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 * 4. 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; /* flag saying that the pageout daemon needs pages */
124 #if !defined(NO_SWAPPING)
125 static int vm_pageout_req_swapout; /* XXX */
126 static int vm_daemon_needed;
128 static int vm_max_launder = 32;
129 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
130 static int vm_pageout_full_stats_interval = 0;
131 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
132 static int defer_swap_pageouts=0;
133 static int disable_swap_pageouts=0;
134 static u_int vm_anonmem_decline = ACT_DECLINE;
135 static u_int vm_filemem_decline = ACT_DECLINE * 2;
137 #if defined(NO_SWAPPING)
138 static int vm_swap_enabled=0;
139 static int vm_swap_idle_enabled=0;
141 static int vm_swap_enabled=1;
142 static int vm_swap_idle_enabled=0;
145 SYSCTL_UINT(_vm, VM_PAGEOUT_ALGORITHM, anonmem_decline,
146 CTLFLAG_RW, &vm_anonmem_decline, 0, "active->inactive anon memory");
148 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, filemem_decline,
149 CTLFLAG_RW, &vm_filemem_decline, 0, "active->inactive file cache");
151 SYSCTL_INT(_vm, OID_AUTO, max_launder,
152 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
154 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
155 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
157 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
158 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
160 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
161 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
163 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
164 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
166 #if defined(NO_SWAPPING)
167 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
168 CTLFLAG_RD, &vm_swap_enabled, 0, "");
169 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
170 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
172 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
173 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
174 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
175 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
178 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
179 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
181 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
182 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
184 static int pageout_lock_miss;
185 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
186 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
188 #define VM_PAGEOUT_PAGE_COUNT 16
189 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
191 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
193 #if !defined(NO_SWAPPING)
194 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
195 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
196 static freeer_fcn_t vm_pageout_object_deactivate_pages;
197 static void vm_req_vmdaemon (void);
199 static void vm_pageout_page_stats(int q);
205 return((n + (PQ_L2_SIZE - 1)) / PQ_L2_SIZE + 1);
207 return((n - (PQ_L2_SIZE - 1)) / PQ_L2_SIZE - 1);
213 * Clean the page and remove it from the laundry. The page must not be
216 * We set the busy bit to cause potential page faults on this page to
217 * block. Note the careful timing, however, the busy bit isn't set till
218 * late and we cannot do anything that will mess with the page.
221 vm_pageout_clean(vm_page_t m)
224 vm_page_t mc[2*vm_pageout_page_count];
227 int ib, is, page_base;
228 vm_pindex_t pindex = m->pindex;
233 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
234 * with the new swapper, but we could have serious problems paging
235 * out other object types if there is insufficient memory.
237 * Unfortunately, checking free memory here is far too late, so the
238 * check has been moved up a procedural level.
242 * Don't mess with the page if it's busy, held, or special
244 * XXX do we really need to check hold_count here? hold_count
245 * isn't supposed to mess with vm_page ops except prevent the
246 * page from being reused.
248 if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
253 mc[vm_pageout_page_count] = m;
255 page_base = vm_pageout_page_count;
260 * Scan object for clusterable pages.
262 * We can cluster ONLY if: ->> the page is NOT
263 * clean, wired, busy, held, or mapped into a
264 * buffer, and one of the following:
265 * 1) The page is inactive, or a seldom used
268 * 2) we force the issue.
270 * During heavy mmap/modification loads the pageout
271 * daemon can really fragment the underlying file
272 * due to flushing pages out of order and not trying
273 * align the clusters (which leave sporatic out-of-order
274 * holes). To solve this problem we do the reverse scan
275 * first and attempt to align our cluster, then do a
276 * forward scan if room remains.
279 vm_object_hold(object);
281 while (ib && pageout_count < vm_pageout_page_count) {
289 p = vm_page_lookup_busy_try(object, pindex - ib, TRUE, &error);
290 if (error || p == NULL) {
294 if ((p->queue - p->pc) == PQ_CACHE ||
295 (p->flags & PG_UNMANAGED)) {
300 vm_page_test_dirty(p);
301 if (((p->dirty & p->valid) == 0 &&
302 (p->flags & PG_NEED_COMMIT) == 0) ||
303 p->queue - p->pc != PQ_INACTIVE ||
304 p->wire_count != 0 || /* may be held by buf cache */
305 p->hold_count != 0) { /* may be undergoing I/O */
314 * alignment boundry, stop here and switch directions. Do
317 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
321 while (pageout_count < vm_pageout_page_count &&
322 pindex + is < object->size) {
325 p = vm_page_lookup_busy_try(object, pindex + is, TRUE, &error);
326 if (error || p == NULL)
328 if (((p->queue - p->pc) == PQ_CACHE) ||
329 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
333 vm_page_test_dirty(p);
334 if (((p->dirty & p->valid) == 0 &&
335 (p->flags & PG_NEED_COMMIT) == 0) ||
336 p->queue - p->pc != PQ_INACTIVE ||
337 p->wire_count != 0 || /* may be held by buf cache */
338 p->hold_count != 0) { /* may be undergoing I/O */
342 mc[page_base + pageout_count] = p;
348 * If we exhausted our forward scan, continue with the reverse scan
349 * when possible, even past a page boundry. This catches boundry
352 if (ib && pageout_count < vm_pageout_page_count)
355 vm_object_drop(object);
358 * we allow reads during pageouts...
360 return vm_pageout_flush(&mc[page_base], pageout_count, 0);
364 * vm_pageout_flush() - launder the given pages
366 * The given pages are laundered. Note that we setup for the start of
367 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
368 * reference count all in here rather then in the parent. If we want
369 * the parent to do more sophisticated things we may have to change
372 * The pages in the array must be busied by the caller and will be
373 * unbusied by this function.
376 vm_pageout_flush(vm_page_t *mc, int count, int flags)
379 int pageout_status[count];
384 * Initiate I/O. Bump the vm_page_t->busy counter.
386 for (i = 0; i < count; i++) {
387 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
388 ("vm_pageout_flush page %p index %d/%d: partially "
389 "invalid page", mc[i], i, count));
390 vm_page_io_start(mc[i]);
394 * We must make the pages read-only. This will also force the
395 * modified bit in the related pmaps to be cleared. The pager
396 * cannot clear the bit for us since the I/O completion code
397 * typically runs from an interrupt. The act of making the page
398 * read-only handles the case for us.
400 * Then we can unbusy the pages, we still hold a reference by virtue
403 for (i = 0; i < count; i++) {
404 vm_page_protect(mc[i], VM_PROT_READ);
405 vm_page_wakeup(mc[i]);
408 object = mc[0]->object;
409 vm_object_pip_add(object, count);
411 vm_pager_put_pages(object, mc, count,
412 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
415 for (i = 0; i < count; i++) {
416 vm_page_t mt = mc[i];
418 switch (pageout_status[i]) {
427 * Page outside of range of object. Right now we
428 * essentially lose the changes by pretending it
431 vm_page_busy_wait(mt, FALSE, "pgbad");
432 pmap_clear_modify(mt);
439 * A page typically cannot be paged out when we
440 * have run out of swap. We leave the page
441 * marked inactive and will try to page it out
444 * Starvation of the active page list is used to
445 * determine when the system is massively memory
454 * If the operation is still going, leave the page busy to
455 * block all other accesses. Also, leave the paging in
456 * progress indicator set so that we don't attempt an object
459 * For any pages which have completed synchronously,
460 * deactivate the page if we are under a severe deficit.
461 * Do not try to enter them into the cache, though, they
462 * might still be read-heavy.
464 if (pageout_status[i] != VM_PAGER_PEND) {
465 vm_page_busy_wait(mt, FALSE, "pgouw");
466 if (vm_page_count_severe())
467 vm_page_deactivate(mt);
469 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
470 vm_page_protect(mt, VM_PROT_READ);
472 vm_page_io_finish(mt);
474 vm_object_pip_wakeup(object);
480 #if !defined(NO_SWAPPING)
482 * deactivate enough pages to satisfy the inactive target
483 * requirements or if vm_page_proc_limit is set, then
484 * deactivate all of the pages in the object and its
487 * The map must be locked.
488 * The caller must hold the vm_object.
490 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
493 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
494 vm_pindex_t desired, int map_remove_only)
496 struct rb_vm_page_scan_info info;
501 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
505 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
507 if (lobject->type == OBJT_DEVICE || lobject->type == OBJT_PHYS)
509 if (lobject->paging_in_progress)
512 remove_mode = map_remove_only;
513 if (lobject->shadow_count > 1)
517 * scan the objects entire memory queue. We hold the
518 * object's token so the scan should not race anything.
520 info.limit = remove_mode;
522 info.desired = desired;
523 vm_page_rb_tree_RB_SCAN(&lobject->rb_memq, NULL,
524 vm_pageout_object_deactivate_pages_callback,
527 while ((tobject = lobject->backing_object) != NULL) {
528 KKASSERT(tobject != object);
529 vm_object_hold(tobject);
530 if (tobject == lobject->backing_object)
532 vm_object_drop(tobject);
534 if (lobject != object) {
536 vm_object_lock_swap();
537 vm_object_drop(lobject);
538 /* leaves tobject locked & at top */
542 if (lobject != object)
543 vm_object_drop(lobject); /* NULL ok */
547 * The caller must hold the vm_object.
550 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
552 struct rb_vm_page_scan_info *info = data;
555 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
558 mycpu->gd_cnt.v_pdpages++;
560 if (vm_page_busy_try(p, TRUE))
562 if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
566 if (!pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
571 actcount = pmap_ts_referenced(p);
573 vm_page_flag_set(p, PG_REFERENCED);
574 } else if (p->flags & PG_REFERENCED) {
578 vm_page_and_queue_spin_lock(p);
579 if (p->queue - p->pc != PQ_ACTIVE && (p->flags & PG_REFERENCED)) {
580 vm_page_and_queue_spin_unlock(p);
582 p->act_count += actcount;
583 vm_page_flag_clear(p, PG_REFERENCED);
584 } else if (p->queue - p->pc == PQ_ACTIVE) {
585 if ((p->flags & PG_REFERENCED) == 0) {
586 p->act_count -= min(p->act_count, ACT_DECLINE);
588 (vm_pageout_algorithm || (p->act_count == 0))) {
589 vm_page_and_queue_spin_unlock(p);
590 vm_page_protect(p, VM_PROT_NONE);
591 vm_page_deactivate(p);
593 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
595 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
597 vm_page_and_queue_spin_unlock(p);
600 vm_page_and_queue_spin_unlock(p);
602 vm_page_flag_clear(p, PG_REFERENCED);
604 vm_page_and_queue_spin_lock(p);
605 if (p->queue - p->pc == PQ_ACTIVE) {
606 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
607 p->act_count += ACT_ADVANCE;
608 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
610 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
613 vm_page_and_queue_spin_unlock(p);
615 } else if (p->queue - p->pc == PQ_INACTIVE) {
616 vm_page_and_queue_spin_unlock(p);
617 vm_page_protect(p, VM_PROT_NONE);
619 vm_page_and_queue_spin_unlock(p);
626 * Deactivate some number of pages in a map, try to do it fairly, but
627 * that is really hard to do.
630 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
633 vm_object_t obj, bigobj;
636 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
644 * first, search out the biggest object, and try to free pages from
647 tmpe = map->header.next;
648 while (tmpe != &map->header) {
649 switch(tmpe->maptype) {
650 case VM_MAPTYPE_NORMAL:
651 case VM_MAPTYPE_VPAGETABLE:
652 obj = tmpe->object.vm_object;
653 if ((obj != NULL) && (obj->shadow_count <= 1) &&
655 (bigobj->resident_page_count < obj->resident_page_count))) {
662 if (tmpe->wired_count > 0)
663 nothingwired = FALSE;
668 vm_object_hold(bigobj);
669 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
670 vm_object_drop(bigobj);
674 * Next, hunt around for other pages to deactivate. We actually
675 * do this search sort of wrong -- .text first is not the best idea.
677 tmpe = map->header.next;
678 while (tmpe != &map->header) {
679 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
681 switch(tmpe->maptype) {
682 case VM_MAPTYPE_NORMAL:
683 case VM_MAPTYPE_VPAGETABLE:
684 obj = tmpe->object.vm_object;
687 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
698 * Remove all mappings if a process is swapped out, this will free page
701 if (desired == 0 && nothingwired)
702 pmap_remove(vm_map_pmap(map),
703 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
709 * Called when the pageout scan wants to free a page. We no longer
710 * try to cycle the vm_object here with a reference & dealloc, which can
711 * cause a non-trivial object collapse in a critical path.
713 * It is unclear why we cycled the ref_count in the past, perhaps to try
714 * to optimize shadow chain collapses but I don't quite see why it would
715 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
716 * synchronously and not have to be kicked-start.
719 vm_pageout_page_free(vm_page_t m)
721 vm_page_protect(m, VM_PROT_NONE);
726 * vm_pageout_scan does the dirty work for the pageout daemon.
728 struct vm_pageout_scan_info {
729 struct proc *bigproc;
733 static int vm_pageout_scan_callback(struct proc *p, void *data);
736 vm_pageout_scan_inactive(int pass, int q, int avail_shortage,
737 int *vnodes_skippedp)
740 struct vm_page marker;
741 struct vnode *vpfailed; /* warning, allowed to be stale */
749 * Start scanning the inactive queue for pages we can move to the
750 * cache or free. The scan will stop when the target is reached or
751 * we have scanned the entire inactive queue. Note that m->act_count
752 * is not used to form decisions for the inactive queue, only for the
755 * maxlaunder limits the number of dirty pages we flush per scan.
756 * For most systems a smaller value (16 or 32) is more robust under
757 * extreme memory and disk pressure because any unnecessary writes
758 * to disk can result in extreme performance degredation. However,
759 * systems with excessive dirty pages (especially when MAP_NOSYNC is
760 * used) will die horribly with limited laundering. If the pageout
761 * daemon cannot clean enough pages in the first pass, we let it go
762 * all out in succeeding passes.
764 if ((maxlaunder = vm_max_launder) <= 1)
770 * Initialize our marker
772 bzero(&marker, sizeof(marker));
773 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
774 marker.queue = PQ_INACTIVE + q;
776 marker.wire_count = 1;
779 * Inactive queue scan.
781 * NOTE: The vm_page must be spinlocked before the queue to avoid
782 * deadlocks, so it is easiest to simply iterate the loop
783 * with the queue unlocked at the top.
787 vm_page_queues_spin_lock(PQ_INACTIVE + q);
788 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
789 maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt;
790 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
792 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
793 maxscan-- > 0 && avail_shortage - delta > 0)
795 vm_page_and_queue_spin_lock(m);
796 if (m != TAILQ_NEXT(&marker, pageq)) {
797 vm_page_and_queue_spin_unlock(m);
801 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
802 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
804 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
806 mycpu->gd_cnt.v_pdpages++;
811 if (m->flags & PG_MARKER) {
812 vm_page_and_queue_spin_unlock(m);
817 * Try to busy the page. Don't mess with pages which are
818 * already busy or reorder them in the queue.
820 if (vm_page_busy_try(m, TRUE)) {
821 vm_page_and_queue_spin_unlock(m);
824 vm_page_and_queue_spin_unlock(m);
825 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
830 * The page has been successfully busied and is now no
831 * longer spinlocked. The queue is no longer spinlocked
836 * It is possible for a page to be busied ad-hoc (e.g. the
837 * pmap_collect() code) and wired and race against the
838 * allocation of a new page. vm_page_alloc() may be forced
839 * to deactivate the wired page in which case it winds up
840 * on the inactive queue and must be handled here. We
841 * correct the problem simply by unqueuing the page.
844 vm_page_unqueue_nowakeup(m);
846 kprintf("WARNING: pagedaemon: wired page on "
847 "inactive queue %p\n", m);
852 * A held page may be undergoing I/O, so skip it.
855 vm_page_and_queue_spin_lock(m);
856 if (m->queue - m->pc == PQ_INACTIVE) {
858 &vm_page_queues[PQ_INACTIVE + q].pl,
861 &vm_page_queues[PQ_INACTIVE + q].pl,
863 ++vm_swapcache_inactive_heuristic;
865 vm_page_and_queue_spin_unlock(m);
870 if (m->object == NULL || m->object->ref_count == 0) {
872 * If the object is not being used, we ignore previous
875 vm_page_flag_clear(m, PG_REFERENCED);
876 pmap_clear_reference(m);
877 /* fall through to end */
878 } else if (((m->flags & PG_REFERENCED) == 0) &&
879 (actcount = pmap_ts_referenced(m))) {
881 * Otherwise, if the page has been referenced while
882 * in the inactive queue, we bump the "activation
883 * count" upwards, making it less likely that the
884 * page will be added back to the inactive queue
885 * prematurely again. Here we check the page tables
886 * (or emulated bits, if any), given the upper level
887 * VM system not knowing anything about existing
891 m->act_count += (actcount + ACT_ADVANCE);
897 * (m) is still busied.
899 * If the upper level VM system knows about any page
900 * references, we activate the page. We also set the
901 * "activation count" higher than normal so that we will less
902 * likely place pages back onto the inactive queue again.
904 if ((m->flags & PG_REFERENCED) != 0) {
905 vm_page_flag_clear(m, PG_REFERENCED);
906 actcount = pmap_ts_referenced(m);
908 m->act_count += (actcount + ACT_ADVANCE + 1);
914 * If the upper level VM system doesn't know anything about
915 * the page being dirty, we have to check for it again. As
916 * far as the VM code knows, any partially dirty pages are
919 * Pages marked PG_WRITEABLE may be mapped into the user
920 * address space of a process running on another cpu. A
921 * user process (without holding the MP lock) running on
922 * another cpu may be able to touch the page while we are
923 * trying to remove it. vm_page_cache() will handle this
927 vm_page_test_dirty(m);
932 if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
934 * Invalid pages can be easily freed
936 vm_pageout_page_free(m);
937 mycpu->gd_cnt.v_dfree++;
939 } else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
941 * Clean pages can be placed onto the cache queue.
942 * This effectively frees them.
946 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
948 * Dirty pages need to be paged out, but flushing
949 * a page is extremely expensive verses freeing
950 * a clean page. Rather then artificially limiting
951 * the number of pages we can flush, we instead give
952 * dirty pages extra priority on the inactive queue
953 * by forcing them to be cycled through the queue
954 * twice before being flushed, after which the
955 * (now clean) page will cycle through once more
956 * before being freed. This significantly extends
957 * the thrash point for a heavily loaded machine.
959 vm_page_flag_set(m, PG_WINATCFLS);
960 vm_page_and_queue_spin_lock(m);
961 if (m->queue - m->pc == PQ_INACTIVE) {
963 &vm_page_queues[PQ_INACTIVE + q].pl,
966 &vm_page_queues[PQ_INACTIVE + q].pl,
968 ++vm_swapcache_inactive_heuristic;
970 vm_page_and_queue_spin_unlock(m);
972 } else if (maxlaunder > 0) {
974 * We always want to try to flush some dirty pages if
975 * we encounter them, to keep the system stable.
976 * Normally this number is small, but under extreme
977 * pressure where there are insufficient clean pages
978 * on the inactive queue, we may have to go all out.
980 int swap_pageouts_ok;
981 struct vnode *vp = NULL;
983 swap_pageouts_ok = 0;
986 (object->type != OBJT_SWAP) &&
987 (object->type != OBJT_DEFAULT)) {
988 swap_pageouts_ok = 1;
990 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
991 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
992 vm_page_count_min(0));
997 * We don't bother paging objects that are "dead".
998 * Those objects are in a "rundown" state.
1000 if (!swap_pageouts_ok ||
1002 (object->flags & OBJ_DEAD)) {
1003 vm_page_and_queue_spin_lock(m);
1004 if (m->queue - m->pc == PQ_INACTIVE) {
1006 &vm_page_queues[PQ_INACTIVE + q].pl,
1009 &vm_page_queues[PQ_INACTIVE + q].pl,
1011 ++vm_swapcache_inactive_heuristic;
1013 vm_page_and_queue_spin_unlock(m);
1019 * (m) is still busied.
1021 * The object is already known NOT to be dead. It
1022 * is possible for the vget() to block the whole
1023 * pageout daemon, but the new low-memory handling
1024 * code should prevent it.
1026 * The previous code skipped locked vnodes and, worse,
1027 * reordered pages in the queue. This results in
1028 * completely non-deterministic operation because,
1029 * quite often, a vm_fault has initiated an I/O and
1030 * is holding a locked vnode at just the point where
1031 * the pageout daemon is woken up.
1033 * We can't wait forever for the vnode lock, we might
1034 * deadlock due to a vn_read() getting stuck in
1035 * vm_wait while holding this vnode. We skip the
1036 * vnode if we can't get it in a reasonable amount
1039 * vpfailed is used to (try to) avoid the case where
1040 * a large number of pages are associated with a
1041 * locked vnode, which could cause the pageout daemon
1042 * to stall for an excessive amount of time.
1044 if (object->type == OBJT_VNODE) {
1047 vp = object->handle;
1048 flags = LK_EXCLUSIVE | LK_NOOBJ;
1052 flags |= LK_TIMELOCK;
1057 * We have unbusied (m) temporarily so we can
1058 * acquire the vp lock without deadlocking.
1059 * (m) is held to prevent destruction.
1061 if (vget(vp, flags) != 0) {
1063 ++pageout_lock_miss;
1064 if (object->flags & OBJ_MIGHTBEDIRTY)
1071 * The page might have been moved to another
1072 * queue during potential blocking in vget()
1073 * above. The page might have been freed and
1074 * reused for another vnode. The object might
1075 * have been reused for another vnode.
1077 if (m->queue - m->pc != PQ_INACTIVE ||
1078 m->object != object ||
1079 object->handle != vp) {
1080 if (object->flags & OBJ_MIGHTBEDIRTY)
1088 * The page may have been busied during the
1089 * blocking in vput(); We don't move the
1090 * page back onto the end of the queue so that
1091 * statistics are more correct if we don't.
1093 if (vm_page_busy_try(m, TRUE)) {
1101 * (m) is busied again
1103 * We own the busy bit and remove our hold
1104 * bit. If the page is still held it
1105 * might be undergoing I/O, so skip it.
1107 if (m->hold_count) {
1108 vm_page_and_queue_spin_lock(m);
1109 if (m->queue - m->pc == PQ_INACTIVE) {
1110 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1111 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1112 ++vm_swapcache_inactive_heuristic;
1114 vm_page_and_queue_spin_unlock(m);
1115 if (object->flags & OBJ_MIGHTBEDIRTY)
1121 /* (m) is left busied as we fall through */
1125 * page is busy and not held here.
1127 * If a page is dirty, then it is either being washed
1128 * (but not yet cleaned) or it is still in the
1129 * laundry. If it is still in the laundry, then we
1130 * start the cleaning operation.
1132 * decrement inactive_shortage on success to account
1133 * for the (future) cleaned page. Otherwise we
1134 * could wind up laundering or cleaning too many
1137 if (vm_pageout_clean(m) != 0) {
1141 /* clean ate busy, page no longer accessible */
1148 vm_page_queues_spin_lock(PQ_INACTIVE + q);
1149 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
1150 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
1155 vm_pageout_scan_active(int pass, int q,
1156 int avail_shortage, int inactive_shortage,
1157 int *recycle_countp)
1159 struct vm_page marker;
1166 * We want to move pages from the active queue to the inactive
1167 * queue to get the inactive queue to the inactive target. If
1168 * we still have a page shortage from above we try to directly free
1169 * clean pages instead of moving them.
1171 * If we do still have a shortage we keep track of the number of
1172 * pages we free or cache (recycle_count) as a measure of thrashing
1173 * between the active and inactive queues.
1175 * If we were able to completely satisfy the free+cache targets
1176 * from the inactive pool we limit the number of pages we move
1177 * from the active pool to the inactive pool to 2x the pages we
1178 * had removed from the inactive pool (with a minimum of 1/5 the
1179 * inactive target). If we were not able to completely satisfy
1180 * the free+cache targets we go for the whole target aggressively.
1182 * NOTE: Both variables can end up negative.
1183 * NOTE: We are still in a critical section.
1186 bzero(&marker, sizeof(marker));
1187 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1188 marker.queue = PQ_ACTIVE + q;
1190 marker.wire_count = 1;
1192 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1193 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1194 maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt;
1195 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1197 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1198 maxscan-- > 0 && (avail_shortage - delta > 0 ||
1199 inactive_shortage > 0))
1201 vm_page_and_queue_spin_lock(m);
1202 if (m != TAILQ_NEXT(&marker, pageq)) {
1203 vm_page_and_queue_spin_unlock(m);
1207 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1208 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1210 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1216 if (m->flags & PG_MARKER) {
1217 vm_page_and_queue_spin_unlock(m);
1222 * Try to busy the page. Don't mess with pages which are
1223 * already busy or reorder them in the queue.
1225 if (vm_page_busy_try(m, TRUE)) {
1226 vm_page_and_queue_spin_unlock(m);
1231 * Don't deactivate pages that are held, even if we can
1232 * busy them. (XXX why not?)
1234 if (m->hold_count != 0) {
1235 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1237 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE + q].pl,
1239 vm_page_and_queue_spin_unlock(m);
1243 vm_page_and_queue_spin_unlock(m);
1247 * The page has been successfully busied and the page and
1248 * queue are no longer locked.
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);
1355 * Clean out our local marker.
1357 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1358 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1359 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1365 * The number of actually free pages can drop down to v_free_reserved,
1366 * we try to build the free count back above v_free_min. Note that
1367 * vm_paging_needed() also returns TRUE if v_free_count is not at
1368 * least v_free_min so that is the minimum we must build the free
1371 * We use a slightly higher target to improve hysteresis,
1372 * ((v_free_target + v_free_min) / 2). Since v_free_target
1373 * is usually the same as v_cache_min this maintains about
1374 * half the pages in the free queue as are in the cache queue,
1375 * providing pretty good pipelining for pageout operation.
1377 * The system operator can manipulate vm.v_cache_min and
1378 * vm.v_free_target to tune the pageout demon. Be sure
1379 * to keep vm.v_free_min < vm.v_free_target.
1381 * Note that the original paging target is to get at least
1382 * (free_min + cache_min) into (free + cache). The slightly
1383 * higher target will shift additional pages from cache to free
1384 * without effecting the original paging target in order to
1385 * maintain better hysteresis and not have the free count always
1386 * be dead-on v_free_min.
1388 * NOTE: we are still in a critical section.
1390 * Pages moved from PQ_CACHE to totally free are not counted in the
1391 * pages_freed counter.
1394 vm_pageout_scan_cache(int avail_shortage, int vnodes_skipped, int recycle_count)
1396 struct vm_pageout_scan_info info;
1399 while (vmstats.v_free_count <
1400 (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1402 * This steals some code from vm/vm_page.c
1404 static int cache_rover = 0;
1406 m = vm_page_list_find(PQ_CACHE, cache_rover & PQ_L2_MASK, FALSE);
1409 /* page is returned removed from its queue and spinlocked */
1410 if (vm_page_busy_try(m, TRUE)) {
1411 vm_page_deactivate_locked(m);
1412 vm_page_spin_unlock(m);
1414 kprintf("Warning: busy page %p found in cache\n", m);
1418 vm_page_spin_unlock(m);
1419 pagedaemon_wakeup();
1423 * Page has been successfully busied and it and its queue
1424 * is no longer spinlocked.
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_second != 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))
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 * The caller must hold proc_token.
1520 vm_pageout_scan_callback(struct proc *p, void *data)
1522 struct vm_pageout_scan_info *info = data;
1526 * Never kill system processes or init. If we have configured swap
1527 * then try to avoid killing low-numbered pids.
1529 if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1530 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1535 * if the process is in a non-running type state,
1538 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1542 * Get the approximate process size. Note that anonymous pages
1543 * with backing swap will be counted twice, but there should not
1544 * be too many such pages due to the stress the VM system is
1545 * under at this point.
1547 size = vmspace_anonymous_count(p->p_vmspace) +
1548 vmspace_swap_count(p->p_vmspace);
1551 * If the this process is bigger than the biggest one
1554 if (info->bigsize < size) {
1556 PRELE(info->bigproc);
1559 info->bigsize = size;
1566 * This routine tries to maintain the pseudo LRU active queue,
1567 * so that during long periods of time where there is no paging,
1568 * that some statistic accumulation still occurs. This code
1569 * helps the situation where paging just starts to occur.
1572 vm_pageout_page_stats(int q)
1574 static int fullintervalcount = 0;
1575 struct vm_page marker;
1577 int pcount, tpcount; /* Number of pages to check */
1580 page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1581 vmstats.v_free_min) -
1582 (vmstats.v_free_count + vmstats.v_inactive_count +
1583 vmstats.v_cache_count);
1585 if (page_shortage <= 0)
1588 pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1589 fullintervalcount += vm_pageout_stats_interval;
1590 if (fullintervalcount < vm_pageout_full_stats_interval) {
1591 tpcount = (vm_pageout_stats_max * pcount) /
1592 vmstats.v_page_count + 1;
1593 if (pcount > tpcount)
1596 fullintervalcount = 0;
1599 bzero(&marker, sizeof(marker));
1600 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1601 marker.queue = PQ_ACTIVE + q;
1603 marker.wire_count = 1;
1605 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1606 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1607 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1609 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1614 vm_page_and_queue_spin_lock(m);
1615 if (m != TAILQ_NEXT(&marker, pageq)) {
1616 vm_page_and_queue_spin_unlock(m);
1620 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1621 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1622 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1628 if (m->flags & PG_MARKER) {
1629 vm_page_and_queue_spin_unlock(m);
1634 * Ignore pages we can't busy
1636 if (vm_page_busy_try(m, TRUE)) {
1637 vm_page_and_queue_spin_unlock(m);
1640 vm_page_and_queue_spin_unlock(m);
1641 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1644 * We now have a safely busied page, the page and queue
1645 * spinlocks have been released.
1649 if (m->hold_count) {
1655 * Calculate activity
1658 if (m->flags & PG_REFERENCED) {
1659 vm_page_flag_clear(m, PG_REFERENCED);
1662 actcount += pmap_ts_referenced(m);
1665 * Update act_count and move page to end of queue.
1668 m->act_count += ACT_ADVANCE + actcount;
1669 if (m->act_count > ACT_MAX)
1670 m->act_count = ACT_MAX;
1671 vm_page_and_queue_spin_lock(m);
1672 if (m->queue - m->pc == PQ_ACTIVE) {
1674 &vm_page_queues[PQ_ACTIVE + q].pl,
1677 &vm_page_queues[PQ_ACTIVE + q].pl,
1680 vm_page_and_queue_spin_unlock(m);
1685 if (m->act_count == 0) {
1687 * We turn off page access, so that we have
1688 * more accurate RSS stats. We don't do this
1689 * in the normal page deactivation when the
1690 * system is loaded VM wise, because the
1691 * cost of the large number of page protect
1692 * operations would be higher than the value
1693 * of doing the operation.
1695 * We use the marker to save our place so
1696 * we can release the spin lock. both (m)
1697 * and (next) will be invalid.
1699 vm_page_protect(m, VM_PROT_NONE);
1700 vm_page_deactivate(m);
1702 m->act_count -= min(m->act_count, ACT_DECLINE);
1703 vm_page_and_queue_spin_lock(m);
1704 if (m->queue - m->pc == PQ_ACTIVE) {
1706 &vm_page_queues[PQ_ACTIVE + q].pl,
1709 &vm_page_queues[PQ_ACTIVE + q].pl,
1712 vm_page_and_queue_spin_unlock(m);
1718 * Remove our local marker
1720 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1721 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1722 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1726 vm_pageout_free_page_calc(vm_size_t count)
1728 if (count < vmstats.v_page_count)
1731 * free_reserved needs to include enough for the largest swap pager
1732 * structures plus enough for any pv_entry structs when paging.
1734 * v_free_min normal allocations
1735 * v_free_reserved system allocations
1736 * v_pageout_free_min allocations by pageout daemon
1737 * v_interrupt_free_min low level allocations (e.g swap structures)
1739 if (vmstats.v_page_count > 1024)
1740 vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1742 vmstats.v_free_min = 64;
1743 vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1744 vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1745 vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1746 vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1753 * vm_pageout is the high level pageout daemon.
1758 vm_pageout_thread(void)
1764 * Initialize some paging parameters.
1766 curthread->td_flags |= TDF_SYSTHREAD;
1768 if (vmstats.v_page_count < 2000)
1769 vm_pageout_page_count = 8;
1771 vm_pageout_free_page_calc(vmstats.v_page_count);
1774 * v_free_target and v_cache_min control pageout hysteresis. Note
1775 * that these are more a measure of the VM cache queue hysteresis
1776 * then the VM free queue. Specifically, v_free_target is the
1777 * high water mark (free+cache pages).
1779 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1780 * low water mark, while v_free_min is the stop. v_cache_min must
1781 * be big enough to handle memory needs while the pageout daemon
1782 * is signalled and run to free more pages.
1784 if (vmstats.v_free_count > 6144)
1785 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1787 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1790 * NOTE: With the new buffer cache b_act_count we want the default
1791 * inactive target to be a percentage of available memory.
1793 * The inactive target essentially determines the minimum
1794 * number of 'temporary' pages capable of caching one-time-use
1795 * files when the VM system is otherwise full of pages
1796 * belonging to multi-time-use files or active program data.
1798 * NOTE: The inactive target is aggressively persued only if the
1799 * inactive queue becomes too small. If the inactive queue
1800 * is large enough to satisfy page movement to free+cache
1801 * then it is repopulated more slowly from the active queue.
1802 * This allows a general inactive_target default to be set.
1804 * There is an issue here for processes which sit mostly idle
1805 * 'overnight', such as sshd, tcsh, and X. Any movement from
1806 * the active queue will eventually cause such pages to
1807 * recycle eventually causing a lot of paging in the morning.
1808 * To reduce the incidence of this pages cycled out of the
1809 * buffer cache are moved directly to the inactive queue if
1810 * they were only used once or twice.
1812 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1813 * Increasing the value (up to 64) increases the number of
1814 * buffer recyclements which go directly to the inactive queue.
1816 if (vmstats.v_free_count > 2048) {
1817 vmstats.v_cache_min = vmstats.v_free_target;
1818 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1820 vmstats.v_cache_min = 0;
1821 vmstats.v_cache_max = 0;
1823 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1825 /* XXX does not really belong here */
1826 if (vm_page_max_wired == 0)
1827 vm_page_max_wired = vmstats.v_free_count / 3;
1829 if (vm_pageout_stats_max == 0)
1830 vm_pageout_stats_max = vmstats.v_free_target;
1833 * Set interval in seconds for stats scan.
1835 if (vm_pageout_stats_interval == 0)
1836 vm_pageout_stats_interval = 5;
1837 if (vm_pageout_full_stats_interval == 0)
1838 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1842 * Set maximum free per pass
1844 if (vm_pageout_stats_free_max == 0)
1845 vm_pageout_stats_free_max = 5;
1847 swap_pager_swap_init();
1851 * The pageout daemon is never done, so loop forever.
1858 int inactive_shortage;
1859 int vnodes_skipped = 0;
1860 int recycle_count = 0;
1864 * Wait for an action request. If we timeout check to
1865 * see if paging is needed (in case the normal wakeup
1868 if (vm_pages_needed == 0) {
1869 error = tsleep(&vm_pages_needed,
1871 vm_pageout_stats_interval * hz);
1873 vm_paging_needed() == 0 &&
1874 vm_pages_needed == 0) {
1875 for (q = 0; q < PQ_L2_SIZE; ++q)
1876 vm_pageout_page_stats(q);
1879 vm_pages_needed = 1;
1882 mycpu->gd_cnt.v_pdwakeups++;
1885 * Do whatever cleanup that the pmap code can.
1890 * Scan for pageout. Try to avoid thrashing the system
1893 * Calculate our target for the number of free+cache pages we
1894 * want to get to. This is higher then the number that causes
1895 * allocations to stall (severe) in order to provide hysteresis,
1896 * and if we don't make it all the way but get to the minimum
1897 * we're happy. Goose it a bit if there are multipler
1898 * requests for memory.
1900 avail_shortage = vm_paging_target() + vm_pageout_deficit;
1901 vm_pageout_deficit = 0;
1903 if (avail_shortage > 0) {
1904 for (q = 0; q < PQ_L2_SIZE; ++q) {
1905 delta1 += vm_pageout_scan_inactive(
1907 PQAVERAGE(avail_shortage),
1910 avail_shortage -= delta1;
1914 * Figure out how many active pages we must deactivate. If
1915 * we were able to reach our target with just the inactive
1916 * scan above we limit the number of active pages we
1917 * deactivate to reduce unnecessary work.
1919 inactive_shortage = vmstats.v_inactive_target -
1920 vmstats.v_inactive_count;
1923 * If we were unable to free sufficient inactive pages to
1924 * satisfy the free/cache queue requirements then simply
1925 * reaching the inactive target may not be good enough.
1926 * Try to deactivate pages in excess of the target based
1929 * However to prevent thrashing the VM system do not
1930 * deactivate more than an additional 1/10 the inactive
1931 * target's worth of active pages.
1933 if (avail_shortage > 0) {
1934 tmp = avail_shortage * 2;
1935 if (tmp > vmstats.v_inactive_target / 10)
1936 tmp = vmstats.v_inactive_target / 10;
1937 inactive_shortage += tmp;
1940 if (avail_shortage > 0 || inactive_shortage > 0) {
1942 for (q = 0; q < PQ_L2_SIZE; ++q) {
1943 delta2 += vm_pageout_scan_active(
1945 PQAVERAGE(avail_shortage),
1946 PQAVERAGE(inactive_shortage),
1949 inactive_shortage -= delta2;
1950 avail_shortage -= delta2;
1954 * Finally free enough cache pages to meet our free page
1955 * requirement and take more drastic measures if we are
1958 vm_pageout_scan_cache(avail_shortage, vnodes_skipped,
1962 * Wait for more work.
1964 if (avail_shortage > 0) {
1966 if (swap_pager_full) {
1968 * Running out of memory, catastrophic back-off
1969 * to one-second intervals.
1971 tsleep(&vm_pages_needed, 0, "pdelay", hz);
1972 } else if (pass < 10 && vm_pages_needed > 1) {
1974 * Normal operation, additional processes
1975 * have already kicked us. Retry immediately.
1977 } else if (pass < 10) {
1979 * Normal operation, fewer processes. Delay
1980 * a bit but allow wakeups.
1982 vm_pages_needed = 0;
1983 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1984 vm_pages_needed = 1;
1987 * We've taken too many passes, forced delay.
1989 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1993 * Interlocked wakeup of waiters (non-optional)
1996 if (vm_pages_needed && !vm_page_count_min(0)) {
1997 wakeup(&vmstats.v_free_count);
1998 vm_pages_needed = 0;
2004 static struct kproc_desc page_kp = {
2009 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
2013 * Called after allocating a page out of the cache or free queue
2014 * to possibly wake the pagedaemon up to replentish our supply.
2016 * We try to generate some hysteresis by waking the pagedaemon up
2017 * when our free+cache pages go below the free_min+cache_min level.
2018 * The pagedaemon tries to get the count back up to at least the
2019 * minimum, and through to the target level if possible.
2021 * If the pagedaemon is already active bump vm_pages_needed as a hint
2022 * that there are even more requests pending.
2028 pagedaemon_wakeup(void)
2030 if (vm_paging_needed() && curthread != pagethread) {
2031 if (vm_pages_needed == 0) {
2032 vm_pages_needed = 1; /* SMP race ok */
2033 wakeup(&vm_pages_needed);
2034 } else if (vm_page_count_min(0)) {
2035 ++vm_pages_needed; /* SMP race ok */
2040 #if !defined(NO_SWAPPING)
2047 vm_req_vmdaemon(void)
2049 static int lastrun = 0;
2051 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2052 wakeup(&vm_daemon_needed);
2057 static int vm_daemon_callback(struct proc *p, void *data __unused);
2066 * XXX vm_daemon_needed specific token?
2069 tsleep(&vm_daemon_needed, 0, "psleep", 0);
2070 if (vm_pageout_req_swapout) {
2071 swapout_procs(vm_pageout_req_swapout);
2072 vm_pageout_req_swapout = 0;
2075 * scan the processes for exceeding their rlimits or if
2076 * process is swapped out -- deactivate pages
2078 allproc_scan(vm_daemon_callback, NULL);
2083 * Caller must hold proc_token.
2086 vm_daemon_callback(struct proc *p, void *data __unused)
2088 vm_pindex_t limit, size;
2091 * if this is a system process or if we have already
2092 * looked at this process, skip it.
2094 if (p->p_flags & (P_SYSTEM | P_WEXIT))
2098 * if the process is in a non-running type state,
2101 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
2107 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2108 p->p_rlimit[RLIMIT_RSS].rlim_max));
2111 * let processes that are swapped out really be
2112 * swapped out. Set the limit to nothing to get as
2113 * many pages out to swap as possible.
2115 if (p->p_flags & P_SWAPPEDOUT)
2118 lwkt_gettoken(&p->p_vmspace->vm_map.token);
2119 size = vmspace_resident_count(p->p_vmspace);
2120 if (limit >= 0 && size >= limit) {
2121 vm_pageout_map_deactivate_pages(&p->p_vmspace->vm_map, limit);
2123 lwkt_reltoken(&p->p_vmspace->vm_map.token);