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 #if !defined(NO_SWAPPING)
145 /* the kernel process "vm_daemon"*/
146 static void vm_daemon(void);
147 static struct proc *vmproc;
149 static struct kproc_desc vm_kp = {
154 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
157 /* Pagedaemon activity rates, in subdivisions of one second. */
158 #define VM_LAUNDER_RATE 10
159 #define VM_INACT_SCAN_RATE 2
161 int vm_pageout_deficit; /* Estimated number of pages deficit */
162 u_int vm_pageout_wakeup_thresh;
163 static int vm_pageout_oom_seq = 12;
164 bool vm_pageout_wanted; /* Event on which pageout daemon sleeps */
165 bool vm_pages_needed; /* Are threads waiting for free pages? */
167 /* Pending request for dirty page laundering. */
170 VM_LAUNDRY_BACKGROUND,
172 } vm_laundry_request = VM_LAUNDRY_IDLE;
174 #if !defined(NO_SWAPPING)
175 static int vm_pageout_req_swapout; /* XXX */
176 static int vm_daemon_needed;
177 static struct mtx vm_daemon_mtx;
178 /* Allow for use by vm_pageout before vm_daemon is initialized. */
179 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
181 static int vm_pageout_update_period;
182 static int disable_swap_pageouts;
183 static int lowmem_period = 10;
184 static time_t lowmem_uptime;
185 static int swapdev_enabled;
187 #if defined(NO_SWAPPING)
188 static int vm_swap_enabled = 0;
189 static int vm_swap_idle_enabled = 0;
191 static int vm_swap_enabled = 1;
192 static int vm_swap_idle_enabled = 0;
195 static int vm_panic_on_oom = 0;
197 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
198 CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
199 "panic on out of memory instead of killing the largest process");
201 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
202 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
203 "free page threshold for waking up the pageout daemon");
205 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
206 CTLFLAG_RW, &vm_pageout_update_period, 0,
207 "Maximum active LRU update period");
209 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
210 "Low memory callback period");
212 #if defined(NO_SWAPPING)
213 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
214 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
215 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
216 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
218 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
219 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
220 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
221 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
224 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
225 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
227 static int pageout_lock_miss;
228 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
229 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
231 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
232 CTLFLAG_RW, &vm_pageout_oom_seq, 0,
233 "back-to-back calls to oom detector to start OOM");
235 static int act_scan_laundry_weight = 3;
236 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RW,
237 &act_scan_laundry_weight, 0,
238 "weight given to clean vs. dirty pages in active queue scans");
240 static u_int vm_background_launder_target;
241 SYSCTL_UINT(_vm, OID_AUTO, background_launder_target, CTLFLAG_RW,
242 &vm_background_launder_target, 0,
243 "background laundering target, in pages");
245 static u_int vm_background_launder_rate = 4096;
246 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RW,
247 &vm_background_launder_rate, 0,
248 "background laundering rate, in kilobytes per second");
250 static u_int vm_background_launder_max = 20 * 1024;
251 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RW,
252 &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
254 #define VM_PAGEOUT_PAGE_COUNT 16
255 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
257 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
258 SYSCTL_INT(_vm, OID_AUTO, max_wired,
259 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
261 static u_int isqrt(u_int num);
262 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
263 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
265 static void vm_pageout_laundry_worker(void *arg);
266 #if !defined(NO_SWAPPING)
267 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
268 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
269 static void vm_req_vmdaemon(int req);
271 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
274 * Initialize a dummy page for marking the caller's place in the specified
275 * paging queue. In principle, this function only needs to set the flag
276 * PG_MARKER. Nonetheless, it write busies and initializes the hold count
277 * to one as safety precautions.
280 vm_pageout_init_marker(vm_page_t marker, u_short queue)
283 bzero(marker, sizeof(*marker));
284 marker->flags = PG_MARKER;
285 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
286 marker->queue = queue;
287 marker->hold_count = 1;
291 * vm_pageout_fallback_object_lock:
293 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
294 * known to have failed and page queue must be either PQ_ACTIVE or
295 * PQ_INACTIVE. To avoid lock order violation, unlock the page queue
296 * while locking the vm object. Use marker page to detect page queue
297 * changes and maintain notion of next page on page queue. Return
298 * TRUE if no changes were detected, FALSE otherwise. vm object is
301 * This function depends on both the lock portion of struct vm_object
302 * and normal struct vm_page being type stable.
305 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
307 struct vm_page marker;
308 struct vm_pagequeue *pq;
314 vm_pageout_init_marker(&marker, queue);
315 pq = vm_page_pagequeue(m);
318 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
319 vm_pagequeue_unlock(pq);
321 VM_OBJECT_WLOCK(object);
323 vm_pagequeue_lock(pq);
326 * The page's object might have changed, and/or the page might
327 * have moved from its original position in the queue. If the
328 * page's object has changed, then the caller should abandon
329 * processing the page because the wrong object lock was
330 * acquired. Use the marker's plinks.q, not the page's, to
331 * determine if the page has been moved. The state of the
332 * page's plinks.q can be indeterminate; whereas, the marker's
333 * plinks.q must be valid.
335 *next = TAILQ_NEXT(&marker, plinks.q);
336 unchanged = m->object == object &&
337 m == TAILQ_PREV(&marker, pglist, plinks.q);
338 KASSERT(!unchanged || m->queue == queue,
339 ("page %p queue %d %d", m, queue, m->queue));
340 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
345 * Lock the page while holding the page queue lock. Use marker page
346 * to detect page queue changes and maintain notion of next page on
347 * page queue. Return TRUE if no changes were detected, FALSE
348 * otherwise. The page is locked on return. The page queue lock might
349 * be dropped and reacquired.
351 * This function depends on normal struct vm_page being type stable.
354 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
356 struct vm_page marker;
357 struct vm_pagequeue *pq;
361 vm_page_lock_assert(m, MA_NOTOWNED);
362 if (vm_page_trylock(m))
366 vm_pageout_init_marker(&marker, queue);
367 pq = vm_page_pagequeue(m);
369 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
370 vm_pagequeue_unlock(pq);
372 vm_pagequeue_lock(pq);
374 /* Page queue might have changed. */
375 *next = TAILQ_NEXT(&marker, plinks.q);
376 unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
377 KASSERT(!unchanged || m->queue == queue,
378 ("page %p queue %d %d", m, queue, m->queue));
379 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
384 * Scan for pages at adjacent offsets within the given page's object that are
385 * eligible for laundering, form a cluster of these pages and the given page,
386 * and launder that cluster.
389 vm_pageout_cluster(vm_page_t m)
392 vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
394 int ib, is, page_base, pageout_count;
396 vm_page_assert_locked(m);
398 VM_OBJECT_ASSERT_WLOCKED(object);
402 * We can't clean the page if it is busy or held.
404 vm_page_assert_unbusied(m);
405 KASSERT(m->hold_count == 0, ("page %p is held", m));
408 mc[vm_pageout_page_count] = pb = ps = m;
410 page_base = vm_pageout_page_count;
415 * We can cluster only if the page is not clean, busy, or held, and
416 * the page is in the laundry queue.
418 * During heavy mmap/modification loads the pageout
419 * daemon can really fragment the underlying file
420 * due to flushing pages out of order and not trying to
421 * align the clusters (which leaves sporadic out-of-order
422 * holes). To solve this problem we do the reverse scan
423 * first and attempt to align our cluster, then do a
424 * forward scan if room remains.
427 while (ib != 0 && pageout_count < vm_pageout_page_count) {
432 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
436 vm_page_test_dirty(p);
442 if (!vm_page_in_laundry(p) ||
443 p->hold_count != 0) { /* may be undergoing I/O */
449 mc[--page_base] = pb = p;
454 * We are at an alignment boundary. Stop here, and switch
455 * directions. Do not clear ib.
457 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
460 while (pageout_count < vm_pageout_page_count &&
461 pindex + is < object->size) {
462 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
464 vm_page_test_dirty(p);
468 if (!vm_page_in_laundry(p) ||
469 p->hold_count != 0) { /* may be undergoing I/O */
474 mc[page_base + pageout_count] = ps = p;
480 * If we exhausted our forward scan, continue with the reverse scan
481 * when possible, even past an alignment boundary. This catches
482 * boundary conditions.
484 if (ib != 0 && pageout_count < vm_pageout_page_count)
487 return (vm_pageout_flush(&mc[page_base], pageout_count,
488 VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
492 * vm_pageout_flush() - launder the given pages
494 * The given pages are laundered. Note that we setup for the start of
495 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
496 * reference count all in here rather then in the parent. If we want
497 * the parent to do more sophisticated things we may have to change
500 * Returned runlen is the count of pages between mreq and first
501 * page after mreq with status VM_PAGER_AGAIN.
502 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
503 * for any page in runlen set.
506 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
509 vm_object_t object = mc[0]->object;
510 int pageout_status[count];
514 VM_OBJECT_ASSERT_WLOCKED(object);
517 * Initiate I/O. Bump the vm_page_t->busy counter and
518 * mark the pages read-only.
520 * We do not have to fixup the clean/dirty bits here... we can
521 * allow the pager to do it after the I/O completes.
523 * NOTE! mc[i]->dirty may be partial or fragmented due to an
524 * edge case with file fragments.
526 for (i = 0; i < count; i++) {
527 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
528 ("vm_pageout_flush: partially invalid page %p index %d/%d",
530 vm_page_sbusy(mc[i]);
531 pmap_remove_write(mc[i]);
533 vm_object_pip_add(object, count);
535 vm_pager_put_pages(object, mc, count, flags, pageout_status);
537 runlen = count - mreq;
540 for (i = 0; i < count; i++) {
541 vm_page_t mt = mc[i];
543 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
544 !pmap_page_is_write_mapped(mt),
545 ("vm_pageout_flush: page %p is not write protected", mt));
546 switch (pageout_status[i]) {
549 if (vm_page_in_laundry(mt))
550 vm_page_deactivate_noreuse(mt);
558 * The page is outside the object's range. We pretend
559 * that the page out worked and clean the page, so the
560 * changes will be lost if the page is reclaimed by
565 if (vm_page_in_laundry(mt))
566 vm_page_deactivate_noreuse(mt);
572 * If the page couldn't be paged out to swap because the
573 * pager wasn't able to find space, place the page in
574 * the PQ_UNSWAPPABLE holding queue. This is an
575 * optimization that prevents the page daemon from
576 * wasting CPU cycles on pages that cannot be reclaimed
577 * becase no swap device is configured.
579 * Otherwise, reactivate the page so that it doesn't
580 * clog the laundry and inactive queues. (We will try
581 * paging it out again later.)
584 if (object->type == OBJT_SWAP &&
585 pageout_status[i] == VM_PAGER_FAIL) {
586 vm_page_unswappable(mt);
589 vm_page_activate(mt);
591 if (eio != NULL && i >= mreq && i - mreq < runlen)
595 if (i >= mreq && i - mreq < runlen)
601 * If the operation is still going, leave the page busy to
602 * block all other accesses. Also, leave the paging in
603 * progress indicator set so that we don't attempt an object
606 if (pageout_status[i] != VM_PAGER_PEND) {
607 vm_object_pip_wakeup(object);
613 return (numpagedout);
617 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
620 atomic_store_rel_int(&swapdev_enabled, 1);
624 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
627 if (swap_pager_nswapdev() == 1)
628 atomic_store_rel_int(&swapdev_enabled, 0);
631 #if !defined(NO_SWAPPING)
633 * vm_pageout_object_deactivate_pages
635 * Deactivate enough pages to satisfy the inactive target
638 * The object and map must be locked.
641 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
644 vm_object_t backing_object, object;
646 int act_delta, remove_mode;
648 VM_OBJECT_ASSERT_LOCKED(first_object);
649 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
651 for (object = first_object;; object = backing_object) {
652 if (pmap_resident_count(pmap) <= desired)
654 VM_OBJECT_ASSERT_LOCKED(object);
655 if ((object->flags & OBJ_UNMANAGED) != 0 ||
656 object->paging_in_progress != 0)
660 if (object->shadow_count > 1)
663 * Scan the object's entire memory queue.
665 TAILQ_FOREACH(p, &object->memq, listq) {
666 if (pmap_resident_count(pmap) <= desired)
668 if (vm_page_busied(p))
670 VM_CNT_INC(v_pdpages);
672 if (p->wire_count != 0 || p->hold_count != 0 ||
673 !pmap_page_exists_quick(pmap, p)) {
677 act_delta = pmap_ts_referenced(p);
678 if ((p->aflags & PGA_REFERENCED) != 0) {
681 vm_page_aflag_clear(p, PGA_REFERENCED);
683 if (!vm_page_active(p) && act_delta != 0) {
685 p->act_count += act_delta;
686 } else if (vm_page_active(p)) {
687 if (act_delta == 0) {
688 p->act_count -= min(p->act_count,
690 if (!remove_mode && p->act_count == 0) {
692 vm_page_deactivate(p);
697 if (p->act_count < ACT_MAX -
699 p->act_count += ACT_ADVANCE;
702 } else if (vm_page_inactive(p))
706 if ((backing_object = object->backing_object) == NULL)
708 VM_OBJECT_RLOCK(backing_object);
709 if (object != first_object)
710 VM_OBJECT_RUNLOCK(object);
713 if (object != first_object)
714 VM_OBJECT_RUNLOCK(object);
718 * deactivate some number of pages in a map, try to do it fairly, but
719 * that is really hard to do.
722 vm_pageout_map_deactivate_pages(map, desired)
727 vm_object_t obj, bigobj;
730 if (!vm_map_trylock(map))
737 * first, search out the biggest object, and try to free pages from
740 tmpe = map->header.next;
741 while (tmpe != &map->header) {
742 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
743 obj = tmpe->object.vm_object;
744 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
745 if (obj->shadow_count <= 1 &&
747 bigobj->resident_page_count < obj->resident_page_count)) {
749 VM_OBJECT_RUNLOCK(bigobj);
752 VM_OBJECT_RUNLOCK(obj);
755 if (tmpe->wired_count > 0)
756 nothingwired = FALSE;
760 if (bigobj != NULL) {
761 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
762 VM_OBJECT_RUNLOCK(bigobj);
765 * Next, hunt around for other pages to deactivate. We actually
766 * do this search sort of wrong -- .text first is not the best idea.
768 tmpe = map->header.next;
769 while (tmpe != &map->header) {
770 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
772 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
773 obj = tmpe->object.vm_object;
775 VM_OBJECT_RLOCK(obj);
776 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
777 VM_OBJECT_RUNLOCK(obj);
784 * Remove all mappings if a process is swapped out, this will free page
787 if (desired == 0 && nothingwired) {
788 pmap_remove(vm_map_pmap(map), vm_map_min(map),
794 #endif /* !defined(NO_SWAPPING) */
797 * Attempt to acquire all of the necessary locks to launder a page and
798 * then call through the clustering layer to PUTPAGES. Wait a short
799 * time for a vnode lock.
801 * Requires the page and object lock on entry, releases both before return.
802 * Returns 0 on success and an errno otherwise.
805 vm_pageout_clean(vm_page_t m, int *numpagedout)
813 vm_page_assert_locked(m);
815 VM_OBJECT_ASSERT_WLOCKED(object);
821 * The object is already known NOT to be dead. It
822 * is possible for the vget() to block the whole
823 * pageout daemon, but the new low-memory handling
824 * code should prevent it.
826 * We can't wait forever for the vnode lock, we might
827 * deadlock due to a vn_read() getting stuck in
828 * vm_wait while holding this vnode. We skip the
829 * vnode if we can't get it in a reasonable amount
832 if (object->type == OBJT_VNODE) {
835 if (vp->v_type == VREG &&
836 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
842 ("vp %p with NULL v_mount", vp));
843 vm_object_reference_locked(object);
845 VM_OBJECT_WUNLOCK(object);
846 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
847 LK_SHARED : LK_EXCLUSIVE;
848 if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
853 VM_OBJECT_WLOCK(object);
856 * While the object and page were unlocked, the page
858 * (1) moved to a different queue,
859 * (2) reallocated to a different object,
860 * (3) reallocated to a different offset, or
863 if (!vm_page_in_laundry(m) || m->object != object ||
864 m->pindex != pindex || m->dirty == 0) {
871 * The page may have been busied or held while the object
872 * and page locks were released.
874 if (vm_page_busied(m) || m->hold_count != 0) {
882 * If a page is dirty, then it is either being washed
883 * (but not yet cleaned) or it is still in the
884 * laundry. If it is still in the laundry, then we
885 * start the cleaning operation.
887 if ((*numpagedout = vm_pageout_cluster(m)) == 0)
891 VM_OBJECT_WUNLOCK(object);
894 vm_page_lock_assert(m, MA_NOTOWNED);
898 vm_object_deallocate(object);
899 vn_finished_write(mp);
906 * Attempt to launder the specified number of pages.
908 * Returns the number of pages successfully laundered.
911 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
913 struct vm_pagequeue *pq;
916 int act_delta, error, maxscan, numpagedout, starting_target;
918 bool pageout_ok, queue_locked;
920 starting_target = launder;
924 * Scan the laundry queues for pages eligible to be laundered. We stop
925 * once the target number of dirty pages have been laundered, or once
926 * we've reached the end of the queue. A single iteration of this loop
927 * may cause more than one page to be laundered because of clustering.
929 * maxscan ensures that we don't re-examine requeued pages. Any
930 * additional pages written as part of a cluster are subtracted from
931 * maxscan since they must be taken from the laundry queue.
933 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
934 * swap devices are configured.
936 if (atomic_load_acq_int(&swapdev_enabled))
937 pq = &vmd->vmd_pagequeues[PQ_UNSWAPPABLE];
939 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
942 vm_pagequeue_lock(pq);
943 maxscan = pq->pq_cnt;
945 for (m = TAILQ_FIRST(&pq->pq_pl);
946 m != NULL && maxscan-- > 0 && launder > 0;
948 vm_pagequeue_assert_locked(pq);
949 KASSERT(queue_locked, ("unlocked laundry queue"));
950 KASSERT(vm_page_in_laundry(m),
951 ("page %p has an inconsistent queue", m));
952 next = TAILQ_NEXT(m, plinks.q);
953 if ((m->flags & PG_MARKER) != 0)
955 KASSERT((m->flags & PG_FICTITIOUS) == 0,
956 ("PG_FICTITIOUS page %p cannot be in laundry queue", m));
957 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
958 ("VPO_UNMANAGED page %p cannot be in laundry queue", m));
959 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
964 if ((!VM_OBJECT_TRYWLOCK(object) &&
965 (!vm_pageout_fallback_object_lock(m, &next) ||
966 m->hold_count != 0)) || vm_page_busied(m)) {
967 VM_OBJECT_WUNLOCK(object);
973 * Unlock the laundry queue, invalidating the 'next' pointer.
974 * Use a marker to remember our place in the laundry queue.
976 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_laundry_marker,
978 vm_pagequeue_unlock(pq);
979 queue_locked = false;
982 * Invalid pages can be easily freed. They cannot be
983 * mapped; vm_page_free() asserts this.
989 * If the page has been referenced and the object is not dead,
990 * reactivate or requeue the page depending on whether the
993 if ((m->aflags & PGA_REFERENCED) != 0) {
994 vm_page_aflag_clear(m, PGA_REFERENCED);
998 if (object->ref_count != 0)
999 act_delta += pmap_ts_referenced(m);
1001 KASSERT(!pmap_page_is_mapped(m),
1002 ("page %p is mapped", m));
1004 if (act_delta != 0) {
1005 if (object->ref_count != 0) {
1006 VM_CNT_INC(v_reactivated);
1007 vm_page_activate(m);
1010 * Increase the activation count if the page
1011 * was referenced while in the laundry queue.
1012 * This makes it less likely that the page will
1013 * be returned prematurely to the inactive
1016 m->act_count += act_delta + ACT_ADVANCE;
1019 * If this was a background laundering, count
1020 * activated pages towards our target. The
1021 * purpose of background laundering is to ensure
1022 * that pages are eventually cycled through the
1023 * laundry queue, and an activation is a valid
1029 } else if ((object->flags & OBJ_DEAD) == 0)
1034 * If the page appears to be clean at the machine-independent
1035 * layer, then remove all of its mappings from the pmap in
1036 * anticipation of freeing it. If, however, any of the page's
1037 * mappings allow write access, then the page may still be
1038 * modified until the last of those mappings are removed.
1040 if (object->ref_count != 0) {
1041 vm_page_test_dirty(m);
1047 * Clean pages are freed, and dirty pages are paged out unless
1048 * they belong to a dead object. Requeueing dirty pages from
1049 * dead objects is pointless, as they are being paged out and
1050 * freed by the thread that destroyed the object.
1052 if (m->dirty == 0) {
1055 VM_CNT_INC(v_dfree);
1056 } else if ((object->flags & OBJ_DEAD) == 0) {
1057 if (object->type != OBJT_SWAP &&
1058 object->type != OBJT_DEFAULT)
1060 else if (disable_swap_pageouts)
1066 vm_pagequeue_lock(pq);
1067 queue_locked = true;
1068 vm_page_requeue_locked(m);
1073 * Form a cluster with adjacent, dirty pages from the
1074 * same object, and page out that entire cluster.
1076 * The adjacent, dirty pages must also be in the
1077 * laundry. However, their mappings are not checked
1078 * for new references. Consequently, a recently
1079 * referenced page may be paged out. However, that
1080 * page will not be prematurely reclaimed. After page
1081 * out, the page will be placed in the inactive queue,
1082 * where any new references will be detected and the
1085 error = vm_pageout_clean(m, &numpagedout);
1087 launder -= numpagedout;
1088 maxscan -= numpagedout - 1;
1089 } else if (error == EDEADLK) {
1090 pageout_lock_miss++;
1097 VM_OBJECT_WUNLOCK(object);
1099 if (!queue_locked) {
1100 vm_pagequeue_lock(pq);
1101 queue_locked = true;
1103 next = TAILQ_NEXT(&vmd->vmd_laundry_marker, plinks.q);
1104 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_laundry_marker, plinks.q);
1106 vm_pagequeue_unlock(pq);
1108 if (launder > 0 && pq == &vmd->vmd_pagequeues[PQ_UNSWAPPABLE]) {
1109 pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1114 * Wakeup the sync daemon if we skipped a vnode in a writeable object
1115 * and we didn't launder enough pages.
1117 if (vnodes_skipped > 0 && launder > 0)
1118 (void)speedup_syncer();
1120 return (starting_target - launder);
1124 * Compute the integer square root.
1129 u_int bit, root, tmp;
1131 bit = 1u << ((NBBY * sizeof(u_int)) - 2);
1148 * Perform the work of the laundry thread: periodically wake up and determine
1149 * whether any pages need to be laundered. If so, determine the number of pages
1150 * that need to be laundered, and launder them.
1153 vm_pageout_laundry_worker(void *arg)
1155 struct vm_domain *domain;
1156 struct vm_pagequeue *pq;
1157 uint64_t nclean, ndirty;
1158 u_int last_launder, wakeups;
1159 int domidx, last_target, launder, shortfall, shortfall_cycle, target;
1162 domidx = (uintptr_t)arg;
1163 domain = &vm_dom[domidx];
1164 pq = &domain->vmd_pagequeues[PQ_LAUNDRY];
1165 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1166 vm_pageout_init_marker(&domain->vmd_laundry_marker, PQ_LAUNDRY);
1169 in_shortfall = false;
1170 shortfall_cycle = 0;
1175 * Calls to these handlers are serialized by the swap syscall lock.
1177 (void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, domain,
1178 EVENTHANDLER_PRI_ANY);
1179 (void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, domain,
1180 EVENTHANDLER_PRI_ANY);
1183 * The pageout laundry worker is never done, so loop forever.
1186 KASSERT(target >= 0, ("negative target %d", target));
1187 KASSERT(shortfall_cycle >= 0,
1188 ("negative cycle %d", shortfall_cycle));
1190 wakeups = VM_CNT_FETCH(v_pdwakeups);
1193 * First determine whether we need to launder pages to meet a
1194 * shortage of free pages.
1196 if (shortfall > 0) {
1197 in_shortfall = true;
1198 shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1200 } else if (!in_shortfall)
1202 else if (shortfall_cycle == 0 || vm_laundry_target() <= 0) {
1204 * We recently entered shortfall and began laundering
1205 * pages. If we have completed that laundering run
1206 * (and we are no longer in shortfall) or we have met
1207 * our laundry target through other activity, then we
1208 * can stop laundering pages.
1210 in_shortfall = false;
1214 last_launder = wakeups;
1215 launder = target / shortfall_cycle--;
1219 * There's no immediate need to launder any pages; see if we
1220 * meet the conditions to perform background laundering:
1222 * 1. The ratio of dirty to clean inactive pages exceeds the
1223 * background laundering threshold and the pagedaemon has
1224 * been woken up to reclaim pages since our last
1226 * 2. we haven't yet reached the target of the current
1227 * background laundering run.
1229 * The background laundering threshold is not a constant.
1230 * Instead, it is a slowly growing function of the number of
1231 * page daemon wakeups since the last laundering. Thus, as the
1232 * ratio of dirty to clean inactive pages grows, the amount of
1233 * memory pressure required to trigger laundering decreases.
1236 nclean = vm_cnt.v_inactive_count + vm_cnt.v_free_count;
1237 ndirty = vm_cnt.v_laundry_count;
1238 if (target == 0 && wakeups != last_launder &&
1239 ndirty * isqrt(wakeups - last_launder) >= nclean) {
1240 target = vm_background_launder_target;
1244 * We have a non-zero background laundering target. If we've
1245 * laundered up to our maximum without observing a page daemon
1246 * wakeup, just stop. This is a safety belt that ensures we
1247 * don't launder an excessive amount if memory pressure is low
1248 * and the ratio of dirty to clean pages is large. Otherwise,
1249 * proceed at the background laundering rate.
1252 if (wakeups != last_launder) {
1253 last_launder = wakeups;
1254 last_target = target;
1255 } else if (last_target - target >=
1256 vm_background_launder_max * PAGE_SIZE / 1024) {
1259 launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1260 launder /= VM_LAUNDER_RATE;
1261 if (launder > target)
1268 * Because of I/O clustering, the number of laundered
1269 * pages could exceed "target" by the maximum size of
1270 * a cluster minus one.
1272 target -= min(vm_pageout_launder(domain, launder,
1273 in_shortfall), target);
1274 pause("laundp", hz / VM_LAUNDER_RATE);
1278 * If we're not currently laundering pages and the page daemon
1279 * hasn't posted a new request, sleep until the page daemon
1282 vm_pagequeue_lock(pq);
1283 if (target == 0 && vm_laundry_request == VM_LAUNDRY_IDLE)
1284 (void)mtx_sleep(&vm_laundry_request,
1285 vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1288 * If the pagedaemon has indicated that it's in shortfall, start
1289 * a shortfall laundering unless we're already in the middle of
1290 * one. This may preempt a background laundering.
1292 if (vm_laundry_request == VM_LAUNDRY_SHORTFALL &&
1293 (!in_shortfall || shortfall_cycle == 0)) {
1294 shortfall = vm_laundry_target() + vm_pageout_deficit;
1300 vm_laundry_request = VM_LAUNDRY_IDLE;
1301 vm_pagequeue_unlock(pq);
1306 * vm_pageout_scan does the dirty work for the pageout daemon.
1308 * pass == 0: Update active LRU/deactivate pages
1309 * pass >= 1: Free inactive pages
1311 * Returns true if pass was zero or enough pages were freed by the inactive
1312 * queue scan to meet the target.
1315 vm_pageout_scan(struct vm_domain *vmd, int pass)
1318 struct vm_pagequeue *pq;
1321 int act_delta, addl_page_shortage, deficit, inactq_shortage, maxscan;
1322 int page_shortage, scan_tick, scanned, starting_page_shortage;
1323 boolean_t queue_locked;
1326 * If we need to reclaim memory ask kernel caches to return
1327 * some. We rate limit to avoid thrashing.
1329 if (vmd == &vm_dom[0] && pass > 0 &&
1330 (time_uptime - lowmem_uptime) >= lowmem_period) {
1332 * Decrease registered cache sizes.
1334 SDT_PROBE0(vm, , , vm__lowmem_scan);
1335 EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1337 * We do this explicitly after the caches have been
1341 lowmem_uptime = time_uptime;
1345 * The addl_page_shortage is the number of temporarily
1346 * stuck pages in the inactive queue. In other words, the
1347 * number of pages from the inactive count that should be
1348 * discounted in setting the target for the active queue scan.
1350 addl_page_shortage = 0;
1353 * Calculate the number of pages that we want to free. This number
1354 * can be negative if many pages are freed between the wakeup call to
1355 * the page daemon and this calculation.
1358 deficit = atomic_readandclear_int(&vm_pageout_deficit);
1359 page_shortage = vm_paging_target() + deficit;
1361 page_shortage = deficit = 0;
1362 starting_page_shortage = page_shortage;
1365 * Start scanning the inactive queue for pages that we can free. The
1366 * scan will stop when we reach the target or we have scanned the
1367 * entire queue. (Note that m->act_count is not used to make
1368 * decisions for the inactive queue, only for the active queue.)
1370 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1371 maxscan = pq->pq_cnt;
1372 vm_pagequeue_lock(pq);
1373 queue_locked = TRUE;
1374 for (m = TAILQ_FIRST(&pq->pq_pl);
1375 m != NULL && maxscan-- > 0 && page_shortage > 0;
1377 vm_pagequeue_assert_locked(pq);
1378 KASSERT(queue_locked, ("unlocked inactive queue"));
1379 KASSERT(vm_page_inactive(m), ("Inactive queue %p", m));
1381 VM_CNT_INC(v_pdpages);
1382 next = TAILQ_NEXT(m, plinks.q);
1387 if (m->flags & PG_MARKER)
1390 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1391 ("Fictitious page %p cannot be in inactive queue", m));
1392 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1393 ("Unmanaged page %p cannot be in inactive queue", m));
1396 * The page or object lock acquisitions fail if the
1397 * page was removed from the queue or moved to a
1398 * different position within the queue. In either
1399 * case, addl_page_shortage should not be incremented.
1401 if (!vm_pageout_page_lock(m, &next))
1403 else if (m->hold_count != 0) {
1405 * Held pages are essentially stuck in the
1406 * queue. So, they ought to be discounted
1407 * from the inactive count. See the
1408 * calculation of inactq_shortage before the
1409 * loop over the active queue below.
1411 addl_page_shortage++;
1415 if (!VM_OBJECT_TRYWLOCK(object)) {
1416 if (!vm_pageout_fallback_object_lock(m, &next))
1418 else if (m->hold_count != 0) {
1419 addl_page_shortage++;
1423 if (vm_page_busied(m)) {
1425 * Don't mess with busy pages. Leave them at
1426 * the front of the queue. Most likely, they
1427 * are being paged out and will leave the
1428 * queue shortly after the scan finishes. So,
1429 * they ought to be discounted from the
1432 addl_page_shortage++;
1434 VM_OBJECT_WUNLOCK(object);
1439 KASSERT(m->hold_count == 0, ("Held page %p", m));
1442 * Dequeue the inactive page and unlock the inactive page
1443 * queue, invalidating the 'next' pointer. Dequeueing the
1444 * page here avoids a later reacquisition (and release) of
1445 * the inactive page queue lock when vm_page_activate(),
1446 * vm_page_free(), or vm_page_launder() is called. Use a
1447 * marker to remember our place in the inactive queue.
1449 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1450 vm_page_dequeue_locked(m);
1451 vm_pagequeue_unlock(pq);
1452 queue_locked = FALSE;
1455 * Invalid pages can be easily freed. They cannot be
1456 * mapped, vm_page_free() asserts this.
1462 * If the page has been referenced and the object is not dead,
1463 * reactivate or requeue the page depending on whether the
1466 if ((m->aflags & PGA_REFERENCED) != 0) {
1467 vm_page_aflag_clear(m, PGA_REFERENCED);
1471 if (object->ref_count != 0) {
1472 act_delta += pmap_ts_referenced(m);
1474 KASSERT(!pmap_page_is_mapped(m),
1475 ("vm_pageout_scan: page %p is mapped", m));
1477 if (act_delta != 0) {
1478 if (object->ref_count != 0) {
1479 VM_CNT_INC(v_reactivated);
1480 vm_page_activate(m);
1483 * Increase the activation count if the page
1484 * was referenced while in the inactive queue.
1485 * This makes it less likely that the page will
1486 * be returned prematurely to the inactive
1489 m->act_count += act_delta + ACT_ADVANCE;
1491 } else if ((object->flags & OBJ_DEAD) == 0) {
1492 vm_pagequeue_lock(pq);
1493 queue_locked = TRUE;
1494 m->queue = PQ_INACTIVE;
1495 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
1496 vm_pagequeue_cnt_inc(pq);
1502 * If the page appears to be clean at the machine-independent
1503 * layer, then remove all of its mappings from the pmap in
1504 * anticipation of freeing it. If, however, any of the page's
1505 * mappings allow write access, then the page may still be
1506 * modified until the last of those mappings are removed.
1508 if (object->ref_count != 0) {
1509 vm_page_test_dirty(m);
1515 * Clean pages can be freed, but dirty pages must be sent back
1516 * to the laundry, unless they belong to a dead object.
1517 * Requeueing dirty pages from dead objects is pointless, as
1518 * they are being paged out and freed by the thread that
1519 * destroyed the object.
1521 if (m->dirty == 0) {
1524 VM_CNT_INC(v_dfree);
1526 } else if ((object->flags & OBJ_DEAD) == 0)
1530 VM_OBJECT_WUNLOCK(object);
1531 if (!queue_locked) {
1532 vm_pagequeue_lock(pq);
1533 queue_locked = TRUE;
1535 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1536 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1538 vm_pagequeue_unlock(pq);
1541 * Wake up the laundry thread so that it can perform any needed
1542 * laundering. If we didn't meet our target, we're in shortfall and
1543 * need to launder more aggressively. If PQ_LAUNDRY is empty and no
1544 * swap devices are configured, the laundry thread has no work to do, so
1545 * don't bother waking it up.
1547 if (vm_laundry_request == VM_LAUNDRY_IDLE &&
1548 starting_page_shortage > 0) {
1549 pq = &vm_dom[0].vmd_pagequeues[PQ_LAUNDRY];
1550 vm_pagequeue_lock(pq);
1551 if (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled)) {
1552 if (page_shortage > 0) {
1553 vm_laundry_request = VM_LAUNDRY_SHORTFALL;
1554 VM_CNT_INC(v_pdshortfalls);
1555 } else if (vm_laundry_request != VM_LAUNDRY_SHORTFALL)
1556 vm_laundry_request = VM_LAUNDRY_BACKGROUND;
1557 wakeup(&vm_laundry_request);
1559 vm_pagequeue_unlock(pq);
1562 #if !defined(NO_SWAPPING)
1564 * Wakeup the swapout daemon if we didn't free the targeted number of
1567 if (vm_swap_enabled && page_shortage > 0)
1568 vm_req_vmdaemon(VM_SWAP_NORMAL);
1572 * If the inactive queue scan fails repeatedly to meet its
1573 * target, kill the largest process.
1575 vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1578 * Compute the number of pages we want to try to move from the
1579 * active queue to either the inactive or laundry queue.
1581 * When scanning active pages, we make clean pages count more heavily
1582 * towards the page shortage than dirty pages. This is because dirty
1583 * pages must be laundered before they can be reused and thus have less
1584 * utility when attempting to quickly alleviate a shortage. However,
1585 * this weighting also causes the scan to deactivate dirty pages more
1586 * more aggressively, improving the effectiveness of clustering and
1587 * ensuring that they can eventually be reused.
1589 inactq_shortage = vm_cnt.v_inactive_target - (vm_cnt.v_inactive_count +
1590 vm_cnt.v_laundry_count / act_scan_laundry_weight) +
1591 vm_paging_target() + deficit + addl_page_shortage;
1592 page_shortage *= act_scan_laundry_weight;
1594 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1595 vm_pagequeue_lock(pq);
1596 maxscan = pq->pq_cnt;
1599 * If we're just idle polling attempt to visit every
1600 * active page within 'update_period' seconds.
1603 if (vm_pageout_update_period != 0) {
1604 min_scan = pq->pq_cnt;
1605 min_scan *= scan_tick - vmd->vmd_last_active_scan;
1606 min_scan /= hz * vm_pageout_update_period;
1609 if (min_scan > 0 || (inactq_shortage > 0 && maxscan > 0))
1610 vmd->vmd_last_active_scan = scan_tick;
1613 * Scan the active queue for pages that can be deactivated. Update
1614 * the per-page activity counter and use it to identify deactivation
1615 * candidates. Held pages may be deactivated.
1617 for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1618 min_scan || (inactq_shortage > 0 && scanned < maxscan)); m = next,
1620 KASSERT(m->queue == PQ_ACTIVE,
1621 ("vm_pageout_scan: page %p isn't active", m));
1622 next = TAILQ_NEXT(m, plinks.q);
1623 if ((m->flags & PG_MARKER) != 0)
1625 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1626 ("Fictitious page %p cannot be in active queue", m));
1627 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1628 ("Unmanaged page %p cannot be in active queue", m));
1629 if (!vm_pageout_page_lock(m, &next)) {
1635 * The count for page daemon pages is updated after checking
1636 * the page for eligibility.
1638 VM_CNT_INC(v_pdpages);
1641 * Check to see "how much" the page has been used.
1643 if ((m->aflags & PGA_REFERENCED) != 0) {
1644 vm_page_aflag_clear(m, PGA_REFERENCED);
1650 * Perform an unsynchronized object ref count check. While
1651 * the page lock ensures that the page is not reallocated to
1652 * another object, in particular, one with unmanaged mappings
1653 * that cannot support pmap_ts_referenced(), two races are,
1654 * nonetheless, possible:
1655 * 1) The count was transitioning to zero, but we saw a non-
1656 * zero value. pmap_ts_referenced() will return zero
1657 * because the page is not mapped.
1658 * 2) The count was transitioning to one, but we saw zero.
1659 * This race delays the detection of a new reference. At
1660 * worst, we will deactivate and reactivate the page.
1662 if (m->object->ref_count != 0)
1663 act_delta += pmap_ts_referenced(m);
1666 * Advance or decay the act_count based on recent usage.
1668 if (act_delta != 0) {
1669 m->act_count += ACT_ADVANCE + act_delta;
1670 if (m->act_count > ACT_MAX)
1671 m->act_count = ACT_MAX;
1673 m->act_count -= min(m->act_count, ACT_DECLINE);
1676 * Move this page to the tail of the active, inactive or laundry
1677 * queue depending on usage.
1679 if (m->act_count == 0) {
1680 /* Dequeue to avoid later lock recursion. */
1681 vm_page_dequeue_locked(m);
1684 * When not short for inactive pages, let dirty pages go
1685 * through the inactive queue before moving to the
1686 * laundry queues. This gives them some extra time to
1687 * be reactivated, potentially avoiding an expensive
1688 * pageout. During a page shortage, the inactive queue
1689 * is necessarily small, so we may move dirty pages
1690 * directly to the laundry queue.
1692 if (inactq_shortage <= 0)
1693 vm_page_deactivate(m);
1696 * Calling vm_page_test_dirty() here would
1697 * require acquisition of the object's write
1698 * lock. However, during a page shortage,
1699 * directing dirty pages into the laundry
1700 * queue is only an optimization and not a
1701 * requirement. Therefore, we simply rely on
1702 * the opportunistic updates to the page's
1703 * dirty field by the pmap.
1705 if (m->dirty == 0) {
1706 vm_page_deactivate(m);
1708 act_scan_laundry_weight;
1715 vm_page_requeue_locked(m);
1718 vm_pagequeue_unlock(pq);
1719 #if !defined(NO_SWAPPING)
1721 * Idle process swapout -- run once per second when we are reclaiming
1724 if (vm_swap_idle_enabled && pass > 0) {
1726 if (time_second != lsec) {
1727 vm_req_vmdaemon(VM_SWAP_IDLE);
1732 return (page_shortage <= 0);
1735 static int vm_pageout_oom_vote;
1738 * The pagedaemon threads randlomly select one to perform the
1739 * OOM. Trying to kill processes before all pagedaemons
1740 * failed to reach free target is premature.
1743 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1744 int starting_page_shortage)
1748 if (starting_page_shortage <= 0 || starting_page_shortage !=
1750 vmd->vmd_oom_seq = 0;
1753 if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1755 vmd->vmd_oom = FALSE;
1756 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1762 * Do not follow the call sequence until OOM condition is
1765 vmd->vmd_oom_seq = 0;
1770 vmd->vmd_oom = TRUE;
1771 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1772 if (old_vote != vm_ndomains - 1)
1776 * The current pagedaemon thread is the last in the quorum to
1777 * start OOM. Initiate the selection and signaling of the
1780 vm_pageout_oom(VM_OOM_MEM);
1783 * After one round of OOM terror, recall our vote. On the
1784 * next pass, current pagedaemon would vote again if the low
1785 * memory condition is still there, due to vmd_oom being
1788 vmd->vmd_oom = FALSE;
1789 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1793 * The OOM killer is the page daemon's action of last resort when
1794 * memory allocation requests have been stalled for a prolonged period
1795 * of time because it cannot reclaim memory. This function computes
1796 * the approximate number of physical pages that could be reclaimed if
1797 * the specified address space is destroyed.
1799 * Private, anonymous memory owned by the address space is the
1800 * principal resource that we expect to recover after an OOM kill.
1801 * Since the physical pages mapped by the address space's COW entries
1802 * are typically shared pages, they are unlikely to be released and so
1803 * they are not counted.
1805 * To get to the point where the page daemon runs the OOM killer, its
1806 * efforts to write-back vnode-backed pages may have stalled. This
1807 * could be caused by a memory allocation deadlock in the write path
1808 * that might be resolved by an OOM kill. Therefore, physical pages
1809 * belonging to vnode-backed objects are counted, because they might
1810 * be freed without being written out first if the address space holds
1811 * the last reference to an unlinked vnode.
1813 * Similarly, physical pages belonging to OBJT_PHYS objects are
1814 * counted because the address space might hold the last reference to
1818 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1821 vm_map_entry_t entry;
1825 map = &vmspace->vm_map;
1826 KASSERT(!map->system_map, ("system map"));
1827 sx_assert(&map->lock, SA_LOCKED);
1829 for (entry = map->header.next; entry != &map->header;
1830 entry = entry->next) {
1831 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1833 obj = entry->object.vm_object;
1836 if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1837 obj->ref_count != 1)
1839 switch (obj->type) {
1844 res += obj->resident_page_count;
1852 vm_pageout_oom(int shortage)
1854 struct proc *p, *bigproc;
1855 vm_offset_t size, bigsize;
1860 * We keep the process bigproc locked once we find it to keep anyone
1861 * from messing with it; however, there is a possibility of
1862 * deadlock if process B is bigproc and one of its child processes
1863 * attempts to propagate a signal to B while we are waiting for A's
1864 * lock while walking this list. To avoid this, we don't block on
1865 * the process lock but just skip a process if it is already locked.
1869 sx_slock(&allproc_lock);
1870 FOREACH_PROC_IN_SYSTEM(p) {
1876 * If this is a system, protected or killed process, skip it.
1878 if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1879 P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1880 p->p_pid == 1 || P_KILLED(p) ||
1881 (p->p_pid < 48 && swap_pager_avail != 0)) {
1886 * If the process is in a non-running type state,
1887 * don't touch it. Check all the threads individually.
1890 FOREACH_THREAD_IN_PROC(p, td) {
1892 if (!TD_ON_RUNQ(td) &&
1893 !TD_IS_RUNNING(td) &&
1894 !TD_IS_SLEEPING(td) &&
1895 !TD_IS_SUSPENDED(td) &&
1896 !TD_IS_SWAPPED(td)) {
1908 * get the process size
1910 vm = vmspace_acquire_ref(p);
1917 sx_sunlock(&allproc_lock);
1918 if (!vm_map_trylock_read(&vm->vm_map)) {
1920 sx_slock(&allproc_lock);
1924 size = vmspace_swap_count(vm);
1925 if (shortage == VM_OOM_MEM)
1926 size += vm_pageout_oom_pagecount(vm);
1927 vm_map_unlock_read(&vm->vm_map);
1929 sx_slock(&allproc_lock);
1932 * If this process is bigger than the biggest one,
1935 if (size > bigsize) {
1936 if (bigproc != NULL)
1944 sx_sunlock(&allproc_lock);
1945 if (bigproc != NULL) {
1946 if (vm_panic_on_oom != 0)
1947 panic("out of swap space");
1949 killproc(bigproc, "out of swap space");
1950 sched_nice(bigproc, PRIO_MIN);
1952 PROC_UNLOCK(bigproc);
1953 wakeup(&vm_cnt.v_free_count);
1958 vm_pageout_worker(void *arg)
1960 struct vm_domain *domain;
1964 domidx = (uintptr_t)arg;
1965 domain = &vm_dom[domidx];
1970 * XXXKIB It could be useful to bind pageout daemon threads to
1971 * the cores belonging to the domain, from which vm_page_array
1975 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1976 domain->vmd_last_active_scan = ticks;
1977 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1978 vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
1979 TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1980 &domain->vmd_inacthead, plinks.q);
1983 * The pageout daemon worker is never done, so loop forever.
1986 mtx_lock(&vm_page_queue_free_mtx);
1989 * Generally, after a level >= 1 scan, if there are enough
1990 * free pages to wakeup the waiters, then they are already
1991 * awake. A call to vm_page_free() during the scan awakened
1992 * them. However, in the following case, this wakeup serves
1993 * to bound the amount of time that a thread might wait.
1994 * Suppose a thread's call to vm_page_alloc() fails, but
1995 * before that thread calls VM_WAIT, enough pages are freed by
1996 * other threads to alleviate the free page shortage. The
1997 * thread will, nonetheless, wait until another page is freed
1998 * or this wakeup is performed.
2000 if (vm_pages_needed && !vm_page_count_min()) {
2001 vm_pages_needed = false;
2002 wakeup(&vm_cnt.v_free_count);
2006 * Do not clear vm_pageout_wanted until we reach our free page
2007 * target. Otherwise, we may be awakened over and over again,
2010 if (vm_pageout_wanted && target_met)
2011 vm_pageout_wanted = false;
2014 * Might the page daemon receive a wakeup call?
2016 if (vm_pageout_wanted) {
2018 * No. Either vm_pageout_wanted was set by another
2019 * thread during the previous scan, which must have
2020 * been a level 0 scan, or vm_pageout_wanted was
2021 * already set and the scan failed to free enough
2022 * pages. If we haven't yet performed a level >= 1
2023 * (page reclamation) scan, then increase the level
2024 * and scan again now. Otherwise, sleep a bit and
2027 mtx_unlock(&vm_page_queue_free_mtx);
2029 pause("psleep", hz / VM_INACT_SCAN_RATE);
2033 * Yes. Sleep until pages need to be reclaimed or
2034 * have their reference stats updated.
2036 if (mtx_sleep(&vm_pageout_wanted,
2037 &vm_page_queue_free_mtx, PDROP | PVM, "psleep",
2039 VM_CNT_INC(v_pdwakeups);
2045 target_met = vm_pageout_scan(domain, pass);
2050 * vm_pageout_init initialises basic pageout daemon settings.
2053 vm_pageout_init(void)
2056 * Initialize some paging parameters.
2058 vm_cnt.v_interrupt_free_min = 2;
2059 if (vm_cnt.v_page_count < 2000)
2060 vm_pageout_page_count = 8;
2063 * v_free_reserved needs to include enough for the largest
2064 * swap pager structures plus enough for any pv_entry structs
2067 if (vm_cnt.v_page_count > 1024)
2068 vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
2070 vm_cnt.v_free_min = 4;
2071 vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
2072 vm_cnt.v_interrupt_free_min;
2073 vm_cnt.v_free_reserved = vm_pageout_page_count +
2074 vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
2075 vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
2076 vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
2077 vm_cnt.v_free_min += vm_cnt.v_free_reserved;
2078 vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
2079 vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
2080 if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
2081 vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
2084 * Set the default wakeup threshold to be 10% above the minimum
2085 * page limit. This keeps the steady state out of shortfall.
2087 vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
2090 * Set interval in seconds for active scan. We want to visit each
2091 * page at least once every ten minutes. This is to prevent worst
2092 * case paging behaviors with stale active LRU.
2094 if (vm_pageout_update_period == 0)
2095 vm_pageout_update_period = 600;
2097 /* XXX does not really belong here */
2098 if (vm_page_max_wired == 0)
2099 vm_page_max_wired = vm_cnt.v_free_count / 3;
2102 * Target amount of memory to move out of the laundry queue during a
2103 * background laundering. This is proportional to the amount of system
2106 vm_background_launder_target = (vm_cnt.v_free_target -
2107 vm_cnt.v_free_min) / 10;
2111 * vm_pageout is the high level pageout daemon.
2117 #ifdef VM_NUMA_ALLOC
2121 swap_pager_swap_init();
2122 error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
2123 0, 0, "laundry: dom0");
2125 panic("starting laundry for domain 0, error %d", error);
2126 #ifdef VM_NUMA_ALLOC
2127 for (i = 1; i < vm_ndomains; i++) {
2128 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
2129 curproc, NULL, 0, 0, "dom%d", i);
2131 panic("starting pageout for domain %d, error %d\n",
2136 error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
2139 panic("starting uma_reclaim helper, error %d\n", error);
2140 vm_pageout_worker((void *)(uintptr_t)0);
2144 * Unless the free page queue lock is held by the caller, this function
2145 * should be regarded as advisory. Specifically, the caller should
2146 * not msleep() on &vm_cnt.v_free_count following this function unless
2147 * the free page queue lock is held until the msleep() is performed.
2150 pagedaemon_wakeup(void)
2153 if (!vm_pageout_wanted && curthread->td_proc != pageproc) {
2154 vm_pageout_wanted = true;
2155 wakeup(&vm_pageout_wanted);
2159 #if !defined(NO_SWAPPING)
2161 vm_req_vmdaemon(int req)
2163 static int lastrun = 0;
2165 mtx_lock(&vm_daemon_mtx);
2166 vm_pageout_req_swapout |= req;
2167 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2168 wakeup(&vm_daemon_needed);
2171 mtx_unlock(&vm_daemon_mtx);
2177 struct rlimit rsslim;
2181 int breakout, swapout_flags, tryagain, attempts;
2183 uint64_t rsize, ravailable;
2187 mtx_lock(&vm_daemon_mtx);
2188 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
2190 racct_enable ? hz : 0
2195 swapout_flags = vm_pageout_req_swapout;
2196 vm_pageout_req_swapout = 0;
2197 mtx_unlock(&vm_daemon_mtx);
2199 swapout_procs(swapout_flags);
2202 * scan the processes for exceeding their rlimits or if
2203 * process is swapped out -- deactivate pages
2209 sx_slock(&allproc_lock);
2210 FOREACH_PROC_IN_SYSTEM(p) {
2211 vm_pindex_t limit, size;
2214 * if this is a system process or if we have already
2215 * looked at this process, skip it.
2218 if (p->p_state != PRS_NORMAL ||
2219 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
2224 * if the process is in a non-running type state,
2228 FOREACH_THREAD_IN_PROC(p, td) {
2230 if (!TD_ON_RUNQ(td) &&
2231 !TD_IS_RUNNING(td) &&
2232 !TD_IS_SLEEPING(td) &&
2233 !TD_IS_SUSPENDED(td)) {
2247 lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
2249 qmin(rsslim.rlim_cur, rsslim.rlim_max));
2252 * let processes that are swapped out really be
2253 * swapped out set the limit to nothing (will force a
2256 if ((p->p_flag & P_INMEM) == 0)
2257 limit = 0; /* XXX */
2258 vm = vmspace_acquire_ref(p);
2265 sx_sunlock(&allproc_lock);
2267 size = vmspace_resident_count(vm);
2268 if (size >= limit) {
2269 vm_pageout_map_deactivate_pages(
2270 &vm->vm_map, limit);
2271 size = vmspace_resident_count(vm);
2275 rsize = IDX_TO_OFF(size);
2277 if (p->p_state == PRS_NORMAL)
2278 racct_set(p, RACCT_RSS, rsize);
2279 ravailable = racct_get_available(p, RACCT_RSS);
2281 if (rsize > ravailable) {
2283 * Don't be overly aggressive; this
2284 * might be an innocent process,
2285 * and the limit could've been exceeded
2286 * by some memory hog. Don't try
2287 * to deactivate more than 1/4th
2288 * of process' resident set size.
2290 if (attempts <= 8) {
2291 if (ravailable < rsize -
2293 ravailable = rsize -
2297 vm_pageout_map_deactivate_pages(
2299 OFF_TO_IDX(ravailable));
2300 /* Update RSS usage after paging out. */
2301 size = vmspace_resident_count(vm);
2302 rsize = IDX_TO_OFF(size);
2304 if (p->p_state == PRS_NORMAL)
2305 racct_set(p, RACCT_RSS, rsize);
2307 if (rsize > ravailable)
2313 sx_slock(&allproc_lock);
2316 sx_sunlock(&allproc_lock);
2317 if (tryagain != 0 && attempts <= 10)
2321 #endif /* !defined(NO_SWAPPING) */