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);
204 * Calculate approximately how many pages on each queue to try to
205 * clean. An exact calculation creates an edge condition when the
206 * queues are unbalanced so add significant slop. The queue scans
207 * will stop early when targets are reached and will start where they
208 * left off on the next pass.
216 avg = ((n + (PQ_L2_SIZE - 1)) / PQ_L2_SIZE + 1);
219 avg = ((n - (PQ_L2_SIZE - 1)) / PQ_L2_SIZE - 1);
228 * Clean the page and remove it from the laundry. The page must not be
231 * We set the busy bit to cause potential page faults on this page to
232 * block. Note the careful timing, however, the busy bit isn't set till
233 * late and we cannot do anything that will mess with the page.
236 vm_pageout_clean(vm_page_t m)
239 vm_page_t mc[BLIST_MAX_ALLOC];
241 int ib, is, page_base;
242 vm_pindex_t pindex = m->pindex;
247 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
248 * with the new swapper, but we could have serious problems paging
249 * out other object types if there is insufficient memory.
251 * Unfortunately, checking free memory here is far too late, so the
252 * check has been moved up a procedural level.
256 * Don't mess with the page if it's busy, held, or special
258 * XXX do we really need to check hold_count here? hold_count
259 * isn't supposed to mess with vm_page ops except prevent the
260 * page from being reused.
262 if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
268 * Place page in cluster. Align cluster for optimal swap space
269 * allocation (whether it is swap or not). This is typically ~16-32
270 * pages, which also tends to align the cluster to multiples of the
271 * filesystem block size if backed by a filesystem.
273 page_base = pindex % BLIST_MAX_ALLOC;
279 * Scan object for clusterable pages.
281 * We can cluster ONLY if: ->> the page is NOT
282 * clean, wired, busy, held, or mapped into a
283 * buffer, and one of the following:
284 * 1) The page is inactive, or a seldom used
287 * 2) we force the issue.
289 * During heavy mmap/modification loads the pageout
290 * daemon can really fragment the underlying file
291 * due to flushing pages out of order and not trying
292 * align the clusters (which leave sporatic out-of-order
293 * holes). To solve this problem we do the reverse scan
294 * first and attempt to align our cluster, then do a
295 * forward scan if room remains.
298 vm_object_hold(object);
302 p = vm_page_lookup_busy_try(object, pindex - page_base + ib,
304 if (error || p == NULL)
306 if ((p->queue - p->pc) == PQ_CACHE ||
307 (p->flags & PG_UNMANAGED)) {
311 vm_page_test_dirty(p);
312 if (((p->dirty & p->valid) == 0 &&
313 (p->flags & PG_NEED_COMMIT) == 0) ||
314 p->queue - p->pc != PQ_INACTIVE ||
315 p->wire_count != 0 || /* may be held by buf cache */
316 p->hold_count != 0) { /* may be undergoing I/O */
325 while (is < BLIST_MAX_ALLOC &&
326 pindex - page_base + is < object->size) {
329 p = vm_page_lookup_busy_try(object, pindex - page_base + is,
331 if (error || p == NULL)
333 if (((p->queue - p->pc) == PQ_CACHE) ||
334 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
338 vm_page_test_dirty(p);
339 if (((p->dirty & p->valid) == 0 &&
340 (p->flags & PG_NEED_COMMIT) == 0) ||
341 p->queue - p->pc != PQ_INACTIVE ||
342 p->wire_count != 0 || /* may be held by buf cache */
343 p->hold_count != 0) { /* may be undergoing I/O */
351 vm_object_drop(object);
354 * we allow reads during pageouts...
356 return vm_pageout_flush(&mc[ib], is - ib, 0);
360 * vm_pageout_flush() - launder the given pages
362 * The given pages are laundered. Note that we setup for the start of
363 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
364 * reference count all in here rather then in the parent. If we want
365 * the parent to do more sophisticated things we may have to change
368 * The pages in the array must be busied by the caller and will be
369 * unbusied by this function.
372 vm_pageout_flush(vm_page_t *mc, int count, int flags)
375 int pageout_status[count];
380 * Initiate I/O. Bump the vm_page_t->busy counter.
382 for (i = 0; i < count; i++) {
383 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
384 ("vm_pageout_flush page %p index %d/%d: partially "
385 "invalid page", mc[i], i, count));
386 vm_page_io_start(mc[i]);
390 * We must make the pages read-only. This will also force the
391 * modified bit in the related pmaps to be cleared. The pager
392 * cannot clear the bit for us since the I/O completion code
393 * typically runs from an interrupt. The act of making the page
394 * read-only handles the case for us.
396 * Then we can unbusy the pages, we still hold a reference by virtue
399 for (i = 0; i < count; i++) {
400 vm_page_protect(mc[i], VM_PROT_READ);
401 vm_page_wakeup(mc[i]);
404 object = mc[0]->object;
405 vm_object_pip_add(object, count);
407 vm_pager_put_pages(object, mc, count,
408 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
411 for (i = 0; i < count; i++) {
412 vm_page_t mt = mc[i];
414 switch (pageout_status[i]) {
423 * Page outside of range of object. Right now we
424 * essentially lose the changes by pretending it
427 vm_page_busy_wait(mt, FALSE, "pgbad");
428 pmap_clear_modify(mt);
435 * A page typically cannot be paged out when we
436 * have run out of swap. We leave the page
437 * marked inactive and will try to page it out
440 * Starvation of the active page list is used to
441 * determine when the system is massively memory
450 * If the operation is still going, leave the page busy to
451 * block all other accesses. Also, leave the paging in
452 * progress indicator set so that we don't attempt an object
455 * For any pages which have completed synchronously,
456 * deactivate the page if we are under a severe deficit.
457 * Do not try to enter them into the cache, though, they
458 * might still be read-heavy.
460 if (pageout_status[i] != VM_PAGER_PEND) {
461 vm_page_busy_wait(mt, FALSE, "pgouw");
462 if (vm_page_count_severe())
463 vm_page_deactivate(mt);
465 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
466 vm_page_protect(mt, VM_PROT_READ);
468 vm_page_io_finish(mt);
470 vm_object_pip_wakeup(object);
476 #if !defined(NO_SWAPPING)
478 * deactivate enough pages to satisfy the inactive target
479 * requirements or if vm_page_proc_limit is set, then
480 * deactivate all of the pages in the object and its
483 * The map must be locked.
484 * The caller must hold the vm_object.
486 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
489 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
490 vm_pindex_t desired, int map_remove_only)
492 struct rb_vm_page_scan_info info;
497 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
501 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
503 if (lobject->type == OBJT_DEVICE ||
504 lobject->type == OBJT_MGTDEVICE ||
505 lobject->type == OBJT_PHYS)
507 if (lobject->paging_in_progress)
510 remove_mode = map_remove_only;
511 if (lobject->shadow_count > 1)
515 * scan the objects entire memory queue. We hold the
516 * object's token so the scan should not race anything.
518 info.limit = remove_mode;
520 info.desired = desired;
521 vm_page_rb_tree_RB_SCAN(&lobject->rb_memq, NULL,
522 vm_pageout_object_deactivate_pages_callback,
525 while ((tobject = lobject->backing_object) != NULL) {
526 KKASSERT(tobject != object);
527 vm_object_hold(tobject);
528 if (tobject == lobject->backing_object)
530 vm_object_drop(tobject);
532 if (lobject != object) {
534 vm_object_lock_swap();
535 vm_object_drop(lobject);
536 /* leaves tobject locked & at top */
540 if (lobject != object)
541 vm_object_drop(lobject); /* NULL ok */
545 * The caller must hold the vm_object.
548 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
550 struct rb_vm_page_scan_info *info = data;
553 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
556 mycpu->gd_cnt.v_pdpages++;
558 if (vm_page_busy_try(p, TRUE))
560 if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
564 if (!pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
569 actcount = pmap_ts_referenced(p);
571 vm_page_flag_set(p, PG_REFERENCED);
572 } else if (p->flags & PG_REFERENCED) {
576 vm_page_and_queue_spin_lock(p);
577 if (p->queue - p->pc != PQ_ACTIVE && (p->flags & PG_REFERENCED)) {
578 vm_page_and_queue_spin_unlock(p);
580 p->act_count += actcount;
581 vm_page_flag_clear(p, PG_REFERENCED);
582 } else if (p->queue - p->pc == PQ_ACTIVE) {
583 if ((p->flags & PG_REFERENCED) == 0) {
584 p->act_count -= min(p->act_count, ACT_DECLINE);
586 (vm_pageout_algorithm || (p->act_count == 0))) {
587 vm_page_and_queue_spin_unlock(p);
588 vm_page_protect(p, VM_PROT_NONE);
589 vm_page_deactivate(p);
591 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
593 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
595 vm_page_and_queue_spin_unlock(p);
598 vm_page_and_queue_spin_unlock(p);
600 vm_page_flag_clear(p, PG_REFERENCED);
602 vm_page_and_queue_spin_lock(p);
603 if (p->queue - p->pc == PQ_ACTIVE) {
604 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
605 p->act_count += ACT_ADVANCE;
606 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
608 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
611 vm_page_and_queue_spin_unlock(p);
613 } else if (p->queue - p->pc == PQ_INACTIVE) {
614 vm_page_and_queue_spin_unlock(p);
615 vm_page_protect(p, VM_PROT_NONE);
617 vm_page_and_queue_spin_unlock(p);
624 * Deactivate some number of pages in a map, try to do it fairly, but
625 * that is really hard to do.
628 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
631 vm_object_t obj, bigobj;
634 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
642 * first, search out the biggest object, and try to free pages from
645 tmpe = map->header.next;
646 while (tmpe != &map->header) {
647 switch(tmpe->maptype) {
648 case VM_MAPTYPE_NORMAL:
649 case VM_MAPTYPE_VPAGETABLE:
650 obj = tmpe->object.vm_object;
651 if ((obj != NULL) && (obj->shadow_count <= 1) &&
653 (bigobj->resident_page_count < obj->resident_page_count))) {
660 if (tmpe->wired_count > 0)
661 nothingwired = FALSE;
666 vm_object_hold(bigobj);
667 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
668 vm_object_drop(bigobj);
672 * Next, hunt around for other pages to deactivate. We actually
673 * do this search sort of wrong -- .text first is not the best idea.
675 tmpe = map->header.next;
676 while (tmpe != &map->header) {
677 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
679 switch(tmpe->maptype) {
680 case VM_MAPTYPE_NORMAL:
681 case VM_MAPTYPE_VPAGETABLE:
682 obj = tmpe->object.vm_object;
685 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
696 * Remove all mappings if a process is swapped out, this will free page
699 if (desired == 0 && nothingwired)
700 pmap_remove(vm_map_pmap(map),
701 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
707 * Called when the pageout scan wants to free a page. We no longer
708 * try to cycle the vm_object here with a reference & dealloc, which can
709 * cause a non-trivial object collapse in a critical path.
711 * It is unclear why we cycled the ref_count in the past, perhaps to try
712 * to optimize shadow chain collapses but I don't quite see why it would
713 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
714 * synchronously and not have to be kicked-start.
717 vm_pageout_page_free(vm_page_t m)
719 vm_page_protect(m, VM_PROT_NONE);
724 * vm_pageout_scan does the dirty work for the pageout daemon.
726 struct vm_pageout_scan_info {
727 struct proc *bigproc;
731 static int vm_pageout_scan_callback(struct proc *p, void *data);
734 vm_pageout_scan_inactive(int pass, int q, int avail_shortage,
735 int *vnodes_skippedp)
738 struct vm_page marker;
739 struct vnode *vpfailed; /* warning, allowed to be stale */
748 * Start scanning the inactive queue for pages we can move to the
749 * cache or free. The scan will stop when the target is reached or
750 * we have scanned the entire inactive queue. Note that m->act_count
751 * is not used to form decisions for the inactive queue, only for the
754 * maxlaunder limits the number of dirty pages we flush per scan.
755 * For most systems a smaller value (16 or 32) is more robust under
756 * extreme memory and disk pressure because any unnecessary writes
757 * to disk can result in extreme performance degredation. However,
758 * systems with excessive dirty pages (especially when MAP_NOSYNC is
759 * used) will die horribly with limited laundering. If the pageout
760 * daemon cannot clean enough pages in the first pass, we let it go
761 * all out in succeeding passes.
763 if ((maxlaunder = vm_max_launder) <= 1)
769 * Initialize our marker
771 bzero(&marker, sizeof(marker));
772 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
773 marker.queue = PQ_INACTIVE + q;
775 marker.wire_count = 1;
778 * Inactive queue scan.
780 * NOTE: The vm_page must be spinlocked before the queue to avoid
781 * deadlocks, so it is easiest to simply iterate the loop
782 * with the queue unlocked at the top.
786 vm_page_queues_spin_lock(PQ_INACTIVE + q);
787 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
788 maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt;
791 * Queue locked at top of loop to avoid stack marker issues.
793 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
794 maxscan-- > 0 && avail_shortage - delta > 0)
796 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
797 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
799 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
801 mycpu->gd_cnt.v_pdpages++;
804 * Skip marker pages (atomic against other markers to avoid
805 * infinite hop-over scans).
807 if (m->flags & PG_MARKER)
811 * Try to busy the page. Don't mess with pages which are
812 * already busy or reorder them in the queue.
814 if (vm_page_busy_try(m, TRUE))
818 * Remaining operations run with the page busy and neither
819 * the page or the queue will be spin-locked.
821 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
822 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
826 * It is possible for a page to be busied ad-hoc (e.g. the
827 * pmap_collect() code) and wired and race against the
828 * allocation of a new page. vm_page_alloc() may be forced
829 * to deactivate the wired page in which case it winds up
830 * on the inactive queue and must be handled here. We
831 * correct the problem simply by unqueuing the page.
834 vm_page_unqueue_nowakeup(m);
836 kprintf("WARNING: pagedaemon: wired page on "
837 "inactive queue %p\n", m);
842 * A held page may be undergoing I/O, so skip it.
845 vm_page_and_queue_spin_lock(m);
846 if (m->queue - m->pc == PQ_INACTIVE) {
848 &vm_page_queues[PQ_INACTIVE + q].pl,
851 &vm_page_queues[PQ_INACTIVE + q].pl,
853 ++vm_swapcache_inactive_heuristic;
855 vm_page_and_queue_spin_unlock(m);
860 if (m->object == NULL || m->object->ref_count == 0) {
862 * If the object is not being used, we ignore previous
865 vm_page_flag_clear(m, PG_REFERENCED);
866 pmap_clear_reference(m);
867 /* fall through to end */
868 } else if (((m->flags & PG_REFERENCED) == 0) &&
869 (actcount = pmap_ts_referenced(m))) {
871 * Otherwise, if the page has been referenced while
872 * in the inactive queue, we bump the "activation
873 * count" upwards, making it less likely that the
874 * page will be added back to the inactive queue
875 * prematurely again. Here we check the page tables
876 * (or emulated bits, if any), given the upper level
877 * VM system not knowing anything about existing
881 m->act_count += (actcount + ACT_ADVANCE);
887 * (m) is still busied.
889 * If the upper level VM system knows about any page
890 * references, we activate the page. We also set the
891 * "activation count" higher than normal so that we will less
892 * likely place pages back onto the inactive queue again.
894 if ((m->flags & PG_REFERENCED) != 0) {
895 vm_page_flag_clear(m, PG_REFERENCED);
896 actcount = pmap_ts_referenced(m);
898 m->act_count += (actcount + ACT_ADVANCE + 1);
904 * If the upper level VM system doesn't know anything about
905 * the page being dirty, we have to check for it again. As
906 * far as the VM code knows, any partially dirty pages are
909 * Pages marked PG_WRITEABLE may be mapped into the user
910 * address space of a process running on another cpu. A
911 * user process (without holding the MP lock) running on
912 * another cpu may be able to touch the page while we are
913 * trying to remove it. vm_page_cache() will handle this
917 vm_page_test_dirty(m);
922 if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
924 * Invalid pages can be easily freed
926 vm_pageout_page_free(m);
927 mycpu->gd_cnt.v_dfree++;
929 } else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
931 * Clean pages can be placed onto the cache queue.
932 * This effectively frees them.
936 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
938 * Dirty pages need to be paged out, but flushing
939 * a page is extremely expensive verses freeing
940 * a clean page. Rather then artificially limiting
941 * the number of pages we can flush, we instead give
942 * dirty pages extra priority on the inactive queue
943 * by forcing them to be cycled through the queue
944 * twice before being flushed, after which the
945 * (now clean) page will cycle through once more
946 * before being freed. This significantly extends
947 * the thrash point for a heavily loaded machine.
949 vm_page_flag_set(m, PG_WINATCFLS);
950 vm_page_and_queue_spin_lock(m);
951 if (m->queue - m->pc == PQ_INACTIVE) {
953 &vm_page_queues[PQ_INACTIVE + q].pl,
956 &vm_page_queues[PQ_INACTIVE + q].pl,
958 ++vm_swapcache_inactive_heuristic;
960 vm_page_and_queue_spin_unlock(m);
962 } else if (maxlaunder > 0) {
964 * We always want to try to flush some dirty pages if
965 * we encounter them, to keep the system stable.
966 * Normally this number is small, but under extreme
967 * pressure where there are insufficient clean pages
968 * on the inactive queue, we may have to go all out.
970 int swap_pageouts_ok;
971 struct vnode *vp = NULL;
973 swap_pageouts_ok = 0;
976 (object->type != OBJT_SWAP) &&
977 (object->type != OBJT_DEFAULT)) {
978 swap_pageouts_ok = 1;
980 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
981 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
982 vm_page_count_min(0));
987 * We don't bother paging objects that are "dead".
988 * Those objects are in a "rundown" state.
990 if (!swap_pageouts_ok ||
992 (object->flags & OBJ_DEAD)) {
993 vm_page_and_queue_spin_lock(m);
994 if (m->queue - m->pc == PQ_INACTIVE) {
996 &vm_page_queues[PQ_INACTIVE + q].pl,
999 &vm_page_queues[PQ_INACTIVE + q].pl,
1001 ++vm_swapcache_inactive_heuristic;
1003 vm_page_and_queue_spin_unlock(m);
1009 * (m) is still busied.
1011 * The object is already known NOT to be dead. It
1012 * is possible for the vget() to block the whole
1013 * pageout daemon, but the new low-memory handling
1014 * code should prevent it.
1016 * The previous code skipped locked vnodes and, worse,
1017 * reordered pages in the queue. This results in
1018 * completely non-deterministic operation because,
1019 * quite often, a vm_fault has initiated an I/O and
1020 * is holding a locked vnode at just the point where
1021 * the pageout daemon is woken up.
1023 * We can't wait forever for the vnode lock, we might
1024 * deadlock due to a vn_read() getting stuck in
1025 * vm_wait while holding this vnode. We skip the
1026 * vnode if we can't get it in a reasonable amount
1029 * vpfailed is used to (try to) avoid the case where
1030 * a large number of pages are associated with a
1031 * locked vnode, which could cause the pageout daemon
1032 * to stall for an excessive amount of time.
1034 if (object->type == OBJT_VNODE) {
1037 vp = object->handle;
1038 flags = LK_EXCLUSIVE;
1042 flags |= LK_TIMELOCK;
1047 * We have unbusied (m) temporarily so we can
1048 * acquire the vp lock without deadlocking.
1049 * (m) is held to prevent destruction.
1051 if (vget(vp, flags) != 0) {
1053 ++pageout_lock_miss;
1054 if (object->flags & OBJ_MIGHTBEDIRTY)
1061 * The page might have been moved to another
1062 * queue during potential blocking in vget()
1063 * above. The page might have been freed and
1064 * reused for another vnode. The object might
1065 * have been reused for another vnode.
1067 if (m->queue - m->pc != PQ_INACTIVE ||
1068 m->object != object ||
1069 object->handle != vp) {
1070 if (object->flags & OBJ_MIGHTBEDIRTY)
1078 * The page may have been busied during the
1079 * blocking in vput(); We don't move the
1080 * page back onto the end of the queue so that
1081 * statistics are more correct if we don't.
1083 if (vm_page_busy_try(m, TRUE)) {
1091 * (m) is busied again
1093 * We own the busy bit and remove our hold
1094 * bit. If the page is still held it
1095 * might be undergoing I/O, so skip it.
1097 if (m->hold_count) {
1098 vm_page_and_queue_spin_lock(m);
1099 if (m->queue - m->pc == PQ_INACTIVE) {
1100 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1101 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1102 ++vm_swapcache_inactive_heuristic;
1104 vm_page_and_queue_spin_unlock(m);
1105 if (object->flags & OBJ_MIGHTBEDIRTY)
1111 /* (m) is left busied as we fall through */
1115 * page is busy and not held here.
1117 * If a page is dirty, then it is either being washed
1118 * (but not yet cleaned) or it is still in the
1119 * laundry. If it is still in the laundry, then we
1120 * start the cleaning operation.
1122 * decrement inactive_shortage on success to account
1123 * for the (future) cleaned page. Otherwise we
1124 * could wind up laundering or cleaning too many
1127 count = vm_pageout_clean(m);
1129 maxlaunder -= count;
1132 * Clean ate busy, page no longer accessible
1142 * Systems with a ton of memory can wind up with huge
1143 * deactivation counts. Because the inactive scan is
1144 * doing a lot of flushing, the combination can result
1145 * in excessive paging even in situations where other
1146 * unrelated threads free up sufficient VM.
1148 * To deal with this we abort the nominal active->inactive
1149 * scan before we hit the inactive target when free+cache
1150 * levels have reached a reasonable target.
1152 * When deciding to stop early we need to add some slop to
1153 * the test and we need to return full completion to the caller
1154 * to prevent the caller from thinking there is something
1155 * wrong and issuing a low-memory+swap warning or pkill.
1157 vm_page_queues_spin_lock(PQ_INACTIVE + q);
1158 if (vm_paging_target() < -vm_max_launder) {
1160 * Stopping early, return full completion to caller.
1162 if (delta < avail_shortage)
1163 delta = avail_shortage;
1168 /* page queue still spin-locked */
1169 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
1170 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
1176 vm_pageout_scan_active(int pass, int q,
1177 int avail_shortage, int inactive_shortage,
1178 int *recycle_countp)
1180 struct vm_page marker;
1187 * We want to move pages from the active queue to the inactive
1188 * queue to get the inactive queue to the inactive target. If
1189 * we still have a page shortage from above we try to directly free
1190 * clean pages instead of moving them.
1192 * If we do still have a shortage we keep track of the number of
1193 * pages we free or cache (recycle_count) as a measure of thrashing
1194 * between the active and inactive queues.
1196 * If we were able to completely satisfy the free+cache targets
1197 * from the inactive pool we limit the number of pages we move
1198 * from the active pool to the inactive pool to 2x the pages we
1199 * had removed from the inactive pool (with a minimum of 1/5 the
1200 * inactive target). If we were not able to completely satisfy
1201 * the free+cache targets we go for the whole target aggressively.
1203 * NOTE: Both variables can end up negative.
1204 * NOTE: We are still in a critical section.
1207 bzero(&marker, sizeof(marker));
1208 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1209 marker.queue = PQ_ACTIVE + q;
1211 marker.wire_count = 1;
1213 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1214 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1215 maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt;
1218 * Queue locked at top of loop to avoid stack marker issues.
1220 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1221 maxscan-- > 0 && (avail_shortage - delta > 0 ||
1222 inactive_shortage > 0))
1224 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1225 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1227 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1231 * Skip marker pages (atomic against other markers to avoid
1232 * infinite hop-over scans).
1234 if (m->flags & PG_MARKER)
1238 * Try to busy the page. Don't mess with pages which are
1239 * already busy or reorder them in the queue.
1241 if (vm_page_busy_try(m, TRUE))
1245 * Remaining operations run with the page busy and neither
1246 * the page or the queue will be spin-locked.
1248 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1249 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1253 * Don't deactivate pages that are held, even if we can
1254 * busy them. (XXX why not?)
1256 if (m->hold_count != 0) {
1257 vm_page_and_queue_spin_lock(m);
1258 if (m->queue - m->pc == PQ_ACTIVE) {
1260 &vm_page_queues[PQ_ACTIVE + q].pl,
1263 &vm_page_queues[PQ_ACTIVE + q].pl,
1266 vm_page_and_queue_spin_unlock(m);
1272 * The count for pagedaemon pages is done after checking the
1273 * page for eligibility...
1275 mycpu->gd_cnt.v_pdpages++;
1278 * Check to see "how much" the page has been used and clear
1279 * the tracking access bits. If the object has no references
1280 * don't bother paying the expense.
1283 if (m->object && m->object->ref_count != 0) {
1284 if (m->flags & PG_REFERENCED)
1286 actcount += pmap_ts_referenced(m);
1288 m->act_count += ACT_ADVANCE + actcount;
1289 if (m->act_count > ACT_MAX)
1290 m->act_count = ACT_MAX;
1293 vm_page_flag_clear(m, PG_REFERENCED);
1296 * actcount is only valid if the object ref_count is non-zero.
1297 * If the page does not have an object, actcount will be zero.
1299 if (actcount && m->object->ref_count != 0) {
1300 vm_page_and_queue_spin_lock(m);
1301 if (m->queue - m->pc == PQ_ACTIVE) {
1303 &vm_page_queues[PQ_ACTIVE + q].pl,
1306 &vm_page_queues[PQ_ACTIVE + q].pl,
1309 vm_page_and_queue_spin_unlock(m);
1312 switch(m->object->type) {
1315 m->act_count -= min(m->act_count,
1316 vm_anonmem_decline);
1319 m->act_count -= min(m->act_count,
1320 vm_filemem_decline);
1323 if (vm_pageout_algorithm ||
1324 (m->object == NULL) ||
1325 (m->object && (m->object->ref_count == 0)) ||
1326 m->act_count < pass + 1
1329 * Deactivate the page. If we had a
1330 * shortage from our inactive scan try to
1331 * free (cache) the page instead.
1333 * Don't just blindly cache the page if
1334 * we do not have a shortage from the
1335 * inactive scan, that could lead to
1336 * gigabytes being moved.
1338 --inactive_shortage;
1339 if (avail_shortage - delta > 0 ||
1340 (m->object && (m->object->ref_count == 0)))
1342 if (avail_shortage - delta > 0)
1344 vm_page_protect(m, VM_PROT_NONE);
1345 if (m->dirty == 0 &&
1346 (m->flags & PG_NEED_COMMIT) == 0 &&
1347 avail_shortage - delta > 0) {
1350 vm_page_deactivate(m);
1354 vm_page_deactivate(m);
1359 vm_page_and_queue_spin_lock(m);
1360 if (m->queue - m->pc == PQ_ACTIVE) {
1362 &vm_page_queues[PQ_ACTIVE + q].pl,
1365 &vm_page_queues[PQ_ACTIVE + q].pl,
1368 vm_page_and_queue_spin_unlock(m);
1373 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1377 * Clean out our local marker.
1379 * Page queue still spin-locked.
1381 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1382 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1388 * The number of actually free pages can drop down to v_free_reserved,
1389 * we try to build the free count back above v_free_min. Note that
1390 * vm_paging_needed() also returns TRUE if v_free_count is not at
1391 * least v_free_min so that is the minimum we must build the free
1394 * We use a slightly higher target to improve hysteresis,
1395 * ((v_free_target + v_free_min) / 2). Since v_free_target
1396 * is usually the same as v_cache_min this maintains about
1397 * half the pages in the free queue as are in the cache queue,
1398 * providing pretty good pipelining for pageout operation.
1400 * The system operator can manipulate vm.v_cache_min and
1401 * vm.v_free_target to tune the pageout demon. Be sure
1402 * to keep vm.v_free_min < vm.v_free_target.
1404 * Note that the original paging target is to get at least
1405 * (free_min + cache_min) into (free + cache). The slightly
1406 * higher target will shift additional pages from cache to free
1407 * without effecting the original paging target in order to
1408 * maintain better hysteresis and not have the free count always
1409 * be dead-on v_free_min.
1411 * NOTE: we are still in a critical section.
1413 * Pages moved from PQ_CACHE to totally free are not counted in the
1414 * pages_freed counter.
1417 vm_pageout_scan_cache(int avail_shortage, int pass,
1418 int vnodes_skipped, int recycle_count)
1420 static int lastkillticks;
1421 struct vm_pageout_scan_info info;
1424 while (vmstats.v_free_count <
1425 (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1427 * This steals some code from vm/vm_page.c
1429 static int cache_rover = 0;
1431 m = vm_page_list_find(PQ_CACHE,
1432 cache_rover & PQ_L2_MASK, FALSE);
1435 /* page is returned removed from its queue and spinlocked */
1436 if (vm_page_busy_try(m, TRUE)) {
1437 vm_page_deactivate_locked(m);
1438 vm_page_spin_unlock(m);
1441 vm_page_spin_unlock(m);
1442 pagedaemon_wakeup();
1446 * Remaining operations run with the page busy and neither
1447 * the page or the queue will be spin-locked.
1449 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1452 vm_page_deactivate(m);
1456 KKASSERT((m->flags & PG_MAPPED) == 0);
1457 KKASSERT(m->dirty == 0);
1458 cache_rover += PQ_PRIME2;
1459 vm_pageout_page_free(m);
1460 mycpu->gd_cnt.v_dfree++;
1463 #if !defined(NO_SWAPPING)
1465 * Idle process swapout -- run once per second.
1467 if (vm_swap_idle_enabled) {
1469 if (time_uptime != lsec) {
1470 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1478 * If we didn't get enough free pages, and we have skipped a vnode
1479 * in a writeable object, wakeup the sync daemon. And kick swapout
1480 * if we did not get enough free pages.
1482 if (vm_paging_target() > 0) {
1483 if (vnodes_skipped && vm_page_count_min(0))
1484 speedup_syncer(NULL);
1485 #if !defined(NO_SWAPPING)
1486 if (vm_swap_enabled && vm_page_count_target()) {
1488 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1494 * Handle catastrophic conditions. Under good conditions we should
1495 * be at the target, well beyond our minimum. If we could not even
1496 * reach our minimum the system is under heavy stress. But just being
1497 * under heavy stress does not trigger process killing.
1499 * We consider ourselves to have run out of memory if the swap pager
1500 * is full and avail_shortage is still positive. The secondary check
1501 * ensures that we do not kill processes if the instantanious
1502 * availability is good, even if the pageout demon pass says it
1503 * couldn't get to the target.
1505 if (swap_pager_almost_full &&
1507 (vm_page_count_min(recycle_count) || avail_shortage > 0)) {
1508 kprintf("Warning: system low on memory+swap "
1509 "shortage %d for %d ticks!\n",
1510 avail_shortage, ticks - swap_fail_ticks);
1512 if (swap_pager_full &&
1514 avail_shortage > 0 &&
1515 vm_paging_target() > 0 &&
1516 (unsigned int)(ticks - lastkillticks) >= hz) {
1518 * Kill something, maximum rate once per second to give
1519 * the process time to free up sufficient memory.
1521 lastkillticks = ticks;
1522 info.bigproc = NULL;
1524 allproc_scan(vm_pageout_scan_callback, &info);
1525 if (info.bigproc != NULL) {
1526 info.bigproc->p_nice = PRIO_MIN;
1527 info.bigproc->p_usched->resetpriority(
1528 FIRST_LWP_IN_PROC(info.bigproc));
1529 killproc(info.bigproc, "out of swap space");
1530 wakeup(&vmstats.v_free_count);
1531 PRELE(info.bigproc);
1537 vm_pageout_scan_callback(struct proc *p, void *data)
1539 struct vm_pageout_scan_info *info = data;
1543 * Never kill system processes or init. If we have configured swap
1544 * then try to avoid killing low-numbered pids.
1546 if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1547 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1551 lwkt_gettoken(&p->p_token);
1554 * if the process is in a non-running type state,
1557 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
1558 lwkt_reltoken(&p->p_token);
1563 * Get the approximate process size. Note that anonymous pages
1564 * with backing swap will be counted twice, but there should not
1565 * be too many such pages due to the stress the VM system is
1566 * under at this point.
1568 size = vmspace_anonymous_count(p->p_vmspace) +
1569 vmspace_swap_count(p->p_vmspace);
1572 * If the this process is bigger than the biggest one
1575 if (info->bigsize < size) {
1577 PRELE(info->bigproc);
1580 info->bigsize = size;
1582 lwkt_reltoken(&p->p_token);
1589 * This routine tries to maintain the pseudo LRU active queue,
1590 * so that during long periods of time where there is no paging,
1591 * that some statistic accumulation still occurs. This code
1592 * helps the situation where paging just starts to occur.
1595 vm_pageout_page_stats(int q)
1597 static int fullintervalcount = 0;
1598 struct vm_page marker;
1600 int pcount, tpcount; /* Number of pages to check */
1603 page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1604 vmstats.v_free_min) -
1605 (vmstats.v_free_count + vmstats.v_inactive_count +
1606 vmstats.v_cache_count);
1608 if (page_shortage <= 0)
1611 pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1612 fullintervalcount += vm_pageout_stats_interval;
1613 if (fullintervalcount < vm_pageout_full_stats_interval) {
1614 tpcount = (vm_pageout_stats_max * pcount) /
1615 vmstats.v_page_count + 1;
1616 if (pcount > tpcount)
1619 fullintervalcount = 0;
1622 bzero(&marker, sizeof(marker));
1623 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1624 marker.queue = PQ_ACTIVE + q;
1626 marker.wire_count = 1;
1628 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1629 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1632 * Queue locked at top of loop to avoid stack marker issues.
1634 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1639 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1640 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1641 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1645 * Skip marker pages (atomic against other markers to avoid
1646 * infinite hop-over scans).
1648 if (m->flags & PG_MARKER)
1652 * Ignore pages we can't busy
1654 if (vm_page_busy_try(m, TRUE))
1658 * Remaining operations run with the page busy and neither
1659 * the page or the queue will be spin-locked.
1661 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1662 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1665 * We now have a safely busied page, the page and queue
1666 * spinlocks have been released.
1670 if (m->hold_count) {
1676 * Calculate activity
1679 if (m->flags & PG_REFERENCED) {
1680 vm_page_flag_clear(m, PG_REFERENCED);
1683 actcount += pmap_ts_referenced(m);
1686 * Update act_count and move page to end of queue.
1689 m->act_count += ACT_ADVANCE + actcount;
1690 if (m->act_count > ACT_MAX)
1691 m->act_count = ACT_MAX;
1692 vm_page_and_queue_spin_lock(m);
1693 if (m->queue - m->pc == PQ_ACTIVE) {
1695 &vm_page_queues[PQ_ACTIVE + q].pl,
1698 &vm_page_queues[PQ_ACTIVE + q].pl,
1701 vm_page_and_queue_spin_unlock(m);
1706 if (m->act_count == 0) {
1708 * We turn off page access, so that we have
1709 * more accurate RSS stats. We don't do this
1710 * in the normal page deactivation when the
1711 * system is loaded VM wise, because the
1712 * cost of the large number of page protect
1713 * operations would be higher than the value
1714 * of doing the operation.
1716 * We use the marker to save our place so
1717 * we can release the spin lock. both (m)
1718 * and (next) will be invalid.
1720 vm_page_protect(m, VM_PROT_NONE);
1721 vm_page_deactivate(m);
1723 m->act_count -= min(m->act_count, ACT_DECLINE);
1724 vm_page_and_queue_spin_lock(m);
1725 if (m->queue - m->pc == PQ_ACTIVE) {
1727 &vm_page_queues[PQ_ACTIVE + q].pl,
1730 &vm_page_queues[PQ_ACTIVE + q].pl,
1733 vm_page_and_queue_spin_unlock(m);
1737 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1741 * Remove our local marker
1743 * Page queue still spin-locked.
1745 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1746 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1750 vm_pageout_free_page_calc(vm_size_t count)
1752 if (count < vmstats.v_page_count)
1755 * free_reserved needs to include enough for the largest swap pager
1756 * structures plus enough for any pv_entry structs when paging.
1758 * v_free_min normal allocations
1759 * v_free_reserved system allocations
1760 * v_pageout_free_min allocations by pageout daemon
1761 * v_interrupt_free_min low level allocations (e.g swap structures)
1763 if (vmstats.v_page_count > 1024)
1764 vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1766 vmstats.v_free_min = 64;
1767 vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1768 vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1769 vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1770 vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1777 * vm_pageout is the high level pageout daemon.
1782 vm_pageout_thread(void)
1790 * Initialize some paging parameters.
1792 curthread->td_flags |= TDF_SYSTHREAD;
1794 vm_pageout_free_page_calc(vmstats.v_page_count);
1797 * v_free_target and v_cache_min control pageout hysteresis. Note
1798 * that these are more a measure of the VM cache queue hysteresis
1799 * then the VM free queue. Specifically, v_free_target is the
1800 * high water mark (free+cache pages).
1802 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1803 * low water mark, while v_free_min is the stop. v_cache_min must
1804 * be big enough to handle memory needs while the pageout daemon
1805 * is signalled and run to free more pages.
1807 if (vmstats.v_free_count > 6144)
1808 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1810 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1813 * NOTE: With the new buffer cache b_act_count we want the default
1814 * inactive target to be a percentage of available memory.
1816 * The inactive target essentially determines the minimum
1817 * number of 'temporary' pages capable of caching one-time-use
1818 * files when the VM system is otherwise full of pages
1819 * belonging to multi-time-use files or active program data.
1821 * NOTE: The inactive target is aggressively persued only if the
1822 * inactive queue becomes too small. If the inactive queue
1823 * is large enough to satisfy page movement to free+cache
1824 * then it is repopulated more slowly from the active queue.
1825 * This allows a general inactive_target default to be set.
1827 * There is an issue here for processes which sit mostly idle
1828 * 'overnight', such as sshd, tcsh, and X. Any movement from
1829 * the active queue will eventually cause such pages to
1830 * recycle eventually causing a lot of paging in the morning.
1831 * To reduce the incidence of this pages cycled out of the
1832 * buffer cache are moved directly to the inactive queue if
1833 * they were only used once or twice.
1835 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1836 * Increasing the value (up to 64) increases the number of
1837 * buffer recyclements which go directly to the inactive queue.
1839 if (vmstats.v_free_count > 2048) {
1840 vmstats.v_cache_min = vmstats.v_free_target;
1841 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1843 vmstats.v_cache_min = 0;
1844 vmstats.v_cache_max = 0;
1846 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1848 /* XXX does not really belong here */
1849 if (vm_page_max_wired == 0)
1850 vm_page_max_wired = vmstats.v_free_count / 3;
1852 if (vm_pageout_stats_max == 0)
1853 vm_pageout_stats_max = vmstats.v_free_target;
1856 * Set interval in seconds for stats scan.
1858 if (vm_pageout_stats_interval == 0)
1859 vm_pageout_stats_interval = 5;
1860 if (vm_pageout_full_stats_interval == 0)
1861 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1865 * Set maximum free per pass
1867 if (vm_pageout_stats_free_max == 0)
1868 vm_pageout_stats_free_max = 5;
1870 swap_pager_swap_init();
1874 * The pageout daemon is never done, so loop forever.
1879 int inactive_shortage;
1880 int vnodes_skipped = 0;
1881 int recycle_count = 0;
1885 * Wait for an action request. If we timeout check to
1886 * see if paging is needed (in case the normal wakeup
1889 if (vm_pages_needed == 0) {
1890 error = tsleep(&vm_pages_needed,
1892 vm_pageout_stats_interval * hz);
1894 vm_paging_needed() == 0 &&
1895 vm_pages_needed == 0) {
1896 for (q = 0; q < PQ_L2_SIZE; ++q)
1897 vm_pageout_page_stats(q);
1900 vm_pages_needed = 1;
1903 mycpu->gd_cnt.v_pdwakeups++;
1906 * Do whatever cleanup that the pmap code can.
1911 * Scan for pageout. Try to avoid thrashing the system
1914 * Calculate our target for the number of free+cache pages we
1915 * want to get to. This is higher then the number that causes
1916 * allocations to stall (severe) in order to provide hysteresis,
1917 * and if we don't make it all the way but get to the minimum
1918 * we're happy. Goose it a bit if there are multiple requests
1921 * Don't reduce avail_shortage inside the loop or the
1922 * PQAVERAGE() calculation will break.
1924 avail_shortage = vm_paging_target() + vm_pageout_deficit;
1925 vm_pageout_deficit = 0;
1927 if (avail_shortage > 0) {
1930 for (q = 0; q < PQ_L2_SIZE; ++q) {
1931 delta += vm_pageout_scan_inactive(
1933 (q + q1iterator) & PQ_L2_MASK,
1934 PQAVERAGE(avail_shortage),
1936 if (avail_shortage - delta <= 0)
1939 avail_shortage -= delta;
1944 * Figure out how many active pages we must deactivate. If
1945 * we were able to reach our target with just the inactive
1946 * scan above we limit the number of active pages we
1947 * deactivate to reduce unnecessary work.
1949 inactive_shortage = vmstats.v_inactive_target -
1950 vmstats.v_inactive_count;
1953 * If we were unable to free sufficient inactive pages to
1954 * satisfy the free/cache queue requirements then simply
1955 * reaching the inactive target may not be good enough.
1956 * Try to deactivate pages in excess of the target based
1959 * However to prevent thrashing the VM system do not
1960 * deactivate more than an additional 1/10 the inactive
1961 * target's worth of active pages.
1963 if (avail_shortage > 0) {
1964 tmp = avail_shortage * 2;
1965 if (tmp > vmstats.v_inactive_target / 10)
1966 tmp = vmstats.v_inactive_target / 10;
1967 inactive_shortage += tmp;
1971 * Only trigger on inactive shortage. Triggering on
1972 * avail_shortage can starve the active queue with
1973 * unnecessary active->inactive transitions and destroy
1976 if (/*avail_shortage > 0 ||*/ inactive_shortage > 0) {
1979 for (q = 0; q < PQ_L2_SIZE; ++q) {
1980 delta += vm_pageout_scan_active(
1982 (q + q2iterator) & PQ_L2_MASK,
1983 PQAVERAGE(avail_shortage),
1984 PQAVERAGE(inactive_shortage),
1986 if (inactive_shortage - delta <= 0 &&
1987 avail_shortage - delta <= 0) {
1991 inactive_shortage -= delta;
1992 avail_shortage -= delta;
1997 * Finally free enough cache pages to meet our free page
1998 * requirement and take more drastic measures if we are
2001 vm_pageout_scan_cache(avail_shortage, pass,
2002 vnodes_skipped, recycle_count);
2005 * Wait for more work.
2007 if (avail_shortage > 0) {
2009 if (pass < 10 && vm_pages_needed > 1) {
2011 * Normal operation, additional processes
2012 * have already kicked us. Retry immediately
2013 * unless swap space is completely full in
2014 * which case delay a bit.
2016 if (swap_pager_full) {
2017 tsleep(&vm_pages_needed, 0, "pdelay",
2019 } /* else immediate retry */
2020 } else if (pass < 10) {
2022 * Normal operation, fewer processes. Delay
2023 * a bit but allow wakeups.
2025 vm_pages_needed = 0;
2026 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2027 vm_pages_needed = 1;
2028 } else if (swap_pager_full == 0) {
2030 * We've taken too many passes, forced delay.
2032 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2035 * Running out of memory, catastrophic
2036 * back-off to one-second intervals.
2038 tsleep(&vm_pages_needed, 0, "pdelay", hz);
2040 } else if (vm_pages_needed) {
2042 * Interlocked wakeup of waiters (non-optional).
2044 * Similar to vm_page_free_wakeup() in vm_page.c,
2048 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2049 !vm_page_count_target()) {
2050 vm_pages_needed = 0;
2051 wakeup(&vmstats.v_free_count);
2059 static struct kproc_desc page_kp = {
2064 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp);
2068 * Called after allocating a page out of the cache or free queue
2069 * to possibly wake the pagedaemon up to replentish our supply.
2071 * We try to generate some hysteresis by waking the pagedaemon up
2072 * when our free+cache pages go below the free_min+cache_min level.
2073 * The pagedaemon tries to get the count back up to at least the
2074 * minimum, and through to the target level if possible.
2076 * If the pagedaemon is already active bump vm_pages_needed as a hint
2077 * that there are even more requests pending.
2083 pagedaemon_wakeup(void)
2085 if (vm_paging_needed() && curthread != pagethread) {
2086 if (vm_pages_needed == 0) {
2087 vm_pages_needed = 1; /* SMP race ok */
2088 wakeup(&vm_pages_needed);
2089 } else if (vm_page_count_min(0)) {
2090 ++vm_pages_needed; /* SMP race ok */
2095 #if !defined(NO_SWAPPING)
2102 vm_req_vmdaemon(void)
2104 static int lastrun = 0;
2106 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2107 wakeup(&vm_daemon_needed);
2112 static int vm_daemon_callback(struct proc *p, void *data __unused);
2121 * XXX vm_daemon_needed specific token?
2124 tsleep(&vm_daemon_needed, 0, "psleep", 0);
2125 if (vm_pageout_req_swapout) {
2126 swapout_procs(vm_pageout_req_swapout);
2127 vm_pageout_req_swapout = 0;
2130 * scan the processes for exceeding their rlimits or if
2131 * process is swapped out -- deactivate pages
2133 allproc_scan(vm_daemon_callback, NULL);
2138 vm_daemon_callback(struct proc *p, void *data __unused)
2141 vm_pindex_t limit, size;
2144 * if this is a system process or if we have already
2145 * looked at this process, skip it.
2147 lwkt_gettoken(&p->p_token);
2149 if (p->p_flags & (P_SYSTEM | P_WEXIT)) {
2150 lwkt_reltoken(&p->p_token);
2155 * if the process is in a non-running type state,
2158 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
2159 lwkt_reltoken(&p->p_token);
2166 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2167 p->p_rlimit[RLIMIT_RSS].rlim_max));
2170 * let processes that are swapped out really be
2171 * swapped out. Set the limit to nothing to get as
2172 * many pages out to swap as possible.
2174 if (p->p_flags & P_SWAPPEDOUT)
2179 size = vmspace_resident_count(vm);
2180 if (limit >= 0 && size >= limit) {
2181 vm_pageout_map_deactivate_pages(&vm->vm_map, limit);
2185 lwkt_reltoken(&p->p_token);