2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
64 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
68 * The proverbial page-out daemon.
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/kernel.h>
76 #include <sys/kthread.h>
77 #include <sys/resourcevar.h>
78 #include <sys/signalvar.h>
79 #include <sys/vnode.h>
80 #include <sys/vmmeter.h>
81 #include <sys/sysctl.h>
84 #include <vm/vm_param.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_pager.h>
91 #include <vm/swap_pager.h>
92 #include <vm/vm_extern.h>
94 #include <sys/thread2.h>
95 #include <sys/spinlock2.h>
96 #include <vm/vm_page2.h>
99 * System initialization
102 /* the kernel process "vm_pageout"*/
103 static int vm_pageout_clean (vm_page_t);
104 static int vm_pageout_free_page_calc (vm_size_t count);
105 struct thread *pagethread;
107 #if !defined(NO_SWAPPING)
108 /* the kernel process "vm_daemon"*/
109 static void vm_daemon (void);
110 static struct thread *vmthread;
112 static struct kproc_desc vm_kp = {
117 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
120 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */
121 int vm_pageout_deficit=0; /* Estimated number of pages deficit */
122 int vm_pageout_pages_needed=0; /* pageout daemon needs pages */
123 int vm_page_free_hysteresis = 16;
125 #if !defined(NO_SWAPPING)
126 static int vm_pageout_req_swapout; /* XXX */
127 static int vm_daemon_needed;
129 static int vm_max_launder = 4096;
130 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
131 static int vm_pageout_full_stats_interval = 0;
132 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
133 static int defer_swap_pageouts=0;
134 static int disable_swap_pageouts=0;
135 static u_int vm_anonmem_decline = ACT_DECLINE;
136 static u_int vm_filemem_decline = ACT_DECLINE * 2;
138 #if defined(NO_SWAPPING)
139 static int vm_swap_enabled=0;
140 static int vm_swap_idle_enabled=0;
142 static int vm_swap_enabled=1;
143 static int vm_swap_idle_enabled=0;
146 SYSCTL_UINT(_vm, VM_PAGEOUT_ALGORITHM, anonmem_decline,
147 CTLFLAG_RW, &vm_anonmem_decline, 0, "active->inactive anon memory");
149 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, filemem_decline,
150 CTLFLAG_RW, &vm_filemem_decline, 0, "active->inactive file cache");
152 SYSCTL_INT(_vm, OID_AUTO, page_free_hysteresis,
153 CTLFLAG_RW, &vm_page_free_hysteresis, 0,
154 "Free more pages than the minimum required");
156 SYSCTL_INT(_vm, OID_AUTO, max_launder,
157 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
159 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
160 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
162 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
163 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
165 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
166 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
168 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
169 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
171 #if defined(NO_SWAPPING)
172 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
173 CTLFLAG_RD, &vm_swap_enabled, 0, "");
174 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
175 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
178 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
180 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
183 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
184 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
186 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
187 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
189 static int pageout_lock_miss;
190 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
191 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
193 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
195 #if !defined(NO_SWAPPING)
196 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
197 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
198 static freeer_fcn_t vm_pageout_object_deactivate_pages;
199 static void vm_req_vmdaemon (void);
201 static void vm_pageout_page_stats(int q);
207 return((n + (PQ_L2_SIZE - 1)) / PQ_L2_SIZE + 1);
209 return((n - (PQ_L2_SIZE - 1)) / PQ_L2_SIZE - 1);
215 * Clean the page and remove it from the laundry. The page must not be
218 * We set the busy bit to cause potential page faults on this page to
219 * block. Note the careful timing, however, the busy bit isn't set till
220 * late and we cannot do anything that will mess with the page.
223 vm_pageout_clean(vm_page_t m)
226 vm_page_t mc[BLIST_MAX_ALLOC];
228 int ib, is, page_base;
229 vm_pindex_t pindex = m->pindex;
234 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
235 * with the new swapper, but we could have serious problems paging
236 * out other object types if there is insufficient memory.
238 * Unfortunately, checking free memory here is far too late, so the
239 * check has been moved up a procedural level.
243 * Don't mess with the page if it's busy, held, or special
245 * XXX do we really need to check hold_count here? hold_count
246 * isn't supposed to mess with vm_page ops except prevent the
247 * page from being reused.
249 if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
255 * Place page in cluster. Align cluster for optimal swap space
256 * allocation (whether it is swap or not). This is typically ~16-32
257 * pages, which also tends to align the cluster to multiples of the
258 * filesystem block size if backed by a filesystem.
260 page_base = pindex % BLIST_MAX_ALLOC;
266 * Scan object for clusterable pages.
268 * We can cluster ONLY if: ->> the page is NOT
269 * clean, wired, busy, held, or mapped into a
270 * buffer, and one of the following:
271 * 1) The page is inactive, or a seldom used
274 * 2) we force the issue.
276 * During heavy mmap/modification loads the pageout
277 * daemon can really fragment the underlying file
278 * due to flushing pages out of order and not trying
279 * align the clusters (which leave sporatic out-of-order
280 * holes). To solve this problem we do the reverse scan
281 * first and attempt to align our cluster, then do a
282 * forward scan if room remains.
285 vm_object_hold(object);
289 p = vm_page_lookup_busy_try(object, pindex - page_base + ib,
291 if (error || p == NULL)
293 if ((p->queue - p->pc) == PQ_CACHE ||
294 (p->flags & PG_UNMANAGED)) {
298 vm_page_test_dirty(p);
299 if (((p->dirty & p->valid) == 0 &&
300 (p->flags & PG_NEED_COMMIT) == 0) ||
301 p->queue - p->pc != PQ_INACTIVE ||
302 p->wire_count != 0 || /* may be held by buf cache */
303 p->hold_count != 0) { /* may be undergoing I/O */
312 while (is < BLIST_MAX_ALLOC &&
313 pindex - page_base + is < object->size) {
316 p = vm_page_lookup_busy_try(object, pindex - page_base + is,
318 if (error || p == NULL)
320 if (((p->queue - p->pc) == PQ_CACHE) ||
321 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
325 vm_page_test_dirty(p);
326 if (((p->dirty & p->valid) == 0 &&
327 (p->flags & PG_NEED_COMMIT) == 0) ||
328 p->queue - p->pc != PQ_INACTIVE ||
329 p->wire_count != 0 || /* may be held by buf cache */
330 p->hold_count != 0) { /* may be undergoing I/O */
338 vm_object_drop(object);
341 * we allow reads during pageouts...
343 return vm_pageout_flush(&mc[ib], is - ib, 0);
347 * vm_pageout_flush() - launder the given pages
349 * The given pages are laundered. Note that we setup for the start of
350 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
351 * reference count all in here rather then in the parent. If we want
352 * the parent to do more sophisticated things we may have to change
355 * The pages in the array must be busied by the caller and will be
356 * unbusied by this function.
359 vm_pageout_flush(vm_page_t *mc, int count, int flags)
362 int pageout_status[count];
367 * Initiate I/O. Bump the vm_page_t->busy counter.
369 for (i = 0; i < count; i++) {
370 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
371 ("vm_pageout_flush page %p index %d/%d: partially "
372 "invalid page", mc[i], i, count));
373 vm_page_io_start(mc[i]);
377 * We must make the pages read-only. This will also force the
378 * modified bit in the related pmaps to be cleared. The pager
379 * cannot clear the bit for us since the I/O completion code
380 * typically runs from an interrupt. The act of making the page
381 * read-only handles the case for us.
383 * Then we can unbusy the pages, we still hold a reference by virtue
386 for (i = 0; i < count; i++) {
387 vm_page_protect(mc[i], VM_PROT_READ);
388 vm_page_wakeup(mc[i]);
391 object = mc[0]->object;
392 vm_object_pip_add(object, count);
394 vm_pager_put_pages(object, mc, count,
395 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
398 for (i = 0; i < count; i++) {
399 vm_page_t mt = mc[i];
401 switch (pageout_status[i]) {
410 * Page outside of range of object. Right now we
411 * essentially lose the changes by pretending it
414 vm_page_busy_wait(mt, FALSE, "pgbad");
415 pmap_clear_modify(mt);
422 * A page typically cannot be paged out when we
423 * have run out of swap. We leave the page
424 * marked inactive and will try to page it out
427 * Starvation of the active page list is used to
428 * determine when the system is massively memory
437 * If the operation is still going, leave the page busy to
438 * block all other accesses. Also, leave the paging in
439 * progress indicator set so that we don't attempt an object
442 * For any pages which have completed synchronously,
443 * deactivate the page if we are under a severe deficit.
444 * Do not try to enter them into the cache, though, they
445 * might still be read-heavy.
447 if (pageout_status[i] != VM_PAGER_PEND) {
448 vm_page_busy_wait(mt, FALSE, "pgouw");
449 if (vm_page_count_severe())
450 vm_page_deactivate(mt);
452 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
453 vm_page_protect(mt, VM_PROT_READ);
455 vm_page_io_finish(mt);
457 vm_object_pip_wakeup(object);
463 #if !defined(NO_SWAPPING)
465 * deactivate enough pages to satisfy the inactive target
466 * requirements or if vm_page_proc_limit is set, then
467 * deactivate all of the pages in the object and its
470 * The map must be locked.
471 * The caller must hold the vm_object.
473 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
476 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
477 vm_pindex_t desired, int map_remove_only)
479 struct rb_vm_page_scan_info info;
484 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
488 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
490 if (lobject->type == OBJT_DEVICE ||
491 lobject->type == OBJT_MGTDEVICE ||
492 lobject->type == OBJT_PHYS)
494 if (lobject->paging_in_progress)
497 remove_mode = map_remove_only;
498 if (lobject->shadow_count > 1)
502 * scan the objects entire memory queue. We hold the
503 * object's token so the scan should not race anything.
505 info.limit = remove_mode;
507 info.desired = desired;
508 vm_page_rb_tree_RB_SCAN(&lobject->rb_memq, NULL,
509 vm_pageout_object_deactivate_pages_callback,
512 while ((tobject = lobject->backing_object) != NULL) {
513 KKASSERT(tobject != object);
514 vm_object_hold(tobject);
515 if (tobject == lobject->backing_object)
517 vm_object_drop(tobject);
519 if (lobject != object) {
521 vm_object_lock_swap();
522 vm_object_drop(lobject);
523 /* leaves tobject locked & at top */
527 if (lobject != object)
528 vm_object_drop(lobject); /* NULL ok */
532 * The caller must hold the vm_object.
535 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
537 struct rb_vm_page_scan_info *info = data;
540 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
543 mycpu->gd_cnt.v_pdpages++;
545 if (vm_page_busy_try(p, TRUE))
547 if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
551 if (!pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
556 actcount = pmap_ts_referenced(p);
558 vm_page_flag_set(p, PG_REFERENCED);
559 } else if (p->flags & PG_REFERENCED) {
563 vm_page_and_queue_spin_lock(p);
564 if (p->queue - p->pc != PQ_ACTIVE && (p->flags & PG_REFERENCED)) {
565 vm_page_and_queue_spin_unlock(p);
567 p->act_count += actcount;
568 vm_page_flag_clear(p, PG_REFERENCED);
569 } else if (p->queue - p->pc == PQ_ACTIVE) {
570 if ((p->flags & PG_REFERENCED) == 0) {
571 p->act_count -= min(p->act_count, ACT_DECLINE);
573 (vm_pageout_algorithm || (p->act_count == 0))) {
574 vm_page_and_queue_spin_unlock(p);
575 vm_page_protect(p, VM_PROT_NONE);
576 vm_page_deactivate(p);
578 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
580 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
582 vm_page_and_queue_spin_unlock(p);
585 vm_page_and_queue_spin_unlock(p);
587 vm_page_flag_clear(p, PG_REFERENCED);
589 vm_page_and_queue_spin_lock(p);
590 if (p->queue - p->pc == PQ_ACTIVE) {
591 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
592 p->act_count += ACT_ADVANCE;
593 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
595 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
598 vm_page_and_queue_spin_unlock(p);
600 } else if (p->queue - p->pc == PQ_INACTIVE) {
601 vm_page_and_queue_spin_unlock(p);
602 vm_page_protect(p, VM_PROT_NONE);
604 vm_page_and_queue_spin_unlock(p);
611 * Deactivate some number of pages in a map, try to do it fairly, but
612 * that is really hard to do.
615 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
618 vm_object_t obj, bigobj;
621 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
629 * first, search out the biggest object, and try to free pages from
632 tmpe = map->header.next;
633 while (tmpe != &map->header) {
634 switch(tmpe->maptype) {
635 case VM_MAPTYPE_NORMAL:
636 case VM_MAPTYPE_VPAGETABLE:
637 obj = tmpe->object.vm_object;
638 if ((obj != NULL) && (obj->shadow_count <= 1) &&
640 (bigobj->resident_page_count < obj->resident_page_count))) {
647 if (tmpe->wired_count > 0)
648 nothingwired = FALSE;
653 vm_object_hold(bigobj);
654 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
655 vm_object_drop(bigobj);
659 * Next, hunt around for other pages to deactivate. We actually
660 * do this search sort of wrong -- .text first is not the best idea.
662 tmpe = map->header.next;
663 while (tmpe != &map->header) {
664 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
666 switch(tmpe->maptype) {
667 case VM_MAPTYPE_NORMAL:
668 case VM_MAPTYPE_VPAGETABLE:
669 obj = tmpe->object.vm_object;
672 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
683 * Remove all mappings if a process is swapped out, this will free page
686 if (desired == 0 && nothingwired)
687 pmap_remove(vm_map_pmap(map),
688 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
694 * Called when the pageout scan wants to free a page. We no longer
695 * try to cycle the vm_object here with a reference & dealloc, which can
696 * cause a non-trivial object collapse in a critical path.
698 * It is unclear why we cycled the ref_count in the past, perhaps to try
699 * to optimize shadow chain collapses but I don't quite see why it would
700 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
701 * synchronously and not have to be kicked-start.
704 vm_pageout_page_free(vm_page_t m)
706 vm_page_protect(m, VM_PROT_NONE);
711 * vm_pageout_scan does the dirty work for the pageout daemon.
713 struct vm_pageout_scan_info {
714 struct proc *bigproc;
718 static int vm_pageout_scan_callback(struct proc *p, void *data);
721 vm_pageout_scan_inactive(int pass, int q, int avail_shortage,
722 int *vnodes_skippedp)
725 struct vm_page marker;
726 struct vnode *vpfailed; /* warning, allowed to be stale */
735 * Start scanning the inactive queue for pages we can move to the
736 * cache or free. The scan will stop when the target is reached or
737 * we have scanned the entire inactive queue. Note that m->act_count
738 * is not used to form decisions for the inactive queue, only for the
741 * maxlaunder limits the number of dirty pages we flush per scan.
742 * For most systems a smaller value (16 or 32) is more robust under
743 * extreme memory and disk pressure because any unnecessary writes
744 * to disk can result in extreme performance degredation. However,
745 * systems with excessive dirty pages (especially when MAP_NOSYNC is
746 * used) will die horribly with limited laundering. If the pageout
747 * daemon cannot clean enough pages in the first pass, we let it go
748 * all out in succeeding passes.
750 if ((maxlaunder = vm_max_launder) <= 1)
756 * Initialize our marker
758 bzero(&marker, sizeof(marker));
759 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
760 marker.queue = PQ_INACTIVE + q;
762 marker.wire_count = 1;
765 * Inactive queue scan.
767 * NOTE: The vm_page must be spinlocked before the queue to avoid
768 * deadlocks, so it is easiest to simply iterate the loop
769 * with the queue unlocked at the top.
773 vm_page_queues_spin_lock(PQ_INACTIVE + q);
774 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
775 maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt;
778 * Queue locked at top of loop to avoid stack marker issues.
780 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
781 maxscan-- > 0 && avail_shortage - delta > 0)
783 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
784 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
786 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
788 mycpu->gd_cnt.v_pdpages++;
791 * Skip marker pages (atomic against other markers to avoid
792 * infinite hop-over scans).
794 if (m->flags & PG_MARKER)
798 * Try to busy the page. Don't mess with pages which are
799 * already busy or reorder them in the queue.
801 if (vm_page_busy_try(m, TRUE))
805 * Remaining operations run with the page busy and neither
806 * the page or the queue will be spin-locked.
808 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
809 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
813 * It is possible for a page to be busied ad-hoc (e.g. the
814 * pmap_collect() code) and wired and race against the
815 * allocation of a new page. vm_page_alloc() may be forced
816 * to deactivate the wired page in which case it winds up
817 * on the inactive queue and must be handled here. We
818 * correct the problem simply by unqueuing the page.
821 vm_page_unqueue_nowakeup(m);
823 kprintf("WARNING: pagedaemon: wired page on "
824 "inactive queue %p\n", m);
829 * A held page may be undergoing I/O, so skip it.
832 vm_page_and_queue_spin_lock(m);
833 if (m->queue - m->pc == PQ_INACTIVE) {
835 &vm_page_queues[PQ_INACTIVE + q].pl,
838 &vm_page_queues[PQ_INACTIVE + q].pl,
840 ++vm_swapcache_inactive_heuristic;
842 vm_page_and_queue_spin_unlock(m);
847 if (m->object == NULL || m->object->ref_count == 0) {
849 * If the object is not being used, we ignore previous
852 vm_page_flag_clear(m, PG_REFERENCED);
853 pmap_clear_reference(m);
854 /* fall through to end */
855 } else if (((m->flags & PG_REFERENCED) == 0) &&
856 (actcount = pmap_ts_referenced(m))) {
858 * Otherwise, if the page has been referenced while
859 * in the inactive queue, we bump the "activation
860 * count" upwards, making it less likely that the
861 * page will be added back to the inactive queue
862 * prematurely again. Here we check the page tables
863 * (or emulated bits, if any), given the upper level
864 * VM system not knowing anything about existing
868 m->act_count += (actcount + ACT_ADVANCE);
874 * (m) is still busied.
876 * If the upper level VM system knows about any page
877 * references, we activate the page. We also set the
878 * "activation count" higher than normal so that we will less
879 * likely place pages back onto the inactive queue again.
881 if ((m->flags & PG_REFERENCED) != 0) {
882 vm_page_flag_clear(m, PG_REFERENCED);
883 actcount = pmap_ts_referenced(m);
885 m->act_count += (actcount + ACT_ADVANCE + 1);
891 * If the upper level VM system doesn't know anything about
892 * the page being dirty, we have to check for it again. As
893 * far as the VM code knows, any partially dirty pages are
896 * Pages marked PG_WRITEABLE may be mapped into the user
897 * address space of a process running on another cpu. A
898 * user process (without holding the MP lock) running on
899 * another cpu may be able to touch the page while we are
900 * trying to remove it. vm_page_cache() will handle this
904 vm_page_test_dirty(m);
909 if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
911 * Invalid pages can be easily freed
913 vm_pageout_page_free(m);
914 mycpu->gd_cnt.v_dfree++;
916 } else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
918 * Clean pages can be placed onto the cache queue.
919 * This effectively frees them.
923 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
925 * Dirty pages need to be paged out, but flushing
926 * a page is extremely expensive verses freeing
927 * a clean page. Rather then artificially limiting
928 * the number of pages we can flush, we instead give
929 * dirty pages extra priority on the inactive queue
930 * by forcing them to be cycled through the queue
931 * twice before being flushed, after which the
932 * (now clean) page will cycle through once more
933 * before being freed. This significantly extends
934 * the thrash point for a heavily loaded machine.
936 vm_page_flag_set(m, PG_WINATCFLS);
937 vm_page_and_queue_spin_lock(m);
938 if (m->queue - m->pc == PQ_INACTIVE) {
940 &vm_page_queues[PQ_INACTIVE + q].pl,
943 &vm_page_queues[PQ_INACTIVE + q].pl,
945 ++vm_swapcache_inactive_heuristic;
947 vm_page_and_queue_spin_unlock(m);
949 } else if (maxlaunder > 0) {
951 * We always want to try to flush some dirty pages if
952 * we encounter them, to keep the system stable.
953 * Normally this number is small, but under extreme
954 * pressure where there are insufficient clean pages
955 * on the inactive queue, we may have to go all out.
957 int swap_pageouts_ok;
958 struct vnode *vp = NULL;
960 swap_pageouts_ok = 0;
963 (object->type != OBJT_SWAP) &&
964 (object->type != OBJT_DEFAULT)) {
965 swap_pageouts_ok = 1;
967 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
968 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
969 vm_page_count_min(0));
974 * We don't bother paging objects that are "dead".
975 * Those objects are in a "rundown" state.
977 if (!swap_pageouts_ok ||
979 (object->flags & OBJ_DEAD)) {
980 vm_page_and_queue_spin_lock(m);
981 if (m->queue - m->pc == PQ_INACTIVE) {
983 &vm_page_queues[PQ_INACTIVE + q].pl,
986 &vm_page_queues[PQ_INACTIVE + q].pl,
988 ++vm_swapcache_inactive_heuristic;
990 vm_page_and_queue_spin_unlock(m);
996 * (m) is still busied.
998 * The object is already known NOT to be dead. It
999 * is possible for the vget() to block the whole
1000 * pageout daemon, but the new low-memory handling
1001 * code should prevent it.
1003 * The previous code skipped locked vnodes and, worse,
1004 * reordered pages in the queue. This results in
1005 * completely non-deterministic operation because,
1006 * quite often, a vm_fault has initiated an I/O and
1007 * is holding a locked vnode at just the point where
1008 * the pageout daemon is woken up.
1010 * We can't wait forever for the vnode lock, we might
1011 * deadlock due to a vn_read() getting stuck in
1012 * vm_wait while holding this vnode. We skip the
1013 * vnode if we can't get it in a reasonable amount
1016 * vpfailed is used to (try to) avoid the case where
1017 * a large number of pages are associated with a
1018 * locked vnode, which could cause the pageout daemon
1019 * to stall for an excessive amount of time.
1021 if (object->type == OBJT_VNODE) {
1024 vp = object->handle;
1025 flags = LK_EXCLUSIVE;
1029 flags |= LK_TIMELOCK;
1034 * We have unbusied (m) temporarily so we can
1035 * acquire the vp lock without deadlocking.
1036 * (m) is held to prevent destruction.
1038 if (vget(vp, flags) != 0) {
1040 ++pageout_lock_miss;
1041 if (object->flags & OBJ_MIGHTBEDIRTY)
1048 * The page might have been moved to another
1049 * queue during potential blocking in vget()
1050 * above. The page might have been freed and
1051 * reused for another vnode. The object might
1052 * have been reused for another vnode.
1054 if (m->queue - m->pc != PQ_INACTIVE ||
1055 m->object != object ||
1056 object->handle != vp) {
1057 if (object->flags & OBJ_MIGHTBEDIRTY)
1065 * The page may have been busied during the
1066 * blocking in vput(); We don't move the
1067 * page back onto the end of the queue so that
1068 * statistics are more correct if we don't.
1070 if (vm_page_busy_try(m, TRUE)) {
1078 * (m) is busied again
1080 * We own the busy bit and remove our hold
1081 * bit. If the page is still held it
1082 * might be undergoing I/O, so skip it.
1084 if (m->hold_count) {
1085 vm_page_and_queue_spin_lock(m);
1086 if (m->queue - m->pc == PQ_INACTIVE) {
1087 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1088 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1089 ++vm_swapcache_inactive_heuristic;
1091 vm_page_and_queue_spin_unlock(m);
1092 if (object->flags & OBJ_MIGHTBEDIRTY)
1098 /* (m) is left busied as we fall through */
1102 * page is busy and not held here.
1104 * If a page is dirty, then it is either being washed
1105 * (but not yet cleaned) or it is still in the
1106 * laundry. If it is still in the laundry, then we
1107 * start the cleaning operation.
1109 * decrement inactive_shortage on success to account
1110 * for the (future) cleaned page. Otherwise we
1111 * could wind up laundering or cleaning too many
1114 count = vm_pageout_clean(m);
1116 maxlaunder -= count;
1119 * Clean ate busy, page no longer accessible
1129 * Systems with a ton of memory can wind up with huge
1130 * deactivation counts. Because the inactive scan is
1131 * doing a lot of flushing, the combination can result
1132 * in excessive paging even in situations where other
1133 * unrelated threads free up sufficient VM.
1135 * To deal with this we abort the nominal active->inactive
1136 * scan before we hit the inactive target when free+cache
1137 * levels have already reached their target.
1139 * Note that nominally the inactive scan is not freeing or
1140 * caching pages, it is deactivating active pages, so it
1141 * will not by itself cause the abort condition.
1143 vm_page_queues_spin_lock(PQ_INACTIVE + q);
1144 if (vm_paging_target() < 0)
1148 /* page queue still spin-locked */
1149 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
1150 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
1156 vm_pageout_scan_active(int pass, int q,
1157 int avail_shortage, int inactive_shortage,
1158 int *recycle_countp)
1160 struct vm_page marker;
1167 * We want to move pages from the active queue to the inactive
1168 * queue to get the inactive queue to the inactive target. If
1169 * we still have a page shortage from above we try to directly free
1170 * clean pages instead of moving them.
1172 * If we do still have a shortage we keep track of the number of
1173 * pages we free or cache (recycle_count) as a measure of thrashing
1174 * between the active and inactive queues.
1176 * If we were able to completely satisfy the free+cache targets
1177 * from the inactive pool we limit the number of pages we move
1178 * from the active pool to the inactive pool to 2x the pages we
1179 * had removed from the inactive pool (with a minimum of 1/5 the
1180 * inactive target). If we were not able to completely satisfy
1181 * the free+cache targets we go for the whole target aggressively.
1183 * NOTE: Both variables can end up negative.
1184 * NOTE: We are still in a critical section.
1187 bzero(&marker, sizeof(marker));
1188 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1189 marker.queue = PQ_ACTIVE + q;
1191 marker.wire_count = 1;
1193 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1194 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1195 maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt;
1198 * Queue locked at top of loop to avoid stack marker issues.
1200 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1201 maxscan-- > 0 && (avail_shortage - delta > 0 ||
1202 inactive_shortage > 0))
1204 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1205 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1207 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1211 * Skip marker pages (atomic against other markers to avoid
1212 * infinite hop-over scans).
1214 if (m->flags & PG_MARKER)
1218 * Try to busy the page. Don't mess with pages which are
1219 * already busy or reorder them in the queue.
1221 if (vm_page_busy_try(m, TRUE))
1225 * Remaining operations run with the page busy and neither
1226 * the page or the queue will be spin-locked.
1228 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1229 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1233 * Don't deactivate pages that are held, even if we can
1234 * busy them. (XXX why not?)
1236 if (m->hold_count != 0) {
1237 vm_page_and_queue_spin_lock(m);
1238 if (m->queue - m->pc == PQ_ACTIVE) {
1240 &vm_page_queues[PQ_ACTIVE + q].pl,
1243 &vm_page_queues[PQ_ACTIVE + q].pl,
1246 vm_page_and_queue_spin_unlock(m);
1252 * The count for pagedaemon pages is done after checking the
1253 * page for eligibility...
1255 mycpu->gd_cnt.v_pdpages++;
1258 * Check to see "how much" the page has been used and clear
1259 * the tracking access bits. If the object has no references
1260 * don't bother paying the expense.
1263 if (m->object && m->object->ref_count != 0) {
1264 if (m->flags & PG_REFERENCED)
1266 actcount += pmap_ts_referenced(m);
1268 m->act_count += ACT_ADVANCE + actcount;
1269 if (m->act_count > ACT_MAX)
1270 m->act_count = ACT_MAX;
1273 vm_page_flag_clear(m, PG_REFERENCED);
1276 * actcount is only valid if the object ref_count is non-zero.
1277 * If the page does not have an object, actcount will be zero.
1279 if (actcount && m->object->ref_count != 0) {
1280 vm_page_and_queue_spin_lock(m);
1281 if (m->queue - m->pc == PQ_ACTIVE) {
1283 &vm_page_queues[PQ_ACTIVE + q].pl,
1286 &vm_page_queues[PQ_ACTIVE + q].pl,
1289 vm_page_and_queue_spin_unlock(m);
1292 switch(m->object->type) {
1295 m->act_count -= min(m->act_count,
1296 vm_anonmem_decline);
1299 m->act_count -= min(m->act_count,
1300 vm_filemem_decline);
1303 if (vm_pageout_algorithm ||
1304 (m->object == NULL) ||
1305 (m->object && (m->object->ref_count == 0)) ||
1306 m->act_count < pass + 1
1309 * Deactivate the page. If we had a
1310 * shortage from our inactive scan try to
1311 * free (cache) the page instead.
1313 * Don't just blindly cache the page if
1314 * we do not have a shortage from the
1315 * inactive scan, that could lead to
1316 * gigabytes being moved.
1318 --inactive_shortage;
1319 if (avail_shortage - delta > 0 ||
1320 (m->object && (m->object->ref_count == 0)))
1322 if (avail_shortage - delta > 0)
1324 vm_page_protect(m, VM_PROT_NONE);
1325 if (m->dirty == 0 &&
1326 (m->flags & PG_NEED_COMMIT) == 0 &&
1327 avail_shortage - delta > 0) {
1330 vm_page_deactivate(m);
1334 vm_page_deactivate(m);
1339 vm_page_and_queue_spin_lock(m);
1340 if (m->queue - m->pc == PQ_ACTIVE) {
1342 &vm_page_queues[PQ_ACTIVE + q].pl,
1345 &vm_page_queues[PQ_ACTIVE + q].pl,
1348 vm_page_and_queue_spin_unlock(m);
1353 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1357 * Clean out our local marker.
1359 * Page queue still spin-locked.
1361 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1362 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1368 * The number of actually free pages can drop down to v_free_reserved,
1369 * we try to build the free count back above v_free_min. Note that
1370 * vm_paging_needed() also returns TRUE if v_free_count is not at
1371 * least v_free_min so that is the minimum we must build the free
1374 * We use a slightly higher target to improve hysteresis,
1375 * ((v_free_target + v_free_min) / 2). Since v_free_target
1376 * is usually the same as v_cache_min this maintains about
1377 * half the pages in the free queue as are in the cache queue,
1378 * providing pretty good pipelining for pageout operation.
1380 * The system operator can manipulate vm.v_cache_min and
1381 * vm.v_free_target to tune the pageout demon. Be sure
1382 * to keep vm.v_free_min < vm.v_free_target.
1384 * Note that the original paging target is to get at least
1385 * (free_min + cache_min) into (free + cache). The slightly
1386 * higher target will shift additional pages from cache to free
1387 * without effecting the original paging target in order to
1388 * maintain better hysteresis and not have the free count always
1389 * be dead-on v_free_min.
1391 * NOTE: we are still in a critical section.
1393 * Pages moved from PQ_CACHE to totally free are not counted in the
1394 * pages_freed counter.
1397 vm_pageout_scan_cache(int avail_shortage, int vnodes_skipped, int recycle_count)
1399 struct vm_pageout_scan_info info;
1402 while (vmstats.v_free_count <
1403 (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1405 * This steals some code from vm/vm_page.c
1407 static int cache_rover = 0;
1409 m = vm_page_list_find(PQ_CACHE, cache_rover & PQ_L2_MASK, FALSE);
1412 /* page is returned removed from its queue and spinlocked */
1413 if (vm_page_busy_try(m, TRUE)) {
1414 vm_page_deactivate_locked(m);
1415 vm_page_spin_unlock(m);
1418 vm_page_spin_unlock(m);
1419 pagedaemon_wakeup();
1423 * Remaining operations run with the page busy and neither
1424 * the page or the queue will be spin-locked.
1426 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1429 vm_page_deactivate(m);
1433 KKASSERT((m->flags & PG_MAPPED) == 0);
1434 KKASSERT(m->dirty == 0);
1435 cache_rover += PQ_PRIME2;
1436 vm_pageout_page_free(m);
1437 mycpu->gd_cnt.v_dfree++;
1440 #if !defined(NO_SWAPPING)
1442 * Idle process swapout -- run once per second.
1444 if (vm_swap_idle_enabled) {
1446 if (time_uptime != lsec) {
1447 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1455 * If we didn't get enough free pages, and we have skipped a vnode
1456 * in a writeable object, wakeup the sync daemon. And kick swapout
1457 * if we did not get enough free pages.
1459 if (vm_paging_target() > 0) {
1460 if (vnodes_skipped && vm_page_count_min(0))
1461 speedup_syncer(NULL);
1462 #if !defined(NO_SWAPPING)
1463 if (vm_swap_enabled && vm_page_count_target()) {
1465 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1471 * Handle catastrophic conditions. Under good conditions we should
1472 * be at the target, well beyond our minimum. If we could not even
1473 * reach our minimum the system is under heavy stress.
1475 * Determine whether we have run out of memory. This occurs when
1476 * swap_pager_full is TRUE and the only pages left in the page
1477 * queues are dirty. We will still likely have page shortages.
1479 * - swap_pager_full is set if insufficient swap was
1480 * available to satisfy a requested pageout.
1482 * - the inactive queue is bloated (4 x size of active queue),
1483 * meaning it is unable to get rid of dirty pages and.
1485 * - vm_page_count_min() without counting pages recycled from the
1486 * active queue (recycle_count) means we could not recover
1487 * enough pages to meet bare minimum needs. This test only
1488 * works if the inactive queue is bloated.
1490 * - due to a positive avail_shortage we shifted the remaining
1491 * dirty pages from the active queue to the inactive queue
1492 * trying to find clean ones to free.
1494 if (swap_pager_full && vm_page_count_min(recycle_count))
1495 kprintf("Warning: system low on memory+swap!\n");
1496 if (swap_pager_full && vm_page_count_min(recycle_count) &&
1497 vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1498 avail_shortage > 0) {
1502 info.bigproc = NULL;
1504 allproc_scan(vm_pageout_scan_callback, &info);
1505 if (info.bigproc != NULL) {
1506 killproc(info.bigproc, "out of swap space");
1507 info.bigproc->p_nice = PRIO_MIN;
1508 info.bigproc->p_usched->resetpriority(
1509 FIRST_LWP_IN_PROC(info.bigproc));
1510 wakeup(&vmstats.v_free_count);
1511 PRELE(info.bigproc);
1517 vm_pageout_scan_callback(struct proc *p, void *data)
1519 struct vm_pageout_scan_info *info = data;
1523 * Never kill system processes or init. If we have configured swap
1524 * then try to avoid killing low-numbered pids.
1526 if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1527 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1531 lwkt_gettoken(&p->p_token);
1534 * if the process is in a non-running type state,
1537 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) {
1538 lwkt_reltoken(&p->p_token);
1543 * Get the approximate process size. Note that anonymous pages
1544 * with backing swap will be counted twice, but there should not
1545 * be too many such pages due to the stress the VM system is
1546 * under at this point.
1548 size = vmspace_anonymous_count(p->p_vmspace) +
1549 vmspace_swap_count(p->p_vmspace);
1552 * If the this process is bigger than the biggest one
1555 if (info->bigsize < size) {
1557 PRELE(info->bigproc);
1560 info->bigsize = size;
1562 lwkt_reltoken(&p->p_token);
1569 * This routine tries to maintain the pseudo LRU active queue,
1570 * so that during long periods of time where there is no paging,
1571 * that some statistic accumulation still occurs. This code
1572 * helps the situation where paging just starts to occur.
1575 vm_pageout_page_stats(int q)
1577 static int fullintervalcount = 0;
1578 struct vm_page marker;
1580 int pcount, tpcount; /* Number of pages to check */
1583 page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1584 vmstats.v_free_min) -
1585 (vmstats.v_free_count + vmstats.v_inactive_count +
1586 vmstats.v_cache_count);
1588 if (page_shortage <= 0)
1591 pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1592 fullintervalcount += vm_pageout_stats_interval;
1593 if (fullintervalcount < vm_pageout_full_stats_interval) {
1594 tpcount = (vm_pageout_stats_max * pcount) /
1595 vmstats.v_page_count + 1;
1596 if (pcount > tpcount)
1599 fullintervalcount = 0;
1602 bzero(&marker, sizeof(marker));
1603 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1604 marker.queue = PQ_ACTIVE + q;
1606 marker.wire_count = 1;
1608 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1609 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1612 * Queue locked at top of loop to avoid stack marker issues.
1614 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1619 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1620 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1621 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1625 * Skip marker pages (atomic against other markers to avoid
1626 * infinite hop-over scans).
1628 if (m->flags & PG_MARKER)
1632 * Ignore pages we can't busy
1634 if (vm_page_busy_try(m, TRUE))
1638 * Remaining operations run with the page busy and neither
1639 * the page or the queue will be spin-locked.
1641 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1642 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1645 * We now have a safely busied page, the page and queue
1646 * spinlocks have been released.
1650 if (m->hold_count) {
1656 * Calculate activity
1659 if (m->flags & PG_REFERENCED) {
1660 vm_page_flag_clear(m, PG_REFERENCED);
1663 actcount += pmap_ts_referenced(m);
1666 * Update act_count and move page to end of queue.
1669 m->act_count += ACT_ADVANCE + actcount;
1670 if (m->act_count > ACT_MAX)
1671 m->act_count = ACT_MAX;
1672 vm_page_and_queue_spin_lock(m);
1673 if (m->queue - m->pc == PQ_ACTIVE) {
1675 &vm_page_queues[PQ_ACTIVE + q].pl,
1678 &vm_page_queues[PQ_ACTIVE + q].pl,
1681 vm_page_and_queue_spin_unlock(m);
1686 if (m->act_count == 0) {
1688 * We turn off page access, so that we have
1689 * more accurate RSS stats. We don't do this
1690 * in the normal page deactivation when the
1691 * system is loaded VM wise, because the
1692 * cost of the large number of page protect
1693 * operations would be higher than the value
1694 * of doing the operation.
1696 * We use the marker to save our place so
1697 * we can release the spin lock. both (m)
1698 * and (next) will be invalid.
1700 vm_page_protect(m, VM_PROT_NONE);
1701 vm_page_deactivate(m);
1703 m->act_count -= min(m->act_count, ACT_DECLINE);
1704 vm_page_and_queue_spin_lock(m);
1705 if (m->queue - m->pc == PQ_ACTIVE) {
1707 &vm_page_queues[PQ_ACTIVE + q].pl,
1710 &vm_page_queues[PQ_ACTIVE + q].pl,
1713 vm_page_and_queue_spin_unlock(m);
1717 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1721 * Remove our local marker
1723 * Page queue still spin-locked.
1725 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1726 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1730 vm_pageout_free_page_calc(vm_size_t count)
1732 if (count < vmstats.v_page_count)
1735 * free_reserved needs to include enough for the largest swap pager
1736 * structures plus enough for any pv_entry structs when paging.
1738 * v_free_min normal allocations
1739 * v_free_reserved system allocations
1740 * v_pageout_free_min allocations by pageout daemon
1741 * v_interrupt_free_min low level allocations (e.g swap structures)
1743 if (vmstats.v_page_count > 1024)
1744 vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1746 vmstats.v_free_min = 64;
1747 vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1748 vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1749 vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1750 vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1757 * vm_pageout is the high level pageout daemon.
1762 vm_pageout_thread(void)
1770 * Initialize some paging parameters.
1772 curthread->td_flags |= TDF_SYSTHREAD;
1774 vm_pageout_free_page_calc(vmstats.v_page_count);
1777 * v_free_target and v_cache_min control pageout hysteresis. Note
1778 * that these are more a measure of the VM cache queue hysteresis
1779 * then the VM free queue. Specifically, v_free_target is the
1780 * high water mark (free+cache pages).
1782 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1783 * low water mark, while v_free_min is the stop. v_cache_min must
1784 * be big enough to handle memory needs while the pageout daemon
1785 * is signalled and run to free more pages.
1787 if (vmstats.v_free_count > 6144)
1788 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1790 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1793 * NOTE: With the new buffer cache b_act_count we want the default
1794 * inactive target to be a percentage of available memory.
1796 * The inactive target essentially determines the minimum
1797 * number of 'temporary' pages capable of caching one-time-use
1798 * files when the VM system is otherwise full of pages
1799 * belonging to multi-time-use files or active program data.
1801 * NOTE: The inactive target is aggressively persued only if the
1802 * inactive queue becomes too small. If the inactive queue
1803 * is large enough to satisfy page movement to free+cache
1804 * then it is repopulated more slowly from the active queue.
1805 * This allows a general inactive_target default to be set.
1807 * There is an issue here for processes which sit mostly idle
1808 * 'overnight', such as sshd, tcsh, and X. Any movement from
1809 * the active queue will eventually cause such pages to
1810 * recycle eventually causing a lot of paging in the morning.
1811 * To reduce the incidence of this pages cycled out of the
1812 * buffer cache are moved directly to the inactive queue if
1813 * they were only used once or twice.
1815 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1816 * Increasing the value (up to 64) increases the number of
1817 * buffer recyclements which go directly to the inactive queue.
1819 if (vmstats.v_free_count > 2048) {
1820 vmstats.v_cache_min = vmstats.v_free_target;
1821 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1823 vmstats.v_cache_min = 0;
1824 vmstats.v_cache_max = 0;
1826 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1828 /* XXX does not really belong here */
1829 if (vm_page_max_wired == 0)
1830 vm_page_max_wired = vmstats.v_free_count / 3;
1832 if (vm_pageout_stats_max == 0)
1833 vm_pageout_stats_max = vmstats.v_free_target;
1836 * Set interval in seconds for stats scan.
1838 if (vm_pageout_stats_interval == 0)
1839 vm_pageout_stats_interval = 5;
1840 if (vm_pageout_full_stats_interval == 0)
1841 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1845 * Set maximum free per pass
1847 if (vm_pageout_stats_free_max == 0)
1848 vm_pageout_stats_free_max = 5;
1850 swap_pager_swap_init();
1854 * The pageout daemon is never done, so loop forever.
1859 int inactive_shortage;
1860 int vnodes_skipped = 0;
1861 int recycle_count = 0;
1865 * Wait for an action request. If we timeout check to
1866 * see if paging is needed (in case the normal wakeup
1869 if (vm_pages_needed == 0) {
1870 error = tsleep(&vm_pages_needed,
1872 vm_pageout_stats_interval * hz);
1874 vm_paging_needed() == 0 &&
1875 vm_pages_needed == 0) {
1876 for (q = 0; q < PQ_L2_SIZE; ++q)
1877 vm_pageout_page_stats(q);
1880 vm_pages_needed = 1;
1883 mycpu->gd_cnt.v_pdwakeups++;
1886 * Do whatever cleanup that the pmap code can.
1891 * Scan for pageout. Try to avoid thrashing the system
1894 * Calculate our target for the number of free+cache pages we
1895 * want to get to. This is higher then the number that causes
1896 * allocations to stall (severe) in order to provide hysteresis,
1897 * and if we don't make it all the way but get to the minimum
1898 * we're happy. Goose it a bit if there are multipler
1899 * requests for memory.
1901 avail_shortage = vm_paging_target() + vm_pageout_deficit;
1902 vm_pageout_deficit = 0;
1904 if (avail_shortage > 0) {
1905 for (q = 0; q < PQ_L2_SIZE; ++q) {
1907 vm_pageout_scan_inactive(
1909 (q + q1iterator) & PQ_L2_MASK,
1910 PQAVERAGE(avail_shortage),
1912 if (avail_shortage <= 0)
1919 * Figure out how many active pages we must deactivate. If
1920 * we were able to reach our target with just the inactive
1921 * scan above we limit the number of active pages we
1922 * deactivate to reduce unnecessary work.
1924 inactive_shortage = vmstats.v_inactive_target -
1925 vmstats.v_inactive_count;
1928 * If we were unable to free sufficient inactive pages to
1929 * satisfy the free/cache queue requirements then simply
1930 * reaching the inactive target may not be good enough.
1931 * Try to deactivate pages in excess of the target based
1934 * However to prevent thrashing the VM system do not
1935 * deactivate more than an additional 1/10 the inactive
1936 * target's worth of active pages.
1938 if (avail_shortage > 0) {
1939 tmp = avail_shortage * 2;
1940 if (tmp > vmstats.v_inactive_target / 10)
1941 tmp = vmstats.v_inactive_target / 10;
1942 inactive_shortage += tmp;
1946 * Only trigger on inactive shortage. Triggering on
1947 * avail_shortage can starve the active queue with
1948 * unnecessary active->inactive transitions and destroy
1951 if (/*avail_shortage > 0 ||*/ inactive_shortage > 0) {
1954 for (q = 0; q < PQ_L2_SIZE; ++q) {
1955 delta = vm_pageout_scan_active(
1957 (q + q2iterator) & PQ_L2_MASK,
1958 PQAVERAGE(avail_shortage),
1959 PQAVERAGE(inactive_shortage),
1961 inactive_shortage -= delta;
1962 avail_shortage -= delta;
1963 if (inactive_shortage <= 0 &&
1964 avail_shortage <= 0) {
1972 * Finally free enough cache pages to meet our free page
1973 * requirement and take more drastic measures if we are
1976 vm_pageout_scan_cache(avail_shortage, vnodes_skipped,
1980 * Wait for more work.
1982 if (avail_shortage > 0) {
1984 if (swap_pager_full) {
1986 * Running out of memory, catastrophic back-off
1987 * to one-second intervals.
1989 tsleep(&vm_pages_needed, 0, "pdelay", hz);
1990 } else if (pass < 10 && vm_pages_needed > 1) {
1992 * Normal operation, additional processes
1993 * have already kicked us. Retry immediately.
1995 } else if (pass < 10) {
1997 * Normal operation, fewer processes. Delay
1998 * a bit but allow wakeups.
2000 vm_pages_needed = 0;
2001 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2002 vm_pages_needed = 1;
2005 * We've taken too many passes, forced delay.
2007 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2009 } else if (vm_pages_needed) {
2011 * Interlocked wakeup of waiters (non-optional).
2013 * Similar to vm_page_free_wakeup() in vm_page.c,
2017 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2018 !vm_page_count_target()) {
2019 vm_pages_needed = 0;
2020 wakeup(&vmstats.v_free_count);
2028 static struct kproc_desc page_kp = {
2033 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
2037 * Called after allocating a page out of the cache or free queue
2038 * to possibly wake the pagedaemon up to replentish our supply.
2040 * We try to generate some hysteresis by waking the pagedaemon up
2041 * when our free+cache pages go below the free_min+cache_min level.
2042 * The pagedaemon tries to get the count back up to at least the
2043 * minimum, and through to the target level if possible.
2045 * If the pagedaemon is already active bump vm_pages_needed as a hint
2046 * that there are even more requests pending.
2052 pagedaemon_wakeup(void)
2054 if (vm_paging_needed() && curthread != pagethread) {
2055 if (vm_pages_needed == 0) {
2056 vm_pages_needed = 1; /* SMP race ok */
2057 wakeup(&vm_pages_needed);
2058 } else if (vm_page_count_min(0)) {
2059 ++vm_pages_needed; /* SMP race ok */
2064 #if !defined(NO_SWAPPING)
2071 vm_req_vmdaemon(void)
2073 static int lastrun = 0;
2075 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2076 wakeup(&vm_daemon_needed);
2081 static int vm_daemon_callback(struct proc *p, void *data __unused);
2090 * XXX vm_daemon_needed specific token?
2093 tsleep(&vm_daemon_needed, 0, "psleep", 0);
2094 if (vm_pageout_req_swapout) {
2095 swapout_procs(vm_pageout_req_swapout);
2096 vm_pageout_req_swapout = 0;
2099 * scan the processes for exceeding their rlimits or if
2100 * process is swapped out -- deactivate pages
2102 allproc_scan(vm_daemon_callback, NULL);
2107 vm_daemon_callback(struct proc *p, void *data __unused)
2110 vm_pindex_t limit, size;
2113 * if this is a system process or if we have already
2114 * looked at this process, skip it.
2116 lwkt_gettoken(&p->p_token);
2118 if (p->p_flags & (P_SYSTEM | P_WEXIT)) {
2119 lwkt_reltoken(&p->p_token);
2124 * if the process is in a non-running type state,
2127 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) {
2128 lwkt_reltoken(&p->p_token);
2135 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2136 p->p_rlimit[RLIMIT_RSS].rlim_max));
2139 * let processes that are swapped out really be
2140 * swapped out. Set the limit to nothing to get as
2141 * many pages out to swap as possible.
2143 if (p->p_flags & P_SWAPPEDOUT)
2148 size = vmspace_resident_count(vm);
2149 if (limit >= 0 && size >= limit) {
2150 vm_pageout_map_deactivate_pages(&vm->vm_map, limit);
2154 lwkt_reltoken(&p->p_token);