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. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
43 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44 * All rights reserved.
46 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48 * Permission to use, copy, modify and distribute this software and
49 * its documentation is hereby granted, provided that both the copyright
50 * notice and this permission notice appear in all copies of the
51 * software, derivative works or modified versions, and any portions
52 * thereof, and that both notices appear in supporting documentation.
54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58 * Carnegie Mellon requests users of this software to return to
60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
61 * School of Computer Science
62 * Carnegie Mellon University
63 * Pittsburgh PA 15213-3890
65 * any improvements or extensions that they make and grant Carnegie the
66 * rights to redistribute these changes.
68 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
69 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.34 2008/04/28 21:16:27 dillon Exp $
73 * The proverbial page-out daemon.
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
81 #include <sys/kthread.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/vnode.h>
85 #include <sys/vmmeter.h>
86 #include <sys/sysctl.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pager.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_extern.h>
99 #include <sys/thread2.h>
100 #include <vm/vm_page2.h>
103 * System initialization
106 /* the kernel process "vm_pageout"*/
107 static void vm_pageout (void);
108 static int vm_pageout_clean (vm_page_t);
109 static void vm_pageout_scan (int pass);
110 static int vm_pageout_free_page_calc (vm_size_t count);
111 struct thread *pagethread;
113 static struct kproc_desc page_kp = {
118 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
120 #if !defined(NO_SWAPPING)
121 /* the kernel process "vm_daemon"*/
122 static void vm_daemon (void);
123 static struct thread *vmthread;
125 static struct kproc_desc vm_kp = {
130 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
134 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */
135 int vm_pageout_deficit=0; /* Estimated number of pages deficit */
136 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */
138 #if !defined(NO_SWAPPING)
139 static int vm_pageout_req_swapout; /* XXX */
140 static int vm_daemon_needed;
142 extern int vm_swap_size;
143 static int vm_max_launder = 32;
144 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
145 static int vm_pageout_full_stats_interval = 0;
146 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
147 static int defer_swap_pageouts=0;
148 static int disable_swap_pageouts=0;
150 #if defined(NO_SWAPPING)
151 static int vm_swap_enabled=0;
152 static int vm_swap_idle_enabled=0;
154 static int vm_swap_enabled=1;
155 static int vm_swap_idle_enabled=0;
158 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
159 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
161 SYSCTL_INT(_vm, OID_AUTO, max_launder,
162 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
164 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
165 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
167 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
168 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
171 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
174 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
176 #if defined(NO_SWAPPING)
177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
178 CTLFLAG_RD, &vm_swap_enabled, 0, "");
179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
180 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
182 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
183 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
184 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
185 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
189 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
191 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
192 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
194 static int pageout_lock_miss;
195 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
196 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
199 SYSCTL_INT(_vm, OID_AUTO, vm_load,
200 CTLFLAG_RD, &vm_load, 0, "load on the VM system");
201 int vm_load_enable = 1;
202 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
203 CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
206 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
207 CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
210 #define VM_PAGEOUT_PAGE_COUNT 16
211 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
213 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
215 #if !defined(NO_SWAPPING)
216 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
217 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
218 static freeer_fcn_t vm_pageout_object_deactivate_pages;
219 static void vm_req_vmdaemon (void);
221 static void vm_pageout_page_stats(void);
227 vm_fault_ratecheck(void)
229 if (vm_pages_needed) {
241 * Clean the page and remove it from the laundry. The page must not be
244 * We set the busy bit to cause potential page faults on this page to
245 * block. Note the careful timing, however, the busy bit isn't set till
246 * late and we cannot do anything that will mess with the page.
250 vm_pageout_clean(vm_page_t m)
253 vm_page_t mc[2*vm_pageout_page_count];
255 int ib, is, page_base;
256 vm_pindex_t pindex = m->pindex;
261 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
262 * with the new swapper, but we could have serious problems paging
263 * out other object types if there is insufficient memory.
265 * Unfortunately, checking free memory here is far too late, so the
266 * check has been moved up a procedural level.
270 * Don't mess with the page if it's busy, held, or special
272 if ((m->hold_count != 0) ||
273 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
277 mc[vm_pageout_page_count] = m;
279 page_base = vm_pageout_page_count;
284 * Scan object for clusterable pages.
286 * We can cluster ONLY if: ->> the page is NOT
287 * clean, wired, busy, held, or mapped into a
288 * buffer, and one of the following:
289 * 1) The page is inactive, or a seldom used
292 * 2) we force the issue.
294 * During heavy mmap/modification loads the pageout
295 * daemon can really fragment the underlying file
296 * due to flushing pages out of order and not trying
297 * align the clusters (which leave sporatic out-of-order
298 * holes). To solve this problem we do the reverse scan
299 * first and attempt to align our cluster, then do a
300 * forward scan if room remains.
304 while (ib && pageout_count < vm_pageout_page_count) {
312 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
316 if (((p->queue - p->pc) == PQ_CACHE) ||
317 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
321 vm_page_test_dirty(p);
322 if ((p->dirty & p->valid) == 0 ||
323 p->queue != PQ_INACTIVE ||
324 p->wire_count != 0 || /* may be held by buf cache */
325 p->hold_count != 0) { /* may be undergoing I/O */
333 * alignment boundry, stop here and switch directions. Do
336 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
340 while (pageout_count < vm_pageout_page_count &&
341 pindex + is < object->size) {
344 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
346 if (((p->queue - p->pc) == PQ_CACHE) ||
347 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
350 vm_page_test_dirty(p);
351 if ((p->dirty & p->valid) == 0 ||
352 p->queue != PQ_INACTIVE ||
353 p->wire_count != 0 || /* may be held by buf cache */
354 p->hold_count != 0) { /* may be undergoing I/O */
357 mc[page_base + pageout_count] = p;
363 * If we exhausted our forward scan, continue with the reverse scan
364 * when possible, even past a page boundry. This catches boundry
367 if (ib && pageout_count < vm_pageout_page_count)
371 * we allow reads during pageouts...
373 return vm_pageout_flush(&mc[page_base], pageout_count, 0);
377 * vm_pageout_flush() - launder the given pages
379 * The given pages are laundered. Note that we setup for the start of
380 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
381 * reference count all in here rather then in the parent. If we want
382 * the parent to do more sophisticated things we may have to change
387 vm_pageout_flush(vm_page_t *mc, int count, int flags)
390 int pageout_status[count];
395 * Initiate I/O. Bump the vm_page_t->busy counter and
396 * mark the pages read-only.
398 * We must make the pages read-only. This will also force the
399 * modified bit in the related pmaps to be cleared. The pager
400 * cannot clear the bit for us since the I/O completion code
401 * typically runs from an interrupt. The act of making the page
402 * read-only handles the case for us.
405 for (i = 0; i < count; i++) {
406 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
407 vm_page_io_start(mc[i]);
408 vm_page_protect(mc[i], VM_PROT_READ);
411 object = mc[0]->object;
412 vm_object_pip_add(object, count);
414 vm_pager_put_pages(object, mc, count,
415 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
418 for (i = 0; i < count; i++) {
419 vm_page_t mt = mc[i];
421 switch (pageout_status[i]) {
430 * Page outside of range of object. Right now we
431 * essentially lose the changes by pretending it
434 pmap_clear_modify(mt);
440 * If page couldn't be paged out, then reactivate the
441 * page so it doesn't clog the inactive list. (We
442 * will try paging out it again later).
444 vm_page_activate(mt);
451 * If the operation is still going, leave the page busy to
452 * block all other accesses. Also, leave the paging in
453 * progress indicator set so that we don't attempt an object
456 * For any pages which have completed synchronously,
457 * deactivate the page if we are under a severe deficit.
458 * Do not try to enter them into the cache, though, they
459 * might still be read-heavy.
461 if (pageout_status[i] != VM_PAGER_PEND) {
462 vm_object_pip_wakeup(object);
463 vm_page_io_finish(mt);
464 if (vm_page_count_severe())
465 vm_page_deactivate(mt);
467 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
468 vm_page_protect(mt, VM_PROT_READ);
475 #if !defined(NO_SWAPPING)
477 * vm_pageout_object_deactivate_pages
479 * deactivate enough pages to satisfy the inactive target
480 * requirements or if vm_page_proc_limit is set, then
481 * deactivate all of the pages in the object and its
484 * The object and map must be locked.
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;
495 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
499 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
501 if (object->paging_in_progress)
504 remove_mode = map_remove_only;
505 if (object->shadow_count > 1)
509 * scan the objects entire memory queue. spl protection is
510 * required to avoid an interrupt unbusy/free race against
514 info.limit = remove_mode;
516 info.desired = desired;
517 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
518 vm_pageout_object_deactivate_pages_callback,
522 object = object->backing_object;
527 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
529 struct rb_vm_page_scan_info *info = data;
532 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
535 mycpu->gd_cnt.v_pdpages++;
536 if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 ||
537 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
538 !pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
542 actcount = pmap_ts_referenced(p);
544 vm_page_flag_set(p, PG_REFERENCED);
545 } else if (p->flags & PG_REFERENCED) {
549 if ((p->queue != PQ_ACTIVE) &&
550 (p->flags & PG_REFERENCED)) {
552 p->act_count += actcount;
553 vm_page_flag_clear(p, PG_REFERENCED);
554 } else if (p->queue == PQ_ACTIVE) {
555 if ((p->flags & PG_REFERENCED) == 0) {
556 p->act_count -= min(p->act_count, ACT_DECLINE);
557 if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) {
558 vm_page_protect(p, VM_PROT_NONE);
559 vm_page_deactivate(p);
561 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
562 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
566 vm_page_flag_clear(p, PG_REFERENCED);
567 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
568 p->act_count += ACT_ADVANCE;
569 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
570 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
572 } else if (p->queue == PQ_INACTIVE) {
573 vm_page_protect(p, VM_PROT_NONE);
579 * deactivate some number of pages in a map, try to do it fairly, but
580 * that is really hard to do.
583 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
586 vm_object_t obj, bigobj;
589 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
597 * first, search out the biggest object, and try to free pages from
600 tmpe = map->header.next;
601 while (tmpe != &map->header) {
602 switch(tmpe->maptype) {
603 case VM_MAPTYPE_NORMAL:
604 case VM_MAPTYPE_VPAGETABLE:
605 obj = tmpe->object.vm_object;
606 if ((obj != NULL) && (obj->shadow_count <= 1) &&
608 (bigobj->resident_page_count < obj->resident_page_count))) {
615 if (tmpe->wired_count > 0)
616 nothingwired = FALSE;
621 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
624 * Next, hunt around for other pages to deactivate. We actually
625 * do this search sort of wrong -- .text first is not the best idea.
627 tmpe = map->header.next;
628 while (tmpe != &map->header) {
629 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
631 switch(tmpe->maptype) {
632 case VM_MAPTYPE_NORMAL:
633 case VM_MAPTYPE_VPAGETABLE:
634 obj = tmpe->object.vm_object;
636 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
645 * Remove all mappings if a process is swapped out, this will free page
648 if (desired == 0 && nothingwired)
649 pmap_remove(vm_map_pmap(map),
650 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
656 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We
657 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
658 * be trivially freed.
661 vm_pageout_page_free(vm_page_t m)
663 vm_object_t object = m->object;
664 int type = object->type;
666 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
667 vm_object_reference(object);
669 vm_page_protect(m, VM_PROT_NONE);
671 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
672 vm_object_deallocate(object);
676 * vm_pageout_scan does the dirty work for the pageout daemon.
679 struct vm_pageout_scan_info {
680 struct proc *bigproc;
684 static int vm_pageout_scan_callback(struct proc *p, void *data);
687 vm_pageout_scan(int pass)
689 struct vm_pageout_scan_info info;
691 struct vm_page marker;
692 int page_shortage, maxscan, pcount;
693 int addl_page_shortage, addl_page_shortage_init;
696 int vnodes_skipped = 0;
700 * Do whatever cleanup that the pmap code can.
704 addl_page_shortage_init = vm_pageout_deficit;
705 vm_pageout_deficit = 0;
708 * Calculate the number of pages we want to either free or move
711 page_shortage = vm_paging_target() + addl_page_shortage_init;
714 * Initialize our marker
716 bzero(&marker, sizeof(marker));
717 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
718 marker.queue = PQ_INACTIVE;
719 marker.wire_count = 1;
722 * Start scanning the inactive queue for pages we can move to the
723 * cache or free. The scan will stop when the target is reached or
724 * we have scanned the entire inactive queue. Note that m->act_count
725 * is not used to form decisions for the inactive queue, only for the
728 * maxlaunder limits the number of dirty pages we flush per scan.
729 * For most systems a smaller value (16 or 32) is more robust under
730 * extreme memory and disk pressure because any unnecessary writes
731 * to disk can result in extreme performance degredation. However,
732 * systems with excessive dirty pages (especially when MAP_NOSYNC is
733 * used) will die horribly with limited laundering. If the pageout
734 * daemon cannot clean enough pages in the first pass, we let it go
735 * all out in succeeding passes.
737 if ((maxlaunder = vm_max_launder) <= 1)
743 * We will generally be in a critical section throughout the
744 * scan, but we can release it temporarily when we are sitting on a
745 * non-busy page without fear. this is required to prevent an
746 * interrupt from unbusying or freeing a page prior to our busy
747 * check, leaving us on the wrong queue or checking the wrong
752 addl_page_shortage = addl_page_shortage_init;
753 maxscan = vmstats.v_inactive_count;
754 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
755 m != NULL && maxscan-- > 0 && page_shortage > 0;
758 mycpu->gd_cnt.v_pdpages++;
761 * Give interrupts a chance
767 * It's easier for some of the conditions below to just loop
768 * and catch queue changes here rather then check everywhere
771 if (m->queue != PQ_INACTIVE)
773 next = TAILQ_NEXT(m, pageq);
778 if (m->flags & PG_MARKER)
782 * A held page may be undergoing I/O, so skip it.
785 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
786 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
787 addl_page_shortage++;
792 * Dont mess with busy pages, keep in the front of the
793 * queue, most likely are being paged out.
795 if (m->busy || (m->flags & PG_BUSY)) {
796 addl_page_shortage++;
800 if (m->object->ref_count == 0) {
802 * If the object is not being used, we ignore previous
805 vm_page_flag_clear(m, PG_REFERENCED);
806 pmap_clear_reference(m);
808 } else if (((m->flags & PG_REFERENCED) == 0) &&
809 (actcount = pmap_ts_referenced(m))) {
811 * Otherwise, if the page has been referenced while
812 * in the inactive queue, we bump the "activation
813 * count" upwards, making it less likely that the
814 * page will be added back to the inactive queue
815 * prematurely again. Here we check the page tables
816 * (or emulated bits, if any), given the upper level
817 * VM system not knowing anything about existing
821 m->act_count += (actcount + ACT_ADVANCE);
826 * If the upper level VM system knows about any page
827 * references, we activate the page. We also set the
828 * "activation count" higher than normal so that we will less
829 * likely place pages back onto the inactive queue again.
831 if ((m->flags & PG_REFERENCED) != 0) {
832 vm_page_flag_clear(m, PG_REFERENCED);
833 actcount = pmap_ts_referenced(m);
835 m->act_count += (actcount + ACT_ADVANCE + 1);
840 * If the upper level VM system doesn't know anything about
841 * the page being dirty, we have to check for it again. As
842 * far as the VM code knows, any partially dirty pages are
845 * Pages marked PG_WRITEABLE may be mapped into the user
846 * address space of a process running on another cpu. A
847 * user process (without holding the MP lock) running on
848 * another cpu may be able to touch the page while we are
849 * trying to remove it. To prevent this from occuring we
850 * must call pmap_remove_all() or otherwise make the page
851 * read-only. If the race occured pmap_remove_all() is
852 * responsible for setting m->dirty.
855 vm_page_test_dirty(m);
857 if (m->dirty == 0 && (m->flags & PG_WRITEABLE) != 0)
866 * Invalid pages can be easily freed
868 vm_pageout_page_free(m);
869 mycpu->gd_cnt.v_dfree++;
871 } else if (m->dirty == 0) {
873 * Clean pages can be placed onto the cache queue.
874 * This effectively frees them.
878 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
880 * Dirty pages need to be paged out, but flushing
881 * a page is extremely expensive verses freeing
882 * a clean page. Rather then artificially limiting
883 * the number of pages we can flush, we instead give
884 * dirty pages extra priority on the inactive queue
885 * by forcing them to be cycled through the queue
886 * twice before being flushed, after which the
887 * (now clean) page will cycle through once more
888 * before being freed. This significantly extends
889 * the thrash point for a heavily loaded machine.
891 vm_page_flag_set(m, PG_WINATCFLS);
892 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
893 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
894 } else if (maxlaunder > 0) {
896 * We always want to try to flush some dirty pages if
897 * we encounter them, to keep the system stable.
898 * Normally this number is small, but under extreme
899 * pressure where there are insufficient clean pages
900 * on the inactive queue, we may have to go all out.
902 int swap_pageouts_ok;
903 struct vnode *vp = NULL;
907 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
908 swap_pageouts_ok = 1;
910 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
911 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
912 vm_page_count_min());
917 * We don't bother paging objects that are "dead".
918 * Those objects are in a "rundown" state.
920 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
921 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
922 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
927 * The object is already known NOT to be dead. It
928 * is possible for the vget() to block the whole
929 * pageout daemon, but the new low-memory handling
930 * code should prevent it.
932 * The previous code skipped locked vnodes and, worse,
933 * reordered pages in the queue. This results in
934 * completely non-deterministic operation because,
935 * quite often, a vm_fault has initiated an I/O and
936 * is holding a locked vnode at just the point where
937 * the pageout daemon is woken up.
939 * We can't wait forever for the vnode lock, we might
940 * deadlock due to a vn_read() getting stuck in
941 * vm_wait while holding this vnode. We skip the
942 * vnode if we can't get it in a reasonable amount
946 if (object->type == OBJT_VNODE) {
949 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) {
951 if (object->flags & OBJ_MIGHTBEDIRTY)
957 * The page might have been moved to another
958 * queue during potential blocking in vget()
959 * above. The page might have been freed and
960 * reused for another vnode. The object might
961 * have been reused for another vnode.
963 if (m->queue != PQ_INACTIVE ||
964 m->object != object ||
965 object->handle != vp) {
966 if (object->flags & OBJ_MIGHTBEDIRTY)
973 * The page may have been busied during the
974 * blocking in vput(); We don't move the
975 * page back onto the end of the queue so that
976 * statistics are more correct if we don't.
978 if (m->busy || (m->flags & PG_BUSY)) {
984 * If the page has become held it might
985 * be undergoing I/O, so skip it
988 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
989 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
990 if (object->flags & OBJ_MIGHTBEDIRTY)
998 * If a page is dirty, then it is either being washed
999 * (but not yet cleaned) or it is still in the
1000 * laundry. If it is still in the laundry, then we
1001 * start the cleaning operation.
1003 * This operation may cluster, invalidating the 'next'
1004 * pointer. To prevent an inordinate number of
1005 * restarts we use our marker to remember our place.
1007 * decrement page_shortage on success to account for
1008 * the (future) cleaned page. Otherwise we could wind
1009 * up laundering or cleaning too many pages.
1011 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1012 if (vm_pageout_clean(m) != 0) {
1016 next = TAILQ_NEXT(&marker, pageq);
1017 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1024 * Compute the number of pages we want to try to move from the
1025 * active queue to the inactive queue.
1027 page_shortage = vm_paging_target() +
1028 vmstats.v_inactive_target - vmstats.v_inactive_count;
1029 page_shortage += addl_page_shortage;
1032 * Scan the active queue for things we can deactivate. We nominally
1033 * track the per-page activity counter and use it to locate
1034 * deactivation candidates.
1036 * NOTE: we are still in a critical section.
1038 pcount = vmstats.v_active_count;
1039 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1041 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1043 * Give interrupts a chance.
1049 * If the page was ripped out from under us, just stop.
1051 if (m->queue != PQ_ACTIVE)
1053 next = TAILQ_NEXT(m, pageq);
1056 * Don't deactivate pages that are busy.
1058 if ((m->busy != 0) ||
1059 (m->flags & PG_BUSY) ||
1060 (m->hold_count != 0)) {
1061 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1062 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1068 * The count for pagedaemon pages is done after checking the
1069 * page for eligibility...
1071 mycpu->gd_cnt.v_pdpages++;
1074 * Check to see "how much" the page has been used.
1077 if (m->object->ref_count != 0) {
1078 if (m->flags & PG_REFERENCED) {
1081 actcount += pmap_ts_referenced(m);
1083 m->act_count += ACT_ADVANCE + actcount;
1084 if (m->act_count > ACT_MAX)
1085 m->act_count = ACT_MAX;
1090 * Since we have "tested" this bit, we need to clear it now.
1092 vm_page_flag_clear(m, PG_REFERENCED);
1095 * Only if an object is currently being used, do we use the
1096 * page activation count stats.
1098 if (actcount && (m->object->ref_count != 0)) {
1099 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1100 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1102 m->act_count -= min(m->act_count, ACT_DECLINE);
1103 if (vm_pageout_algorithm ||
1104 m->object->ref_count == 0 ||
1105 m->act_count < pass) {
1107 if (m->object->ref_count == 0) {
1108 vm_page_protect(m, VM_PROT_NONE);
1112 vm_page_deactivate(m);
1114 vm_page_deactivate(m);
1117 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1118 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1125 * We try to maintain some *really* free pages, this allows interrupt
1126 * code to be guaranteed space. Since both cache and free queues
1127 * are considered basically 'free', moving pages from cache to free
1128 * does not effect other calculations.
1130 * NOTE: we are still in a critical section.
1133 while (vmstats.v_free_count < vmstats.v_free_reserved) {
1134 static int cache_rover = 0;
1135 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1138 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1143 kprintf("Warning: busy page %p found in cache\n", m);
1145 vm_page_deactivate(m);
1148 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1149 vm_pageout_page_free(m);
1150 mycpu->gd_cnt.v_dfree++;
1155 #if !defined(NO_SWAPPING)
1157 * Idle process swapout -- run once per second.
1159 if (vm_swap_idle_enabled) {
1161 if (time_second != lsec) {
1162 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1170 * If we didn't get enough free pages, and we have skipped a vnode
1171 * in a writeable object, wakeup the sync daemon. And kick swapout
1172 * if we did not get enough free pages.
1174 if (vm_paging_target() > 0) {
1175 if (vnodes_skipped && vm_page_count_min())
1177 #if !defined(NO_SWAPPING)
1178 if (vm_swap_enabled && vm_page_count_target()) {
1180 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1186 * If we are out of swap and were not able to reach our paging
1187 * target, kill the largest process.
1189 if ((vm_swap_size < 64 && vm_page_count_min()) ||
1190 (swap_pager_full && vm_paging_target() > 0)) {
1192 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) {
1194 info.bigproc = NULL;
1196 allproc_scan(vm_pageout_scan_callback, &info);
1197 if (info.bigproc != NULL) {
1198 killproc(info.bigproc, "out of swap space");
1199 info.bigproc->p_nice = PRIO_MIN;
1200 info.bigproc->p_usched->resetpriority(
1201 FIRST_LWP_IN_PROC(info.bigproc));
1202 wakeup(&vmstats.v_free_count);
1203 PRELE(info.bigproc);
1209 vm_pageout_scan_callback(struct proc *p, void *data)
1211 struct vm_pageout_scan_info *info = data;
1215 * if this is a system process, skip it
1217 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1218 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1223 * if the process is in a non-running type state,
1226 if (p->p_stat != SACTIVE && p->p_stat != SSTOP) {
1231 * get the process size
1233 size = vmspace_resident_count(p->p_vmspace) +
1234 vmspace_swap_count(p->p_vmspace);
1237 * If the this process is bigger than the biggest one
1240 if (size > info->bigsize) {
1242 PRELE(info->bigproc);
1245 info->bigsize = size;
1251 * This routine tries to maintain the pseudo LRU active queue,
1252 * so that during long periods of time where there is no paging,
1253 * that some statistic accumulation still occurs. This code
1254 * helps the situation where paging just starts to occur.
1257 vm_pageout_page_stats(void)
1260 int pcount,tpcount; /* Number of pages to check */
1261 static int fullintervalcount = 0;
1265 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1266 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1268 if (page_shortage <= 0)
1273 pcount = vmstats.v_active_count;
1274 fullintervalcount += vm_pageout_stats_interval;
1275 if (fullintervalcount < vm_pageout_full_stats_interval) {
1276 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1277 if (pcount > tpcount)
1280 fullintervalcount = 0;
1283 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1284 while ((m != NULL) && (pcount-- > 0)) {
1287 if (m->queue != PQ_ACTIVE) {
1291 next = TAILQ_NEXT(m, pageq);
1293 * Don't deactivate pages that are busy.
1295 if ((m->busy != 0) ||
1296 (m->flags & PG_BUSY) ||
1297 (m->hold_count != 0)) {
1298 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1299 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1305 if (m->flags & PG_REFERENCED) {
1306 vm_page_flag_clear(m, PG_REFERENCED);
1310 actcount += pmap_ts_referenced(m);
1312 m->act_count += ACT_ADVANCE + actcount;
1313 if (m->act_count > ACT_MAX)
1314 m->act_count = ACT_MAX;
1315 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1316 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1318 if (m->act_count == 0) {
1320 * We turn off page access, so that we have
1321 * more accurate RSS stats. We don't do this
1322 * in the normal page deactivation when the
1323 * system is loaded VM wise, because the
1324 * cost of the large number of page protect
1325 * operations would be higher than the value
1326 * of doing the operation.
1328 vm_page_protect(m, VM_PROT_NONE);
1329 vm_page_deactivate(m);
1331 m->act_count -= min(m->act_count, ACT_DECLINE);
1332 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1333 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1343 vm_pageout_free_page_calc(vm_size_t count)
1345 if (count < vmstats.v_page_count)
1348 * free_reserved needs to include enough for the largest swap pager
1349 * structures plus enough for any pv_entry structs when paging.
1351 if (vmstats.v_page_count > 1024)
1352 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1354 vmstats.v_free_min = 4;
1355 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1356 vmstats.v_interrupt_free_min;
1357 vmstats.v_free_reserved = vm_pageout_page_count +
1358 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1359 vmstats.v_free_severe = vmstats.v_free_min / 2;
1360 vmstats.v_free_min += vmstats.v_free_reserved;
1361 vmstats.v_free_severe += vmstats.v_free_reserved;
1367 * vm_pageout is the high level pageout daemon.
1375 * Initialize some paging parameters.
1378 vmstats.v_interrupt_free_min = 2;
1379 if (vmstats.v_page_count < 2000)
1380 vm_pageout_page_count = 8;
1382 vm_pageout_free_page_calc(vmstats.v_page_count);
1384 * v_free_target and v_cache_min control pageout hysteresis. Note
1385 * that these are more a measure of the VM cache queue hysteresis
1386 * then the VM free queue. Specifically, v_free_target is the
1387 * high water mark (free+cache pages).
1389 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1390 * low water mark, while v_free_min is the stop. v_cache_min must
1391 * be big enough to handle memory needs while the pageout daemon
1392 * is signalled and run to free more pages.
1394 if (vmstats.v_free_count > 6144)
1395 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1397 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1399 if (vmstats.v_free_count > 2048) {
1400 vmstats.v_cache_min = vmstats.v_free_target;
1401 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1402 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2;
1404 vmstats.v_cache_min = 0;
1405 vmstats.v_cache_max = 0;
1406 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1408 if (vmstats.v_inactive_target > vmstats.v_free_count / 3)
1409 vmstats.v_inactive_target = vmstats.v_free_count / 3;
1411 /* XXX does not really belong here */
1412 if (vm_page_max_wired == 0)
1413 vm_page_max_wired = vmstats.v_free_count / 3;
1415 if (vm_pageout_stats_max == 0)
1416 vm_pageout_stats_max = vmstats.v_free_target;
1419 * Set interval in seconds for stats scan.
1421 if (vm_pageout_stats_interval == 0)
1422 vm_pageout_stats_interval = 5;
1423 if (vm_pageout_full_stats_interval == 0)
1424 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1428 * Set maximum free per pass
1430 if (vm_pageout_stats_free_max == 0)
1431 vm_pageout_stats_free_max = 5;
1433 swap_pager_swap_init();
1436 * The pageout daemon is never done, so loop forever.
1442 * If we have enough free memory, wakeup waiters. Do
1443 * not clear vm_pages_needed until we reach our target,
1444 * otherwise we may be woken up over and over again and
1445 * waste a lot of cpu.
1448 if (vm_pages_needed && !vm_page_count_min()) {
1449 if (vm_paging_needed() <= 0)
1450 vm_pages_needed = 0;
1451 wakeup(&vmstats.v_free_count);
1453 if (vm_pages_needed) {
1455 * Still not done, take a second pass without waiting
1456 * (unlimited dirty cleaning), otherwise sleep a bit
1461 tsleep(&vm_pages_needed, 0, "psleep", hz/2);
1464 * Good enough, sleep & handle stats. Prime the pass
1471 error = tsleep(&vm_pages_needed,
1472 0, "psleep", vm_pageout_stats_interval * hz);
1473 if (error && !vm_pages_needed) {
1476 vm_pageout_page_stats();
1481 if (vm_pages_needed)
1482 mycpu->gd_cnt.v_pdwakeups++;
1484 vm_pageout_scan(pass);
1485 vm_pageout_deficit = 0;
1490 pagedaemon_wakeup(void)
1492 if (!vm_pages_needed && curthread != pagethread) {
1494 wakeup(&vm_pages_needed);
1498 #if !defined(NO_SWAPPING)
1500 vm_req_vmdaemon(void)
1502 static int lastrun = 0;
1504 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1505 wakeup(&vm_daemon_needed);
1510 static int vm_daemon_callback(struct proc *p, void *data __unused);
1516 tsleep(&vm_daemon_needed, 0, "psleep", 0);
1517 if (vm_pageout_req_swapout) {
1518 swapout_procs(vm_pageout_req_swapout);
1519 vm_pageout_req_swapout = 0;
1522 * scan the processes for exceeding their rlimits or if
1523 * process is swapped out -- deactivate pages
1525 allproc_scan(vm_daemon_callback, NULL);
1530 vm_daemon_callback(struct proc *p, void *data __unused)
1532 vm_pindex_t limit, size;
1535 * if this is a system process or if we have already
1536 * looked at this process, skip it.
1538 if (p->p_flag & (P_SYSTEM | P_WEXIT))
1542 * if the process is in a non-running type state,
1545 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1551 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1552 p->p_rlimit[RLIMIT_RSS].rlim_max));
1555 * let processes that are swapped out really be
1556 * swapped out. Set the limit to nothing to get as
1557 * many pages out to swap as possible.
1559 if (p->p_flag & P_SWAPPEDOUT)
1562 size = vmspace_resident_count(p->p_vmspace);
1563 if (limit >= 0 && size >= limit) {
1564 vm_pageout_map_deactivate_pages(
1565 &p->p_vmspace->vm_map, limit);