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>
82 #include <sys/sysctl.h>
85 #include <vm/vm_param.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_extern.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, long *max_launderp,
104 long *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 *emergpager;
110 struct thread *pagethread;
111 static int sequence_emerg_pager;
113 #if !defined(NO_SWAPPING)
114 /* the kernel process "vm_daemon"*/
115 static void vm_daemon (void);
116 static struct thread *vmthread;
118 static struct kproc_desc vm_kp = {
123 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
126 int vm_pages_needed = 0; /* Event on which pageout daemon sleeps */
127 int vm_pageout_deficit = 0; /* Estimated number of pages deficit */
128 int vm_pageout_pages_needed = 0;/* pageout daemon needs pages */
129 int vm_page_free_hysteresis = 16;
130 static int vm_pagedaemon_time;
132 #if !defined(NO_SWAPPING)
133 static int vm_pageout_req_swapout;
134 static int vm_daemon_needed;
136 static int vm_max_launder = 4096;
137 static int vm_emerg_launder = 100;
138 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
139 static int vm_pageout_full_stats_interval = 0;
140 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
141 static int defer_swap_pageouts=0;
142 static int disable_swap_pageouts=0;
143 static u_int vm_anonmem_decline = ACT_DECLINE;
144 static u_int vm_filemem_decline = ACT_DECLINE * 2;
146 #if defined(NO_SWAPPING)
147 static int vm_swap_enabled=0;
148 static int vm_swap_idle_enabled=0;
150 static int vm_swap_enabled=1;
151 static int vm_swap_idle_enabled=0;
153 int vm_pageout_memuse_mode=1; /* 0-disable, 1-passive, 2-active swp*/
155 SYSCTL_UINT(_vm, VM_PAGEOUT_ALGORITHM, anonmem_decline,
156 CTLFLAG_RW, &vm_anonmem_decline, 0, "active->inactive anon memory");
158 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, filemem_decline,
159 CTLFLAG_RW, &vm_filemem_decline, 0, "active->inactive file cache");
161 SYSCTL_INT(_vm, OID_AUTO, page_free_hysteresis,
162 CTLFLAG_RW, &vm_page_free_hysteresis, 0,
163 "Free more pages than the minimum required");
165 SYSCTL_INT(_vm, OID_AUTO, max_launder,
166 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
167 SYSCTL_INT(_vm, OID_AUTO, emerg_launder,
168 CTLFLAG_RW, &vm_emerg_launder, 0, "Emergency pager minimum");
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
171 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
173 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
174 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
176 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
177 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
180 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
181 SYSCTL_INT(_vm, OID_AUTO, pageout_memuse_mode,
182 CTLFLAG_RW, &vm_pageout_memuse_mode, 0, "memoryuse resource mode");
184 #if defined(NO_SWAPPING)
185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
186 CTLFLAG_RD, &vm_swap_enabled, 0, "");
187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
188 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
190 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
191 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
192 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
193 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
196 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
197 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
199 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
200 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
202 static int pageout_lock_miss;
203 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
204 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
206 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
208 #if !defined(NO_SWAPPING)
209 static void vm_req_vmdaemon (void);
211 static void vm_pageout_page_stats(int q);
214 * Calculate approximately how many pages on each queue to try to
215 * clean. An exact calculation creates an edge condition when the
216 * queues are unbalanced so add significant slop. The queue scans
217 * will stop early when targets are reached and will start where they
218 * left off on the next pass.
220 * We need to be generous here because there are all sorts of loading
221 * conditions that can cause edge cases if try to average over all queues.
222 * In particular, storage subsystems have become so fast that paging
223 * activity can become quite frantic. Eventually we will probably need
224 * two paging threads, one for dirty pages and one for clean, to deal
225 * with the bandwidth requirements.
227 * So what we do is calculate a value that can be satisfied nominally by
228 * only having to scan half the queues.
236 avg = ((n + (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) + 1);
238 avg = ((n - (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) - 1);
244 * vm_pageout_clean_helper:
246 * Clean the page and remove it from the laundry. The page must be busied
247 * by the caller and will be disposed of (put away, flushed) by this routine.
250 vm_pageout_clean_helper(vm_page_t m, int vmflush_flags)
253 vm_page_t mc[BLIST_MAX_ALLOC];
255 int ib, is, page_base;
256 vm_pindex_t pindex = m->pindex;
261 * Don't mess with the page if it's held or special.
263 * XXX do we really need to check hold_count here? hold_count
264 * isn't supposed to mess with vm_page ops except prevent the
265 * page from being reused.
267 if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
273 * Place page in cluster. Align cluster for optimal swap space
274 * allocation (whether it is swap or not). This is typically ~16-32
275 * pages, which also tends to align the cluster to multiples of the
276 * filesystem block size if backed by a filesystem.
278 page_base = pindex % BLIST_MAX_ALLOC;
284 * Scan object for clusterable pages.
286 * We can cluster ONLY if: ->> the page is NOT
287 * clean, wired, busy, held, or mapped into a
288 * buffer, and one of the following:
289 * 1) The page is inactive, or a seldom used
292 * 2) we force the issue.
294 * During heavy mmap/modification loads the pageout
295 * daemon can really fragment the underlying file
296 * due to flushing pages out of order and not trying
297 * align the clusters (which leave sporatic out-of-order
298 * holes). To solve this problem we do the reverse scan
299 * first and attempt to align our cluster, then do a
300 * forward scan if room remains.
302 vm_object_hold(object);
307 p = vm_page_lookup_busy_try(object, pindex - page_base + ib,
309 if (error || p == NULL)
311 if ((p->queue - p->pc) == PQ_CACHE ||
312 (p->flags & PG_UNMANAGED)) {
316 vm_page_test_dirty(p);
317 if (((p->dirty & p->valid) == 0 &&
318 (p->flags & PG_NEED_COMMIT) == 0) ||
319 p->wire_count != 0 || /* may be held by buf cache */
320 p->hold_count != 0) { /* may be undergoing I/O */
324 if (p->queue - p->pc != PQ_INACTIVE) {
325 if (p->queue - p->pc != PQ_ACTIVE ||
326 (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
333 * Try to maintain page groupings in the cluster.
335 if (m->flags & PG_WINATCFLS)
336 vm_page_flag_set(p, PG_WINATCFLS);
338 vm_page_flag_clear(p, PG_WINATCFLS);
339 p->act_count = m->act_count;
346 while (is < BLIST_MAX_ALLOC &&
347 pindex - page_base + is < object->size) {
350 p = vm_page_lookup_busy_try(object, pindex - page_base + is,
352 if (error || p == NULL)
354 if (((p->queue - p->pc) == PQ_CACHE) ||
355 (p->flags & PG_UNMANAGED)) {
359 vm_page_test_dirty(p);
360 if (((p->dirty & p->valid) == 0 &&
361 (p->flags & PG_NEED_COMMIT) == 0) ||
362 p->wire_count != 0 || /* may be held by buf cache */
363 p->hold_count != 0) { /* may be undergoing I/O */
367 if (p->queue - p->pc != PQ_INACTIVE) {
368 if (p->queue - p->pc != PQ_ACTIVE ||
369 (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
376 * Try to maintain page groupings in the cluster.
378 if (m->flags & PG_WINATCFLS)
379 vm_page_flag_set(p, PG_WINATCFLS);
381 vm_page_flag_clear(p, PG_WINATCFLS);
382 p->act_count = m->act_count;
388 vm_object_drop(object);
391 * we allow reads during pageouts...
393 return vm_pageout_flush(&mc[ib], is - ib, vmflush_flags);
397 * vm_pageout_flush() - launder the given pages
399 * The given pages are laundered. Note that we setup for the start of
400 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
401 * reference count all in here rather then in the parent. If we want
402 * the parent to do more sophisticated things we may have to change
405 * The pages in the array must be busied by the caller and will be
406 * unbusied by this function.
409 vm_pageout_flush(vm_page_t *mc, int count, int vmflush_flags)
412 int pageout_status[count];
417 * Initiate I/O. Bump the vm_page_t->busy counter.
419 for (i = 0; i < count; i++) {
420 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
421 ("vm_pageout_flush page %p index %d/%d: partially "
422 "invalid page", mc[i], i, count));
423 vm_page_io_start(mc[i]);
427 * We must make the pages read-only. This will also force the
428 * modified bit in the related pmaps to be cleared. The pager
429 * cannot clear the bit for us since the I/O completion code
430 * typically runs from an interrupt. The act of making the page
431 * read-only handles the case for us.
433 * Then we can unbusy the pages, we still hold a reference by virtue
436 for (i = 0; i < count; i++) {
437 if (vmflush_flags & VM_PAGER_TRY_TO_CACHE)
438 vm_page_protect(mc[i], VM_PROT_NONE);
440 vm_page_protect(mc[i], VM_PROT_READ);
441 vm_page_wakeup(mc[i]);
444 object = mc[0]->object;
445 vm_object_pip_add(object, count);
447 vm_pager_put_pages(object, mc, count,
449 ((object == &kernel_object) ?
450 VM_PAGER_PUT_SYNC : 0)),
453 for (i = 0; i < count; i++) {
454 vm_page_t mt = mc[i];
456 switch (pageout_status[i]) {
465 * Page outside of range of object. Right now we
466 * essentially lose the changes by pretending it
469 vm_page_busy_wait(mt, FALSE, "pgbad");
470 pmap_clear_modify(mt);
477 * A page typically cannot be paged out when we
478 * have run out of swap. We leave the page
479 * marked inactive and will try to page it out
482 * Starvation of the active page list is used to
483 * determine when the system is massively memory
492 * If not PENDing this was a synchronous operation and we
493 * clean up after the I/O. If it is PENDing the mess is
494 * cleaned up asynchronously.
496 * Also nominally act on the caller's wishes if the caller
497 * wants to try to really clean (cache or free) the page.
499 * Also nominally deactivate the page if the system is
502 if (pageout_status[i] != VM_PAGER_PEND) {
503 vm_page_busy_wait(mt, FALSE, "pgouw");
504 vm_page_io_finish(mt);
505 if (vmflush_flags & VM_PAGER_TRY_TO_CACHE) {
506 vm_page_try_to_cache(mt);
507 } else if (vm_page_count_severe()) {
508 vm_page_deactivate(mt);
513 vm_object_pip_wakeup(object);
519 #if !defined(NO_SWAPPING)
522 * Callback function, page busied for us. We must dispose of the busy
523 * condition. Any related pmap pages may be held but will not be locked.
527 vm_pageout_mdp_callback(struct pmap_pgscan_info *info, vm_offset_t va,
534 * Basic tests - There should never be a marker, and we can stop
535 * once the RSS is below the required level.
537 KKASSERT((p->flags & PG_MARKER) == 0);
538 if (pmap_resident_tlnw_count(info->pmap) <= info->limit) {
543 mycpu->gd_cnt.v_pdpages++;
545 if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
553 * Check if the page has been referened recently. If it has,
554 * activate it and skip.
556 actcount = pmap_ts_referenced(p);
558 vm_page_flag_set(p, PG_REFERENCED);
559 } else if (p->flags & PG_REFERENCED) {
564 if (p->queue - p->pc != PQ_ACTIVE) {
565 vm_page_and_queue_spin_lock(p);
566 if (p->queue - p->pc != PQ_ACTIVE) {
567 vm_page_and_queue_spin_unlock(p);
570 vm_page_and_queue_spin_unlock(p);
573 p->act_count += actcount;
574 if (p->act_count > ACT_MAX)
575 p->act_count = ACT_MAX;
577 vm_page_flag_clear(p, PG_REFERENCED);
583 * Remove the page from this particular pmap. Once we do this, our
584 * pmap scans will not see it again (unless it gets faulted in), so
585 * we must actively dispose of or deal with the page.
587 pmap_remove_specific(info->pmap, p);
590 * If the page is not mapped to another process (i.e. as would be
591 * typical if this were a shared page from a library) then deactivate
592 * the page and clean it in two passes only.
594 * If the page hasn't been referenced since the last check, remove it
595 * from the pmap. If it is no longer mapped, deactivate it
596 * immediately, accelerating the normal decline.
598 * Once the page has been removed from the pmap the RSS code no
599 * longer tracks it so we have to make sure that it is staged for
600 * potential flush action.
602 if ((p->flags & PG_MAPPED) == 0) {
603 if (p->queue - p->pc == PQ_ACTIVE) {
604 vm_page_deactivate(p);
606 if (p->queue - p->pc == PQ_INACTIVE) {
612 * Ok, try to fully clean the page and any nearby pages such that at
613 * least the requested page is freed or moved to the cache queue.
615 * We usually do this synchronously to allow us to get the page into
616 * the CACHE queue quickly, which will prevent memory exhaustion if
617 * a process with a memoryuse limit is running away. However, the
618 * sysadmin may desire to set vm.swap_user_async which relaxes this
619 * and improves write performance.
622 long max_launder = 0x7FFF;
623 long vnodes_skipped = 0;
625 struct vnode *vpfailed = NULL;
629 if (vm_pageout_memuse_mode >= 2) {
630 vmflush_flags = VM_PAGER_TRY_TO_CACHE |
631 VM_PAGER_ALLOW_ACTIVE;
632 if (swap_user_async == 0)
633 vmflush_flags |= VM_PAGER_PUT_SYNC;
634 vm_page_flag_set(p, PG_WINATCFLS);
636 vm_pageout_page(p, &max_launder,
638 &vpfailed, 1, vmflush_flags);
648 * Must be at end to avoid SMP races.
656 * Deactivate some number of pages in a map due to set RLIMIT_RSS limits.
657 * that is relatively difficult to do. We try to keep track of where we
658 * left off last time to reduce scan overhead.
660 * Called when vm_pageout_memuse_mode is >= 1.
663 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t limit)
665 vm_offset_t pgout_offset;
666 struct pmap_pgscan_info info;
669 pgout_offset = map->pgout_offset;
672 kprintf("%016jx ", pgout_offset);
674 if (pgout_offset < VM_MIN_USER_ADDRESS)
675 pgout_offset = VM_MIN_USER_ADDRESS;
676 if (pgout_offset >= VM_MAX_USER_ADDRESS)
678 info.pmap = vm_map_pmap(map);
680 info.beg_addr = pgout_offset;
681 info.end_addr = VM_MAX_USER_ADDRESS;
682 info.callback = vm_pageout_mdp_callback;
684 info.actioncount = 0;
688 pgout_offset = info.offset;
690 kprintf("%016jx %08lx %08lx\n", pgout_offset,
691 info.cleancount, info.actioncount);
694 if (pgout_offset != VM_MAX_USER_ADDRESS &&
695 pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
697 } else if (retries &&
698 pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
702 map->pgout_offset = pgout_offset;
707 * Called when the pageout scan wants to free a page. We no longer
708 * try to cycle the vm_object here with a reference & dealloc, which can
709 * cause a non-trivial object collapse in a critical path.
711 * It is unclear why we cycled the ref_count in the past, perhaps to try
712 * to optimize shadow chain collapses but I don't quite see why it would
713 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
714 * synchronously and not have to be kicked-start.
717 vm_pageout_page_free(vm_page_t m)
719 vm_page_protect(m, VM_PROT_NONE);
724 * vm_pageout_scan does the dirty work for the pageout daemon.
726 struct vm_pageout_scan_info {
727 struct proc *bigproc;
731 static int vm_pageout_scan_callback(struct proc *p, void *data);
734 * Scan inactive queue
736 * WARNING! Can be called from two pagedaemon threads simultaneously.
739 vm_pageout_scan_inactive(int pass, int q, long avail_shortage,
740 long *vnodes_skipped)
743 struct vm_page marker;
744 struct vnode *vpfailed; /* warning, allowed to be stale */
750 isep = (curthread == emergpager);
753 * Start scanning the inactive queue for pages we can move to the
754 * cache or free. The scan will stop when the target is reached or
755 * we have scanned the entire inactive queue. Note that m->act_count
756 * is not used to form decisions for the inactive queue, only for the
759 * max_launder limits the number of dirty pages we flush per scan.
760 * For most systems a smaller value (16 or 32) is more robust under
761 * extreme memory and disk pressure because any unnecessary writes
762 * to disk can result in extreme performance degredation. However,
763 * systems with excessive dirty pages (especially when MAP_NOSYNC is
764 * used) will die horribly with limited laundering. If the pageout
765 * daemon cannot clean enough pages in the first pass, we let it go
766 * all out in succeeding passes.
768 * NOTE! THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
771 if ((max_launder = vm_max_launder) <= 1)
777 * Initialize our marker
779 bzero(&marker, sizeof(marker));
780 marker.flags = PG_FICTITIOUS | PG_MARKER;
781 marker.busy_count = PBUSY_LOCKED;
782 marker.queue = PQ_INACTIVE + q;
784 marker.wire_count = 1;
787 * Inactive queue scan.
789 * NOTE: The vm_page must be spinlocked before the queue to avoid
790 * deadlocks, so it is easiest to simply iterate the loop
791 * with the queue unlocked at the top.
795 vm_page_queues_spin_lock(PQ_INACTIVE + q);
796 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
797 maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt;
800 * Queue locked at top of loop to avoid stack marker issues.
802 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
803 maxscan-- > 0 && avail_shortage - delta > 0)
807 KKASSERT(m->queue == PQ_INACTIVE + q);
808 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
810 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
812 mycpu->gd_cnt.v_pdpages++;
815 * Skip marker pages (atomic against other markers to avoid
816 * infinite hop-over scans).
818 if (m->flags & PG_MARKER)
822 * Try to busy the page. Don't mess with pages which are
823 * already busy or reorder them in the queue.
825 if (vm_page_busy_try(m, TRUE))
829 * Remaining operations run with the page busy and neither
830 * the page or the queue will be spin-locked.
832 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
833 KKASSERT(m->queue == PQ_INACTIVE + q);
836 * The emergency pager runs when the primary pager gets
837 * stuck, which typically means the primary pager deadlocked
838 * on a vnode-backed page. Therefore, the emergency pager
839 * must skip any complex objects.
841 * We disallow VNODEs unless they are VCHR whos device ops
842 * does not flag D_NOEMERGPGR.
844 if (isep && m->object) {
847 switch(m->object->type) {
851 * Allow anonymous memory and assume that
852 * swap devices are not complex, since its
853 * kinda worthless if we can't swap out dirty
859 * Allow VCHR device if the D_NOEMERGPGR
860 * flag is not set, deny other vnode types
861 * as being too complex.
863 vp = m->object->handle;
864 if (vp && vp->v_type == VCHR &&
865 vp->v_rdev && vp->v_rdev->si_ops &&
866 (vp->v_rdev->si_ops->head.flags &
867 D_NOEMERGPGR) == 0) {
870 /* Deny - fall through */
876 vm_page_queues_spin_lock(PQ_INACTIVE + q);
883 * Try to pageout the page and perhaps other nearby pages.
885 count = vm_pageout_page(m, &max_launder, vnodes_skipped,
890 * Systems with a ton of memory can wind up with huge
891 * deactivation counts. Because the inactive scan is
892 * doing a lot of flushing, the combination can result
893 * in excessive paging even in situations where other
894 * unrelated threads free up sufficient VM.
896 * To deal with this we abort the nominal active->inactive
897 * scan before we hit the inactive target when free+cache
898 * levels have reached a reasonable target.
900 * When deciding to stop early we need to add some slop to
901 * the test and we need to return full completion to the caller
902 * to prevent the caller from thinking there is something
903 * wrong and issuing a low-memory+swap warning or pkill.
905 * A deficit forces paging regardless of the state of the
906 * VM page queues (used for RSS enforcement).
909 vm_page_queues_spin_lock(PQ_INACTIVE + q);
910 if (vm_paging_target() < -vm_max_launder) {
912 * Stopping early, return full completion to caller.
914 if (delta < avail_shortage)
915 delta = avail_shortage;
920 /* page queue still spin-locked */
921 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
922 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
928 * Pageout the specified page, return the total number of pages paged out
929 * (this routine may cluster).
931 * The page must be busied and soft-busied by the caller and will be disposed
932 * of by this function.
935 vm_pageout_page(vm_page_t m, long *max_launderp, long *vnodes_skippedp,
936 struct vnode **vpfailedp, int pass, int vmflush_flags)
943 * It is possible for a page to be busied ad-hoc (e.g. the
944 * pmap_collect() code) and wired and race against the
945 * allocation of a new page. vm_page_alloc() may be forced
946 * to deactivate the wired page in which case it winds up
947 * on the inactive queue and must be handled here. We
948 * correct the problem simply by unqueuing the page.
951 vm_page_unqueue_nowakeup(m);
953 kprintf("WARNING: pagedaemon: wired page on "
954 "inactive queue %p\n", m);
959 * A held page may be undergoing I/O, so skip it.
962 vm_page_and_queue_spin_lock(m);
963 if (m->queue - m->pc == PQ_INACTIVE) {
965 &vm_page_queues[m->queue].pl, m, pageq);
967 &vm_page_queues[m->queue].pl, m, pageq);
968 ++vm_swapcache_inactive_heuristic;
970 vm_page_and_queue_spin_unlock(m);
975 if (m->object == NULL || m->object->ref_count == 0) {
977 * If the object is not being used, we ignore previous
980 vm_page_flag_clear(m, PG_REFERENCED);
981 pmap_clear_reference(m);
982 /* fall through to end */
983 } else if (((m->flags & PG_REFERENCED) == 0) &&
984 (actcount = pmap_ts_referenced(m))) {
986 * Otherwise, if the page has been referenced while
987 * in the inactive queue, we bump the "activation
988 * count" upwards, making it less likely that the
989 * page will be added back to the inactive queue
990 * prematurely again. Here we check the page tables
991 * (or emulated bits, if any), given the upper level
992 * VM system not knowing anything about existing
996 m->act_count += (actcount + ACT_ADVANCE);
1002 * (m) is still busied.
1004 * If the upper level VM system knows about any page
1005 * references, we activate the page. We also set the
1006 * "activation count" higher than normal so that we will less
1007 * likely place pages back onto the inactive queue again.
1009 if ((m->flags & PG_REFERENCED) != 0) {
1010 vm_page_flag_clear(m, PG_REFERENCED);
1011 actcount = pmap_ts_referenced(m);
1012 vm_page_activate(m);
1013 m->act_count += (actcount + ACT_ADVANCE + 1);
1019 * If the upper level VM system doesn't know anything about
1020 * the page being dirty, we have to check for it again. As
1021 * far as the VM code knows, any partially dirty pages are
1024 * Pages marked PG_WRITEABLE may be mapped into the user
1025 * address space of a process running on another cpu. A
1026 * user process (without holding the MP lock) running on
1027 * another cpu may be able to touch the page while we are
1028 * trying to remove it. vm_page_cache() will handle this
1031 if (m->dirty == 0) {
1032 vm_page_test_dirty(m);
1037 if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
1039 * Invalid pages can be easily freed
1041 vm_pageout_page_free(m);
1042 mycpu->gd_cnt.v_dfree++;
1044 } else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
1046 * Clean pages can be placed onto the cache queue.
1047 * This effectively frees them.
1051 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
1053 * Dirty pages need to be paged out, but flushing
1054 * a page is extremely expensive verses freeing
1055 * a clean page. Rather then artificially limiting
1056 * the number of pages we can flush, we instead give
1057 * dirty pages extra priority on the inactive queue
1058 * by forcing them to be cycled through the queue
1059 * twice before being flushed, after which the
1060 * (now clean) page will cycle through once more
1061 * before being freed. This significantly extends
1062 * the thrash point for a heavily loaded machine.
1064 vm_page_flag_set(m, PG_WINATCFLS);
1065 vm_page_and_queue_spin_lock(m);
1066 if (m->queue - m->pc == PQ_INACTIVE) {
1068 &vm_page_queues[m->queue].pl, m, pageq);
1070 &vm_page_queues[m->queue].pl, m, pageq);
1071 ++vm_swapcache_inactive_heuristic;
1073 vm_page_and_queue_spin_unlock(m);
1075 } else if (*max_launderp > 0) {
1077 * We always want to try to flush some dirty pages if
1078 * we encounter them, to keep the system stable.
1079 * Normally this number is small, but under extreme
1080 * pressure where there are insufficient clean pages
1081 * on the inactive queue, we may have to go all out.
1083 int swap_pageouts_ok;
1084 struct vnode *vp = NULL;
1086 swap_pageouts_ok = 0;
1089 (object->type != OBJT_SWAP) &&
1090 (object->type != OBJT_DEFAULT)) {
1091 swap_pageouts_ok = 1;
1093 swap_pageouts_ok = !(defer_swap_pageouts ||
1094 disable_swap_pageouts);
1095 swap_pageouts_ok |= (!disable_swap_pageouts &&
1096 defer_swap_pageouts &&
1097 vm_page_count_min(0));
1101 * We don't bother paging objects that are "dead".
1102 * Those objects are in a "rundown" state.
1104 if (!swap_pageouts_ok ||
1106 (object->flags & OBJ_DEAD)) {
1107 vm_page_and_queue_spin_lock(m);
1108 if (m->queue - m->pc == PQ_INACTIVE) {
1110 &vm_page_queues[m->queue].pl,
1113 &vm_page_queues[m->queue].pl,
1115 ++vm_swapcache_inactive_heuristic;
1117 vm_page_and_queue_spin_unlock(m);
1123 * (m) is still busied.
1125 * The object is already known NOT to be dead. It
1126 * is possible for the vget() to block the whole
1127 * pageout daemon, but the new low-memory handling
1128 * code should prevent it.
1130 * The previous code skipped locked vnodes and, worse,
1131 * reordered pages in the queue. This results in
1132 * completely non-deterministic operation because,
1133 * quite often, a vm_fault has initiated an I/O and
1134 * is holding a locked vnode at just the point where
1135 * the pageout daemon is woken up.
1137 * We can't wait forever for the vnode lock, we might
1138 * deadlock due to a vn_read() getting stuck in
1139 * vm_wait while holding this vnode. We skip the
1140 * vnode if we can't get it in a reasonable amount
1143 * vpfailed is used to (try to) avoid the case where
1144 * a large number of pages are associated with a
1145 * locked vnode, which could cause the pageout daemon
1146 * to stall for an excessive amount of time.
1148 if (object->type == OBJT_VNODE) {
1151 vp = object->handle;
1152 flags = LK_EXCLUSIVE;
1153 if (vp == *vpfailedp)
1156 flags |= LK_TIMELOCK;
1161 * We have unbusied (m) temporarily so we can
1162 * acquire the vp lock without deadlocking.
1163 * (m) is held to prevent destruction.
1165 if (vget(vp, flags) != 0) {
1167 ++pageout_lock_miss;
1168 if (object->flags & OBJ_MIGHTBEDIRTY)
1175 * The page might have been moved to another
1176 * queue during potential blocking in vget()
1177 * above. The page might have been freed and
1178 * reused for another vnode. The object might
1179 * have been reused for another vnode.
1181 if (m->queue - m->pc != PQ_INACTIVE ||
1182 m->object != object ||
1183 object->handle != vp) {
1184 if (object->flags & OBJ_MIGHTBEDIRTY)
1192 * The page may have been busied during the
1193 * blocking in vput(); We don't move the
1194 * page back onto the end of the queue so that
1195 * statistics are more correct if we don't.
1197 if (vm_page_busy_try(m, TRUE)) {
1205 * (m) is busied again
1207 * We own the busy bit and remove our hold
1208 * bit. If the page is still held it
1209 * might be undergoing I/O, so skip it.
1211 if (m->hold_count) {
1212 vm_page_and_queue_spin_lock(m);
1213 if (m->queue - m->pc == PQ_INACTIVE) {
1214 TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq);
1215 TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1216 ++vm_swapcache_inactive_heuristic;
1218 vm_page_and_queue_spin_unlock(m);
1219 if (object->flags & OBJ_MIGHTBEDIRTY)
1225 /* (m) is left busied as we fall through */
1229 * page is busy and not held here.
1231 * If a page is dirty, then it is either being washed
1232 * (but not yet cleaned) or it is still in the
1233 * laundry. If it is still in the laundry, then we
1234 * start the cleaning operation.
1236 * decrement inactive_shortage on success to account
1237 * for the (future) cleaned page. Otherwise we
1238 * could wind up laundering or cleaning too many
1241 * NOTE: Cleaning the page here does not cause
1242 * force_deficit to be adjusted, because the
1243 * page is not being freed or moved to the
1246 count = vm_pageout_clean_helper(m, vmflush_flags);
1247 *max_launderp -= count;
1250 * Clean ate busy, page no longer accessible
1263 * WARNING! Can be called from two pagedaemon threads simultaneously.
1266 vm_pageout_scan_active(int pass, int q,
1267 long avail_shortage, long inactive_shortage,
1268 long *recycle_countp)
1270 struct vm_page marker;
1277 isep = (curthread == emergpager);
1280 * We want to move pages from the active queue to the inactive
1281 * queue to get the inactive queue to the inactive target. If
1282 * we still have a page shortage from above we try to directly free
1283 * clean pages instead of moving them.
1285 * If we do still have a shortage we keep track of the number of
1286 * pages we free or cache (recycle_count) as a measure of thrashing
1287 * between the active and inactive queues.
1289 * If we were able to completely satisfy the free+cache targets
1290 * from the inactive pool we limit the number of pages we move
1291 * from the active pool to the inactive pool to 2x the pages we
1292 * had removed from the inactive pool (with a minimum of 1/5 the
1293 * inactive target). If we were not able to completely satisfy
1294 * the free+cache targets we go for the whole target aggressively.
1296 * NOTE: Both variables can end up negative.
1297 * NOTE: We are still in a critical section.
1299 * NOTE! THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
1303 bzero(&marker, sizeof(marker));
1304 marker.flags = PG_FICTITIOUS | PG_MARKER;
1305 marker.busy_count = PBUSY_LOCKED;
1306 marker.queue = PQ_ACTIVE + q;
1308 marker.wire_count = 1;
1310 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1311 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1312 maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt;
1315 * Queue locked at top of loop to avoid stack marker issues.
1317 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1318 maxscan-- > 0 && (avail_shortage - delta > 0 ||
1319 inactive_shortage > 0))
1321 KKASSERT(m->queue == PQ_ACTIVE + q);
1322 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1324 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1328 * Skip marker pages (atomic against other markers to avoid
1329 * infinite hop-over scans).
1331 if (m->flags & PG_MARKER)
1335 * Try to busy the page. Don't mess with pages which are
1336 * already busy or reorder them in the queue.
1338 if (vm_page_busy_try(m, TRUE))
1342 * Remaining operations run with the page busy and neither
1343 * the page or the queue will be spin-locked.
1345 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1346 KKASSERT(m->queue == PQ_ACTIVE + q);
1349 * Don't deactivate pages that are held, even if we can
1350 * busy them. (XXX why not?)
1352 if (m->hold_count != 0) {
1353 vm_page_and_queue_spin_lock(m);
1354 if (m->queue - m->pc == PQ_ACTIVE) {
1356 &vm_page_queues[PQ_ACTIVE + q].pl,
1359 &vm_page_queues[PQ_ACTIVE + q].pl,
1362 vm_page_and_queue_spin_unlock(m);
1368 * The emergency pager ignores vnode-backed pages as these
1369 * are the pages that probably bricked the main pager.
1371 if (isep && m->object && m->object->type == OBJT_VNODE) {
1372 vm_page_and_queue_spin_lock(m);
1373 if (m->queue - m->pc == PQ_ACTIVE) {
1375 &vm_page_queues[PQ_ACTIVE + q].pl,
1378 &vm_page_queues[PQ_ACTIVE + q].pl,
1381 vm_page_and_queue_spin_unlock(m);
1387 * The count for pagedaemon pages is done after checking the
1388 * page for eligibility...
1390 mycpu->gd_cnt.v_pdpages++;
1393 * Check to see "how much" the page has been used and clear
1394 * the tracking access bits. If the object has no references
1395 * don't bother paying the expense.
1398 if (m->object && m->object->ref_count != 0) {
1399 if (m->flags & PG_REFERENCED)
1401 actcount += pmap_ts_referenced(m);
1403 m->act_count += ACT_ADVANCE + actcount;
1404 if (m->act_count > ACT_MAX)
1405 m->act_count = ACT_MAX;
1408 vm_page_flag_clear(m, PG_REFERENCED);
1411 * actcount is only valid if the object ref_count is non-zero.
1412 * If the page does not have an object, actcount will be zero.
1414 if (actcount && m->object->ref_count != 0) {
1415 vm_page_and_queue_spin_lock(m);
1416 if (m->queue - m->pc == PQ_ACTIVE) {
1418 &vm_page_queues[PQ_ACTIVE + q].pl,
1421 &vm_page_queues[PQ_ACTIVE + q].pl,
1424 vm_page_and_queue_spin_unlock(m);
1427 switch(m->object->type) {
1430 m->act_count -= min(m->act_count,
1431 vm_anonmem_decline);
1434 m->act_count -= min(m->act_count,
1435 vm_filemem_decline);
1438 if (vm_pageout_algorithm ||
1439 (m->object == NULL) ||
1440 (m->object && (m->object->ref_count == 0)) ||
1441 m->act_count < pass + 1
1444 * Deactivate the page. If we had a
1445 * shortage from our inactive scan try to
1446 * free (cache) the page instead.
1448 * Don't just blindly cache the page if
1449 * we do not have a shortage from the
1450 * inactive scan, that could lead to
1451 * gigabytes being moved.
1453 --inactive_shortage;
1454 if (avail_shortage - delta > 0 ||
1455 (m->object && (m->object->ref_count == 0)))
1457 if (avail_shortage - delta > 0)
1459 vm_page_protect(m, VM_PROT_NONE);
1460 if (m->dirty == 0 &&
1461 (m->flags & PG_NEED_COMMIT) == 0 &&
1462 avail_shortage - delta > 0) {
1465 vm_page_deactivate(m);
1469 vm_page_deactivate(m);
1474 vm_page_and_queue_spin_lock(m);
1475 if (m->queue - m->pc == PQ_ACTIVE) {
1477 &vm_page_queues[PQ_ACTIVE + q].pl,
1480 &vm_page_queues[PQ_ACTIVE + q].pl,
1483 vm_page_and_queue_spin_unlock(m);
1489 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1493 * Clean out our local marker.
1495 * Page queue still spin-locked.
1497 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1498 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1504 * The number of actually free pages can drop down to v_free_reserved,
1505 * we try to build the free count back above v_free_min. Note that
1506 * vm_paging_needed() also returns TRUE if v_free_count is not at
1507 * least v_free_min so that is the minimum we must build the free
1510 * We use a slightly higher target to improve hysteresis,
1511 * ((v_free_target + v_free_min) / 2). Since v_free_target
1512 * is usually the same as v_cache_min this maintains about
1513 * half the pages in the free queue as are in the cache queue,
1514 * providing pretty good pipelining for pageout operation.
1516 * The system operator can manipulate vm.v_cache_min and
1517 * vm.v_free_target to tune the pageout demon. Be sure
1518 * to keep vm.v_free_min < vm.v_free_target.
1520 * Note that the original paging target is to get at least
1521 * (free_min + cache_min) into (free + cache). The slightly
1522 * higher target will shift additional pages from cache to free
1523 * without effecting the original paging target in order to
1524 * maintain better hysteresis and not have the free count always
1525 * be dead-on v_free_min.
1527 * NOTE: we are still in a critical section.
1529 * Pages moved from PQ_CACHE to totally free are not counted in the
1530 * pages_freed counter.
1532 * WARNING! Can be called from two pagedaemon threads simultaneously.
1535 vm_pageout_scan_cache(long avail_shortage, int pass,
1536 long vnodes_skipped, long recycle_count)
1538 static int lastkillticks;
1539 struct vm_pageout_scan_info info;
1543 isep = (curthread == emergpager);
1545 while (vmstats.v_free_count <
1546 (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1548 * This steals some code from vm/vm_page.c
1550 * Create two rovers and adjust the code to reduce
1551 * chances of them winding up at the same index (which
1552 * can cause a lot of contention).
1554 static int cache_rover[2] = { 0, PQ_L2_MASK / 2 };
1556 if (((cache_rover[0] ^ cache_rover[1]) & PQ_L2_MASK) == 0)
1559 m = vm_page_list_find(PQ_CACHE, cache_rover[isep] & PQ_L2_MASK);
1562 /* page is returned removed from its queue and spinlocked */
1563 if (vm_page_busy_try(m, TRUE)) {
1564 vm_page_deactivate_locked(m);
1565 vm_page_spin_unlock(m);
1568 vm_page_spin_unlock(m);
1569 pagedaemon_wakeup();
1573 * Remaining operations run with the page busy and neither
1574 * the page or the queue will be spin-locked.
1576 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1579 vm_page_deactivate(m);
1583 KKASSERT((m->flags & PG_MAPPED) == 0);
1584 KKASSERT(m->dirty == 0);
1585 vm_pageout_page_free(m);
1586 mycpu->gd_cnt.v_dfree++;
1589 cache_rover[1] -= PQ_PRIME2;
1591 cache_rover[0] += PQ_PRIME2;
1594 #if !defined(NO_SWAPPING)
1596 * Idle process swapout -- run once per second.
1598 if (vm_swap_idle_enabled) {
1600 if (time_uptime != lsec) {
1601 atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_IDLE);
1609 * If we didn't get enough free pages, and we have skipped a vnode
1610 * in a writeable object, wakeup the sync daemon. And kick swapout
1611 * if we did not get enough free pages.
1613 if (vm_paging_target() > 0) {
1614 if (vnodes_skipped && vm_page_count_min(0))
1615 speedup_syncer(NULL);
1616 #if !defined(NO_SWAPPING)
1617 if (vm_swap_enabled && vm_page_count_target()) {
1618 atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_NORMAL);
1625 * Handle catastrophic conditions. Under good conditions we should
1626 * be at the target, well beyond our minimum. If we could not even
1627 * reach our minimum the system is under heavy stress. But just being
1628 * under heavy stress does not trigger process killing.
1630 * We consider ourselves to have run out of memory if the swap pager
1631 * is full and avail_shortage is still positive. The secondary check
1632 * ensures that we do not kill processes if the instantanious
1633 * availability is good, even if the pageout demon pass says it
1634 * couldn't get to the target.
1636 * NOTE! THE EMERGENCY PAGER (isep) DOES NOT HANDLE SWAP FULL
1639 if (swap_pager_almost_full &&
1642 (vm_page_count_min(recycle_count) || avail_shortage > 0)) {
1643 kprintf("Warning: system low on memory+swap "
1644 "shortage %ld for %d ticks!\n",
1645 avail_shortage, ticks - swap_fail_ticks);
1647 kprintf("Metrics: spaf=%d spf=%d pass=%d "
1648 "avail=%ld target=%ld last=%u\n",
1649 swap_pager_almost_full,
1654 (unsigned int)(ticks - lastkillticks));
1656 if (swap_pager_full &&
1659 avail_shortage > 0 &&
1660 vm_paging_target() > 0 &&
1661 (unsigned int)(ticks - lastkillticks) >= hz) {
1663 * Kill something, maximum rate once per second to give
1664 * the process time to free up sufficient memory.
1666 lastkillticks = ticks;
1667 info.bigproc = NULL;
1669 allproc_scan(vm_pageout_scan_callback, &info, 0);
1670 if (info.bigproc != NULL) {
1671 kprintf("Try to kill process %d %s\n",
1672 info.bigproc->p_pid, info.bigproc->p_comm);
1673 info.bigproc->p_nice = PRIO_MIN;
1674 info.bigproc->p_usched->resetpriority(
1675 FIRST_LWP_IN_PROC(info.bigproc));
1676 atomic_set_int(&info.bigproc->p_flags, P_LOWMEMKILL);
1677 killproc(info.bigproc, "out of swap space");
1678 wakeup(&vmstats.v_free_count);
1679 PRELE(info.bigproc);
1685 vm_pageout_scan_callback(struct proc *p, void *data)
1687 struct vm_pageout_scan_info *info = data;
1691 * Never kill system processes or init. If we have configured swap
1692 * then try to avoid killing low-numbered pids.
1694 if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1695 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1699 lwkt_gettoken(&p->p_token);
1702 * if the process is in a non-running type state,
1705 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
1706 lwkt_reltoken(&p->p_token);
1711 * Get the approximate process size. Note that anonymous pages
1712 * with backing swap will be counted twice, but there should not
1713 * be too many such pages due to the stress the VM system is
1714 * under at this point.
1716 size = vmspace_anonymous_count(p->p_vmspace) +
1717 vmspace_swap_count(p->p_vmspace);
1720 * If the this process is bigger than the biggest one
1723 if (info->bigsize < size) {
1725 PRELE(info->bigproc);
1728 info->bigsize = size;
1730 lwkt_reltoken(&p->p_token);
1737 * This routine tries to maintain the pseudo LRU active queue,
1738 * so that during long periods of time where there is no paging,
1739 * that some statistic accumulation still occurs. This code
1740 * helps the situation where paging just starts to occur.
1743 vm_pageout_page_stats(int q)
1745 static int fullintervalcount = 0;
1746 struct vm_page marker;
1748 long pcount, tpcount; /* Number of pages to check */
1751 page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1752 vmstats.v_free_min) -
1753 (vmstats.v_free_count + vmstats.v_inactive_count +
1754 vmstats.v_cache_count);
1756 if (page_shortage <= 0)
1759 pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1760 fullintervalcount += vm_pageout_stats_interval;
1761 if (fullintervalcount < vm_pageout_full_stats_interval) {
1762 tpcount = (vm_pageout_stats_max * pcount) /
1763 vmstats.v_page_count + 1;
1764 if (pcount > tpcount)
1767 fullintervalcount = 0;
1770 bzero(&marker, sizeof(marker));
1771 marker.flags = PG_FICTITIOUS | PG_MARKER;
1772 marker.busy_count = PBUSY_LOCKED;
1773 marker.queue = PQ_ACTIVE + q;
1775 marker.wire_count = 1;
1777 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1778 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1781 * Queue locked at top of loop to avoid stack marker issues.
1783 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1788 KKASSERT(m->queue == PQ_ACTIVE + q);
1789 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1790 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1794 * Skip marker pages (atomic against other markers to avoid
1795 * infinite hop-over scans).
1797 if (m->flags & PG_MARKER)
1801 * Ignore pages we can't busy
1803 if (vm_page_busy_try(m, TRUE))
1807 * Remaining operations run with the page busy and neither
1808 * the page or the queue will be spin-locked.
1810 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1811 KKASSERT(m->queue == PQ_ACTIVE + q);
1814 * We now have a safely busied page, the page and queue
1815 * spinlocks have been released.
1819 if (m->hold_count) {
1825 * Calculate activity
1828 if (m->flags & PG_REFERENCED) {
1829 vm_page_flag_clear(m, PG_REFERENCED);
1832 actcount += pmap_ts_referenced(m);
1835 * Update act_count and move page to end of queue.
1838 m->act_count += ACT_ADVANCE + actcount;
1839 if (m->act_count > ACT_MAX)
1840 m->act_count = ACT_MAX;
1841 vm_page_and_queue_spin_lock(m);
1842 if (m->queue - m->pc == PQ_ACTIVE) {
1844 &vm_page_queues[PQ_ACTIVE + q].pl,
1847 &vm_page_queues[PQ_ACTIVE + q].pl,
1850 vm_page_and_queue_spin_unlock(m);
1855 if (m->act_count == 0) {
1857 * We turn off page access, so that we have
1858 * more accurate RSS stats. We don't do this
1859 * in the normal page deactivation when the
1860 * system is loaded VM wise, because the
1861 * cost of the large number of page protect
1862 * operations would be higher than the value
1863 * of doing the operation.
1865 * We use the marker to save our place so
1866 * we can release the spin lock. both (m)
1867 * and (next) will be invalid.
1869 vm_page_protect(m, VM_PROT_NONE);
1870 vm_page_deactivate(m);
1872 m->act_count -= min(m->act_count, ACT_DECLINE);
1873 vm_page_and_queue_spin_lock(m);
1874 if (m->queue - m->pc == PQ_ACTIVE) {
1876 &vm_page_queues[PQ_ACTIVE + q].pl,
1879 &vm_page_queues[PQ_ACTIVE + q].pl,
1882 vm_page_and_queue_spin_unlock(m);
1886 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1890 * Remove our local marker
1892 * Page queue still spin-locked.
1894 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1895 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1899 vm_pageout_free_page_calc(vm_size_t count)
1901 if (count < vmstats.v_page_count)
1904 * free_reserved needs to include enough for the largest swap pager
1905 * structures plus enough for any pv_entry structs when paging.
1907 * v_free_min normal allocations
1908 * v_free_reserved system allocations
1909 * v_pageout_free_min allocations by pageout daemon
1910 * v_interrupt_free_min low level allocations (e.g swap structures)
1912 if (vmstats.v_page_count > 1024)
1913 vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1915 vmstats.v_free_min = 64;
1918 * Make sure the vmmeter slop can't blow out our global minimums.
1920 * However, to accomodate weird configurations (vkernels with many
1921 * cpus and little memory, or artifically reduced hw.physmem), do
1922 * not allow v_free_min to exceed 1/20 of ram or the pageout demon
1923 * will go out of control.
1925 if (vmstats.v_free_min < VMMETER_SLOP_COUNT * ncpus * 10)
1926 vmstats.v_free_min = VMMETER_SLOP_COUNT * ncpus * 10;
1927 if (vmstats.v_free_min > vmstats.v_page_count / 20)
1928 vmstats.v_free_min = vmstats.v_page_count / 20;
1930 vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1931 vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1932 vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1933 vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1940 * vm_pageout is the high level pageout daemon. TWO kernel threads run
1941 * this daemon, the primary pageout daemon and the emergency pageout daemon.
1943 * The emergency pageout daemon takes over when the primary pageout daemon
1944 * deadlocks. The emergency pageout daemon ONLY pages out to swap, thus
1945 * avoiding the many low-memory deadlocks which can occur when paging out
1949 vm_pageout_thread(void)
1957 curthread->td_flags |= TDF_SYSTHREAD;
1960 * We only need to setup once.
1963 if (curthread == emergpager) {
1969 * Initialize some paging parameters.
1971 vm_pageout_free_page_calc(vmstats.v_page_count);
1974 * v_free_target and v_cache_min control pageout hysteresis. Note
1975 * that these are more a measure of the VM cache queue hysteresis
1976 * then the VM free queue. Specifically, v_free_target is the
1977 * high water mark (free+cache pages).
1979 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1980 * low water mark, while v_free_min is the stop. v_cache_min must
1981 * be big enough to handle memory needs while the pageout daemon
1982 * is signalled and run to free more pages.
1984 if (vmstats.v_free_count > 6144)
1985 vmstats.v_free_target = 4 * vmstats.v_free_min +
1986 vmstats.v_free_reserved;
1988 vmstats.v_free_target = 2 * vmstats.v_free_min +
1989 vmstats.v_free_reserved;
1992 * NOTE: With the new buffer cache b_act_count we want the default
1993 * inactive target to be a percentage of available memory.
1995 * The inactive target essentially determines the minimum
1996 * number of 'temporary' pages capable of caching one-time-use
1997 * files when the VM system is otherwise full of pages
1998 * belonging to multi-time-use files or active program data.
2000 * NOTE: The inactive target is aggressively persued only if the
2001 * inactive queue becomes too small. If the inactive queue
2002 * is large enough to satisfy page movement to free+cache
2003 * then it is repopulated more slowly from the active queue.
2004 * This allows a general inactive_target default to be set.
2006 * There is an issue here for processes which sit mostly idle
2007 * 'overnight', such as sshd, tcsh, and X. Any movement from
2008 * the active queue will eventually cause such pages to
2009 * recycle eventually causing a lot of paging in the morning.
2010 * To reduce the incidence of this pages cycled out of the
2011 * buffer cache are moved directly to the inactive queue if
2012 * they were only used once or twice.
2014 * The vfs.vm_cycle_point sysctl can be used to adjust this.
2015 * Increasing the value (up to 64) increases the number of
2016 * buffer recyclements which go directly to the inactive queue.
2018 if (vmstats.v_free_count > 2048) {
2019 vmstats.v_cache_min = vmstats.v_free_target;
2020 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
2022 vmstats.v_cache_min = 0;
2023 vmstats.v_cache_max = 0;
2025 vmstats.v_inactive_target = vmstats.v_free_count / 4;
2027 /* XXX does not really belong here */
2028 if (vm_page_max_wired == 0)
2029 vm_page_max_wired = vmstats.v_free_count / 3;
2031 if (vm_pageout_stats_max == 0)
2032 vm_pageout_stats_max = vmstats.v_free_target;
2035 * Set interval in seconds for stats scan.
2037 if (vm_pageout_stats_interval == 0)
2038 vm_pageout_stats_interval = 5;
2039 if (vm_pageout_full_stats_interval == 0)
2040 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
2044 * Set maximum free per pass
2046 if (vm_pageout_stats_free_max == 0)
2047 vm_pageout_stats_free_max = 5;
2049 swap_pager_swap_init();
2052 atomic_swap_int(&sequence_emerg_pager, 1);
2053 wakeup(&sequence_emerg_pager);
2057 * Sequence emergency pager startup
2060 while (sequence_emerg_pager == 0)
2061 tsleep(&sequence_emerg_pager, 0, "pstartup", hz);
2065 * The pageout daemon is never done, so loop forever.
2067 * WARNING! This code is being executed by two kernel threads
2068 * potentially simultaneously.
2072 long avail_shortage;
2073 long inactive_shortage;
2074 long vnodes_skipped = 0;
2075 long recycle_count = 0;
2079 * Wait for an action request. If we timeout check to
2080 * see if paging is needed (in case the normal wakeup
2085 * Emergency pagedaemon monitors the primary
2086 * pagedaemon while vm_pages_needed != 0.
2088 * The emergency pagedaemon only runs if VM paging
2089 * is needed and the primary pagedaemon has not
2090 * updated vm_pagedaemon_time for more than 2 seconds.
2092 if (vm_pages_needed)
2093 tsleep(&vm_pagedaemon_time, 0, "psleep", hz);
2095 tsleep(&vm_pagedaemon_time, 0, "psleep", hz*10);
2096 if (vm_pages_needed == 0) {
2100 if ((int)(ticks - vm_pagedaemon_time) < hz * 2) {
2106 * Primary pagedaemon
2108 if (vm_pages_needed == 0) {
2109 error = tsleep(&vm_pages_needed,
2111 vm_pageout_stats_interval * hz);
2113 vm_paging_needed() == 0 &&
2114 vm_pages_needed == 0) {
2115 for (q = 0; q < PQ_L2_SIZE; ++q)
2116 vm_pageout_page_stats(q);
2119 vm_pagedaemon_time = ticks;
2120 vm_pages_needed = 1;
2123 * Wake the emergency pagedaemon up so it
2124 * can monitor us. It will automatically
2125 * go back into a long sleep when
2126 * vm_pages_needed returns to 0.
2128 wakeup(&vm_pagedaemon_time);
2132 mycpu->gd_cnt.v_pdwakeups++;
2135 * Scan for INACTIVE->CLEAN/PAGEOUT
2137 * This routine tries to avoid thrashing the system with
2138 * unnecessary activity.
2140 * Calculate our target for the number of free+cache pages we
2141 * want to get to. This is higher then the number that causes
2142 * allocations to stall (severe) in order to provide hysteresis,
2143 * and if we don't make it all the way but get to the minimum
2144 * we're happy. Goose it a bit if there are multiple requests
2147 * Don't reduce avail_shortage inside the loop or the
2148 * PQAVERAGE() calculation will break.
2150 * NOTE! deficit is differentiated from avail_shortage as
2151 * REQUIRING at least (deficit) pages to be cleaned,
2152 * even if the page queues are in good shape. This
2153 * is used primarily for handling per-process
2154 * RLIMIT_RSS and may also see small values when
2155 * processes block due to low memory.
2159 vm_pagedaemon_time = ticks;
2160 avail_shortage = vm_paging_target() + vm_pageout_deficit;
2161 vm_pageout_deficit = 0;
2163 if (avail_shortage > 0) {
2168 for (q = 0; q < PQ_L2_SIZE; ++q) {
2169 delta += vm_pageout_scan_inactive(
2172 PQAVERAGE(avail_shortage),
2178 if (avail_shortage - delta <= 0)
2181 avail_shortage -= delta;
2186 * Figure out how many active pages we must deactivate. If
2187 * we were able to reach our target with just the inactive
2188 * scan above we limit the number of active pages we
2189 * deactivate to reduce unnecessary work.
2193 vm_pagedaemon_time = ticks;
2194 inactive_shortage = vmstats.v_inactive_target -
2195 vmstats.v_inactive_count;
2198 * If we were unable to free sufficient inactive pages to
2199 * satisfy the free/cache queue requirements then simply
2200 * reaching the inactive target may not be good enough.
2201 * Try to deactivate pages in excess of the target based
2204 * However to prevent thrashing the VM system do not
2205 * deactivate more than an additional 1/10 the inactive
2206 * target's worth of active pages.
2208 if (avail_shortage > 0) {
2209 tmp = avail_shortage * 2;
2210 if (tmp > vmstats.v_inactive_target / 10)
2211 tmp = vmstats.v_inactive_target / 10;
2212 inactive_shortage += tmp;
2216 * Only trigger a pmap cleanup on inactive shortage.
2218 if (isep == 0 && inactive_shortage > 0) {
2223 * Scan for ACTIVE->INACTIVE
2225 * Only trigger on inactive shortage. Triggering on
2226 * avail_shortage can starve the active queue with
2227 * unnecessary active->inactive transitions and destroy
2230 * If this is the emergency pager, always try to move
2231 * a few pages from active to inactive because the inactive
2232 * queue might have enough pages, but not enough anonymous
2235 if (isep && inactive_shortage < vm_emerg_launder)
2236 inactive_shortage = vm_emerg_launder;
2238 if (/*avail_shortage > 0 ||*/ inactive_shortage > 0) {
2243 for (q = 0; q < PQ_L2_SIZE; ++q) {
2244 delta += vm_pageout_scan_active(
2247 PQAVERAGE(avail_shortage),
2248 PQAVERAGE(inactive_shortage),
2254 if (inactive_shortage - delta <= 0 &&
2255 avail_shortage - delta <= 0) {
2259 inactive_shortage -= delta;
2260 avail_shortage -= delta;
2265 * Scan for CACHE->FREE
2267 * Finally free enough cache pages to meet our free page
2268 * requirement and take more drastic measures if we are
2273 vm_pagedaemon_time = ticks;
2274 vm_pageout_scan_cache(avail_shortage, pass,
2275 vnodes_skipped, recycle_count);
2278 * Wait for more work.
2280 if (avail_shortage > 0) {
2282 if (pass < 10 && vm_pages_needed > 1) {
2284 * Normal operation, additional processes
2285 * have already kicked us. Retry immediately
2286 * unless swap space is completely full in
2287 * which case delay a bit.
2289 if (swap_pager_full) {
2290 tsleep(&vm_pages_needed, 0, "pdelay",
2292 } /* else immediate retry */
2293 } else if (pass < 10) {
2295 * Normal operation, fewer processes. Delay
2296 * a bit but allow wakeups. vm_pages_needed
2297 * is only adjusted against the primary
2301 vm_pages_needed = 0;
2302 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2304 vm_pages_needed = 1;
2305 } else if (swap_pager_full == 0) {
2307 * We've taken too many passes, forced delay.
2309 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2312 * Running out of memory, catastrophic
2313 * back-off to one-second intervals.
2315 tsleep(&vm_pages_needed, 0, "pdelay", hz);
2317 } else if (vm_pages_needed) {
2319 * Interlocked wakeup of waiters (non-optional).
2321 * Similar to vm_page_free_wakeup() in vm_page.c,
2325 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2326 !vm_page_count_target()) {
2327 vm_pages_needed = 0;
2328 wakeup(&vmstats.v_free_count);
2336 static struct kproc_desc pg1_kp = {
2341 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &pg1_kp);
2343 static struct kproc_desc pg2_kp = {
2348 SYSINIT(emergpager, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, kproc_start, &pg2_kp);
2352 * Called after allocating a page out of the cache or free queue
2353 * to possibly wake the pagedaemon up to replentish our supply.
2355 * We try to generate some hysteresis by waking the pagedaemon up
2356 * when our free+cache pages go below the free_min+cache_min level.
2357 * The pagedaemon tries to get the count back up to at least the
2358 * minimum, and through to the target level if possible.
2360 * If the pagedaemon is already active bump vm_pages_needed as a hint
2361 * that there are even more requests pending.
2367 pagedaemon_wakeup(void)
2369 if (vm_paging_needed() && curthread != pagethread) {
2370 if (vm_pages_needed == 0) {
2371 vm_pages_needed = 1; /* SMP race ok */
2372 wakeup(&vm_pages_needed);
2373 } else if (vm_page_count_min(0)) {
2374 ++vm_pages_needed; /* SMP race ok */
2379 #if !defined(NO_SWAPPING)
2386 vm_req_vmdaemon(void)
2388 static int lastrun = 0;
2390 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2391 wakeup(&vm_daemon_needed);
2396 static int vm_daemon_callback(struct proc *p, void *data __unused);
2407 tsleep(&vm_daemon_needed, 0, "psleep", 0);
2408 req_swapout = atomic_swap_int(&vm_pageout_req_swapout, 0);
2414 swapout_procs(vm_pageout_req_swapout);
2417 * scan the processes for exceeding their rlimits or if
2418 * process is swapped out -- deactivate pages
2420 allproc_scan(vm_daemon_callback, NULL, 0);
2425 vm_daemon_callback(struct proc *p, void *data __unused)
2428 vm_pindex_t limit, size;
2431 * if this is a system process or if we have already
2432 * looked at this process, skip it.
2434 lwkt_gettoken(&p->p_token);
2436 if (p->p_flags & (P_SYSTEM | P_WEXIT)) {
2437 lwkt_reltoken(&p->p_token);
2442 * if the process is in a non-running type state,
2445 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
2446 lwkt_reltoken(&p->p_token);
2453 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2454 p->p_rlimit[RLIMIT_RSS].rlim_max));
2457 * let processes that are swapped out really be
2458 * swapped out. Set the limit to nothing to get as
2459 * many pages out to swap as possible.
2461 if (p->p_flags & P_SWAPPEDOUT)
2466 size = pmap_resident_tlnw_count(&vm->vm_pmap);
2467 if (limit >= 0 && size > 4096 &&
2468 size - 4096 >= limit && vm_pageout_memuse_mode >= 1) {
2469 vm_pageout_map_deactivate_pages(&vm->vm_map, limit);
2473 lwkt_reltoken(&p->p_token);