2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
4 * Copyright (c) 1991 Regents of the University of California.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
10 * Copyright (c) 2005 Yahoo! Technologies Norway AS
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * The Mach Operating System project at Carnegie-Mellon University.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
44 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
47 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
48 * All rights reserved.
50 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
52 * Permission to use, copy, modify and distribute this software and
53 * its documentation is hereby granted, provided that both the copyright
54 * notice and this permission notice appear in all copies of the
55 * software, derivative works or modified versions, and any portions
56 * thereof, and that both notices appear in supporting documentation.
58 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
59 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
60 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
62 * Carnegie Mellon requests users of this software to return to
64 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
65 * School of Computer Science
66 * Carnegie Mellon University
67 * Pittsburgh PA 15213-3890
69 * any improvements or extensions that they make and grant Carnegie the
70 * rights to redistribute these changes.
74 * The proverbial page-out daemon.
77 #include <sys/cdefs.h>
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
83 #include <sys/blockcount.h>
84 #include <sys/eventhandler.h>
86 #include <sys/mutex.h>
88 #include <sys/kthread.h>
90 #include <sys/mount.h>
91 #include <sys/racct.h>
92 #include <sys/resourcevar.h>
93 #include <sys/sched.h>
95 #include <sys/signalvar.h>
98 #include <sys/vnode.h>
99 #include <sys/vmmeter.h>
100 #include <sys/rwlock.h>
102 #include <sys/sysctl.h>
105 #include <vm/vm_param.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_map.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_phys.h>
112 #include <vm/vm_pagequeue.h>
113 #include <vm/swap_pager.h>
114 #include <vm/vm_extern.h>
118 * System initialization
121 /* the kernel process "vm_pageout"*/
122 static void vm_pageout(void);
123 static void vm_pageout_init(void);
124 static int vm_pageout_clean(vm_page_t m, int *numpagedout);
125 static int vm_pageout_cluster(vm_page_t m);
126 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
127 int starting_page_shortage);
129 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
132 struct proc *pageproc;
134 static struct kproc_desc page_kp = {
139 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
142 SDT_PROVIDER_DEFINE(vm);
143 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
145 /* Pagedaemon activity rates, in subdivisions of one second. */
146 #define VM_LAUNDER_RATE 10
147 #define VM_INACT_SCAN_RATE 10
149 static int swapdev_enabled;
150 int vm_pageout_page_count = 32;
152 static int vm_panic_on_oom = 0;
153 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
154 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
155 "Panic on the given number of out-of-memory errors instead of "
156 "killing the largest process");
158 static int vm_pageout_update_period;
159 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
160 CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
161 "Maximum active LRU update period");
163 static int pageout_cpus_per_thread = 16;
164 SYSCTL_INT(_vm, OID_AUTO, pageout_cpus_per_thread, CTLFLAG_RDTUN,
165 &pageout_cpus_per_thread, 0,
166 "Number of CPUs per pagedaemon worker thread");
168 static int lowmem_period = 10;
169 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
170 "Low memory callback period");
172 static int disable_swap_pageouts;
173 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
174 CTLFLAG_RWTUN, &disable_swap_pageouts, 0,
175 "Disallow swapout of dirty pages");
177 static int pageout_lock_miss;
178 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
179 CTLFLAG_RD, &pageout_lock_miss, 0,
180 "vget() lock misses during pageout");
182 static int vm_pageout_oom_seq = 12;
183 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
184 CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
185 "back-to-back calls to oom detector to start OOM");
187 static int act_scan_laundry_weight = 3;
188 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
189 &act_scan_laundry_weight, 0,
190 "weight given to clean vs. dirty pages in active queue scans");
192 static u_int vm_background_launder_rate = 4096;
193 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
194 &vm_background_launder_rate, 0,
195 "background laundering rate, in kilobytes per second");
197 static u_int vm_background_launder_max = 20 * 1024;
198 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
199 &vm_background_launder_max, 0,
200 "background laundering cap, in kilobytes");
202 u_long vm_page_max_user_wired;
203 SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW,
204 &vm_page_max_user_wired, 0,
205 "system-wide limit to user-wired page count");
207 static u_int isqrt(u_int num);
208 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
210 static void vm_pageout_laundry_worker(void *arg);
213 struct vm_batchqueue bq;
214 struct vm_pagequeue *pq;
221 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
222 vm_page_t marker, vm_page_t after, int maxscan)
225 vm_pagequeue_assert_locked(pq);
226 KASSERT((marker->a.flags & PGA_ENQUEUED) == 0,
227 ("marker %p already enqueued", marker));
230 TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
232 TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
233 vm_page_aflag_set(marker, PGA_ENQUEUED);
235 vm_batchqueue_init(&ss->bq);
238 ss->maxscan = maxscan;
240 vm_pagequeue_unlock(pq);
244 vm_pageout_end_scan(struct scan_state *ss)
246 struct vm_pagequeue *pq;
249 vm_pagequeue_assert_locked(pq);
250 KASSERT((ss->marker->a.flags & PGA_ENQUEUED) != 0,
251 ("marker %p not enqueued", ss->marker));
253 TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
254 vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
255 pq->pq_pdpages += ss->scanned;
259 * Add a small number of queued pages to a batch queue for later processing
260 * without the corresponding queue lock held. The caller must have enqueued a
261 * marker page at the desired start point for the scan. Pages will be
262 * physically dequeued if the caller so requests. Otherwise, the returned
263 * batch may contain marker pages, and it is up to the caller to handle them.
265 * When processing the batch queue, vm_pageout_defer() must be used to
266 * determine whether the page has been logically dequeued since the batch was
269 static __always_inline void
270 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
272 struct vm_pagequeue *pq;
273 vm_page_t m, marker, n;
278 KASSERT((marker->a.flags & PGA_ENQUEUED) != 0,
279 ("marker %p not enqueued", ss->marker));
281 vm_pagequeue_lock(pq);
282 for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
283 ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
284 m = n, ss->scanned++) {
285 n = TAILQ_NEXT(m, plinks.q);
286 if ((m->flags & PG_MARKER) == 0) {
287 KASSERT((m->a.flags & PGA_ENQUEUED) != 0,
288 ("page %p not enqueued", m));
289 KASSERT((m->flags & PG_FICTITIOUS) == 0,
290 ("Fictitious page %p cannot be in page queue", m));
291 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
292 ("Unmanaged page %p cannot be in page queue", m));
296 (void)vm_batchqueue_insert(&ss->bq, m);
298 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
299 vm_page_aflag_clear(m, PGA_ENQUEUED);
302 TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
303 if (__predict_true(m != NULL))
304 TAILQ_INSERT_BEFORE(m, marker, plinks.q);
306 TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
308 vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
309 vm_pagequeue_unlock(pq);
313 * Return the next page to be scanned, or NULL if the scan is complete.
315 static __always_inline vm_page_t
316 vm_pageout_next(struct scan_state *ss, const bool dequeue)
319 if (ss->bq.bq_cnt == 0)
320 vm_pageout_collect_batch(ss, dequeue);
321 return (vm_batchqueue_pop(&ss->bq));
325 * Determine whether processing of a page should be deferred and ensure that any
326 * outstanding queue operations are processed.
328 static __always_inline bool
329 vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued)
333 as = vm_page_astate_load(m);
334 if (__predict_false(as.queue != queue ||
335 ((as.flags & PGA_ENQUEUED) != 0) != enqueued))
337 if ((as.flags & PGA_QUEUE_OP_MASK) != 0) {
338 vm_page_pqbatch_submit(m, queue);
345 * Scan for pages at adjacent offsets within the given page's object that are
346 * eligible for laundering, form a cluster of these pages and the given page,
347 * and launder that cluster.
350 vm_pageout_cluster(vm_page_t m)
353 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
355 int ib, is, page_base, pageout_count;
358 VM_OBJECT_ASSERT_WLOCKED(object);
361 vm_page_assert_xbusied(m);
363 mc[vm_pageout_page_count] = pb = ps = m;
365 page_base = vm_pageout_page_count;
370 * We can cluster only if the page is not clean, busy, or held, and
371 * the page is in the laundry queue.
373 * During heavy mmap/modification loads the pageout
374 * daemon can really fragment the underlying file
375 * due to flushing pages out of order and not trying to
376 * align the clusters (which leaves sporadic out-of-order
377 * holes). To solve this problem we do the reverse scan
378 * first and attempt to align our cluster, then do a
379 * forward scan if room remains.
382 while (ib != 0 && pageout_count < vm_pageout_page_count) {
387 if ((p = vm_page_prev(pb)) == NULL ||
388 vm_page_tryxbusy(p) == 0) {
392 if (vm_page_wired(p)) {
397 vm_page_test_dirty(p);
403 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
408 mc[--page_base] = pb = p;
413 * We are at an alignment boundary. Stop here, and switch
414 * directions. Do not clear ib.
416 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
419 while (pageout_count < vm_pageout_page_count &&
420 pindex + is < object->size) {
421 if ((p = vm_page_next(ps)) == NULL ||
422 vm_page_tryxbusy(p) == 0)
424 if (vm_page_wired(p)) {
428 vm_page_test_dirty(p);
433 if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
437 mc[page_base + pageout_count] = ps = p;
443 * If we exhausted our forward scan, continue with the reverse scan
444 * when possible, even past an alignment boundary. This catches
445 * boundary conditions.
447 if (ib != 0 && pageout_count < vm_pageout_page_count)
450 return (vm_pageout_flush(&mc[page_base], pageout_count,
451 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
455 * vm_pageout_flush() - launder the given pages
457 * The given pages are laundered. Note that we setup for the start of
458 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
459 * reference count all in here rather then in the parent. If we want
460 * the parent to do more sophisticated things we may have to change
463 * Returned runlen is the count of pages between mreq and first
464 * page after mreq with status VM_PAGER_AGAIN.
465 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
466 * for any page in runlen set.
469 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
472 vm_object_t object = mc[0]->object;
473 int pageout_status[count];
477 VM_OBJECT_ASSERT_WLOCKED(object);
480 * Initiate I/O. Mark the pages shared busy and verify that they're
481 * valid and read-only.
483 * We do not have to fixup the clean/dirty bits here... we can
484 * allow the pager to do it after the I/O completes.
486 * NOTE! mc[i]->dirty may be partial or fragmented due to an
487 * edge case with file fragments.
489 for (i = 0; i < count; i++) {
490 KASSERT(vm_page_all_valid(mc[i]),
491 ("vm_pageout_flush: partially invalid page %p index %d/%d",
493 KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
494 ("vm_pageout_flush: writeable page %p", mc[i]));
495 vm_page_busy_downgrade(mc[i]);
497 vm_object_pip_add(object, count);
499 vm_pager_put_pages(object, mc, count, flags, pageout_status);
501 runlen = count - mreq;
504 for (i = 0; i < count; i++) {
505 vm_page_t mt = mc[i];
507 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
508 !pmap_page_is_write_mapped(mt),
509 ("vm_pageout_flush: page %p is not write protected", mt));
510 switch (pageout_status[i]) {
513 * The page may have moved since laundering started, in
514 * which case it should be left alone.
516 if (vm_page_in_laundry(mt))
517 vm_page_deactivate_noreuse(mt);
524 * The page is outside the object's range. We pretend
525 * that the page out worked and clean the page, so the
526 * changes will be lost if the page is reclaimed by
530 if (vm_page_in_laundry(mt))
531 vm_page_deactivate_noreuse(mt);
536 * If the page couldn't be paged out to swap because the
537 * pager wasn't able to find space, place the page in
538 * the PQ_UNSWAPPABLE holding queue. This is an
539 * optimization that prevents the page daemon from
540 * wasting CPU cycles on pages that cannot be reclaimed
541 * because no swap device is configured.
543 * Otherwise, reactivate the page so that it doesn't
544 * clog the laundry and inactive queues. (We will try
545 * paging it out again later.)
547 if ((object->flags & OBJ_SWAP) != 0 &&
548 pageout_status[i] == VM_PAGER_FAIL) {
549 vm_page_unswappable(mt);
552 vm_page_activate(mt);
553 if (eio != NULL && i >= mreq && i - mreq < runlen)
557 if (i >= mreq && i - mreq < runlen)
563 * If the operation is still going, leave the page busy to
564 * block all other accesses. Also, leave the paging in
565 * progress indicator set so that we don't attempt an object
568 if (pageout_status[i] != VM_PAGER_PEND) {
569 vm_object_pip_wakeup(object);
575 return (numpagedout);
579 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
582 atomic_store_rel_int(&swapdev_enabled, 1);
586 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
589 if (swap_pager_nswapdev() == 1)
590 atomic_store_rel_int(&swapdev_enabled, 0);
594 * Attempt to acquire all of the necessary locks to launder a page and
595 * then call through the clustering layer to PUTPAGES. Wait a short
596 * time for a vnode lock.
598 * Requires the page and object lock on entry, releases both before return.
599 * Returns 0 on success and an errno otherwise.
602 vm_pageout_clean(vm_page_t m, int *numpagedout)
611 VM_OBJECT_ASSERT_WLOCKED(object);
617 * The object is already known NOT to be dead. It
618 * is possible for the vget() to block the whole
619 * pageout daemon, but the new low-memory handling
620 * code should prevent it.
622 * We can't wait forever for the vnode lock, we might
623 * deadlock due to a vn_read() getting stuck in
624 * vm_wait while holding this vnode. We skip the
625 * vnode if we can't get it in a reasonable amount
628 if (object->type == OBJT_VNODE) {
631 if (vp->v_type == VREG &&
632 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
638 ("vp %p with NULL v_mount", vp));
639 vm_object_reference_locked(object);
641 VM_OBJECT_WUNLOCK(object);
642 if (vget(vp, vn_lktype_write(NULL, vp) | LK_TIMELOCK) != 0) {
647 VM_OBJECT_WLOCK(object);
650 * Ensure that the object and vnode were not disassociated
651 * while locks were dropped.
653 if (vp->v_object != object) {
659 * While the object was unlocked, the page may have been:
660 * (1) moved to a different queue,
661 * (2) reallocated to a different object,
662 * (3) reallocated to a different offset, or
665 if (!vm_page_in_laundry(m) || m->object != object ||
666 m->pindex != pindex || m->dirty == 0) {
672 * The page may have been busied while the object lock was
675 if (vm_page_tryxbusy(m) == 0) {
682 * Remove all writeable mappings, failing if the page is wired.
684 if (!vm_page_try_remove_write(m)) {
691 * If a page is dirty, then it is either being washed
692 * (but not yet cleaned) or it is still in the
693 * laundry. If it is still in the laundry, then we
694 * start the cleaning operation.
696 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
700 VM_OBJECT_WUNLOCK(object);
706 vm_object_deallocate(object);
707 vn_finished_write(mp);
714 * Attempt to launder the specified number of pages.
716 * Returns the number of pages successfully laundered.
719 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
721 struct scan_state ss;
722 struct vm_pagequeue *pq;
725 vm_page_astate_t new, old;
726 int act_delta, error, numpagedout, queue, refs, starting_target;
731 starting_target = launder;
735 * Scan the laundry queues for pages eligible to be laundered. We stop
736 * once the target number of dirty pages have been laundered, or once
737 * we've reached the end of the queue. A single iteration of this loop
738 * may cause more than one page to be laundered because of clustering.
740 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
741 * swap devices are configured.
743 if (atomic_load_acq_int(&swapdev_enabled))
744 queue = PQ_UNSWAPPABLE;
749 marker = &vmd->vmd_markers[queue];
750 pq = &vmd->vmd_pagequeues[queue];
751 vm_pagequeue_lock(pq);
752 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
753 while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
754 if (__predict_false((m->flags & PG_MARKER) != 0))
758 * Don't touch a page that was removed from the queue after the
759 * page queue lock was released. Otherwise, ensure that any
760 * pending queue operations, such as dequeues for wired pages,
763 if (vm_pageout_defer(m, queue, true))
767 * Lock the page's object.
769 if (object == NULL || object != m->object) {
771 VM_OBJECT_WUNLOCK(object);
772 object = atomic_load_ptr(&m->object);
773 if (__predict_false(object == NULL))
774 /* The page is being freed by another thread. */
777 /* Depends on type-stability. */
778 VM_OBJECT_WLOCK(object);
779 if (__predict_false(m->object != object)) {
780 VM_OBJECT_WUNLOCK(object);
786 if (vm_page_tryxbusy(m) == 0)
790 * Check for wirings now that we hold the object lock and have
791 * exclusively busied the page. If the page is mapped, it may
792 * still be wired by pmap lookups. The call to
793 * vm_page_try_remove_all() below atomically checks for such
794 * wirings and removes mappings. If the page is unmapped, the
795 * wire count is guaranteed not to increase after this check.
797 if (__predict_false(vm_page_wired(m)))
801 * Invalid pages can be easily freed. They cannot be
802 * mapped; vm_page_free() asserts this.
804 if (vm_page_none_valid(m))
807 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
809 for (old = vm_page_astate_load(m);;) {
811 * Check to see if the page has been removed from the
812 * queue since the first such check. Leave it alone if
813 * so, discarding any references collected by
814 * pmap_ts_referenced().
816 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
821 if ((old.flags & PGA_REFERENCED) != 0) {
822 new.flags &= ~PGA_REFERENCED;
825 if (act_delta == 0) {
827 } else if (object->ref_count != 0) {
829 * Increase the activation count if the page was
830 * referenced while in the laundry queue. This
831 * makes it less likely that the page will be
832 * returned prematurely to the laundry queue.
834 new.act_count += ACT_ADVANCE +
836 if (new.act_count > ACT_MAX)
837 new.act_count = ACT_MAX;
839 new.flags &= ~PGA_QUEUE_OP_MASK;
840 new.flags |= PGA_REQUEUE;
841 new.queue = PQ_ACTIVE;
842 if (!vm_page_pqstate_commit(m, &old, new))
846 * If this was a background laundering, count
847 * activated pages towards our target. The
848 * purpose of background laundering is to ensure
849 * that pages are eventually cycled through the
850 * laundry queue, and an activation is a valid
855 VM_CNT_INC(v_reactivated);
857 } else if ((object->flags & OBJ_DEAD) == 0) {
858 new.flags |= PGA_REQUEUE;
859 if (!vm_page_pqstate_commit(m, &old, new))
867 * If the page appears to be clean at the machine-independent
868 * layer, then remove all of its mappings from the pmap in
869 * anticipation of freeing it. If, however, any of the page's
870 * mappings allow write access, then the page may still be
871 * modified until the last of those mappings are removed.
873 if (object->ref_count != 0) {
874 vm_page_test_dirty(m);
875 if (m->dirty == 0 && !vm_page_try_remove_all(m))
880 * Clean pages are freed, and dirty pages are paged out unless
881 * they belong to a dead object. Requeueing dirty pages from
882 * dead objects is pointless, as they are being paged out and
883 * freed by the thread that destroyed the object.
888 * Now we are guaranteed that no other threads are
889 * manipulating the page, check for a last-second
892 if (vm_pageout_defer(m, queue, true))
896 } else if ((object->flags & OBJ_DEAD) == 0) {
897 if ((object->flags & OBJ_SWAP) == 0 &&
898 object->type != OBJT_DEFAULT)
900 else if (disable_swap_pageouts)
910 * Form a cluster with adjacent, dirty pages from the
911 * same object, and page out that entire cluster.
913 * The adjacent, dirty pages must also be in the
914 * laundry. However, their mappings are not checked
915 * for new references. Consequently, a recently
916 * referenced page may be paged out. However, that
917 * page will not be prematurely reclaimed. After page
918 * out, the page will be placed in the inactive queue,
919 * where any new references will be detected and the
922 error = vm_pageout_clean(m, &numpagedout);
924 launder -= numpagedout;
925 ss.scanned += numpagedout;
926 } else if (error == EDEADLK) {
936 if (object != NULL) {
937 VM_OBJECT_WUNLOCK(object);
940 vm_pagequeue_lock(pq);
941 vm_pageout_end_scan(&ss);
942 vm_pagequeue_unlock(pq);
944 if (launder > 0 && queue == PQ_UNSWAPPABLE) {
950 * Wakeup the sync daemon if we skipped a vnode in a writeable object
951 * and we didn't launder enough pages.
953 if (vnodes_skipped > 0 && launder > 0)
954 (void)speedup_syncer();
956 return (starting_target - launder);
960 * Compute the integer square root.
965 u_int bit, root, tmp;
967 bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
982 * Perform the work of the laundry thread: periodically wake up and determine
983 * whether any pages need to be laundered. If so, determine the number of pages
984 * that need to be laundered, and launder them.
987 vm_pageout_laundry_worker(void *arg)
989 struct vm_domain *vmd;
990 struct vm_pagequeue *pq;
991 uint64_t nclean, ndirty, nfreed;
992 int domain, last_target, launder, shortfall, shortfall_cycle, target;
995 domain = (uintptr_t)arg;
996 vmd = VM_DOMAIN(domain);
997 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
998 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1001 in_shortfall = false;
1002 shortfall_cycle = 0;
1003 last_target = target = 0;
1007 * Calls to these handlers are serialized by the swap syscall lock.
1009 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
1010 EVENTHANDLER_PRI_ANY);
1011 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
1012 EVENTHANDLER_PRI_ANY);
1015 * The pageout laundry worker is never done, so loop forever.
1018 KASSERT(target >= 0, ("negative target %d", target));
1019 KASSERT(shortfall_cycle >= 0,
1020 ("negative cycle %d", shortfall_cycle));
1024 * First determine whether we need to launder pages to meet a
1025 * shortage of free pages.
1027 if (shortfall > 0) {
1028 in_shortfall = true;
1029 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1031 } else if (!in_shortfall)
1033 else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1035 * We recently entered shortfall and began laundering
1036 * pages. If we have completed that laundering run
1037 * (and we are no longer in shortfall) or we have met
1038 * our laundry target through other activity, then we
1039 * can stop laundering pages.
1041 in_shortfall = false;
1045 launder = target / shortfall_cycle--;
1049 * There's no immediate need to launder any pages; see if we
1050 * meet the conditions to perform background laundering:
1052 * 1. The ratio of dirty to clean inactive pages exceeds the
1053 * background laundering threshold, or
1054 * 2. we haven't yet reached the target of the current
1055 * background laundering run.
1057 * The background laundering threshold is not a constant.
1058 * Instead, it is a slowly growing function of the number of
1059 * clean pages freed by the page daemon since the last
1060 * background laundering. Thus, as the ratio of dirty to
1061 * clean inactive pages grows, the amount of memory pressure
1062 * required to trigger laundering decreases. We ensure
1063 * that the threshold is non-zero after an inactive queue
1064 * scan, even if that scan failed to free a single clean page.
1067 nclean = vmd->vmd_free_count +
1068 vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1069 ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1070 if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1071 vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1072 target = vmd->vmd_background_launder_target;
1076 * We have a non-zero background laundering target. If we've
1077 * laundered up to our maximum without observing a page daemon
1078 * request, just stop. This is a safety belt that ensures we
1079 * don't launder an excessive amount if memory pressure is low
1080 * and the ratio of dirty to clean pages is large. Otherwise,
1081 * proceed at the background laundering rate.
1086 last_target = target;
1087 } else if (last_target - target >=
1088 vm_background_launder_max * PAGE_SIZE / 1024) {
1091 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1092 launder /= VM_LAUNDER_RATE;
1093 if (launder > target)
1100 * Because of I/O clustering, the number of laundered
1101 * pages could exceed "target" by the maximum size of
1102 * a cluster minus one.
1104 target -= min(vm_pageout_launder(vmd, launder,
1105 in_shortfall), target);
1106 pause("laundp", hz / VM_LAUNDER_RATE);
1110 * If we're not currently laundering pages and the page daemon
1111 * hasn't posted a new request, sleep until the page daemon
1114 vm_pagequeue_lock(pq);
1115 if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1116 (void)mtx_sleep(&vmd->vmd_laundry_request,
1117 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1120 * If the pagedaemon has indicated that it's in shortfall, start
1121 * a shortfall laundering unless we're already in the middle of
1122 * one. This may preempt a background laundering.
1124 if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1125 (!in_shortfall || shortfall_cycle == 0)) {
1126 shortfall = vm_laundry_target(vmd) +
1127 vmd->vmd_pageout_deficit;
1133 vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1134 nfreed += vmd->vmd_clean_pages_freed;
1135 vmd->vmd_clean_pages_freed = 0;
1136 vm_pagequeue_unlock(pq);
1141 * Compute the number of pages we want to try to move from the
1142 * active queue to either the inactive or laundry queue.
1144 * When scanning active pages during a shortage, we make clean pages
1145 * count more heavily towards the page shortage than dirty pages.
1146 * This is because dirty pages must be laundered before they can be
1147 * reused and thus have less utility when attempting to quickly
1148 * alleviate a free page shortage. However, this weighting also
1149 * causes the scan to deactivate dirty pages more aggressively,
1150 * improving the effectiveness of clustering.
1153 vm_pageout_active_target(struct vm_domain *vmd)
1157 shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1158 (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1159 vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1160 shortage *= act_scan_laundry_weight;
1165 * Scan the active queue. If there is no shortage of inactive pages, scan a
1166 * small portion of the queue in order to maintain quasi-LRU.
1169 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1171 struct scan_state ss;
1173 vm_page_t m, marker;
1174 struct vm_pagequeue *pq;
1175 vm_page_astate_t old, new;
1177 int act_delta, max_scan, ps_delta, refs, scan_tick;
1180 marker = &vmd->vmd_markers[PQ_ACTIVE];
1181 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1182 vm_pagequeue_lock(pq);
1185 * If we're just idle polling attempt to visit every
1186 * active page within 'update_period' seconds.
1189 if (vm_pageout_update_period != 0) {
1190 min_scan = pq->pq_cnt;
1191 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1192 min_scan /= hz * vm_pageout_update_period;
1195 if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1196 vmd->vmd_last_active_scan = scan_tick;
1199 * Scan the active queue for pages that can be deactivated. Update
1200 * the per-page activity counter and use it to identify deactivation
1201 * candidates. Held pages may be deactivated.
1203 * To avoid requeuing each page that remains in the active queue, we
1204 * implement the CLOCK algorithm. To keep the implementation of the
1205 * enqueue operation consistent for all page queues, we use two hands,
1206 * represented by marker pages. Scans begin at the first hand, which
1207 * precedes the second hand in the queue. When the two hands meet,
1208 * they are moved back to the head and tail of the queue, respectively,
1209 * and scanning resumes.
1211 max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1213 vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1214 while ((m = vm_pageout_next(&ss, false)) != NULL) {
1215 if (__predict_false(m == &vmd->vmd_clock[1])) {
1216 vm_pagequeue_lock(pq);
1217 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1218 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1219 TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1221 TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1223 max_scan -= ss.scanned;
1224 vm_pageout_end_scan(&ss);
1227 if (__predict_false((m->flags & PG_MARKER) != 0))
1231 * Don't touch a page that was removed from the queue after the
1232 * page queue lock was released. Otherwise, ensure that any
1233 * pending queue operations, such as dequeues for wired pages,
1236 if (vm_pageout_defer(m, PQ_ACTIVE, true))
1240 * A page's object pointer may be set to NULL before
1241 * the object lock is acquired.
1243 object = atomic_load_ptr(&m->object);
1244 if (__predict_false(object == NULL))
1246 * The page has been removed from its object.
1250 /* Deferred free of swap space. */
1251 if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
1252 VM_OBJECT_TRYWLOCK(object)) {
1253 if (m->object == object)
1254 vm_pager_page_unswapped(m);
1255 VM_OBJECT_WUNLOCK(object);
1259 * Check to see "how much" the page has been used.
1261 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1262 * that a reference from a concurrently destroyed mapping is
1263 * observed here and now.
1265 * Perform an unsynchronized object ref count check. While
1266 * the page lock ensures that the page is not reallocated to
1267 * another object, in particular, one with unmanaged mappings
1268 * that cannot support pmap_ts_referenced(), two races are,
1269 * nonetheless, possible:
1270 * 1) The count was transitioning to zero, but we saw a non-
1271 * zero value. pmap_ts_referenced() will return zero
1272 * because the page is not mapped.
1273 * 2) The count was transitioning to one, but we saw zero.
1274 * This race delays the detection of a new reference. At
1275 * worst, we will deactivate and reactivate the page.
1277 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1279 old = vm_page_astate_load(m);
1282 * Check to see if the page has been removed from the
1283 * queue since the first such check. Leave it alone if
1284 * so, discarding any references collected by
1285 * pmap_ts_referenced().
1287 if (__predict_false(_vm_page_queue(old) == PQ_NONE)) {
1293 * Advance or decay the act_count based on recent usage.
1297 if ((old.flags & PGA_REFERENCED) != 0) {
1298 new.flags &= ~PGA_REFERENCED;
1301 if (act_delta != 0) {
1302 new.act_count += ACT_ADVANCE + act_delta;
1303 if (new.act_count > ACT_MAX)
1304 new.act_count = ACT_MAX;
1306 new.act_count -= min(new.act_count,
1310 if (new.act_count > 0) {
1312 * Adjust the activation count and keep the page
1313 * in the active queue. The count might be left
1314 * unchanged if it is saturated. The page may
1315 * have been moved to a different queue since we
1316 * started the scan, in which case we move it
1320 if (old.queue != PQ_ACTIVE) {
1321 new.flags &= ~PGA_QUEUE_OP_MASK;
1322 new.flags |= PGA_REQUEUE;
1323 new.queue = PQ_ACTIVE;
1327 * When not short for inactive pages, let dirty
1328 * pages go through the inactive queue before
1329 * moving to the laundry queue. This gives them
1330 * some extra time to be reactivated,
1331 * potentially avoiding an expensive pageout.
1332 * However, during a page shortage, the inactive
1333 * queue is necessarily small, and so dirty
1334 * pages would only spend a trivial amount of
1335 * time in the inactive queue. Therefore, we
1336 * might as well place them directly in the
1337 * laundry queue to reduce queuing overhead.
1339 * Calling vm_page_test_dirty() here would
1340 * require acquisition of the object's write
1341 * lock. However, during a page shortage,
1342 * directing dirty pages into the laundry queue
1343 * is only an optimization and not a
1344 * requirement. Therefore, we simply rely on
1345 * the opportunistic updates to the page's dirty
1346 * field by the pmap.
1348 if (page_shortage <= 0) {
1349 nqueue = PQ_INACTIVE;
1351 } else if (m->dirty == 0) {
1352 nqueue = PQ_INACTIVE;
1353 ps_delta = act_scan_laundry_weight;
1355 nqueue = PQ_LAUNDRY;
1359 new.flags &= ~PGA_QUEUE_OP_MASK;
1360 new.flags |= PGA_REQUEUE;
1363 } while (!vm_page_pqstate_commit(m, &old, new));
1365 page_shortage -= ps_delta;
1367 vm_pagequeue_lock(pq);
1368 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1369 TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1370 vm_pageout_end_scan(&ss);
1371 vm_pagequeue_unlock(pq);
1375 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker,
1378 vm_page_astate_t as;
1380 vm_pagequeue_assert_locked(pq);
1382 as = vm_page_astate_load(m);
1383 if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
1385 vm_page_aflag_set(m, PGA_ENQUEUED);
1386 TAILQ_INSERT_BEFORE(marker, m, plinks.q);
1391 * Re-add stuck pages to the inactive queue. We will examine them again
1392 * during the next scan. If the queue state of a page has changed since
1393 * it was physically removed from the page queue in
1394 * vm_pageout_collect_batch(), don't do anything with that page.
1397 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1400 struct vm_pagequeue *pq;
1405 marker = ss->marker;
1409 if (vm_batchqueue_insert(bq, m))
1411 vm_pagequeue_lock(pq);
1412 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1414 vm_pagequeue_lock(pq);
1415 while ((m = vm_batchqueue_pop(bq)) != NULL)
1416 delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1417 vm_pagequeue_cnt_add(pq, delta);
1418 vm_pagequeue_unlock(pq);
1419 vm_batchqueue_init(bq);
1423 vm_pageout_scan_inactive(struct vm_domain *vmd, int page_shortage)
1425 struct timeval start, end;
1426 struct scan_state ss;
1427 struct vm_batchqueue rq;
1428 struct vm_page marker_page;
1429 vm_page_t m, marker;
1430 struct vm_pagequeue *pq;
1432 vm_page_astate_t old, new;
1433 int act_delta, addl_page_shortage, starting_page_shortage, refs;
1436 vm_batchqueue_init(&rq);
1437 getmicrouptime(&start);
1440 * The addl_page_shortage is an estimate of the number of temporarily
1441 * stuck pages in the inactive queue. In other words, the
1442 * number of pages from the inactive count that should be
1443 * discounted in setting the target for the active queue scan.
1445 addl_page_shortage = 0;
1448 * Start scanning the inactive queue for pages that we can free. The
1449 * scan will stop when we reach the target or we have scanned the
1450 * entire queue. (Note that m->a.act_count is not used to make
1451 * decisions for the inactive queue, only for the active queue.)
1453 starting_page_shortage = page_shortage;
1454 marker = &marker_page;
1455 vm_page_init_marker(marker, PQ_INACTIVE, 0);
1456 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1457 vm_pagequeue_lock(pq);
1458 vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1459 while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1460 KASSERT((m->flags & PG_MARKER) == 0,
1461 ("marker page %p was dequeued", m));
1464 * Don't touch a page that was removed from the queue after the
1465 * page queue lock was released. Otherwise, ensure that any
1466 * pending queue operations, such as dequeues for wired pages,
1469 if (vm_pageout_defer(m, PQ_INACTIVE, false))
1473 * Lock the page's object.
1475 if (object == NULL || object != m->object) {
1477 VM_OBJECT_WUNLOCK(object);
1478 object = atomic_load_ptr(&m->object);
1479 if (__predict_false(object == NULL))
1480 /* The page is being freed by another thread. */
1483 /* Depends on type-stability. */
1484 VM_OBJECT_WLOCK(object);
1485 if (__predict_false(m->object != object)) {
1486 VM_OBJECT_WUNLOCK(object);
1492 if (vm_page_tryxbusy(m) == 0) {
1494 * Don't mess with busy pages. Leave them at
1495 * the front of the queue. Most likely, they
1496 * are being paged out and will leave the
1497 * queue shortly after the scan finishes. So,
1498 * they ought to be discounted from the
1501 addl_page_shortage++;
1505 /* Deferred free of swap space. */
1506 if ((m->a.flags & PGA_SWAP_FREE) != 0)
1507 vm_pager_page_unswapped(m);
1510 * Check for wirings now that we hold the object lock and have
1511 * exclusively busied the page. If the page is mapped, it may
1512 * still be wired by pmap lookups. The call to
1513 * vm_page_try_remove_all() below atomically checks for such
1514 * wirings and removes mappings. If the page is unmapped, the
1515 * wire count is guaranteed not to increase after this check.
1517 if (__predict_false(vm_page_wired(m)))
1521 * Invalid pages can be easily freed. They cannot be
1522 * mapped, vm_page_free() asserts this.
1524 if (vm_page_none_valid(m))
1527 refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1529 for (old = vm_page_astate_load(m);;) {
1531 * Check to see if the page has been removed from the
1532 * queue since the first such check. Leave it alone if
1533 * so, discarding any references collected by
1534 * pmap_ts_referenced().
1536 if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1541 if ((old.flags & PGA_REFERENCED) != 0) {
1542 new.flags &= ~PGA_REFERENCED;
1545 if (act_delta == 0) {
1547 } else if (object->ref_count != 0) {
1549 * Increase the activation count if the
1550 * page was referenced while in the
1551 * inactive queue. This makes it less
1552 * likely that the page will be returned
1553 * prematurely to the inactive queue.
1555 new.act_count += ACT_ADVANCE +
1557 if (new.act_count > ACT_MAX)
1558 new.act_count = ACT_MAX;
1560 new.flags &= ~PGA_QUEUE_OP_MASK;
1561 new.flags |= PGA_REQUEUE;
1562 new.queue = PQ_ACTIVE;
1563 if (!vm_page_pqstate_commit(m, &old, new))
1566 VM_CNT_INC(v_reactivated);
1568 } else if ((object->flags & OBJ_DEAD) == 0) {
1569 new.queue = PQ_INACTIVE;
1570 new.flags |= PGA_REQUEUE;
1571 if (!vm_page_pqstate_commit(m, &old, new))
1579 * If the page appears to be clean at the machine-independent
1580 * layer, then remove all of its mappings from the pmap in
1581 * anticipation of freeing it. If, however, any of the page's
1582 * mappings allow write access, then the page may still be
1583 * modified until the last of those mappings are removed.
1585 if (object->ref_count != 0) {
1586 vm_page_test_dirty(m);
1587 if (m->dirty == 0 && !vm_page_try_remove_all(m))
1592 * Clean pages can be freed, but dirty pages must be sent back
1593 * to the laundry, unless they belong to a dead object.
1594 * Requeueing dirty pages from dead objects is pointless, as
1595 * they are being paged out and freed by the thread that
1596 * destroyed the object.
1598 if (m->dirty == 0) {
1601 * Now we are guaranteed that no other threads are
1602 * manipulating the page, check for a last-second
1603 * reference that would save it from doom.
1605 if (vm_pageout_defer(m, PQ_INACTIVE, false))
1609 * Because we dequeued the page and have already checked
1610 * for pending dequeue and enqueue requests, we can
1611 * safely disassociate the page from the inactive queue
1612 * without holding the queue lock.
1614 m->a.queue = PQ_NONE;
1619 if ((object->flags & OBJ_DEAD) == 0)
1625 vm_pageout_reinsert_inactive(&ss, &rq, m);
1628 VM_OBJECT_WUNLOCK(object);
1629 vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1630 vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1631 vm_pagequeue_lock(pq);
1632 vm_pageout_end_scan(&ss);
1633 vm_pagequeue_unlock(pq);
1636 * Record the remaining shortage and the progress and rate it was made.
1638 atomic_add_int(&vmd->vmd_addl_shortage, addl_page_shortage);
1639 getmicrouptime(&end);
1640 timevalsub(&end, &start);
1641 atomic_add_int(&vmd->vmd_inactive_us,
1642 end.tv_sec * 1000000 + end.tv_usec);
1643 atomic_add_int(&vmd->vmd_inactive_freed,
1644 starting_page_shortage - page_shortage);
1648 * Dispatch a number of inactive threads according to load and collect the
1649 * results to present a coherent view of paging activity on this domain.
1652 vm_pageout_inactive_dispatch(struct vm_domain *vmd, int shortage)
1654 u_int freed, pps, slop, threads, us;
1656 vmd->vmd_inactive_shortage = shortage;
1660 * If we have more work than we can do in a quarter of our interval, we
1661 * fire off multiple threads to process it.
1663 threads = vmd->vmd_inactive_threads;
1664 if (threads > 1 && vmd->vmd_inactive_pps != 0 &&
1665 shortage > vmd->vmd_inactive_pps / VM_INACT_SCAN_RATE / 4) {
1666 vmd->vmd_inactive_shortage /= threads;
1667 slop = shortage % threads;
1668 vm_domain_pageout_lock(vmd);
1669 blockcount_acquire(&vmd->vmd_inactive_starting, threads - 1);
1670 blockcount_acquire(&vmd->vmd_inactive_running, threads - 1);
1671 wakeup(&vmd->vmd_inactive_shortage);
1672 vm_domain_pageout_unlock(vmd);
1675 /* Run the local thread scan. */
1676 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage + slop);
1679 * Block until helper threads report results and then accumulate
1682 blockcount_wait(&vmd->vmd_inactive_running, NULL, "vmpoid", PVM);
1683 freed = atomic_readandclear_int(&vmd->vmd_inactive_freed);
1684 VM_CNT_ADD(v_dfree, freed);
1687 * Calculate the per-thread paging rate with an exponential decay of
1688 * prior results. Careful to avoid integer rounding errors with large
1691 us = max(atomic_readandclear_int(&vmd->vmd_inactive_us), 1);
1693 /* Keep rounding to tenths */
1694 pps = (freed * 10) / ((us * 10) / 1000000);
1696 pps = (1000000 / us) * freed;
1697 vmd->vmd_inactive_pps = (vmd->vmd_inactive_pps / 2) + (pps / 2);
1699 return (shortage - freed);
1703 * Attempt to reclaim the requested number of pages from the inactive queue.
1704 * Returns true if the shortage was addressed.
1707 vm_pageout_inactive(struct vm_domain *vmd, int shortage, int *addl_shortage)
1709 struct vm_pagequeue *pq;
1710 u_int addl_page_shortage, deficit, page_shortage;
1711 u_int starting_page_shortage;
1714 * vmd_pageout_deficit counts the number of pages requested in
1715 * allocations that failed because of a free page shortage. We assume
1716 * that the allocations will be reattempted and thus include the deficit
1717 * in our scan target.
1719 deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1720 starting_page_shortage = shortage + deficit;
1723 * Run the inactive scan on as many threads as is necessary.
1725 page_shortage = vm_pageout_inactive_dispatch(vmd, starting_page_shortage);
1726 addl_page_shortage = atomic_readandclear_int(&vmd->vmd_addl_shortage);
1729 * Wake up the laundry thread so that it can perform any needed
1730 * laundering. If we didn't meet our target, we're in shortfall and
1731 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1732 * swap devices are configured, the laundry thread has no work to do, so
1733 * don't bother waking it up.
1735 * The laundry thread uses the number of inactive queue scans elapsed
1736 * since the last laundering to determine whether to launder again, so
1739 if (starting_page_shortage > 0) {
1740 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1741 vm_pagequeue_lock(pq);
1742 if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1743 (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1744 if (page_shortage > 0) {
1745 vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1746 VM_CNT_INC(v_pdshortfalls);
1747 } else if (vmd->vmd_laundry_request !=
1748 VM_LAUNDRY_SHORTFALL)
1749 vmd->vmd_laundry_request =
1750 VM_LAUNDRY_BACKGROUND;
1751 wakeup(&vmd->vmd_laundry_request);
1753 vmd->vmd_clean_pages_freed +=
1754 starting_page_shortage - page_shortage;
1755 vm_pagequeue_unlock(pq);
1759 * Wakeup the swapout daemon if we didn't free the targeted number of
1762 if (page_shortage > 0)
1766 * If the inactive queue scan fails repeatedly to meet its
1767 * target, kill the largest process.
1769 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1772 * Reclaim pages by swapping out idle processes, if configured to do so.
1774 vm_swapout_run_idle();
1777 * See the description of addl_page_shortage above.
1779 *addl_shortage = addl_page_shortage + deficit;
1781 return (page_shortage <= 0);
1784 static int vm_pageout_oom_vote;
1787 * The pagedaemon threads randlomly select one to perform the
1788 * OOM. Trying to kill processes before all pagedaemons
1789 * failed to reach free target is premature.
1792 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1793 int starting_page_shortage)
1797 if (starting_page_shortage <= 0 || starting_page_shortage !=
1799 vmd->vmd_oom_seq = 0;
1802 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1804 vmd->vmd_oom = FALSE;
1805 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1811 * Do not follow the call sequence until OOM condition is
1814 vmd->vmd_oom_seq = 0;
1819 vmd->vmd_oom = TRUE;
1820 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1821 if (old_vote != vm_ndomains - 1)
1825 * The current pagedaemon thread is the last in the quorum to
1826 * start OOM. Initiate the selection and signaling of the
1829 vm_pageout_oom(VM_OOM_MEM);
1832 * After one round of OOM terror, recall our vote. On the
1833 * next pass, current pagedaemon would vote again if the low
1834 * memory condition is still there, due to vmd_oom being
1837 vmd->vmd_oom = FALSE;
1838 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1842 * The OOM killer is the page daemon's action of last resort when
1843 * memory allocation requests have been stalled for a prolonged period
1844 * of time because it cannot reclaim memory. This function computes
1845 * the approximate number of physical pages that could be reclaimed if
1846 * the specified address space is destroyed.
1848 * Private, anonymous memory owned by the address space is the
1849 * principal resource that we expect to recover after an OOM kill.
1850 * Since the physical pages mapped by the address space's COW entries
1851 * are typically shared pages, they are unlikely to be released and so
1852 * they are not counted.
1854 * To get to the point where the page daemon runs the OOM killer, its
1855 * efforts to write-back vnode-backed pages may have stalled. This
1856 * could be caused by a memory allocation deadlock in the write path
1857 * that might be resolved by an OOM kill. Therefore, physical pages
1858 * belonging to vnode-backed objects are counted, because they might
1859 * be freed without being written out first if the address space holds
1860 * the last reference to an unlinked vnode.
1862 * Similarly, physical pages belonging to OBJT_PHYS objects are
1863 * counted because the address space might hold the last reference to
1867 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1870 vm_map_entry_t entry;
1874 map = &vmspace->vm_map;
1875 KASSERT(!map->system_map, ("system map"));
1876 sx_assert(&map->lock, SA_LOCKED);
1878 VM_MAP_ENTRY_FOREACH(entry, map) {
1879 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1881 obj = entry->object.vm_object;
1884 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1885 obj->ref_count != 1)
1887 if (obj->type == OBJT_DEFAULT || obj->type == OBJT_SWAP ||
1888 obj->type == OBJT_PHYS || obj->type == OBJT_VNODE ||
1889 (obj->flags & OBJ_SWAP) != 0)
1890 res += obj->resident_page_count;
1895 static int vm_oom_ratelim_last;
1896 static int vm_oom_pf_secs = 10;
1897 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1899 static struct mtx vm_oom_ratelim_mtx;
1902 vm_pageout_oom(int shortage)
1905 struct proc *p, *bigproc;
1906 vm_offset_t size, bigsize;
1913 * For OOM requests originating from vm_fault(), there is a high
1914 * chance that a single large process faults simultaneously in
1915 * several threads. Also, on an active system running many
1916 * processes of middle-size, like buildworld, all of them
1917 * could fault almost simultaneously as well.
1919 * To avoid killing too many processes, rate-limit OOMs
1920 * initiated by vm_fault() time-outs on the waits for free
1923 mtx_lock(&vm_oom_ratelim_mtx);
1925 if (shortage == VM_OOM_MEM_PF &&
1926 (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1927 mtx_unlock(&vm_oom_ratelim_mtx);
1930 vm_oom_ratelim_last = now;
1931 mtx_unlock(&vm_oom_ratelim_mtx);
1934 * We keep the process bigproc locked once we find it to keep anyone
1935 * from messing with it; however, there is a possibility of
1936 * deadlock if process B is bigproc and one of its child processes
1937 * attempts to propagate a signal to B while we are waiting for A's
1938 * lock while walking this list. To avoid this, we don't block on
1939 * the process lock but just skip a process if it is already locked.
1943 sx_slock(&allproc_lock);
1944 FOREACH_PROC_IN_SYSTEM(p) {
1948 * If this is a system, protected or killed process, skip it.
1950 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1951 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1952 p->p_pid == 1 || P_KILLED(p) ||
1953 (p->p_pid < 48 && swap_pager_avail != 0)) {
1958 * If the process is in a non-running type state,
1959 * don't touch it. Check all the threads individually.
1962 FOREACH_THREAD_IN_PROC(p, td) {
1964 if (!TD_ON_RUNQ(td) &&
1965 !TD_IS_RUNNING(td) &&
1966 !TD_IS_SLEEPING(td) &&
1967 !TD_IS_SUSPENDED(td) &&
1968 !TD_IS_SWAPPED(td)) {
1980 * get the process size
1982 vm = vmspace_acquire_ref(p);
1989 sx_sunlock(&allproc_lock);
1990 if (!vm_map_trylock_read(&vm->vm_map)) {
1992 sx_slock(&allproc_lock);
1996 size = vmspace_swap_count(vm);
1997 if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
1998 size += vm_pageout_oom_pagecount(vm);
1999 vm_map_unlock_read(&vm->vm_map);
2001 sx_slock(&allproc_lock);
2004 * If this process is bigger than the biggest one,
2007 if (size > bigsize) {
2008 if (bigproc != NULL)
2016 sx_sunlock(&allproc_lock);
2018 if (bigproc != NULL) {
2021 reason = "failed to reclaim memory";
2024 reason = "a thread waited too long to allocate a page";
2027 reason = "out of swap space";
2030 panic("unknown OOM reason %d", shortage);
2032 if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0)
2033 panic("%s", reason);
2035 killproc(bigproc, reason);
2036 sched_nice(bigproc, PRIO_MIN);
2038 PROC_UNLOCK(bigproc);
2043 * Signal a free page shortage to subsystems that have registered an event
2044 * handler. Reclaim memory from UMA in the event of a severe shortage.
2045 * Return true if the free page count should be re-evaluated.
2048 vm_pageout_lowmem(void)
2050 static int lowmem_ticks = 0;
2056 last = atomic_load_int(&lowmem_ticks);
2057 while ((u_int)(ticks - last) / hz >= lowmem_period) {
2058 if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
2062 * Decrease registered cache sizes.
2064 SDT_PROBE0(vm, , , vm__lowmem_scan);
2065 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
2068 * We do this explicitly after the caches have been
2071 uma_reclaim(UMA_RECLAIM_TRIM);
2077 * Kick off an asynchronous reclaim of cached memory if one of the
2078 * page daemons is failing to keep up with demand. Use the "severe"
2079 * threshold instead of "min" to ensure that we do not blow away the
2080 * caches if a subset of the NUMA domains are depleted by kernel memory
2081 * allocations; the domainset iterators automatically skip domains
2082 * below the "min" threshold on the first pass.
2084 * UMA reclaim worker has its own rate-limiting mechanism, so don't
2085 * worry about kicking it too often.
2087 if (vm_page_count_severe())
2088 uma_reclaim_wakeup();
2094 vm_pageout_worker(void *arg)
2096 struct vm_domain *vmd;
2098 int addl_shortage, domain, shortage;
2101 domain = (uintptr_t)arg;
2102 vmd = VM_DOMAIN(domain);
2107 * XXXKIB It could be useful to bind pageout daemon threads to
2108 * the cores belonging to the domain, from which vm_page_array
2112 KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
2113 vmd->vmd_last_active_scan = ticks;
2116 * The pageout daemon worker is never done, so loop forever.
2119 vm_domain_pageout_lock(vmd);
2122 * We need to clear wanted before we check the limits. This
2123 * prevents races with wakers who will check wanted after they
2126 atomic_store_int(&vmd->vmd_pageout_wanted, 0);
2129 * Might the page daemon need to run again?
2131 if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
2133 * Yes. If the scan failed to produce enough free
2134 * pages, sleep uninterruptibly for some time in the
2135 * hope that the laundry thread will clean some pages.
2137 vm_domain_pageout_unlock(vmd);
2139 pause("pwait", hz / VM_INACT_SCAN_RATE);
2142 * No, sleep until the next wakeup or until pages
2143 * need to have their reference stats updated.
2145 if (mtx_sleep(&vmd->vmd_pageout_wanted,
2146 vm_domain_pageout_lockptr(vmd), PDROP | PVM,
2147 "psleep", hz / VM_INACT_SCAN_RATE) == 0)
2148 VM_CNT_INC(v_pdwakeups);
2151 /* Prevent spurious wakeups by ensuring that wanted is set. */
2152 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2155 * Use the controller to calculate how many pages to free in
2156 * this interval, and scan the inactive queue. If the lowmem
2157 * handlers appear to have freed up some pages, subtract the
2158 * difference from the inactive queue scan target.
2160 shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
2162 ofree = vmd->vmd_free_count;
2163 if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
2164 shortage -= min(vmd->vmd_free_count - ofree,
2166 target_met = vm_pageout_inactive(vmd, shortage,
2172 * Scan the active queue. A positive value for shortage
2173 * indicates that we must aggressively deactivate pages to avoid
2176 shortage = vm_pageout_active_target(vmd) + addl_shortage;
2177 vm_pageout_scan_active(vmd, shortage);
2182 * vm_pageout_helper runs additional pageout daemons in times of high paging
2186 vm_pageout_helper(void *arg)
2188 struct vm_domain *vmd;
2191 domain = (uintptr_t)arg;
2192 vmd = VM_DOMAIN(domain);
2194 vm_domain_pageout_lock(vmd);
2196 msleep(&vmd->vmd_inactive_shortage,
2197 vm_domain_pageout_lockptr(vmd), PVM, "psleep", 0);
2198 blockcount_release(&vmd->vmd_inactive_starting, 1);
2200 vm_domain_pageout_unlock(vmd);
2201 vm_pageout_scan_inactive(vmd, vmd->vmd_inactive_shortage);
2202 vm_domain_pageout_lock(vmd);
2205 * Release the running count while the pageout lock is held to
2206 * prevent wakeup races.
2208 blockcount_release(&vmd->vmd_inactive_running, 1);
2213 get_pageout_threads_per_domain(const struct vm_domain *vmd)
2215 unsigned total_pageout_threads, eligible_cpus, domain_cpus;
2217 if (VM_DOMAIN_EMPTY(vmd->vmd_domain))
2221 * Semi-arbitrarily constrain pagedaemon threads to less than half the
2222 * total number of CPUs in the system as an upper limit.
2224 if (pageout_cpus_per_thread < 2)
2225 pageout_cpus_per_thread = 2;
2226 else if (pageout_cpus_per_thread > mp_ncpus)
2227 pageout_cpus_per_thread = mp_ncpus;
2229 total_pageout_threads = howmany(mp_ncpus, pageout_cpus_per_thread);
2230 domain_cpus = CPU_COUNT(&cpuset_domain[vmd->vmd_domain]);
2232 /* Pagedaemons are not run in empty domains. */
2233 eligible_cpus = mp_ncpus;
2234 for (unsigned i = 0; i < vm_ndomains; i++)
2235 if (VM_DOMAIN_EMPTY(i))
2236 eligible_cpus -= CPU_COUNT(&cpuset_domain[i]);
2239 * Assign a portion of the total pageout threads to this domain
2240 * corresponding to the fraction of pagedaemon-eligible CPUs in the
2241 * domain. In asymmetric NUMA systems, domains with more CPUs may be
2242 * allocated more threads than domains with fewer CPUs.
2244 return (howmany(total_pageout_threads * domain_cpus, eligible_cpus));
2248 * Initialize basic pageout daemon settings. See the comment above the
2249 * definition of vm_domain for some explanation of how these thresholds are
2253 vm_pageout_init_domain(int domain)
2255 struct vm_domain *vmd;
2256 struct sysctl_oid *oid;
2258 vmd = VM_DOMAIN(domain);
2259 vmd->vmd_interrupt_free_min = 2;
2262 * v_free_reserved needs to include enough for the largest
2263 * swap pager structures plus enough for any pv_entry structs
2266 vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE +
2267 vmd->vmd_interrupt_free_min;
2268 vmd->vmd_free_reserved = vm_pageout_page_count +
2269 vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768;
2270 vmd->vmd_free_min = vmd->vmd_page_count / 200;
2271 vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2272 vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2273 vmd->vmd_free_min += vmd->vmd_free_reserved;
2274 vmd->vmd_free_severe += vmd->vmd_free_reserved;
2275 vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2276 if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2277 vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2280 * Set the default wakeup threshold to be 10% below the paging
2281 * target. This keeps the steady state out of shortfall.
2283 vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2286 * Target amount of memory to move out of the laundry queue during a
2287 * background laundering. This is proportional to the amount of system
2290 vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2291 vmd->vmd_free_min) / 10;
2293 /* Initialize the pageout daemon pid controller. */
2294 pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2295 vmd->vmd_free_target, PIDCTRL_BOUND,
2296 PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2297 oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2298 "pidctrl", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2299 pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2301 vmd->vmd_inactive_threads = get_pageout_threads_per_domain(vmd);
2305 vm_pageout_init(void)
2311 * Initialize some paging parameters.
2313 if (vm_cnt.v_page_count < 2000)
2314 vm_pageout_page_count = 8;
2317 for (i = 0; i < vm_ndomains; i++) {
2318 struct vm_domain *vmd;
2320 vm_pageout_init_domain(i);
2322 vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2323 vm_cnt.v_free_target += vmd->vmd_free_target;
2324 vm_cnt.v_free_min += vmd->vmd_free_min;
2325 vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2326 vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2327 vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2328 vm_cnt.v_free_severe += vmd->vmd_free_severe;
2329 freecount += vmd->vmd_free_count;
2333 * Set interval in seconds for active scan. We want to visit each
2334 * page at least once every ten minutes. This is to prevent worst
2335 * case paging behaviors with stale active LRU.
2337 if (vm_pageout_update_period == 0)
2338 vm_pageout_update_period = 600;
2341 * Set the maximum number of user-wired virtual pages. Historically the
2342 * main source of such pages was mlock(2) and mlockall(2). Hypervisors
2343 * may also request user-wired memory.
2345 if (vm_page_max_user_wired == 0)
2346 vm_page_max_user_wired = 4 * freecount / 5;
2350 * vm_pageout is the high level pageout daemon.
2357 int error, first, i, j, pageout_threads;
2362 mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2363 swap_pager_swap_init();
2364 for (first = -1, i = 0; i < vm_ndomains; i++) {
2365 if (VM_DOMAIN_EMPTY(i)) {
2367 printf("domain %d empty; skipping pageout\n",
2374 error = kthread_add(vm_pageout_worker,
2375 (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2377 panic("starting pageout for domain %d: %d\n",
2380 pageout_threads = VM_DOMAIN(i)->vmd_inactive_threads;
2381 for (j = 0; j < pageout_threads - 1; j++) {
2382 error = kthread_add(vm_pageout_helper,
2383 (void *)(uintptr_t)i, p, NULL, 0, 0,
2384 "dom%d helper%d", i, j);
2386 panic("starting pageout helper %d for domain "
2387 "%d: %d\n", j, i, error);
2389 error = kthread_add(vm_pageout_laundry_worker,
2390 (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2392 panic("starting laundry for domain %d: %d", i, error);
2394 error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2396 panic("starting uma_reclaim helper, error %d\n", error);
2398 snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2399 vm_pageout_worker((void *)(uintptr_t)first);
2403 * Perform an advisory wakeup of the page daemon.
2406 pagedaemon_wakeup(int domain)
2408 struct vm_domain *vmd;
2410 vmd = VM_DOMAIN(domain);
2411 vm_domain_pageout_assert_unlocked(vmd);
2412 if (curproc == pageproc)
2415 if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2416 vm_domain_pageout_lock(vmd);
2417 atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2418 wakeup(&vmd->vmd_pageout_wanted);
2419 vm_domain_pageout_unlock(vmd);