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_page(vm_page_t m, int *max_launderp,
104 int *vnodes_skippedp, struct vnode **vpfailedp,
105 int pass, int vmflush_flags);
106 static int vm_pageout_clean_helper (vm_page_t, int);
107 static int vm_pageout_free_page_calc (vm_size_t count);
108 static void vm_pageout_page_free(vm_page_t m) ;
109 struct thread *pagethread;
111 #if !defined(NO_SWAPPING)
112 /* the kernel process "vm_daemon"*/
113 static void vm_daemon (void);
114 static struct thread *vmthread;
116 static struct kproc_desc vm_kp = {
121 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
124 int vm_pages_needed = 0; /* Event on which pageout daemon sleeps */
125 int vm_pageout_deficit = 0; /* Estimated number of pages deficit */
126 int vm_pageout_pages_needed = 0;/* pageout daemon needs pages */
127 int vm_page_free_hysteresis = 16;
129 #if !defined(NO_SWAPPING)
130 static int vm_pageout_req_swapout;
131 static int vm_daemon_needed;
133 static int vm_max_launder = 4096;
134 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
135 static int vm_pageout_full_stats_interval = 0;
136 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
137 static int defer_swap_pageouts=0;
138 static int disable_swap_pageouts=0;
139 static u_int vm_anonmem_decline = ACT_DECLINE;
140 static u_int vm_filemem_decline = ACT_DECLINE * 2;
142 #if defined(NO_SWAPPING)
143 static int vm_swap_enabled=0;
144 static int vm_swap_idle_enabled=0;
146 static int vm_swap_enabled=1;
147 static int vm_swap_idle_enabled=0;
149 int vm_pageout_memuse_mode=1; /* 0-disable, 1-passive, 2-active swp*/
151 SYSCTL_UINT(_vm, VM_PAGEOUT_ALGORITHM, anonmem_decline,
152 CTLFLAG_RW, &vm_anonmem_decline, 0, "active->inactive anon memory");
154 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, filemem_decline,
155 CTLFLAG_RW, &vm_filemem_decline, 0, "active->inactive file cache");
157 SYSCTL_INT(_vm, OID_AUTO, page_free_hysteresis,
158 CTLFLAG_RW, &vm_page_free_hysteresis, 0,
159 "Free more pages than the minimum required");
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");
175 SYSCTL_INT(_vm, OID_AUTO, pageout_memuse_mode,
176 CTLFLAG_RW, &vm_pageout_memuse_mode, 0, "memoryuse resource mode");
178 #if defined(NO_SWAPPING)
179 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
180 CTLFLAG_RD, &vm_swap_enabled, 0, "");
181 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
182 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
190 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
191 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
193 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
194 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
196 static int pageout_lock_miss;
197 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
198 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
200 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
202 #if !defined(NO_SWAPPING)
203 static void vm_req_vmdaemon (void);
205 static void vm_pageout_page_stats(int q);
208 * Calculate approximately how many pages on each queue to try to
209 * clean. An exact calculation creates an edge condition when the
210 * queues are unbalanced so add significant slop. The queue scans
211 * will stop early when targets are reached and will start where they
212 * left off on the next pass.
214 * We need to be generous here because there are all sorts of loading
215 * conditions that can cause edge cases if try to average over all queues.
216 * In particular, storage subsystems have become so fast that paging
217 * activity can become quite frantic. Eventually we will probably need
218 * two paging threads, one for dirty pages and one for clean, to deal
219 * with the bandwidth requirements.
221 * So what we do is calculate a value that can be satisfied nominally by
222 * only having to scan half the queues.
230 avg = ((n + (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) + 1);
232 avg = ((n - (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) - 1);
238 * vm_pageout_clean_helper:
240 * Clean the page and remove it from the laundry. The page must not be
243 * We set the busy bit to cause potential page faults on this page to
244 * block. Note the careful timing, however, the busy bit isn't set till
245 * late and we cannot do anything that will mess with the page.
248 vm_pageout_clean_helper(vm_page_t m, int vmflush_flags)
251 vm_page_t mc[BLIST_MAX_ALLOC];
253 int ib, is, page_base;
254 vm_pindex_t pindex = m->pindex;
259 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
260 * with the new swapper, but we could have serious problems paging
261 * out other object types if there is insufficient memory.
263 * Unfortunately, checking free memory here is far too late, so the
264 * check has been moved up a procedural level.
268 * Don't mess with the page if it's busy, held, or special
270 * XXX do we really need to check hold_count here? hold_count
271 * isn't supposed to mess with vm_page ops except prevent the
272 * page from being reused.
274 if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
280 * Place page in cluster. Align cluster for optimal swap space
281 * allocation (whether it is swap or not). This is typically ~16-32
282 * pages, which also tends to align the cluster to multiples of the
283 * filesystem block size if backed by a filesystem.
285 page_base = pindex % BLIST_MAX_ALLOC;
291 * Scan object for clusterable pages.
293 * We can cluster ONLY if: ->> the page is NOT
294 * clean, wired, busy, held, or mapped into a
295 * buffer, and one of the following:
296 * 1) The page is inactive, or a seldom used
299 * 2) we force the issue.
301 * During heavy mmap/modification loads the pageout
302 * daemon can really fragment the underlying file
303 * due to flushing pages out of order and not trying
304 * align the clusters (which leave sporatic out-of-order
305 * holes). To solve this problem we do the reverse scan
306 * first and attempt to align our cluster, then do a
307 * forward scan if room remains.
309 vm_object_hold(object);
314 p = vm_page_lookup_busy_try(object, pindex - page_base + ib,
316 if (error || p == NULL)
318 if ((p->queue - p->pc) == PQ_CACHE ||
319 (p->flags & PG_UNMANAGED)) {
323 vm_page_test_dirty(p);
324 if (((p->dirty & p->valid) == 0 &&
325 (p->flags & PG_NEED_COMMIT) == 0) ||
326 p->wire_count != 0 || /* may be held by buf cache */
327 p->hold_count != 0) { /* may be undergoing I/O */
331 if (p->queue - p->pc != PQ_INACTIVE) {
332 if (p->queue - p->pc != PQ_ACTIVE ||
333 (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
340 * Try to maintain page groupings in the cluster.
342 if (m->flags & PG_WINATCFLS)
343 vm_page_flag_set(p, PG_WINATCFLS);
345 vm_page_flag_clear(p, PG_WINATCFLS);
346 p->act_count = m->act_count;
353 while (is < BLIST_MAX_ALLOC &&
354 pindex - page_base + is < object->size) {
357 p = vm_page_lookup_busy_try(object, pindex - page_base + is,
359 if (error || p == NULL)
361 if (((p->queue - p->pc) == PQ_CACHE) ||
362 (p->flags & PG_UNMANAGED)) {
366 vm_page_test_dirty(p);
367 if (((p->dirty & p->valid) == 0 &&
368 (p->flags & PG_NEED_COMMIT) == 0) ||
369 p->wire_count != 0 || /* may be held by buf cache */
370 p->hold_count != 0) { /* may be undergoing I/O */
374 if (p->queue - p->pc != PQ_INACTIVE) {
375 if (p->queue - p->pc != PQ_ACTIVE ||
376 (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
383 * Try to maintain page groupings in the cluster.
385 if (m->flags & PG_WINATCFLS)
386 vm_page_flag_set(p, PG_WINATCFLS);
388 vm_page_flag_clear(p, PG_WINATCFLS);
389 p->act_count = m->act_count;
395 vm_object_drop(object);
398 * we allow reads during pageouts...
400 return vm_pageout_flush(&mc[ib], is - ib, vmflush_flags);
404 * vm_pageout_flush() - launder the given pages
406 * The given pages are laundered. Note that we setup for the start of
407 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
408 * reference count all in here rather then in the parent. If we want
409 * the parent to do more sophisticated things we may have to change
412 * The pages in the array must be busied by the caller and will be
413 * unbusied by this function.
416 vm_pageout_flush(vm_page_t *mc, int count, int vmflush_flags)
419 int pageout_status[count];
424 * Initiate I/O. Bump the vm_page_t->busy counter.
426 for (i = 0; i < count; i++) {
427 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
428 ("vm_pageout_flush page %p index %d/%d: partially "
429 "invalid page", mc[i], i, count));
430 vm_page_io_start(mc[i]);
434 * We must make the pages read-only. This will also force the
435 * modified bit in the related pmaps to be cleared. The pager
436 * cannot clear the bit for us since the I/O completion code
437 * typically runs from an interrupt. The act of making the page
438 * read-only handles the case for us.
440 * Then we can unbusy the pages, we still hold a reference by virtue
443 for (i = 0; i < count; i++) {
444 if (vmflush_flags & VM_PAGER_TRY_TO_CACHE)
445 vm_page_protect(mc[i], VM_PROT_NONE);
447 vm_page_protect(mc[i], VM_PROT_READ);
448 vm_page_wakeup(mc[i]);
451 object = mc[0]->object;
452 vm_object_pip_add(object, count);
454 vm_pager_put_pages(object, mc, count,
456 ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
459 for (i = 0; i < count; i++) {
460 vm_page_t mt = mc[i];
462 switch (pageout_status[i]) {
471 * Page outside of range of object. Right now we
472 * essentially lose the changes by pretending it
475 vm_page_busy_wait(mt, FALSE, "pgbad");
476 pmap_clear_modify(mt);
483 * A page typically cannot be paged out when we
484 * have run out of swap. We leave the page
485 * marked inactive and will try to page it out
488 * Starvation of the active page list is used to
489 * determine when the system is massively memory
498 * If not PENDing this was a synchronous operation and we
499 * clean up after the I/O. If it is PENDing the mess is
500 * cleaned up asynchronously.
502 * Also nominally act on the caller's wishes if the caller
503 * wants to try to really clean (cache or free) the page.
505 * Also nominally deactivate the page if the system is
508 if (pageout_status[i] != VM_PAGER_PEND) {
509 vm_page_busy_wait(mt, FALSE, "pgouw");
510 vm_page_io_finish(mt);
511 if (vmflush_flags & VM_PAGER_TRY_TO_CACHE) {
512 vm_page_try_to_cache(mt);
513 } else if (vm_page_count_severe()) {
514 vm_page_deactivate(mt);
519 vm_object_pip_wakeup(object);
525 #if !defined(NO_SWAPPING)
528 * Callback function, page busied for us. We must dispose of the busy
529 * condition. Any related pmap pages may be held but will not be locked.
533 vm_pageout_mdp_callback(struct pmap_pgscan_info *info, vm_offset_t va,
540 * Basic tests - There should never be a marker, and we can stop
541 * once the RSS is below the required level.
543 KKASSERT((p->flags & PG_MARKER) == 0);
544 if (pmap_resident_tlnw_count(info->pmap) <= info->limit) {
549 mycpu->gd_cnt.v_pdpages++;
551 if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
559 * Check if the page has been referened recently. If it has,
560 * activate it and skip.
562 actcount = pmap_ts_referenced(p);
564 vm_page_flag_set(p, PG_REFERENCED);
565 } else if (p->flags & PG_REFERENCED) {
570 if (p->queue - p->pc != PQ_ACTIVE) {
571 vm_page_and_queue_spin_lock(p);
572 if (p->queue - p->pc != PQ_ACTIVE) {
573 vm_page_and_queue_spin_unlock(p);
576 vm_page_and_queue_spin_unlock(p);
579 p->act_count += actcount;
580 if (p->act_count > ACT_MAX)
581 p->act_count = ACT_MAX;
583 vm_page_flag_clear(p, PG_REFERENCED);
589 * Remove the page from this particular pmap. Once we do this, our
590 * pmap scans will not see it again (unless it gets faulted in), so
591 * we must actively dispose of or deal with the page.
593 pmap_remove_specific(info->pmap, p);
596 * If the page is not mapped to another process (i.e. as would be
597 * typical if this were a shared page from a library) then deactivate
598 * the page and clean it in two passes only.
600 * If the page hasn't been referenced since the last check, remove it
601 * from the pmap. If it is no longer mapped, deactivate it
602 * immediately, accelerating the normal decline.
604 * Once the page has been removed from the pmap the RSS code no
605 * longer tracks it so we have to make sure that it is staged for
606 * potential flush action.
608 if ((p->flags & PG_MAPPED) == 0) {
609 if (p->queue - p->pc == PQ_ACTIVE) {
610 vm_page_deactivate(p);
612 if (p->queue - p->pc == PQ_INACTIVE) {
618 * Ok, try to fully clean the page and any nearby pages such that at
619 * least the requested page is freed or moved to the cache queue.
621 * We usually do this synchronously to allow us to get the page into
622 * the CACHE queue quickly, which will prevent memory exhaustion if
623 * a process with a memoryuse limit is running away. However, the
624 * sysadmin may desire to set vm.swap_user_async which relaxes this
625 * and improves write performance.
628 int max_launder = 0x7FFF;
629 int vnodes_skipped = 0;
631 struct vnode *vpfailed = NULL;
635 if (vm_pageout_memuse_mode >= 2) {
636 vmflush_flags = VM_PAGER_TRY_TO_CACHE |
637 VM_PAGER_ALLOW_ACTIVE;
638 if (swap_user_async == 0)
639 vmflush_flags |= VM_PAGER_PUT_SYNC;
640 vm_page_flag_set(p, PG_WINATCFLS);
642 vm_pageout_page(p, &max_launder,
644 &vpfailed, 1, vmflush_flags);
654 * Must be at end to avoid SMP races.
662 * Deactivate some number of pages in a map due to set RLIMIT_RSS limits.
663 * that is relatively difficult to do. We try to keep track of where we
664 * left off last time to reduce scan overhead.
666 * Called when vm_pageout_memuse_mode is >= 1.
669 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t limit)
671 vm_offset_t pgout_offset;
672 struct pmap_pgscan_info info;
675 pgout_offset = map->pgout_offset;
678 kprintf("%016jx ", pgout_offset);
680 if (pgout_offset < VM_MIN_USER_ADDRESS)
681 pgout_offset = VM_MIN_USER_ADDRESS;
682 if (pgout_offset >= VM_MAX_USER_ADDRESS)
684 info.pmap = vm_map_pmap(map);
686 info.beg_addr = pgout_offset;
687 info.end_addr = VM_MAX_USER_ADDRESS;
688 info.callback = vm_pageout_mdp_callback;
690 info.actioncount = 0;
694 pgout_offset = info.offset;
696 kprintf("%016jx %08lx %08lx\n", pgout_offset,
697 info.cleancount, info.actioncount);
700 if (pgout_offset != VM_MAX_USER_ADDRESS &&
701 pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
703 } else if (retries &&
704 pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
708 map->pgout_offset = pgout_offset;
713 * Called when the pageout scan wants to free a page. We no longer
714 * try to cycle the vm_object here with a reference & dealloc, which can
715 * cause a non-trivial object collapse in a critical path.
717 * It is unclear why we cycled the ref_count in the past, perhaps to try
718 * to optimize shadow chain collapses but I don't quite see why it would
719 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
720 * synchronously and not have to be kicked-start.
723 vm_pageout_page_free(vm_page_t m)
725 vm_page_protect(m, VM_PROT_NONE);
730 * vm_pageout_scan does the dirty work for the pageout daemon.
732 struct vm_pageout_scan_info {
733 struct proc *bigproc;
737 static int vm_pageout_scan_callback(struct proc *p, void *data);
740 vm_pageout_scan_inactive(int pass, int q, int avail_shortage,
744 struct vm_page marker;
745 struct vnode *vpfailed; /* warning, allowed to be stale */
751 * Start scanning the inactive queue for pages we can move to the
752 * cache or free. The scan will stop when the target is reached or
753 * we have scanned the entire inactive queue. Note that m->act_count
754 * is not used to form decisions for the inactive queue, only for the
757 * max_launder limits the number of dirty pages we flush per scan.
758 * For most systems a smaller value (16 or 32) is more robust under
759 * extreme memory and disk pressure because any unnecessary writes
760 * to disk can result in extreme performance degredation. However,
761 * systems with excessive dirty pages (especially when MAP_NOSYNC is
762 * used) will die horribly with limited laundering. If the pageout
763 * daemon cannot clean enough pages in the first pass, we let it go
764 * all out in succeeding passes.
766 if ((max_launder = vm_max_launder) <= 1)
772 * Initialize our marker
774 bzero(&marker, sizeof(marker));
775 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
776 marker.queue = PQ_INACTIVE + q;
778 marker.wire_count = 1;
781 * Inactive queue scan.
783 * NOTE: The vm_page must be spinlocked before the queue to avoid
784 * deadlocks, so it is easiest to simply iterate the loop
785 * with the queue unlocked at the top.
789 vm_page_queues_spin_lock(PQ_INACTIVE + q);
790 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
791 maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt;
794 * Queue locked at top of loop to avoid stack marker issues.
796 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
797 maxscan-- > 0 && avail_shortage - delta > 0)
801 KKASSERT(m->queue == PQ_INACTIVE + q);
802 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
804 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
806 mycpu->gd_cnt.v_pdpages++;
809 * Skip marker pages (atomic against other markers to avoid
810 * infinite hop-over scans).
812 if (m->flags & PG_MARKER)
816 * Try to busy the page. Don't mess with pages which are
817 * already busy or reorder them in the queue.
819 if (vm_page_busy_try(m, TRUE))
823 * Remaining operations run with the page busy and neither
824 * the page or the queue will be spin-locked.
826 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
827 KKASSERT(m->queue == PQ_INACTIVE + q);
829 count = vm_pageout_page(m, &max_launder, vnodes_skipped,
834 * Systems with a ton of memory can wind up with huge
835 * deactivation counts. Because the inactive scan is
836 * doing a lot of flushing, the combination can result
837 * in excessive paging even in situations where other
838 * unrelated threads free up sufficient VM.
840 * To deal with this we abort the nominal active->inactive
841 * scan before we hit the inactive target when free+cache
842 * levels have reached a reasonable target.
844 * When deciding to stop early we need to add some slop to
845 * the test and we need to return full completion to the caller
846 * to prevent the caller from thinking there is something
847 * wrong and issuing a low-memory+swap warning or pkill.
849 * A deficit forces paging regardless of the state of the
850 * VM page queues (used for RSS enforcement).
853 vm_page_queues_spin_lock(PQ_INACTIVE + q);
854 if (vm_paging_target() < -vm_max_launder) {
856 * Stopping early, return full completion to caller.
858 if (delta < avail_shortage)
859 delta = avail_shortage;
864 /* page queue still spin-locked */
865 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
866 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
872 * Pageout the specified page, return the total number of pages paged out
873 * (this routine may cluster).
875 * The page must be busied and soft-busied by the caller and will be disposed
876 * of by this function.
879 vm_pageout_page(vm_page_t m, int *max_launderp, int *vnodes_skippedp,
880 struct vnode **vpfailedp, int pass, int vmflush_flags)
887 * It is possible for a page to be busied ad-hoc (e.g. the
888 * pmap_collect() code) and wired and race against the
889 * allocation of a new page. vm_page_alloc() may be forced
890 * to deactivate the wired page in which case it winds up
891 * on the inactive queue and must be handled here. We
892 * correct the problem simply by unqueuing the page.
895 vm_page_unqueue_nowakeup(m);
897 kprintf("WARNING: pagedaemon: wired page on "
898 "inactive queue %p\n", m);
903 * A held page may be undergoing I/O, so skip it.
906 vm_page_and_queue_spin_lock(m);
907 if (m->queue - m->pc == PQ_INACTIVE) {
909 &vm_page_queues[m->queue].pl, m, pageq);
911 &vm_page_queues[m->queue].pl, m, pageq);
912 ++vm_swapcache_inactive_heuristic;
914 vm_page_and_queue_spin_unlock(m);
919 if (m->object == NULL || m->object->ref_count == 0) {
921 * If the object is not being used, we ignore previous
924 vm_page_flag_clear(m, PG_REFERENCED);
925 pmap_clear_reference(m);
926 /* fall through to end */
927 } else if (((m->flags & PG_REFERENCED) == 0) &&
928 (actcount = pmap_ts_referenced(m))) {
930 * Otherwise, if the page has been referenced while
931 * in the inactive queue, we bump the "activation
932 * count" upwards, making it less likely that the
933 * page will be added back to the inactive queue
934 * prematurely again. Here we check the page tables
935 * (or emulated bits, if any), given the upper level
936 * VM system not knowing anything about existing
940 m->act_count += (actcount + ACT_ADVANCE);
946 * (m) is still busied.
948 * If the upper level VM system knows about any page
949 * references, we activate the page. We also set the
950 * "activation count" higher than normal so that we will less
951 * likely place pages back onto the inactive queue again.
953 if ((m->flags & PG_REFERENCED) != 0) {
954 vm_page_flag_clear(m, PG_REFERENCED);
955 actcount = pmap_ts_referenced(m);
957 m->act_count += (actcount + ACT_ADVANCE + 1);
963 * If the upper level VM system doesn't know anything about
964 * the page being dirty, we have to check for it again. As
965 * far as the VM code knows, any partially dirty pages are
968 * Pages marked PG_WRITEABLE may be mapped into the user
969 * address space of a process running on another cpu. A
970 * user process (without holding the MP lock) running on
971 * another cpu may be able to touch the page while we are
972 * trying to remove it. vm_page_cache() will handle this
976 vm_page_test_dirty(m);
981 if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
983 * Invalid pages can be easily freed
985 vm_pageout_page_free(m);
986 mycpu->gd_cnt.v_dfree++;
988 } else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
990 * Clean pages can be placed onto the cache queue.
991 * This effectively frees them.
995 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
997 * Dirty pages need to be paged out, but flushing
998 * a page is extremely expensive verses freeing
999 * a clean page. Rather then artificially limiting
1000 * the number of pages we can flush, we instead give
1001 * dirty pages extra priority on the inactive queue
1002 * by forcing them to be cycled through the queue
1003 * twice before being flushed, after which the
1004 * (now clean) page will cycle through once more
1005 * before being freed. This significantly extends
1006 * the thrash point for a heavily loaded machine.
1008 vm_page_flag_set(m, PG_WINATCFLS);
1009 vm_page_and_queue_spin_lock(m);
1010 if (m->queue - m->pc == PQ_INACTIVE) {
1012 &vm_page_queues[m->queue].pl, m, pageq);
1014 &vm_page_queues[m->queue].pl, m, pageq);
1015 ++vm_swapcache_inactive_heuristic;
1017 vm_page_and_queue_spin_unlock(m);
1019 } else if (*max_launderp > 0) {
1021 * We always want to try to flush some dirty pages if
1022 * we encounter them, to keep the system stable.
1023 * Normally this number is small, but under extreme
1024 * pressure where there are insufficient clean pages
1025 * on the inactive queue, we may have to go all out.
1027 int swap_pageouts_ok;
1028 struct vnode *vp = NULL;
1030 swap_pageouts_ok = 0;
1033 (object->type != OBJT_SWAP) &&
1034 (object->type != OBJT_DEFAULT)) {
1035 swap_pageouts_ok = 1;
1037 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
1038 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
1039 vm_page_count_min(0));
1043 * We don't bother paging objects that are "dead".
1044 * Those objects are in a "rundown" state.
1046 if (!swap_pageouts_ok ||
1048 (object->flags & OBJ_DEAD)) {
1049 vm_page_and_queue_spin_lock(m);
1050 if (m->queue - m->pc == PQ_INACTIVE) {
1052 &vm_page_queues[m->queue].pl,
1055 &vm_page_queues[m->queue].pl,
1057 ++vm_swapcache_inactive_heuristic;
1059 vm_page_and_queue_spin_unlock(m);
1065 * (m) is still busied.
1067 * The object is already known NOT to be dead. It
1068 * is possible for the vget() to block the whole
1069 * pageout daemon, but the new low-memory handling
1070 * code should prevent it.
1072 * The previous code skipped locked vnodes and, worse,
1073 * reordered pages in the queue. This results in
1074 * completely non-deterministic operation because,
1075 * quite often, a vm_fault has initiated an I/O and
1076 * is holding a locked vnode at just the point where
1077 * the pageout daemon is woken up.
1079 * We can't wait forever for the vnode lock, we might
1080 * deadlock due to a vn_read() getting stuck in
1081 * vm_wait while holding this vnode. We skip the
1082 * vnode if we can't get it in a reasonable amount
1085 * vpfailed is used to (try to) avoid the case where
1086 * a large number of pages are associated with a
1087 * locked vnode, which could cause the pageout daemon
1088 * to stall for an excessive amount of time.
1090 if (object->type == OBJT_VNODE) {
1093 vp = object->handle;
1094 flags = LK_EXCLUSIVE;
1095 if (vp == *vpfailedp)
1098 flags |= LK_TIMELOCK;
1103 * We have unbusied (m) temporarily so we can
1104 * acquire the vp lock without deadlocking.
1105 * (m) is held to prevent destruction.
1107 if (vget(vp, flags) != 0) {
1109 ++pageout_lock_miss;
1110 if (object->flags & OBJ_MIGHTBEDIRTY)
1117 * The page might have been moved to another
1118 * queue during potential blocking in vget()
1119 * above. The page might have been freed and
1120 * reused for another vnode. The object might
1121 * have been reused for another vnode.
1123 if (m->queue - m->pc != PQ_INACTIVE ||
1124 m->object != object ||
1125 object->handle != vp) {
1126 if (object->flags & OBJ_MIGHTBEDIRTY)
1134 * The page may have been busied during the
1135 * blocking in vput(); We don't move the
1136 * page back onto the end of the queue so that
1137 * statistics are more correct if we don't.
1139 if (vm_page_busy_try(m, TRUE)) {
1147 * (m) is busied again
1149 * We own the busy bit and remove our hold
1150 * bit. If the page is still held it
1151 * might be undergoing I/O, so skip it.
1153 if (m->hold_count) {
1154 vm_page_and_queue_spin_lock(m);
1155 if (m->queue - m->pc == PQ_INACTIVE) {
1156 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq);
1157 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1158 ++vm_swapcache_inactive_heuristic;
1160 vm_page_and_queue_spin_unlock(m);
1161 if (object->flags & OBJ_MIGHTBEDIRTY)
1167 /* (m) is left busied as we fall through */
1171 * page is busy and not held here.
1173 * If a page is dirty, then it is either being washed
1174 * (but not yet cleaned) or it is still in the
1175 * laundry. If it is still in the laundry, then we
1176 * start the cleaning operation.
1178 * decrement inactive_shortage on success to account
1179 * for the (future) cleaned page. Otherwise we
1180 * could wind up laundering or cleaning too many
1183 * NOTE: Cleaning the page here does not cause
1184 * force_deficit to be adjusted, because the
1185 * page is not being freed or moved to the
1188 count = vm_pageout_clean_helper(m, vmflush_flags);
1189 *max_launderp -= count;
1192 * Clean ate busy, page no longer accessible
1203 vm_pageout_scan_active(int pass, int q,
1204 int avail_shortage, int inactive_shortage,
1205 int *recycle_countp)
1207 struct vm_page marker;
1214 * We want to move pages from the active queue to the inactive
1215 * queue to get the inactive queue to the inactive target. If
1216 * we still have a page shortage from above we try to directly free
1217 * clean pages instead of moving them.
1219 * If we do still have a shortage we keep track of the number of
1220 * pages we free or cache (recycle_count) as a measure of thrashing
1221 * between the active and inactive queues.
1223 * If we were able to completely satisfy the free+cache targets
1224 * from the inactive pool we limit the number of pages we move
1225 * from the active pool to the inactive pool to 2x the pages we
1226 * had removed from the inactive pool (with a minimum of 1/5 the
1227 * inactive target). If we were not able to completely satisfy
1228 * the free+cache targets we go for the whole target aggressively.
1230 * NOTE: Both variables can end up negative.
1231 * NOTE: We are still in a critical section.
1234 bzero(&marker, sizeof(marker));
1235 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1236 marker.queue = PQ_ACTIVE + q;
1238 marker.wire_count = 1;
1240 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1241 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1242 maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt;
1245 * Queue locked at top of loop to avoid stack marker issues.
1247 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1248 maxscan-- > 0 && (avail_shortage - delta > 0 ||
1249 inactive_shortage > 0))
1251 KKASSERT(m->queue == PQ_ACTIVE + q);
1252 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1254 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1258 * Skip marker pages (atomic against other markers to avoid
1259 * infinite hop-over scans).
1261 if (m->flags & PG_MARKER)
1265 * Try to busy the page. Don't mess with pages which are
1266 * already busy or reorder them in the queue.
1268 if (vm_page_busy_try(m, TRUE))
1272 * Remaining operations run with the page busy and neither
1273 * the page or the queue will be spin-locked.
1275 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1276 KKASSERT(m->queue == PQ_ACTIVE + q);
1279 * Don't deactivate pages that are held, even if we can
1280 * busy them. (XXX why not?)
1282 if (m->hold_count != 0) {
1283 vm_page_and_queue_spin_lock(m);
1284 if (m->queue - m->pc == PQ_ACTIVE) {
1286 &vm_page_queues[PQ_ACTIVE + q].pl,
1289 &vm_page_queues[PQ_ACTIVE + q].pl,
1292 vm_page_and_queue_spin_unlock(m);
1298 * The count for pagedaemon pages is done after checking the
1299 * page for eligibility...
1301 mycpu->gd_cnt.v_pdpages++;
1304 * Check to see "how much" the page has been used and clear
1305 * the tracking access bits. If the object has no references
1306 * don't bother paying the expense.
1309 if (m->object && m->object->ref_count != 0) {
1310 if (m->flags & PG_REFERENCED)
1312 actcount += pmap_ts_referenced(m);
1314 m->act_count += ACT_ADVANCE + actcount;
1315 if (m->act_count > ACT_MAX)
1316 m->act_count = ACT_MAX;
1319 vm_page_flag_clear(m, PG_REFERENCED);
1322 * actcount is only valid if the object ref_count is non-zero.
1323 * If the page does not have an object, actcount will be zero.
1325 if (actcount && m->object->ref_count != 0) {
1326 vm_page_and_queue_spin_lock(m);
1327 if (m->queue - m->pc == PQ_ACTIVE) {
1329 &vm_page_queues[PQ_ACTIVE + q].pl,
1332 &vm_page_queues[PQ_ACTIVE + q].pl,
1335 vm_page_and_queue_spin_unlock(m);
1338 switch(m->object->type) {
1341 m->act_count -= min(m->act_count,
1342 vm_anonmem_decline);
1345 m->act_count -= min(m->act_count,
1346 vm_filemem_decline);
1349 if (vm_pageout_algorithm ||
1350 (m->object == NULL) ||
1351 (m->object && (m->object->ref_count == 0)) ||
1352 m->act_count < pass + 1
1355 * Deactivate the page. If we had a
1356 * shortage from our inactive scan try to
1357 * free (cache) the page instead.
1359 * Don't just blindly cache the page if
1360 * we do not have a shortage from the
1361 * inactive scan, that could lead to
1362 * gigabytes being moved.
1364 --inactive_shortage;
1365 if (avail_shortage - delta > 0 ||
1366 (m->object && (m->object->ref_count == 0)))
1368 if (avail_shortage - delta > 0)
1370 vm_page_protect(m, VM_PROT_NONE);
1371 if (m->dirty == 0 &&
1372 (m->flags & PG_NEED_COMMIT) == 0 &&
1373 avail_shortage - delta > 0) {
1376 vm_page_deactivate(m);
1380 vm_page_deactivate(m);
1385 vm_page_and_queue_spin_lock(m);
1386 if (m->queue - m->pc == PQ_ACTIVE) {
1388 &vm_page_queues[PQ_ACTIVE + q].pl,
1391 &vm_page_queues[PQ_ACTIVE + q].pl,
1394 vm_page_and_queue_spin_unlock(m);
1400 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1404 * Clean out our local marker.
1406 * Page queue still spin-locked.
1408 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1409 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1415 * The number of actually free pages can drop down to v_free_reserved,
1416 * we try to build the free count back above v_free_min. Note that
1417 * vm_paging_needed() also returns TRUE if v_free_count is not at
1418 * least v_free_min so that is the minimum we must build the free
1421 * We use a slightly higher target to improve hysteresis,
1422 * ((v_free_target + v_free_min) / 2). Since v_free_target
1423 * is usually the same as v_cache_min this maintains about
1424 * half the pages in the free queue as are in the cache queue,
1425 * providing pretty good pipelining for pageout operation.
1427 * The system operator can manipulate vm.v_cache_min and
1428 * vm.v_free_target to tune the pageout demon. Be sure
1429 * to keep vm.v_free_min < vm.v_free_target.
1431 * Note that the original paging target is to get at least
1432 * (free_min + cache_min) into (free + cache). The slightly
1433 * higher target will shift additional pages from cache to free
1434 * without effecting the original paging target in order to
1435 * maintain better hysteresis and not have the free count always
1436 * be dead-on v_free_min.
1438 * NOTE: we are still in a critical section.
1440 * Pages moved from PQ_CACHE to totally free are not counted in the
1441 * pages_freed counter.
1444 vm_pageout_scan_cache(int avail_shortage, int pass,
1445 int vnodes_skipped, int recycle_count)
1447 static int lastkillticks;
1448 struct vm_pageout_scan_info info;
1451 while (vmstats.v_free_count <
1452 (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1454 * This steals some code from vm/vm_page.c
1456 static int cache_rover = 0;
1458 m = vm_page_list_find(PQ_CACHE,
1459 cache_rover & PQ_L2_MASK, FALSE);
1462 /* page is returned removed from its queue and spinlocked */
1463 if (vm_page_busy_try(m, TRUE)) {
1464 vm_page_deactivate_locked(m);
1465 vm_page_spin_unlock(m);
1468 vm_page_spin_unlock(m);
1469 pagedaemon_wakeup();
1473 * Remaining operations run with the page busy and neither
1474 * the page or the queue will be spin-locked.
1476 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1479 vm_page_deactivate(m);
1483 KKASSERT((m->flags & PG_MAPPED) == 0);
1484 KKASSERT(m->dirty == 0);
1485 cache_rover += PQ_PRIME2;
1486 vm_pageout_page_free(m);
1487 mycpu->gd_cnt.v_dfree++;
1490 #if !defined(NO_SWAPPING)
1492 * Idle process swapout -- run once per second.
1494 if (vm_swap_idle_enabled) {
1496 if (time_uptime != lsec) {
1497 atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_IDLE);
1505 * If we didn't get enough free pages, and we have skipped a vnode
1506 * in a writeable object, wakeup the sync daemon. And kick swapout
1507 * if we did not get enough free pages.
1509 if (vm_paging_target() > 0) {
1510 if (vnodes_skipped && vm_page_count_min(0))
1511 speedup_syncer(NULL);
1512 #if !defined(NO_SWAPPING)
1513 if (vm_swap_enabled && vm_page_count_target()) {
1514 atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_NORMAL);
1521 * Handle catastrophic conditions. Under good conditions we should
1522 * be at the target, well beyond our minimum. If we could not even
1523 * reach our minimum the system is under heavy stress. But just being
1524 * under heavy stress does not trigger process killing.
1526 * We consider ourselves to have run out of memory if the swap pager
1527 * is full and avail_shortage is still positive. The secondary check
1528 * ensures that we do not kill processes if the instantanious
1529 * availability is good, even if the pageout demon pass says it
1530 * couldn't get to the target.
1532 if (swap_pager_almost_full &&
1534 (vm_page_count_min(recycle_count) || avail_shortage > 0)) {
1535 kprintf("Warning: system low on memory+swap "
1536 "shortage %d for %d ticks!\n",
1537 avail_shortage, ticks - swap_fail_ticks);
1539 if (swap_pager_full &&
1541 avail_shortage > 0 &&
1542 vm_paging_target() > 0 &&
1543 (unsigned int)(ticks - lastkillticks) >= hz) {
1545 * Kill something, maximum rate once per second to give
1546 * the process time to free up sufficient memory.
1548 lastkillticks = ticks;
1549 info.bigproc = NULL;
1551 allproc_scan(vm_pageout_scan_callback, &info);
1552 if (info.bigproc != NULL) {
1553 info.bigproc->p_nice = PRIO_MIN;
1554 info.bigproc->p_usched->resetpriority(
1555 FIRST_LWP_IN_PROC(info.bigproc));
1556 atomic_set_int(&info.bigproc->p_flags, P_LOWMEMKILL);
1557 killproc(info.bigproc, "out of swap space");
1558 wakeup(&vmstats.v_free_count);
1559 PRELE(info.bigproc);
1565 vm_pageout_scan_callback(struct proc *p, void *data)
1567 struct vm_pageout_scan_info *info = data;
1571 * Never kill system processes or init. If we have configured swap
1572 * then try to avoid killing low-numbered pids.
1574 if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1575 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1579 lwkt_gettoken(&p->p_token);
1582 * if the process is in a non-running type state,
1585 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
1586 lwkt_reltoken(&p->p_token);
1591 * Get the approximate process size. Note that anonymous pages
1592 * with backing swap will be counted twice, but there should not
1593 * be too many such pages due to the stress the VM system is
1594 * under at this point.
1596 size = vmspace_anonymous_count(p->p_vmspace) +
1597 vmspace_swap_count(p->p_vmspace);
1600 * If the this process is bigger than the biggest one
1603 if (info->bigsize < size) {
1605 PRELE(info->bigproc);
1608 info->bigsize = size;
1610 lwkt_reltoken(&p->p_token);
1617 * This routine tries to maintain the pseudo LRU active queue,
1618 * so that during long periods of time where there is no paging,
1619 * that some statistic accumulation still occurs. This code
1620 * helps the situation where paging just starts to occur.
1623 vm_pageout_page_stats(int q)
1625 static int fullintervalcount = 0;
1626 struct vm_page marker;
1628 int pcount, tpcount; /* Number of pages to check */
1631 page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1632 vmstats.v_free_min) -
1633 (vmstats.v_free_count + vmstats.v_inactive_count +
1634 vmstats.v_cache_count);
1636 if (page_shortage <= 0)
1639 pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1640 fullintervalcount += vm_pageout_stats_interval;
1641 if (fullintervalcount < vm_pageout_full_stats_interval) {
1642 tpcount = (vm_pageout_stats_max * pcount) /
1643 vmstats.v_page_count + 1;
1644 if (pcount > tpcount)
1647 fullintervalcount = 0;
1650 bzero(&marker, sizeof(marker));
1651 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1652 marker.queue = PQ_ACTIVE + q;
1654 marker.wire_count = 1;
1656 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1657 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1660 * Queue locked at top of loop to avoid stack marker issues.
1662 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1667 KKASSERT(m->queue == PQ_ACTIVE + q);
1668 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1669 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1673 * Skip marker pages (atomic against other markers to avoid
1674 * infinite hop-over scans).
1676 if (m->flags & PG_MARKER)
1680 * Ignore pages we can't busy
1682 if (vm_page_busy_try(m, TRUE))
1686 * Remaining operations run with the page busy and neither
1687 * the page or the queue will be spin-locked.
1689 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1690 KKASSERT(m->queue == PQ_ACTIVE + q);
1693 * We now have a safely busied page, the page and queue
1694 * spinlocks have been released.
1698 if (m->hold_count) {
1704 * Calculate activity
1707 if (m->flags & PG_REFERENCED) {
1708 vm_page_flag_clear(m, PG_REFERENCED);
1711 actcount += pmap_ts_referenced(m);
1714 * Update act_count and move page to end of queue.
1717 m->act_count += ACT_ADVANCE + actcount;
1718 if (m->act_count > ACT_MAX)
1719 m->act_count = ACT_MAX;
1720 vm_page_and_queue_spin_lock(m);
1721 if (m->queue - m->pc == PQ_ACTIVE) {
1723 &vm_page_queues[PQ_ACTIVE + q].pl,
1726 &vm_page_queues[PQ_ACTIVE + q].pl,
1729 vm_page_and_queue_spin_unlock(m);
1734 if (m->act_count == 0) {
1736 * We turn off page access, so that we have
1737 * more accurate RSS stats. We don't do this
1738 * in the normal page deactivation when the
1739 * system is loaded VM wise, because the
1740 * cost of the large number of page protect
1741 * operations would be higher than the value
1742 * of doing the operation.
1744 * We use the marker to save our place so
1745 * we can release the spin lock. both (m)
1746 * and (next) will be invalid.
1748 vm_page_protect(m, VM_PROT_NONE);
1749 vm_page_deactivate(m);
1751 m->act_count -= min(m->act_count, ACT_DECLINE);
1752 vm_page_and_queue_spin_lock(m);
1753 if (m->queue - m->pc == PQ_ACTIVE) {
1755 &vm_page_queues[PQ_ACTIVE + q].pl,
1758 &vm_page_queues[PQ_ACTIVE + q].pl,
1761 vm_page_and_queue_spin_unlock(m);
1765 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1769 * Remove our local marker
1771 * Page queue still spin-locked.
1773 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1774 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1778 vm_pageout_free_page_calc(vm_size_t count)
1780 if (count < vmstats.v_page_count)
1783 * free_reserved needs to include enough for the largest swap pager
1784 * structures plus enough for any pv_entry structs when paging.
1786 * v_free_min normal allocations
1787 * v_free_reserved system allocations
1788 * v_pageout_free_min allocations by pageout daemon
1789 * v_interrupt_free_min low level allocations (e.g swap structures)
1791 if (vmstats.v_page_count > 1024)
1792 vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1794 vmstats.v_free_min = 64;
1797 * Make sure the vmmeter slop can't blow out our global minimums.
1799 * However, to accomodate weird configurations (vkernels with many
1800 * cpus and little memory, or artifically reduced hw.physmem), do
1801 * not allow v_free_min to exceed 1/20 of ram or the pageout demon
1802 * will go out of control.
1804 if (vmstats.v_free_min < VMMETER_SLOP_COUNT * ncpus * 10)
1805 vmstats.v_free_min = VMMETER_SLOP_COUNT * ncpus * 10;
1806 if (vmstats.v_free_min > vmstats.v_page_count / 20)
1807 vmstats.v_free_min = vmstats.v_page_count / 20;
1809 vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1810 vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1811 vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1812 vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1819 * vm_pageout is the high level pageout daemon.
1824 vm_pageout_thread(void)
1832 * Initialize some paging parameters.
1834 curthread->td_flags |= TDF_SYSTHREAD;
1836 vm_pageout_free_page_calc(vmstats.v_page_count);
1839 * v_free_target and v_cache_min control pageout hysteresis. Note
1840 * that these are more a measure of the VM cache queue hysteresis
1841 * then the VM free queue. Specifically, v_free_target is the
1842 * high water mark (free+cache pages).
1844 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1845 * low water mark, while v_free_min is the stop. v_cache_min must
1846 * be big enough to handle memory needs while the pageout daemon
1847 * is signalled and run to free more pages.
1849 if (vmstats.v_free_count > 6144)
1850 vmstats.v_free_target = 4 * vmstats.v_free_min +
1851 vmstats.v_free_reserved;
1853 vmstats.v_free_target = 2 * vmstats.v_free_min +
1854 vmstats.v_free_reserved;
1857 * NOTE: With the new buffer cache b_act_count we want the default
1858 * inactive target to be a percentage of available memory.
1860 * The inactive target essentially determines the minimum
1861 * number of 'temporary' pages capable of caching one-time-use
1862 * files when the VM system is otherwise full of pages
1863 * belonging to multi-time-use files or active program data.
1865 * NOTE: The inactive target is aggressively persued only if the
1866 * inactive queue becomes too small. If the inactive queue
1867 * is large enough to satisfy page movement to free+cache
1868 * then it is repopulated more slowly from the active queue.
1869 * This allows a general inactive_target default to be set.
1871 * There is an issue here for processes which sit mostly idle
1872 * 'overnight', such as sshd, tcsh, and X. Any movement from
1873 * the active queue will eventually cause such pages to
1874 * recycle eventually causing a lot of paging in the morning.
1875 * To reduce the incidence of this pages cycled out of the
1876 * buffer cache are moved directly to the inactive queue if
1877 * they were only used once or twice.
1879 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1880 * Increasing the value (up to 64) increases the number of
1881 * buffer recyclements which go directly to the inactive queue.
1883 if (vmstats.v_free_count > 2048) {
1884 vmstats.v_cache_min = vmstats.v_free_target;
1885 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1887 vmstats.v_cache_min = 0;
1888 vmstats.v_cache_max = 0;
1890 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1892 /* XXX does not really belong here */
1893 if (vm_page_max_wired == 0)
1894 vm_page_max_wired = vmstats.v_free_count / 3;
1896 if (vm_pageout_stats_max == 0)
1897 vm_pageout_stats_max = vmstats.v_free_target;
1900 * Set interval in seconds for stats scan.
1902 if (vm_pageout_stats_interval == 0)
1903 vm_pageout_stats_interval = 5;
1904 if (vm_pageout_full_stats_interval == 0)
1905 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1909 * Set maximum free per pass
1911 if (vm_pageout_stats_free_max == 0)
1912 vm_pageout_stats_free_max = 5;
1914 swap_pager_swap_init();
1918 * The pageout daemon is never done, so loop forever.
1923 int inactive_shortage;
1924 int vnodes_skipped = 0;
1925 int recycle_count = 0;
1929 * Wait for an action request. If we timeout check to
1930 * see if paging is needed (in case the normal wakeup
1933 if (vm_pages_needed == 0) {
1934 error = tsleep(&vm_pages_needed,
1936 vm_pageout_stats_interval * hz);
1938 vm_paging_needed() == 0 &&
1939 vm_pages_needed == 0) {
1940 for (q = 0; q < PQ_L2_SIZE; ++q)
1941 vm_pageout_page_stats(q);
1944 vm_pages_needed = 1;
1947 mycpu->gd_cnt.v_pdwakeups++;
1950 * Scan for INACTIVE->CLEAN/PAGEOUT
1952 * This routine tries to avoid thrashing the system with
1953 * unnecessary activity.
1955 * Calculate our target for the number of free+cache pages we
1956 * want to get to. This is higher then the number that causes
1957 * allocations to stall (severe) in order to provide hysteresis,
1958 * and if we don't make it all the way but get to the minimum
1959 * we're happy. Goose it a bit if there are multiple requests
1962 * Don't reduce avail_shortage inside the loop or the
1963 * PQAVERAGE() calculation will break.
1965 * NOTE! deficit is differentiated from avail_shortage as
1966 * REQUIRING at least (deficit) pages to be cleaned,
1967 * even if the page queues are in good shape. This
1968 * is used primarily for handling per-process
1969 * RLIMIT_RSS and may also see small values when
1970 * processes block due to low memory.
1973 avail_shortage = vm_paging_target() + vm_pageout_deficit;
1974 vm_pageout_deficit = 0;
1976 if (avail_shortage > 0) {
1979 for (q = 0; q < PQ_L2_SIZE; ++q) {
1980 delta += vm_pageout_scan_inactive(
1982 (q + q1iterator) & PQ_L2_MASK,
1983 PQAVERAGE(avail_shortage),
1985 if (avail_shortage - delta <= 0)
1988 avail_shortage -= delta;
1993 * Figure out how many active pages we must deactivate. If
1994 * we were able to reach our target with just the inactive
1995 * scan above we limit the number of active pages we
1996 * deactivate to reduce unnecessary work.
1999 inactive_shortage = vmstats.v_inactive_target -
2000 vmstats.v_inactive_count;
2003 * If we were unable to free sufficient inactive pages to
2004 * satisfy the free/cache queue requirements then simply
2005 * reaching the inactive target may not be good enough.
2006 * Try to deactivate pages in excess of the target based
2009 * However to prevent thrashing the VM system do not
2010 * deactivate more than an additional 1/10 the inactive
2011 * target's worth of active pages.
2013 if (avail_shortage > 0) {
2014 tmp = avail_shortage * 2;
2015 if (tmp > vmstats.v_inactive_target / 10)
2016 tmp = vmstats.v_inactive_target / 10;
2017 inactive_shortage += tmp;
2021 * Only trigger a pmap cleanup on inactive shortage.
2023 if (inactive_shortage > 0) {
2028 * Scan for ACTIVE->INACTIVE
2030 * Only trigger on inactive shortage. Triggering on
2031 * avail_shortage can starve the active queue with
2032 * unnecessary active->inactive transitions and destroy
2035 if (/*avail_shortage > 0 ||*/ inactive_shortage > 0) {
2038 for (q = 0; q < PQ_L2_SIZE; ++q) {
2039 delta += vm_pageout_scan_active(
2041 (q + q2iterator) & PQ_L2_MASK,
2042 PQAVERAGE(avail_shortage),
2043 PQAVERAGE(inactive_shortage),
2045 if (inactive_shortage - delta <= 0 &&
2046 avail_shortage - delta <= 0) {
2050 inactive_shortage -= delta;
2051 avail_shortage -= delta;
2056 * Scan for CACHE->FREE
2058 * Finally free enough cache pages to meet our free page
2059 * requirement and take more drastic measures if we are
2063 vm_pageout_scan_cache(avail_shortage, pass,
2064 vnodes_skipped, recycle_count);
2067 * Wait for more work.
2069 if (avail_shortage > 0) {
2071 if (pass < 10 && vm_pages_needed > 1) {
2073 * Normal operation, additional processes
2074 * have already kicked us. Retry immediately
2075 * unless swap space is completely full in
2076 * which case delay a bit.
2078 if (swap_pager_full) {
2079 tsleep(&vm_pages_needed, 0, "pdelay",
2081 } /* else immediate retry */
2082 } else if (pass < 10) {
2084 * Normal operation, fewer processes. Delay
2085 * a bit but allow wakeups.
2087 vm_pages_needed = 0;
2088 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2089 vm_pages_needed = 1;
2090 } else if (swap_pager_full == 0) {
2092 * We've taken too many passes, forced delay.
2094 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2097 * Running out of memory, catastrophic
2098 * back-off to one-second intervals.
2100 tsleep(&vm_pages_needed, 0, "pdelay", hz);
2102 } else if (vm_pages_needed) {
2104 * Interlocked wakeup of waiters (non-optional).
2106 * Similar to vm_page_free_wakeup() in vm_page.c,
2110 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2111 !vm_page_count_target()) {
2112 vm_pages_needed = 0;
2113 wakeup(&vmstats.v_free_count);
2121 static struct kproc_desc page_kp = {
2126 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp);
2130 * Called after allocating a page out of the cache or free queue
2131 * to possibly wake the pagedaemon up to replentish our supply.
2133 * We try to generate some hysteresis by waking the pagedaemon up
2134 * when our free+cache pages go below the free_min+cache_min level.
2135 * The pagedaemon tries to get the count back up to at least the
2136 * minimum, and through to the target level if possible.
2138 * If the pagedaemon is already active bump vm_pages_needed as a hint
2139 * that there are even more requests pending.
2145 pagedaemon_wakeup(void)
2147 if (vm_paging_needed() && curthread != pagethread) {
2148 if (vm_pages_needed == 0) {
2149 vm_pages_needed = 1; /* SMP race ok */
2150 wakeup(&vm_pages_needed);
2151 } else if (vm_page_count_min(0)) {
2152 ++vm_pages_needed; /* SMP race ok */
2157 #if !defined(NO_SWAPPING)
2164 vm_req_vmdaemon(void)
2166 static int lastrun = 0;
2168 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2169 wakeup(&vm_daemon_needed);
2174 static int vm_daemon_callback(struct proc *p, void *data __unused);
2185 tsleep(&vm_daemon_needed, 0, "psleep", 0);
2186 req_swapout = atomic_swap_int(&vm_pageout_req_swapout, 0);
2192 swapout_procs(vm_pageout_req_swapout);
2195 * scan the processes for exceeding their rlimits or if
2196 * process is swapped out -- deactivate pages
2198 allproc_scan(vm_daemon_callback, NULL);
2203 vm_daemon_callback(struct proc *p, void *data __unused)
2206 vm_pindex_t limit, size;
2209 * if this is a system process or if we have already
2210 * looked at this process, skip it.
2212 lwkt_gettoken(&p->p_token);
2214 if (p->p_flags & (P_SYSTEM | P_WEXIT)) {
2215 lwkt_reltoken(&p->p_token);
2220 * if the process is in a non-running type state,
2223 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
2224 lwkt_reltoken(&p->p_token);
2231 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2232 p->p_rlimit[RLIMIT_RSS].rlim_max));
2235 * let processes that are swapped out really be
2236 * swapped out. Set the limit to nothing to get as
2237 * many pages out to swap as possible.
2239 if (p->p_flags & P_SWAPPEDOUT)
2244 size = pmap_resident_tlnw_count(&vm->vm_pmap);
2245 if (limit >= 0 && size > 4096 &&
2246 size - 4096 >= limit && vm_pageout_memuse_mode >= 1) {
2247 vm_pageout_map_deactivate_pages(&vm->vm_map, limit);
2251 lwkt_reltoken(&p->p_token);