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 1.8 2003/08/20 08:03:01 rob 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>
98 #include <vm/vm_page2.h>
101 * System initialization
104 /* the kernel process "vm_pageout"*/
105 static void vm_pageout (void);
106 static int vm_pageout_clean (vm_page_t);
107 static void vm_pageout_scan (int pass);
108 static int vm_pageout_free_page_calc (vm_size_t count);
109 struct thread *pagethread;
111 static struct kproc_desc page_kp = {
116 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
118 #if !defined(NO_SWAPPING)
119 /* the kernel process "vm_daemon"*/
120 static void vm_daemon (void);
121 static struct thread *vmthread;
123 static struct kproc_desc vm_kp = {
128 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
132 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */
133 int vm_pageout_deficit=0; /* Estimated number of pages deficit */
134 int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */
136 #if !defined(NO_SWAPPING)
137 static int vm_pageout_req_swapout; /* XXX */
138 static int vm_daemon_needed;
140 extern int vm_swap_size;
141 static int vm_max_launder = 32;
142 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
143 static int vm_pageout_full_stats_interval = 0;
144 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
145 static int defer_swap_pageouts=0;
146 static int disable_swap_pageouts=0;
148 #if defined(NO_SWAPPING)
149 static int vm_swap_enabled=0;
150 static int vm_swap_idle_enabled=0;
152 static int vm_swap_enabled=1;
153 static int vm_swap_idle_enabled=0;
156 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
157 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
159 SYSCTL_INT(_vm, OID_AUTO, max_launder,
160 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
162 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
163 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
165 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
166 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
168 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
169 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
171 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
172 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
174 #if defined(NO_SWAPPING)
175 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
176 CTLFLAG_RD, &vm_swap_enabled, 0, "");
177 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
178 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
180 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
181 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
182 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
183 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
186 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
187 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
189 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
190 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
192 static int pageout_lock_miss;
193 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
194 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
196 #define VM_PAGEOUT_PAGE_COUNT 16
197 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
199 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
201 #if !defined(NO_SWAPPING)
202 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
203 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
204 static freeer_fcn_t vm_pageout_object_deactivate_pages;
205 static void vm_req_vmdaemon (void);
207 static void vm_pageout_page_stats(void);
212 * Clean the page and remove it from the laundry.
214 * We set the busy bit to cause potential page faults on this page to
215 * block. Note the careful timing, however, the busy bit isn't set till
216 * late and we cannot do anything that will mess with the page.
224 vm_page_t mc[2*vm_pageout_page_count];
226 int ib, is, page_base;
227 vm_pindex_t pindex = m->pindex;
232 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
233 * with the new swapper, but we could have serious problems paging
234 * out other object types if there is insufficient memory.
236 * Unfortunately, checking free memory here is far too late, so the
237 * check has been moved up a procedural level.
241 * Don't mess with the page if it's busy, held, or special
243 if ((m->hold_count != 0) ||
244 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
248 mc[vm_pageout_page_count] = m;
250 page_base = vm_pageout_page_count;
255 * Scan object for clusterable pages.
257 * We can cluster ONLY if: ->> the page is NOT
258 * clean, wired, busy, held, or mapped into a
259 * buffer, and one of the following:
260 * 1) The page is inactive, or a seldom used
263 * 2) we force the issue.
265 * During heavy mmap/modification loads the pageout
266 * daemon can really fragment the underlying file
267 * due to flushing pages out of order and not trying
268 * align the clusters (which leave sporatic out-of-order
269 * holes). To solve this problem we do the reverse scan
270 * first and attempt to align our cluster, then do a
271 * forward scan if room remains.
275 while (ib && pageout_count < vm_pageout_page_count) {
283 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
287 if (((p->queue - p->pc) == PQ_CACHE) ||
288 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
292 vm_page_test_dirty(p);
293 if ((p->dirty & p->valid) == 0 ||
294 p->queue != PQ_INACTIVE ||
295 p->wire_count != 0 || /* may be held by buf cache */
296 p->hold_count != 0) { /* may be undergoing I/O */
304 * alignment boundry, stop here and switch directions. Do
307 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
311 while (pageout_count < vm_pageout_page_count &&
312 pindex + is < object->size) {
315 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
317 if (((p->queue - p->pc) == PQ_CACHE) ||
318 (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 */
328 mc[page_base + pageout_count] = p;
334 * If we exhausted our forward scan, continue with the reverse scan
335 * when possible, even past a page boundry. This catches boundry
338 if (ib && pageout_count < vm_pageout_page_count)
342 * we allow reads during pageouts...
344 return vm_pageout_flush(&mc[page_base], pageout_count, 0);
348 * vm_pageout_flush() - launder the given pages
350 * The given pages are laundered. Note that we setup for the start of
351 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
352 * reference count all in here rather then in the parent. If we want
353 * the parent to do more sophisticated things we may have to change
358 vm_pageout_flush(mc, count, flags)
364 int pageout_status[count];
369 * Initiate I/O. Bump the vm_page_t->busy counter and
370 * mark the pages read-only.
372 * We do not have to fixup the clean/dirty bits here... we can
373 * allow the pager to do it after the I/O completes.
376 for (i = 0; i < count; i++) {
377 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
378 vm_page_io_start(mc[i]);
379 vm_page_protect(mc[i], VM_PROT_READ);
382 object = mc[0]->object;
383 vm_object_pip_add(object, count);
385 vm_pager_put_pages(object, mc, count,
386 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
389 for (i = 0; i < count; i++) {
390 vm_page_t mt = mc[i];
392 switch (pageout_status[i]) {
401 * Page outside of range of object. Right now we
402 * essentially lose the changes by pretending it
405 pmap_clear_modify(mt);
411 * If page couldn't be paged out, then reactivate the
412 * page so it doesn't clog the inactive list. (We
413 * will try paging out it again later).
415 vm_page_activate(mt);
422 * If the operation is still going, leave the page busy to
423 * block all other accesses. Also, leave the paging in
424 * progress indicator set so that we don't attempt an object
427 if (pageout_status[i] != VM_PAGER_PEND) {
428 vm_object_pip_wakeup(object);
429 vm_page_io_finish(mt);
430 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
431 vm_page_protect(mt, VM_PROT_READ);
437 #if !defined(NO_SWAPPING)
439 * vm_pageout_object_deactivate_pages
441 * deactivate enough pages to satisfy the inactive target
442 * requirements or if vm_page_proc_limit is set, then
443 * deactivate all of the pages in the object and its
446 * The object and map must be locked.
449 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only)
460 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
464 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
466 if (object->paging_in_progress)
469 remove_mode = map_remove_only;
470 if (object->shadow_count > 1)
473 * scan the objects entire memory queue
475 rcount = object->resident_page_count;
476 p = TAILQ_FIRST(&object->memq);
477 while (p && (rcount-- > 0)) {
479 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
481 next = TAILQ_NEXT(p, listq);
482 mycpu->gd_cnt.v_pdpages++;
483 if (p->wire_count != 0 ||
484 p->hold_count != 0 ||
486 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
487 !pmap_page_exists_quick(vm_map_pmap(map), p)) {
492 actcount = pmap_ts_referenced(p);
494 vm_page_flag_set(p, PG_REFERENCED);
495 } else if (p->flags & PG_REFERENCED) {
499 if ((p->queue != PQ_ACTIVE) &&
500 (p->flags & PG_REFERENCED)) {
502 p->act_count += actcount;
503 vm_page_flag_clear(p, PG_REFERENCED);
504 } else if (p->queue == PQ_ACTIVE) {
505 if ((p->flags & PG_REFERENCED) == 0) {
506 p->act_count -= min(p->act_count, ACT_DECLINE);
507 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
508 vm_page_protect(p, VM_PROT_NONE);
509 vm_page_deactivate(p);
512 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
513 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
518 vm_page_flag_clear(p, PG_REFERENCED);
519 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
520 p->act_count += ACT_ADVANCE;
522 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
523 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
526 } else if (p->queue == PQ_INACTIVE) {
527 vm_page_protect(p, VM_PROT_NONE);
531 object = object->backing_object;
537 * deactivate some number of pages in a map, try to do it fairly, but
538 * that is really hard to do.
541 vm_pageout_map_deactivate_pages(map, desired)
546 vm_object_t obj, bigobj;
549 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curthread)) {
557 * first, search out the biggest object, and try to free pages from
560 tmpe = map->header.next;
561 while (tmpe != &map->header) {
562 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
563 obj = tmpe->object.vm_object;
564 if ((obj != NULL) && (obj->shadow_count <= 1) &&
566 (bigobj->resident_page_count < obj->resident_page_count))) {
570 if (tmpe->wired_count > 0)
571 nothingwired = FALSE;
576 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
579 * Next, hunt around for other pages to deactivate. We actually
580 * do this search sort of wrong -- .text first is not the best idea.
582 tmpe = map->header.next;
583 while (tmpe != &map->header) {
584 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
586 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
587 obj = tmpe->object.vm_object;
589 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
595 * Remove all mappings if a process is swapped out, this will free page
598 if (desired == 0 && nothingwired)
599 pmap_remove(vm_map_pmap(map),
600 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
607 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore
608 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects
609 * which we know can be trivially freed.
613 vm_pageout_page_free(vm_page_t m) {
614 vm_object_t object = m->object;
615 int type = object->type;
617 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
618 vm_object_reference(object);
620 vm_page_protect(m, VM_PROT_NONE);
622 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
623 vm_object_deallocate(object);
627 * vm_pageout_scan does the dirty work for the pageout daemon.
630 vm_pageout_scan(int pass)
633 struct vm_page marker;
634 int page_shortage, maxscan, pcount;
635 int addl_page_shortage, addl_page_shortage_init;
636 struct proc *p, *bigproc;
637 vm_offset_t size, bigsize;
640 int vnodes_skipped = 0;
645 * Do whatever cleanup that the pmap code can.
649 addl_page_shortage_init = vm_pageout_deficit;
650 vm_pageout_deficit = 0;
653 * Calculate the number of pages we want to either free or move
656 page_shortage = vm_paging_target() + addl_page_shortage_init;
659 * Initialize our marker
661 bzero(&marker, sizeof(marker));
662 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
663 marker.queue = PQ_INACTIVE;
664 marker.wire_count = 1;
667 * Start scanning the inactive queue for pages we can move to the
668 * cache or free. The scan will stop when the target is reached or
669 * we have scanned the entire inactive queue. Note that m->act_count
670 * is not used to form decisions for the inactive queue, only for the
673 * maxlaunder limits the number of dirty pages we flush per scan.
674 * For most systems a smaller value (16 or 32) is more robust under
675 * extreme memory and disk pressure because any unnecessary writes
676 * to disk can result in extreme performance degredation. However,
677 * systems with excessive dirty pages (especially when MAP_NOSYNC is
678 * used) will die horribly with limited laundering. If the pageout
679 * daemon cannot clean enough pages in the first pass, we let it go
680 * all out in succeeding passes.
682 if ((maxlaunder = vm_max_launder) <= 1)
688 addl_page_shortage = addl_page_shortage_init;
689 maxscan = vmstats.v_inactive_count;
690 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
691 m != NULL && maxscan-- > 0 && page_shortage > 0;
694 mycpu->gd_cnt.v_pdpages++;
696 if (m->queue != PQ_INACTIVE) {
700 next = TAILQ_NEXT(m, pageq);
705 if (m->flags & PG_MARKER)
709 * A held page may be undergoing I/O, so skip it.
713 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
714 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
716 addl_page_shortage++;
720 * Dont mess with busy pages, keep in the front of the
721 * queue, most likely are being paged out.
723 if (m->busy || (m->flags & PG_BUSY)) {
724 addl_page_shortage++;
729 * If the object is not being used, we ignore previous
732 if (m->object->ref_count == 0) {
733 vm_page_flag_clear(m, PG_REFERENCED);
734 pmap_clear_reference(m);
737 * Otherwise, if the page has been referenced while in the
738 * inactive queue, we bump the "activation count" upwards,
739 * making it less likely that the page will be added back to
740 * the inactive queue prematurely again. Here we check the
741 * page tables (or emulated bits, if any), given the upper
742 * level VM system not knowing anything about existing
745 } else if (((m->flags & PG_REFERENCED) == 0) &&
746 (actcount = pmap_ts_referenced(m))) {
748 m->act_count += (actcount + ACT_ADVANCE);
753 * If the upper level VM system knows about any page
754 * references, we activate the page. We also set the
755 * "activation count" higher than normal so that we will less
756 * likely place pages back onto the inactive queue again.
758 if ((m->flags & PG_REFERENCED) != 0) {
759 vm_page_flag_clear(m, PG_REFERENCED);
760 actcount = pmap_ts_referenced(m);
762 m->act_count += (actcount + ACT_ADVANCE + 1);
767 * If the upper level VM system doesn't know anything about
768 * the page being dirty, we have to check for it again. As
769 * far as the VM code knows, any partially dirty pages are
773 vm_page_test_dirty(m);
779 * Invalid pages can be easily freed
782 vm_pageout_page_free(m);
783 mycpu->gd_cnt.v_dfree++;
787 * Clean pages can be placed onto the cache queue. This
788 * effectively frees them.
790 } else if (m->dirty == 0) {
792 * Clean pages can be immediately freed to the cache.
796 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
798 * Dirty pages need to be paged out, but flushing
799 * a page is extremely expensive verses freeing
800 * a clean page. Rather then artificially limiting
801 * the number of pages we can flush, we instead give
802 * dirty pages extra priority on the inactive queue
803 * by forcing them to be cycled through the queue
804 * twice before being flushed, after which the
805 * (now clean) page will cycle through once more
806 * before being freed. This significantly extends
807 * the thrash point for a heavily loaded machine.
810 vm_page_flag_set(m, PG_WINATCFLS);
811 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
812 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
814 } else if (maxlaunder > 0) {
816 * We always want to try to flush some dirty pages if
817 * we encounter them, to keep the system stable.
818 * Normally this number is small, but under extreme
819 * pressure where there are insufficient clean pages
820 * on the inactive queue, we may have to go all out.
822 int swap_pageouts_ok;
823 struct vnode *vp = NULL;
827 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
828 swap_pageouts_ok = 1;
830 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
831 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
832 vm_page_count_min());
837 * We don't bother paging objects that are "dead".
838 * Those objects are in a "rundown" state.
840 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
842 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
843 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
849 * The object is already known NOT to be dead. It
850 * is possible for the vget() to block the whole
851 * pageout daemon, but the new low-memory handling
852 * code should prevent it.
854 * The previous code skipped locked vnodes and, worse,
855 * reordered pages in the queue. This results in
856 * completely non-deterministic operation because,
857 * quite often, a vm_fault has initiated an I/O and
858 * is holding a locked vnode at just the point where
859 * the pageout daemon is woken up.
861 * We can't wait forever for the vnode lock, we might
862 * deadlock due to a vn_read() getting stuck in
863 * vm_wait while holding this vnode. We skip the
864 * vnode if we can't get it in a reasonable amount
868 if (object->type == OBJT_VNODE) {
871 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK, curthread)) {
873 if (object->flags & OBJ_MIGHTBEDIRTY)
879 * The page might have been moved to another
880 * queue during potential blocking in vget()
881 * above. The page might have been freed and
882 * reused for another vnode. The object might
883 * have been reused for another vnode.
885 if (m->queue != PQ_INACTIVE ||
886 m->object != object ||
887 object->handle != vp) {
888 if (object->flags & OBJ_MIGHTBEDIRTY)
895 * The page may have been busied during the
896 * blocking in vput(); We don't move the
897 * page back onto the end of the queue so that
898 * statistics are more correct if we don't.
900 if (m->busy || (m->flags & PG_BUSY)) {
906 * If the page has become held it might
907 * be undergoing I/O, so skip it
911 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
912 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
914 if (object->flags & OBJ_MIGHTBEDIRTY)
922 * If a page is dirty, then it is either being washed
923 * (but not yet cleaned) or it is still in the
924 * laundry. If it is still in the laundry, then we
925 * start the cleaning operation.
927 * This operation may cluster, invalidating the 'next'
928 * pointer. To prevent an inordinate number of
929 * restarts we use our marker to remember our place.
931 * decrement page_shortage on success to account for
932 * the (future) cleaned page. Otherwise we could wind
933 * up laundering or cleaning too many pages.
936 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
938 if (vm_pageout_clean(m) != 0) {
943 next = TAILQ_NEXT(&marker, pageq);
944 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
952 * Compute the number of pages we want to try to move from the
953 * active queue to the inactive queue.
955 page_shortage = vm_paging_target() +
956 vmstats.v_inactive_target - vmstats.v_inactive_count;
957 page_shortage += addl_page_shortage;
960 * Scan the active queue for things we can deactivate. We nominally
961 * track the per-page activity counter and use it to locate
962 * deactivation candidates.
965 pcount = vmstats.v_active_count;
966 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
968 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
971 * This is a consistency check, and should likely be a panic
974 if (m->queue != PQ_ACTIVE) {
978 next = TAILQ_NEXT(m, pageq);
980 * Don't deactivate pages that are busy.
982 if ((m->busy != 0) ||
983 (m->flags & PG_BUSY) ||
984 (m->hold_count != 0)) {
986 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
987 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
994 * The count for pagedaemon pages is done after checking the
995 * page for eligibility...
997 mycpu->gd_cnt.v_pdpages++;
1000 * Check to see "how much" the page has been used.
1003 if (m->object->ref_count != 0) {
1004 if (m->flags & PG_REFERENCED) {
1007 actcount += pmap_ts_referenced(m);
1009 m->act_count += ACT_ADVANCE + actcount;
1010 if (m->act_count > ACT_MAX)
1011 m->act_count = ACT_MAX;
1016 * Since we have "tested" this bit, we need to clear it now.
1018 vm_page_flag_clear(m, PG_REFERENCED);
1021 * Only if an object is currently being used, do we use the
1022 * page activation count stats.
1024 if (actcount && (m->object->ref_count != 0)) {
1026 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1027 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1030 m->act_count -= min(m->act_count, ACT_DECLINE);
1031 if (vm_pageout_algorithm ||
1032 m->object->ref_count == 0 ||
1033 m->act_count == 0) {
1035 if (m->object->ref_count == 0) {
1036 vm_page_protect(m, VM_PROT_NONE);
1040 vm_page_deactivate(m);
1042 vm_page_deactivate(m);
1046 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1047 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1057 * We try to maintain some *really* free pages, this allows interrupt
1058 * code to be guaranteed space. Since both cache and free queues
1059 * are considered basically 'free', moving pages from cache to free
1060 * does not effect other calculations.
1063 while (vmstats.v_free_count < vmstats.v_free_reserved) {
1064 static int cache_rover = 0;
1065 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1068 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1073 printf("Warning: busy page %p found in cache\n", m);
1075 vm_page_deactivate(m);
1078 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1079 vm_pageout_page_free(m);
1080 mycpu->gd_cnt.v_dfree++;
1084 #if !defined(NO_SWAPPING)
1086 * Idle process swapout -- run once per second.
1088 if (vm_swap_idle_enabled) {
1090 if (time_second != lsec) {
1091 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1099 * If we didn't get enough free pages, and we have skipped a vnode
1100 * in a writeable object, wakeup the sync daemon. And kick swapout
1101 * if we did not get enough free pages.
1103 if (vm_paging_target() > 0) {
1104 if (vnodes_skipped && vm_page_count_min())
1105 (void) speedup_syncer();
1106 #if !defined(NO_SWAPPING)
1107 if (vm_swap_enabled && vm_page_count_target()) {
1109 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1115 * If we are out of swap and were not able to reach our paging
1116 * target, kill the largest process.
1118 if ((vm_swap_size < 64 && vm_page_count_min()) ||
1119 (swap_pager_full && vm_paging_target() > 0)) {
1121 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) {
1125 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1127 * if this is a system process, skip it
1129 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1130 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1134 * if the process is in a non-running type state,
1137 if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1141 * get the process size
1143 size = vmspace_resident_count(p->p_vmspace) +
1144 vmspace_swap_count(p->p_vmspace);
1146 * if the this process is bigger than the biggest one
1149 if (size > bigsize) {
1154 if (bigproc != NULL) {
1155 killproc(bigproc, "out of swap space");
1156 bigproc->p_estcpu = 0;
1157 bigproc->p_nice = PRIO_MIN;
1158 resetpriority(bigproc);
1159 wakeup(&vmstats.v_free_count);
1165 * This routine tries to maintain the pseudo LRU active queue,
1166 * so that during long periods of time where there is no paging,
1167 * that some statistic accumulation still occurs. This code
1168 * helps the situation where paging just starts to occur.
1171 vm_pageout_page_stats()
1175 int pcount,tpcount; /* Number of pages to check */
1176 static int fullintervalcount = 0;
1181 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1182 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1184 if (page_shortage <= 0)
1189 pcount = vmstats.v_active_count;
1190 fullintervalcount += vm_pageout_stats_interval;
1191 if (fullintervalcount < vm_pageout_full_stats_interval) {
1192 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1193 if (pcount > tpcount)
1196 fullintervalcount = 0;
1199 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1200 while ((m != NULL) && (pcount-- > 0)) {
1203 if (m->queue != PQ_ACTIVE) {
1207 next = TAILQ_NEXT(m, pageq);
1209 * Don't deactivate pages that are busy.
1211 if ((m->busy != 0) ||
1212 (m->flags & PG_BUSY) ||
1213 (m->hold_count != 0)) {
1215 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1216 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1223 if (m->flags & PG_REFERENCED) {
1224 vm_page_flag_clear(m, PG_REFERENCED);
1228 actcount += pmap_ts_referenced(m);
1230 m->act_count += ACT_ADVANCE + actcount;
1231 if (m->act_count > ACT_MAX)
1232 m->act_count = ACT_MAX;
1234 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1235 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1238 if (m->act_count == 0) {
1240 * We turn off page access, so that we have
1241 * more accurate RSS stats. We don't do this
1242 * in the normal page deactivation when the
1243 * system is loaded VM wise, because the
1244 * cost of the large number of page protect
1245 * operations would be higher than the value
1246 * of doing the operation.
1248 vm_page_protect(m, VM_PROT_NONE);
1249 vm_page_deactivate(m);
1251 m->act_count -= min(m->act_count, ACT_DECLINE);
1253 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1254 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1265 vm_pageout_free_page_calc(count)
1268 if (count < vmstats.v_page_count)
1271 * free_reserved needs to include enough for the largest swap pager
1272 * structures plus enough for any pv_entry structs when paging.
1274 if (vmstats.v_page_count > 1024)
1275 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1277 vmstats.v_free_min = 4;
1278 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1279 vmstats.v_interrupt_free_min;
1280 vmstats.v_free_reserved = vm_pageout_page_count +
1281 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1282 vmstats.v_free_severe = vmstats.v_free_min / 2;
1283 vmstats.v_free_min += vmstats.v_free_reserved;
1284 vmstats.v_free_severe += vmstats.v_free_reserved;
1290 * vm_pageout is the high level pageout daemon.
1298 * Initialize some paging parameters.
1301 vmstats.v_interrupt_free_min = 2;
1302 if (vmstats.v_page_count < 2000)
1303 vm_pageout_page_count = 8;
1305 vm_pageout_free_page_calc(vmstats.v_page_count);
1307 * v_free_target and v_cache_min control pageout hysteresis. Note
1308 * that these are more a measure of the VM cache queue hysteresis
1309 * then the VM free queue. Specifically, v_free_target is the
1310 * high water mark (free+cache pages).
1312 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1313 * low water mark, while v_free_min is the stop. v_cache_min must
1314 * be big enough to handle memory needs while the pageout daemon
1315 * is signalled and run to free more pages.
1317 if (vmstats.v_free_count > 6144)
1318 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1320 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1322 if (vmstats.v_free_count > 2048) {
1323 vmstats.v_cache_min = vmstats.v_free_target;
1324 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1325 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2;
1327 vmstats.v_cache_min = 0;
1328 vmstats.v_cache_max = 0;
1329 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1331 if (vmstats.v_inactive_target > vmstats.v_free_count / 3)
1332 vmstats.v_inactive_target = vmstats.v_free_count / 3;
1334 /* XXX does not really belong here */
1335 if (vm_page_max_wired == 0)
1336 vm_page_max_wired = vmstats.v_free_count / 3;
1338 if (vm_pageout_stats_max == 0)
1339 vm_pageout_stats_max = vmstats.v_free_target;
1342 * Set interval in seconds for stats scan.
1344 if (vm_pageout_stats_interval == 0)
1345 vm_pageout_stats_interval = 5;
1346 if (vm_pageout_full_stats_interval == 0)
1347 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1351 * Set maximum free per pass
1353 if (vm_pageout_stats_free_max == 0)
1354 vm_pageout_stats_free_max = 5;
1356 swap_pager_swap_init();
1359 * The pageout daemon is never done, so loop forever.
1366 * If we have enough free memory, wakeup waiters. Do
1367 * not clear vm_pages_needed until we reach our target,
1368 * otherwise we may be woken up over and over again and
1369 * waste a lot of cpu.
1371 if (vm_pages_needed && !vm_page_count_min()) {
1372 if (vm_paging_needed() <= 0)
1373 vm_pages_needed = 0;
1374 wakeup(&vmstats.v_free_count);
1376 if (vm_pages_needed) {
1378 * Still not done, take a second pass without waiting
1379 * (unlimited dirty cleaning), otherwise sleep a bit
1384 tsleep(&vm_pages_needed, 0, "psleep", hz/2);
1387 * Good enough, sleep & handle stats. Prime the pass
1394 error = tsleep(&vm_pages_needed,
1395 0, "psleep", vm_pageout_stats_interval * hz);
1396 if (error && !vm_pages_needed) {
1399 vm_pageout_page_stats();
1404 if (vm_pages_needed)
1405 mycpu->gd_cnt.v_pdwakeups++;
1407 vm_pageout_scan(pass);
1408 vm_pageout_deficit = 0;
1415 if (!vm_pages_needed && curthread != pagethread) {
1417 wakeup(&vm_pages_needed);
1421 #if !defined(NO_SWAPPING)
1425 static int lastrun = 0;
1427 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1428 wakeup(&vm_daemon_needed);
1439 tsleep(&vm_daemon_needed, 0, "psleep", 0);
1440 if (vm_pageout_req_swapout) {
1441 swapout_procs(vm_pageout_req_swapout);
1442 vm_pageout_req_swapout = 0;
1445 * scan the processes for exceeding their rlimits or if
1446 * process is swapped out -- deactivate pages
1449 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1450 vm_pindex_t limit, size;
1453 * if this is a system process or if we have already
1454 * looked at this process, skip it.
1456 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1460 * if the process is in a non-running type state,
1463 if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1470 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1471 p->p_rlimit[RLIMIT_RSS].rlim_max));
1474 * let processes that are swapped out really be
1475 * swapped out set the limit to nothing (will force a
1478 if ((p->p_flag & P_INMEM) == 0)
1479 limit = 0; /* XXX */
1481 size = vmspace_resident_count(p->p_vmspace);
1482 if (limit >= 0 && size >= limit) {
1483 vm_pageout_map_deactivate_pages(
1484 &p->p_vmspace->vm_map, limit);