 pq = &vmd->vmd_pagequeues[PQ_UNSWAPPABLE];
733 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
736 vm_pagequeue_lock(pq);
737 maxscan = pq->pq_cnt;
739 for (m = TAILQ_FIRST(&pq->pq_pl);
740 m != NULL && maxscan-- > 0 && launder > 0;
742 vm_pagequeue_assert_locked(pq);
743 KASSERT(queue_locked, ("unlocked laundry queue"));
744 KASSERT(vm_page_in_laundry(m),
745 ("page %p has an inconsistent queue", m));
746 next = TAILQ_NEXT(m, plinks.q);
747 if ((m->flags & PG_MARKER) != 0)
749 KASSERT((m->flags & PG_FICTITIOUS) == 0,
750 ("PG_FICTITIOUS page %p cannot be in laundry queue", m));
751 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
752 ("VPO_UNMANAGED page %p cannot be in laundry queue", m));
753 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
758 if ((!VM_OBJECT_TRYWLOCK(object) &&
759 (!vm_pageout_fallback_object_lock(m, &next) ||
760 m->hold_count != 0)) || vm_page_busied(m)) {
761 VM_OBJECT_WUNLOCK(object);
767 * Unlock the laundry queue, invalidating the 'next' pointer.
768 * Use a marker to remember our place in the laundry queue.
770 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_laundry_marker,
772 vm_pagequeue_unlock(pq);
773 queue_locked = false;
776 * Invalid pages can be easily freed. They cannot be
777 * mapped; vm_page_free() asserts this.
783 * If the page has been referenced and the object is not dead,
784 * reactivate or requeue the page depending on whether the
787 if ((m->aflags & PGA_REFERENCED) != 0) {
788 vm_page_aflag_clear(m, PGA_REFERENCED);
792 if (object->ref_count != 0)
793 act_delta += pmap_ts_referenced(m);
795 KASSERT(!pmap_page_is_mapped(m),
796 ("page %p is mapped", m));
798 if (act_delta != 0) {
799 if (object->ref_count != 0) {
800 VM_CNT_INC(v_reactivated);
804 * Increase the activation count if the page
805 * was referenced while in the laundry queue.
806 * This makes it less likely that the page will
807 * be returned prematurely to the inactive
810 m->act_count += act_delta + ACT_ADVANCE;
813 * If this was a background laundering, count
814 * activated pages towards our target. The
815 * purpose of background laundering is to ensure
816 * that pages are eventually cycled through the
817 * laundry queue, and an activation is a valid
823 } else if ((object->flags & OBJ_DEAD) == 0)
828 * If the page appears to be clean at the machine-independent
829 * layer, then remove all of its mappings from the pmap in
830 * anticipation of freeing it. If, however, any of the page's
831 * mappings allow write access, then the page may still be
832 * modified until the last of those mappings are removed.
834 if (object->ref_count != 0) {
835 vm_page_test_dirty(m);
841 * Clean pages are freed, and dirty pages are paged out unless
842 * they belong to a dead object. Requeueing dirty pages from
843 * dead objects is pointless, as they are being paged out and
844 * freed by the thread that destroyed the object.
850 } else if ((object->flags & OBJ_DEAD) == 0) {
851 if (object->type != OBJT_SWAP &&
852 object->type != OBJT_DEFAULT)
854 else if (disable_swap_pageouts)
860 vm_pagequeue_lock(pq);
862 vm_page_requeue_locked(m);
867 * Form a cluster with adjacent, dirty pages from the
868 * same object, and page out that entire cluster.
870 * The adjacent, dirty pages must also be in the
871 * laundry. However, their mappings are not checked
872 * for new references. Consequently, a recently
873 * referenced page may be paged out. However, that
874 * page will not be prematurely reclaimed. After page
875 * out, the page will be placed in the inactive queue,
876 * where any new references will be detected and the
879 error = vm_pageout_clean(m, &numpagedout);
881 launder -= numpagedout;
882 maxscan -= numpagedout - 1;
883 } else if (error == EDEADLK) {
891 VM_OBJECT_WUNLOCK(object);
894 vm_pagequeue_lock(pq);
897 next = TAILQ_NEXT(&vmd->vmd_laundry_marker, plinks.q);
898 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_laundry_marker, plinks.q);
900 vm_pagequeue_unlock(pq);
902 if (launder > 0 && pq == &vmd->vmd_pagequeues[PQ_UNSWAPPABLE]) {
903 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
908 * Wakeup the sync daemon if we skipped a vnode in a writeable object
909 * and we didn't launder enough pages.
911 if (vnodes_skipped > 0 && launder > 0)
912 (void)speedup_syncer();
914 return (starting_target - launder);
918 * Compute the integer square root.
923 u_int bit, root, tmp;
925 bit = 1u << ((NBBY * sizeof(u_int)) - 2);
942 * Perform the work of the laundry thread: periodically wake up and determine
943 * whether any pages need to be laundered. If so, determine the number of pages
944 * that need to be laundered, and launder them.
947 vm_pageout_laundry_worker(void *arg)
949 struct vm_domain *domain;
950 struct vm_pagequeue *pq;
951 uint64_t nclean, ndirty;
952 u_int last_launder, wakeups;
953 int domidx, last_target, launder, shortfall, shortfall_cycle, target;
956 domidx = (uintptr_t)arg;
957 domain = &vm_dom[domidx];
958 pq = &domain->vmd_pagequeues[PQ_LAUNDRY];
959 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
960 vm_pageout_init_marker(&domain->vmd_laundry_marker, PQ_LAUNDRY);
963 in_shortfall = false;
969 * Calls to these handlers are serialized by the swap syscall lock.
971 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, domain,
972 EVENTHANDLER_PRI_ANY);
973 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, domain,
974 EVENTHANDLER_PRI_ANY);
977 * The pageout laundry worker is never done, so loop forever.
980 KASSERT(target >= 0, ("negative target %d", target));
981 KASSERT(shortfall_cycle >= 0,
982 ("negative cycle %d", shortfall_cycle));
984 wakeups = VM_CNT_FETCH(v_pdwakeups);
987 * First determine whether we need to launder pages to meet a
988 * shortage of free pages.
992 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
994 } else if (!in_shortfall)
996 else if (shortfall_cycle == 0 || vm_laundry_target() <= 0) {
998 * We recently entered shortfall and began laundering
999 * pages. If we have completed that laundering run
1000 * (and we are no longer in shortfall) or we have met
1001 * our laundry target through other activity, then we
1002 * can stop laundering pages.
1004 in_shortfall = false;
1008 last_launder = wakeups;
1009 launder = target / shortfall_cycle--;
1013 * There's no immediate need to launder any pages; see if we
1014 * meet the conditions to perform background laundering:
1016 * 1. The ratio of dirty to clean inactive pages exceeds the
1017 * background laundering threshold and the pagedaemon has
1018 * been woken up to reclaim pages since our last
1020 * 2. we haven't yet reached the target of the current
1021 * background laundering run.
1023 * The background laundering threshold is not a constant.
1024 * Instead, it is a slowly growing function of the number of
1025 * page daemon wakeups since the last laundering. Thus, as the
1026 * ratio of dirty to clean inactive pages grows, the amount of
1027 * memory pressure required to trigger laundering decreases.
1030 nclean = vm_cnt.v_inactive_count + vm_cnt.v_free_count;
1031 ndirty = vm_cnt.v_laundry_count;
1032 if (target == 0 && wakeups != last_launder &&
1033 ndirty * isqrt(wakeups - last_launder) >= nclean) {
1034 target = vm_background_launder_target;
1038 * We have a non-zero background laundering target. If we've
1039 * laundered up to our maximum without observing a page daemon
1040 * wakeup, just stop. This is a safety belt that ensures we
1041 * don't launder an excessive amount if memory pressure is low
1042 * and the ratio of dirty to clean pages is large. Otherwise,
1043 * proceed at the background laundering rate.
1046 if (wakeups != last_launder) {
1047 last_launder = wakeups;
1048 last_target = target;
1049 } else if (last_target - target >=
1050 vm_background_launder_max * PAGE_SIZE / 1024) {
1053 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1054 launder /= VM_LAUNDER_RATE;
1055 if (launder > target)
1062 * Because of I/O clustering, the number of laundered
1063 * pages could exceed "target" by the maximum size of
1064 * a cluster minus one.
1066 target -= min(vm_pageout_launder(domain, launder,
1067 in_shortfall), target);
1068 pause("laundp", hz / VM_LAUNDER_RATE);
1072 * If we're not currently laundering pages and the page daemon
1073 * hasn't posted a new request, sleep until the page daemon
1076 vm_pagequeue_lock(pq);
1077 if (target == 0 && vm_laundry_request == VM_LAUNDRY_IDLE)
1078 (void)mtx_sleep(&vm_laundry_request,
1079 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1082 * If the pagedaemon has indicated that it's in shortfall, start
1083 * a shortfall laundering unless we're already in the middle of
1084 * one. This may preempt a background laundering.
1086 if (vm_laundry_request == VM_LAUNDRY_SHORTFALL &&
1087 (!in_shortfall || shortfall_cycle == 0)) {
1088 shortfall = vm_laundry_target() + vm_pageout_deficit;
1094 vm_laundry_request = VM_LAUNDRY_IDLE;
1095 vm_pagequeue_unlock(pq);
1100 * vm_pageout_scan does the dirty work for the pageout daemon.
1102 * pass == 0: Update active LRU/deactivate pages
1103 * pass >= 1: Free inactive pages
1105 * Returns true if pass was zero or enough pages were freed by the inactive
1106 * queue scan to meet the target.
1109 vm_pageout_scan(struct vm_domain *vmd, int pass)
1112 struct vm_pagequeue *pq;
1115 int act_delta, addl_page_shortage, deficit, inactq_shortage, maxscan;
1116 int page_shortage, scan_tick, scanned, starting_page_shortage;
1117 boolean_t queue_locked;
1120 * If we need to reclaim memory ask kernel caches to return
1121 * some. We rate limit to avoid thrashing.
1123 if (vmd == &vm_dom[0] && pass > 0 &&
1124 (time_uptime - lowmem_uptime) >= lowmem_period) {
1126 * Decrease registered cache sizes.
1128 SDT_PROBE0(vm, , , vm__lowmem_scan);
1129 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1131 * We do this explicitly after the caches have been
1135 lowmem_uptime = time_uptime;
1139 * The addl_page_shortage is the number of temporarily
1140 * stuck pages in the inactive queue. In other words, the
1141 * number of pages from the inactive count that should be
1142 * discounted in setting the target for the active queue scan.
1144 addl_page_shortage = 0;
1147 * Calculate the number of pages that we want to free. This number
1148 * can be negative if many pages are freed between the wakeup call to
1149 * the page daemon and this calculation.
1152 deficit = atomic_readandclear_int(&vm_pageout_deficit);
1153 page_shortage = vm_paging_target() + deficit;
1155 page_shortage = deficit = 0;
1156 starting_page_shortage = page_shortage;
1159 * Start scanning the inactive queue for pages that we can free. The
1160 * scan will stop when we reach the target or we have scanned the
1161 * entire queue. (Note that m->act_count is not used to make
1162 * decisions for the inactive queue, only for the active queue.)
1164 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1165 maxscan = pq->pq_cnt;
1166 vm_pagequeue_lock(pq);
1167 queue_locked = TRUE;
1168 for (m = TAILQ_FIRST(&pq->pq_pl);
1169 m != NULL && maxscan-- > 0 && page_shortage > 0;
1171 vm_pagequeue_assert_locked(pq);
1172 KASSERT(queue_locked, ("unlocked inactive queue"));
1173 KASSERT(vm_page_inactive(m), ("Inactive queue %p", m));
1175 VM_CNT_INC(v_pdpages);
1176 next = TAILQ_NEXT(m, plinks.q);
1181 if (m->flags & PG_MARKER)
1184 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1185 ("Fictitious page %p cannot be in inactive queue", m));
1186 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1187 ("Unmanaged page %p cannot be in inactive queue", m));
1190 * The page or object lock acquisitions fail if the
1191 * page was removed from the queue or moved to a
1192 * different position within the queue. In either
1193 * case, addl_page_shortage should not be incremented.
1195 if (!vm_pageout_page_lock(m, &next))
1197 else if (m->hold_count != 0) {
1199 * Held pages are essentially stuck in the
1200 * queue. So, they ought to be discounted
1201 * from the inactive count. See the
1202 * calculation of inactq_shortage before the
1203 * loop over the active queue below.
1205 addl_page_shortage++;
1209 if (!VM_OBJECT_TRYWLOCK(object)) {
1210 if (!vm_pageout_fallback_object_lock(m, &next))
1212 else if (m->hold_count != 0) {
1213 addl_page_shortage++;
1217 if (vm_page_busied(m)) {
1219 * Don't mess with busy pages. Leave them at
1220 * the front of the queue. Most likely, they
1221 * are being paged out and will leave the
1222 * queue shortly after the scan finishes. So,
1223 * they ought to be discounted from the
1226 addl_page_shortage++;
1228 VM_OBJECT_WUNLOCK(object);
1233 KASSERT(m->hold_count == 0, ("Held page %p", m));
1236 * Dequeue the inactive page and unlock the inactive page
1237 * queue, invalidating the 'next' pointer. Dequeueing the
1238 * page here avoids a later reacquisition (and release) of
1239 * the inactive page queue lock when vm_page_activate(),
1240 * vm_page_free(), or vm_page_launder() is called. Use a
1241 * marker to remember our place in the inactive queue.
1243 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1244 vm_page_dequeue_locked(m);
1245 vm_pagequeue_unlock(pq);
1246 queue_locked = FALSE;
1249 * Invalid pages can be easily freed. They cannot be
1250 * mapped, vm_page_free() asserts this.
1256 * If the page has been referenced and the object is not dead,
1257 * reactivate or requeue the page depending on whether the
1260 if ((m->aflags & PGA_REFERENCED) != 0) {
1261 vm_page_aflag_clear(m, PGA_REFERENCED);
1265 if (object->ref_count != 0) {
1266 act_delta += pmap_ts_referenced(m);
1268 KASSERT(!pmap_page_is_mapped(m),
1269 ("vm_pageout_scan: page %p is mapped", m));
1271 if (act_delta != 0) {
1272 if (object->ref_count != 0) {
1273 VM_CNT_INC(v_reactivated);
1274 vm_page_activate(m);
1277 * Increase the activation count if the page
1278 * was referenced while in the inactive queue.
1279 * This makes it less likely that the page will
1280 * be returned prematurely to the inactive
1283 m->act_count += act_delta + ACT_ADVANCE;
1285 } else if ((object->flags & OBJ_DEAD) == 0) {
1286 vm_pagequeue_lock(pq);
1287 queue_locked = TRUE;
1288 m->queue = PQ_INACTIVE;
1289 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
1290 vm_pagequeue_cnt_inc(pq);
1296 * If the page appears to be clean at the machine-independent
1297 * layer, then remove all of its mappings from the pmap in
1298 * anticipation of freeing it. If, however, any of the page's
1299 * mappings allow write access, then the page may still be
1300 * modified until the last of those mappings are removed.
1302 if (object->ref_count != 0) {
1303 vm_page_test_dirty(m);
1309 * Clean pages can be freed, but dirty pages must be sent back
1310 * to the laundry, unless they belong to a dead object.
1311 * Requeueing dirty pages from dead objects is pointless, as
1312 * they are being paged out and freed by the thread that
1313 * destroyed the object.
1315 if (m->dirty == 0) {
1318 VM_CNT_INC(v_dfree);
1320 } else if ((object->flags & OBJ_DEAD) == 0)
1324 VM_OBJECT_WUNLOCK(object);
1325 if (!queue_locked) {
1326 vm_pagequeue_lock(pq);
1327 queue_locked = TRUE;
1329 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1330 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1332 vm_pagequeue_unlock(pq);
1335 * Wake up the laundry thread so that it can perform any needed
1336 * laundering. If we didn't meet our target, we're in shortfall and
1337 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1338 * swap devices are configured, the laundry thread has no work to do, so
1339 * don't bother waking it up.
1341 if (vm_laundry_request == VM_LAUNDRY_IDLE &&
1342 starting_page_shortage > 0) {
1343 pq = &vm_dom[0].vmd_pagequeues[PQ_LAUNDRY];
1344 vm_pagequeue_lock(pq);
1345 if (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled)) {
1346 if (page_shortage > 0) {
1347 vm_laundry_request = VM_LAUNDRY_SHORTFALL;
1348 VM_CNT_INC(v_pdshortfalls);
1349 } else if (vm_laundry_request != VM_LAUNDRY_SHORTFALL)
1350 vm_laundry_request = VM_LAUNDRY_BACKGROUND;
1351 wakeup(&vm_laundry_request);
1353 vm_pagequeue_unlock(pq);
1357 * Wakeup the swapout daemon if we didn't free the targeted number of
1360 if (page_shortage > 0)
1364 * If the inactive queue scan fails repeatedly to meet its
1365 * target, kill the largest process.
1367 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1370 * Compute the number of pages we want to try to move from the
1371 * active queue to either the inactive or laundry queue.
1373 * When scanning active pages, we make clean pages count more heavily
1374 * towards the page shortage than dirty pages. This is because dirty
1375 * pages must be laundered before they can be reused and thus have less
1376 * utility when attempting to quickly alleviate a shortage. However,
1377 * this weighting also causes the scan to deactivate dirty pages more
1378 * more aggressively, improving the effectiveness of clustering and
1379 * ensuring that they can eventually be reused.
1381 inactq_shortage = vm_cnt.v_inactive_target - (vm_cnt.v_inactive_count +
1382 vm_cnt.v_laundry_count / act_scan_laundry_weight) +
1383 vm_paging_target() + deficit + addl_page_shortage;
1384 page_shortage *= act_scan_laundry_weight;
1386 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1387 vm_pagequeue_lock(pq);
1388 maxscan = pq->pq_cnt;
1391 * If we're just idle polling attempt to visit every
1392 * active page within 'update_period' seconds.
1395 if (vm_pageout_update_period != 0) {
1396 min_scan = pq->pq_cnt;
1397 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1398 min_scan /= hz * vm_pageout_update_period;
1401 if (min_scan > 0 || (inactq_shortage > 0 && maxscan > 0))
1402 vmd->vmd_last_active_scan = scan_tick;
1405 * Scan the active queue for pages that can be deactivated. Update
1406 * the per-page activity counter and use it to identify deactivation
1407 * candidates. Held pages may be deactivated.
1409 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1410 min_scan || (inactq_shortage > 0 && scanned < maxscan)); m = next,
1412 KASSERT(m->queue == PQ_ACTIVE,
1413 ("vm_pageout_scan: page %p isn't active", m));
1414 next = TAILQ_NEXT(m, plinks.q);
1415 if ((m->flags & PG_MARKER) != 0)
1417 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1418 ("Fictitious page %p cannot be in active queue", m));
1419 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1420 ("Unmanaged page %p cannot be in active queue", m));
1421 if (!vm_pageout_page_lock(m, &next)) {
1427 * The count for page daemon pages is updated after checking
1428 * the page for eligibility.
1430 VM_CNT_INC(v_pdpages);
1433 * Check to see "how much" the page has been used.
1435 if ((m->aflags & PGA_REFERENCED) != 0) {
1436 vm_page_aflag_clear(m, PGA_REFERENCED);
1442 * Perform an unsynchronized object ref count check. While
1443 * the page lock ensures that the page is not reallocated to
1444 * another object, in particular, one with unmanaged mappings
1445 * that cannot support pmap_ts_referenced(), two races are,
1446 * nonetheless, possible:
1447 * 1) The count was transitioning to zero, but we saw a non-
1448 * zero value. pmap_ts_referenced() will return zero
1449 * because the page is not mapped.
1450 * 2) The count was transitioning to one, but we saw zero.
1451 * This race delays the detection of a new reference. At
1452 * worst, we will deactivate and reactivate the page.
1454 if (m->object->ref_count != 0)
1455 act_delta += pmap_ts_referenced(m);
1458 * Advance or decay the act_count based on recent usage.
1460 if (act_delta != 0) {
1461 m->act_count += ACT_ADVANCE + act_delta;
1462 if (m->act_count > ACT_MAX)
1463 m->act_count = ACT_MAX;
1465 m->act_count -= min(m->act_count, ACT_DECLINE);
1468 * Move this page to the tail of the active, inactive or laundry
1469 * queue depending on usage.
1471 if (m->act_count == 0) {
1472 /* Dequeue to avoid later lock recursion. */
1473 vm_page_dequeue_locked(m);
1476 * When not short for inactive pages, let dirty pages go
1477 * through the inactive queue before moving to the
1478 * laundry queues. This gives them some extra time to
1479 * be reactivated, potentially avoiding an expensive
1480 * pageout. During a page shortage, the inactive queue
1481 * is necessarily small, so we may move dirty pages
1482 * directly to the laundry queue.
1484 if (inactq_shortage <= 0)
1485 vm_page_deactivate(m);
1488 * Calling vm_page_test_dirty() here would
1489 * require acquisition of the object's write
1490 * lock. However, during a page shortage,
1491 * directing dirty pages into the laundry
1492 * queue is only an optimization and not a
1493 * requirement. Therefore, we simply rely on
1494 * the opportunistic updates to the page's
1495 * dirty field by the pmap.
1497 if (m->dirty == 0) {
1498 vm_page_deactivate(m);
1500 act_scan_laundry_weight;
1507 vm_page_requeue_locked(m);
1510 vm_pagequeue_unlock(pq);
1512 vm_swapout_run_idle();
1513 return (page_shortage <= 0);
1516 static int vm_pageout_oom_vote;
1519 * The pagedaemon threads randlomly select one to perform the
1520 * OOM. Trying to kill processes before all pagedaemons
1521 * failed to reach free target is premature.
1524 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1525 int starting_page_shortage)
1529 if (starting_page_shortage <= 0 || starting_page_shortage !=
1531 vmd->vmd_oom_seq = 0;
1534 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1536 vmd->vmd_oom = FALSE;
1537 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1543 * Do not follow the call sequence until OOM condition is
1546 vmd->vmd_oom_seq = 0;
1551 vmd->vmd_oom = TRUE;
1552 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1553 if (old_vote != vm_ndomains - 1)
1557 * The current pagedaemon thread is the last in the quorum to
1558 * start OOM. Initiate the selection and signaling of the
1561 vm_pageout_oom(VM_OOM_MEM);
1564 * After one round of OOM terror, recall our vote. On the
1565 * next pass, current pagedaemon would vote again if the low
1566 * memory condition is still there, due to vmd_oom being
1569 vmd->vmd_oom = FALSE;
1570 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1574 * The OOM killer is the page daemon's action of last resort when
1575 * memory allocation requests have been stalled for a prolonged period
1576 * of time because it cannot reclaim memory. This function computes
1577 * the approximate number of physical pages that could be reclaimed if
1578 * the specified address space is destroyed.
1580 * Private, anonymous memory owned by the address space is the
1581 * principal resource that we expect to recover after an OOM kill.
1582 * Since the physical pages mapped by the address space's COW entries
1583 * are typically shared pages, they are unlikely to be released and so
1584 * they are not counted.
1586 * To get to the point where the page daemon runs the OOM killer, its
1587 * efforts to write-back vnode-backed pages may have stalled. This
1588 * could be caused by a memory allocation deadlock in the write path
1589 * that might be resolved by an OOM kill. Therefore, physical pages
1590 * belonging to vnode-backed objects are counted, because they might
1591 * be freed without being written out first if the address space holds
1592 * the last reference to an unlinked vnode.
1594 * Similarly, physical pages belonging to OBJT_PHYS objects are
1595 * counted because the address space might hold the last reference to
1599 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1602 vm_map_entry_t entry;
1606 map = &vmspace->vm_map;
1607 KASSERT(!map->system_map, ("system map"));
1608 sx_assert(&map->lock, SA_LOCKED);
1610 for (entry = map->header.next; entry != &map->header;
1611 entry = entry->next) {
1612 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1614 obj = entry->object.vm_object;
1617 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1618 obj->ref_count != 1)
1620 switch (obj->type) {
1625 res += obj->resident_page_count;
1633 vm_pageout_oom(int shortage)
1635 struct proc *p, *bigproc;
1636 vm_offset_t size, bigsize;
1642 * We keep the process bigproc locked once we find it to keep anyone
1643 * from messing with it; however, there is a possibility of
1644 * deadlock if process B is bigproc and one of its child processes
1645 * attempts to propagate a signal to B while we are waiting for A's
1646 * lock while walking this list. To avoid this, we don't block on
1647 * the process lock but just skip a process if it is already locked.
1651 sx_slock(&allproc_lock);
1652 FOREACH_PROC_IN_SYSTEM(p) {
1656 * If this is a system, protected or killed process, skip it.
1658 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1659 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1660 p->p_pid == 1 || P_KILLED(p) ||
1661 (p->p_pid < 48 && swap_pager_avail != 0)) {
1666 * If the process is in a non-running type state,
1667 * don't touch it. Check all the threads individually.
1670 FOREACH_THREAD_IN_PROC(p, td) {
1672 if (!TD_ON_RUNQ(td) &&
1673 !TD_IS_RUNNING(td) &&
1674 !TD_IS_SLEEPING(td) &&
1675 !TD_IS_SUSPENDED(td) &&
1676 !TD_IS_SWAPPED(td)) {
1688 * get the process size
1690 vm = vmspace_acquire_ref(p);
1697 sx_sunlock(&allproc_lock);
1698 if (!vm_map_trylock_read(&vm->vm_map)) {
1700 sx_slock(&allproc_lock);
1704 size = vmspace_swap_count(vm);
1705 if (shortage == VM_OOM_MEM)
1706 size += vm_pageout_oom_pagecount(vm);
1707 vm_map_unlock_read(&vm->vm_map);
1709 sx_slock(&allproc_lock);
1712 * If this process is bigger than the biggest one,
1715 if (size > bigsize) {
1716 if (bigproc != NULL)
1724 sx_sunlock(&allproc_lock);
1725 if (bigproc != NULL) {
1726 if (vm_panic_on_oom != 0)
1727 panic("out of swap space");
1729 killproc(bigproc, "out of swap space");
1730 sched_nice(bigproc, PRIO_MIN);
1732 PROC_UNLOCK(bigproc);
1733 wakeup(&vm_cnt.v_free_count);
1738 vm_pageout_worker(void *arg)
1740 struct vm_domain *domain;
1744 domidx = (uintptr_t)arg;
1745 domain = &vm_dom[domidx];
1750 * XXXKIB It could be useful to bind pageout daemon threads to
1751 * the cores belonging to the domain, from which vm_page_array
1755 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1756 domain->vmd_last_active_scan = ticks;
1757 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1758 vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
1759 TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1760 &domain->vmd_inacthead, plinks.q);
1763 * The pageout daemon worker is never done, so loop forever.
1766 mtx_lock(&vm_page_queue_free_mtx);
1769 * Generally, after a level >= 1 scan, if there are enough
1770 * free pages to wakeup the waiters, then they are already
1771 * awake. A call to vm_page_free() during the scan awakened
1772 * them. However, in the following case, this wakeup serves
1773 * to bound the amount of time that a thread might wait.
1774 * Suppose a thread's call to vm_page_alloc() fails, but
1775 * before that thread calls VM_WAIT, enough pages are freed by
1776 * other threads to alleviate the free page shortage. The
1777 * thread will, nonetheless, wait until another page is freed
1778 * or this wakeup is performed.
1780 if (vm_pages_needed && !vm_page_count_min()) {
1781 vm_pages_needed = false;
1782 wakeup(&vm_cnt.v_free_count);
1786 * Do not clear vm_pageout_wanted until we reach our free page
1787 * target. Otherwise, we may be awakened over and over again,
1790 if (vm_pageout_wanted && target_met)
1791 vm_pageout_wanted = false;
1794 * Might the page daemon receive a wakeup call?
1796 if (vm_pageout_wanted) {
1798 * No. Either vm_pageout_wanted was set by another
1799 * thread during the previous scan, which must have
1800 * been a level 0 scan, or vm_pageout_wanted was
1801 * already set and the scan failed to free enough
1802 * pages. If we haven't yet performed a level >= 1
1803 * (page reclamation) scan, then increase the level
1804 * and scan again now. Otherwise, sleep a bit and
1807 mtx_unlock(&vm_page_queue_free_mtx);
1809 pause("psleep", hz / VM_INACT_SCAN_RATE);
1813 * Yes. Sleep until pages need to be reclaimed or
1814 * have their reference stats updated.
1816 if (mtx_sleep(&vm_pageout_wanted,
1817 &vm_page_queue_free_mtx, PDROP | PVM, "psleep",
1819 VM_CNT_INC(v_pdwakeups);
1825 target_met = vm_pageout_scan(domain, pass);
1830 * vm_pageout_init initialises basic pageout daemon settings.
1833 vm_pageout_init(void)
1836 * Initialize some paging parameters.
1838 vm_cnt.v_interrupt_free_min = 2;
1839 if (vm_cnt.v_page_count < 2000)
1840 vm_pageout_page_count = 8;
1843 * v_free_reserved needs to include enough for the largest
1844 * swap pager structures plus enough for any pv_entry structs
1847 if (vm_cnt.v_page_count > 1024)
1848 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1850 vm_cnt.v_free_min = 4;
1851 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1852 vm_cnt.v_interrupt_free_min;
1853 vm_cnt.v_free_reserved = vm_pageout_page_count +
1854 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1855 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1856 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1857 vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1858 vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1859 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1860 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1861 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1864 * Set the default wakeup threshold to be 10% above the minimum
1865 * page limit. This keeps the steady state out of shortfall.
1867 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1870 * Set interval in seconds for active scan. We want to visit each
1871 * page at least once every ten minutes. This is to prevent worst
1872 * case paging behaviors with stale active LRU.
1874 if (vm_pageout_update_period == 0)
1875 vm_pageout_update_period = 600;
1877 /* XXX does not really belong here */
1878 if (vm_page_max_wired == 0)
1879 vm_page_max_wired = vm_cnt.v_free_count / 3;
1882 * Target amount of memory to move out of the laundry queue during a
1883 * background laundering. This is proportional to the amount of system
1886 vm_background_launder_target = (vm_cnt.v_free_target -
1887 vm_cnt.v_free_min) / 10;
1891 * vm_pageout is the high level pageout daemon.
1897 #ifdef VM_NUMA_ALLOC
1901 swap_pager_swap_init();
1902 error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
1903 0, 0, "laundry: dom0");
1905 panic("starting laundry for domain 0, error %d", error);
1906 #ifdef VM_NUMA_ALLOC
1907 for (i = 1; i < vm_ndomains; i++) {
1908 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1909 curproc, NULL, 0, 0, "dom%d", i);
1911 panic("starting pageout for domain %d, error %d\n",
1916 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1919 panic("starting uma_reclaim helper, error %d\n", error);
1920 vm_pageout_worker((void *)(uintptr_t)0);
1924 * Unless the free page queue lock is held by the caller, this function
1925 * should be regarded as advisory. Specifically, the caller should
1926 * not msleep() on &vm_cnt.v_free_count following this function unless
1927 * the free page queue lock is held until the msleep() is performed.
1930 pagedaemon_wakeup(void)
1933 if (!vm_pageout_wanted && curthread->td_proc != pageproc) {
1934 vm_pageout_wanted = true;
1935 wakeup(&vm_pageout_wanted);