2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
64 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
68 * The proverbial page-out daemon.
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/kernel.h>
76 #include <sys/kthread.h>
77 #include <sys/resourcevar.h>
78 #include <sys/signalvar.h>
79 #include <sys/vnode.h>
80 #include <sys/vmmeter.h>
81 #include <sys/sysctl.h>
84 #include <vm/vm_param.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_pager.h>
91 #include <vm/swap_pager.h>
92 #include <vm/vm_extern.h>
94 #include <sys/thread2.h>
95 #include <sys/spinlock2.h>
96 #include <vm/vm_page2.h>
99 * System initialization
102 /* the kernel process "vm_pageout"*/
103 static int vm_pageout_clean (vm_page_t);
104 static int vm_pageout_free_page_calc (vm_size_t count);
105 struct thread *pagethread;
107 #if !defined(NO_SWAPPING)
108 /* the kernel process "vm_daemon"*/
109 static void vm_daemon (void);
110 static struct thread *vmthread;
112 static struct kproc_desc vm_kp = {
117 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
120 int vm_pages_needed=0; /* Event on which pageout daemon sleeps */
121 int vm_pageout_deficit=0; /* Estimated number of pages deficit */
122 int vm_pageout_pages_needed=0; /* pageout daemon needs pages */
123 int vm_page_free_hysteresis = 16;
125 #if !defined(NO_SWAPPING)
126 static int vm_pageout_req_swapout; /* XXX */
127 static int vm_daemon_needed;
129 static int vm_max_launder = 4096;
130 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
131 static int vm_pageout_full_stats_interval = 0;
132 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
133 static int defer_swap_pageouts=0;
134 static int disable_swap_pageouts=0;
135 static u_int vm_anonmem_decline = ACT_DECLINE;
136 static u_int vm_filemem_decline = ACT_DECLINE * 2;
138 #if defined(NO_SWAPPING)
139 static int vm_swap_enabled=0;
140 static int vm_swap_idle_enabled=0;
142 static int vm_swap_enabled=1;
143 static int vm_swap_idle_enabled=0;
146 SYSCTL_UINT(_vm, VM_PAGEOUT_ALGORITHM, anonmem_decline,
147 CTLFLAG_RW, &vm_anonmem_decline, 0, "active->inactive anon memory");
149 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, filemem_decline,
150 CTLFLAG_RW, &vm_filemem_decline, 0, "active->inactive file cache");
152 SYSCTL_INT(_vm, OID_AUTO, page_free_hysteresis,
153 CTLFLAG_RW, &vm_page_free_hysteresis, 0,
154 "Free more pages than the minimum required");
156 SYSCTL_INT(_vm, OID_AUTO, max_launder,
157 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
159 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
160 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
162 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
163 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
165 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
166 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
168 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
169 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
171 #if defined(NO_SWAPPING)
172 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
173 CTLFLAG_RD, &vm_swap_enabled, 0, "");
174 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
175 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
178 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
180 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
183 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
184 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
186 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
187 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
189 static int pageout_lock_miss;
190 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
191 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
193 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
195 #if !defined(NO_SWAPPING)
196 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
197 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
198 static freeer_fcn_t vm_pageout_object_deactivate_pages;
199 static void vm_req_vmdaemon (void);
201 static void vm_pageout_page_stats(int q);
204 * Calculate approximately how many pages on each queue to try to
205 * clean. An exact calculation creates an edge condition when the
206 * queues are unbalanced so add significant slop. The queue scans
207 * will stop early when targets are reached and will start where they
208 * left off on the next pass.
210 * We need to be generous here because there are all sorts of loading
211 * conditions that can cause edge cases if try to average over all queues.
212 * In particular, storage subsystems have become so fast that paging
213 * activity can become quite frantic. Eventually we will probably need
214 * two paging threads, one for dirty pages and one for clean, to deal
215 * with the bandwidth requirements.
217 * So what we do is calculate a value that can be satisfied nominally by
218 * only having to scan half the queues.
226 avg = ((n + (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) + 1);
228 avg = ((n - (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) - 1);
236 * Clean the page and remove it from the laundry. The page must not be
239 * We set the busy bit to cause potential page faults on this page to
240 * block. Note the careful timing, however, the busy bit isn't set till
241 * late and we cannot do anything that will mess with the page.
244 vm_pageout_clean(vm_page_t m)
247 vm_page_t mc[BLIST_MAX_ALLOC];
249 int ib, is, page_base;
250 vm_pindex_t pindex = m->pindex;
255 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
256 * with the new swapper, but we could have serious problems paging
257 * out other object types if there is insufficient memory.
259 * Unfortunately, checking free memory here is far too late, so the
260 * check has been moved up a procedural level.
264 * Don't mess with the page if it's busy, held, or special
266 * XXX do we really need to check hold_count here? hold_count
267 * isn't supposed to mess with vm_page ops except prevent the
268 * page from being reused.
270 if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
276 * Place page in cluster. Align cluster for optimal swap space
277 * allocation (whether it is swap or not). This is typically ~16-32
278 * pages, which also tends to align the cluster to multiples of the
279 * filesystem block size if backed by a filesystem.
281 page_base = pindex % BLIST_MAX_ALLOC;
287 * Scan object for clusterable pages.
289 * We can cluster ONLY if: ->> the page is NOT
290 * clean, wired, busy, held, or mapped into a
291 * buffer, and one of the following:
292 * 1) The page is inactive, or a seldom used
295 * 2) we force the issue.
297 * During heavy mmap/modification loads the pageout
298 * daemon can really fragment the underlying file
299 * due to flushing pages out of order and not trying
300 * align the clusters (which leave sporatic out-of-order
301 * holes). To solve this problem we do the reverse scan
302 * first and attempt to align our cluster, then do a
303 * forward scan if room remains.
306 vm_object_hold(object);
310 p = vm_page_lookup_busy_try(object, pindex - page_base + ib,
312 if (error || p == NULL)
314 if ((p->queue - p->pc) == PQ_CACHE ||
315 (p->flags & PG_UNMANAGED)) {
319 vm_page_test_dirty(p);
320 if (((p->dirty & p->valid) == 0 &&
321 (p->flags & PG_NEED_COMMIT) == 0) ||
322 p->queue - p->pc != PQ_INACTIVE ||
323 p->wire_count != 0 || /* may be held by buf cache */
324 p->hold_count != 0) { /* may be undergoing I/O */
333 while (is < BLIST_MAX_ALLOC &&
334 pindex - page_base + is < object->size) {
337 p = vm_page_lookup_busy_try(object, pindex - page_base + is,
339 if (error || p == NULL)
341 if (((p->queue - p->pc) == PQ_CACHE) ||
342 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
346 vm_page_test_dirty(p);
347 if (((p->dirty & p->valid) == 0 &&
348 (p->flags & PG_NEED_COMMIT) == 0) ||
349 p->queue - p->pc != PQ_INACTIVE ||
350 p->wire_count != 0 || /* may be held by buf cache */
351 p->hold_count != 0) { /* may be undergoing I/O */
359 vm_object_drop(object);
362 * we allow reads during pageouts...
364 return vm_pageout_flush(&mc[ib], is - ib, 0);
368 * vm_pageout_flush() - launder the given pages
370 * The given pages are laundered. Note that we setup for the start of
371 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
372 * reference count all in here rather then in the parent. If we want
373 * the parent to do more sophisticated things we may have to change
376 * The pages in the array must be busied by the caller and will be
377 * unbusied by this function.
380 vm_pageout_flush(vm_page_t *mc, int count, int flags)
383 int pageout_status[count];
388 * Initiate I/O. Bump the vm_page_t->busy counter.
390 for (i = 0; i < count; i++) {
391 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
392 ("vm_pageout_flush page %p index %d/%d: partially "
393 "invalid page", mc[i], i, count));
394 vm_page_io_start(mc[i]);
398 * We must make the pages read-only. This will also force the
399 * modified bit in the related pmaps to be cleared. The pager
400 * cannot clear the bit for us since the I/O completion code
401 * typically runs from an interrupt. The act of making the page
402 * read-only handles the case for us.
404 * Then we can unbusy the pages, we still hold a reference by virtue
407 for (i = 0; i < count; i++) {
408 vm_page_protect(mc[i], VM_PROT_READ);
409 vm_page_wakeup(mc[i]);
412 object = mc[0]->object;
413 vm_object_pip_add(object, count);
415 vm_pager_put_pages(object, mc, count,
416 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
419 for (i = 0; i < count; i++) {
420 vm_page_t mt = mc[i];
422 switch (pageout_status[i]) {
431 * Page outside of range of object. Right now we
432 * essentially lose the changes by pretending it
435 vm_page_busy_wait(mt, FALSE, "pgbad");
436 pmap_clear_modify(mt);
443 * A page typically cannot be paged out when we
444 * have run out of swap. We leave the page
445 * marked inactive and will try to page it out
448 * Starvation of the active page list is used to
449 * determine when the system is massively memory
458 * If the operation is still going, leave the page busy to
459 * block all other accesses. Also, leave the paging in
460 * progress indicator set so that we don't attempt an object
463 * For any pages which have completed synchronously,
464 * deactivate the page if we are under a severe deficit.
465 * Do not try to enter them into the cache, though, they
466 * might still be read-heavy.
468 if (pageout_status[i] != VM_PAGER_PEND) {
469 vm_page_busy_wait(mt, FALSE, "pgouw");
470 if (vm_page_count_severe())
471 vm_page_deactivate(mt);
473 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
474 vm_page_protect(mt, VM_PROT_READ);
476 vm_page_io_finish(mt);
478 vm_object_pip_wakeup(object);
484 #if !defined(NO_SWAPPING)
486 * deactivate enough pages to satisfy the inactive target
487 * requirements or if vm_page_proc_limit is set, then
488 * deactivate all of the pages in the object and its
491 * The map must be locked.
492 * The caller must hold the vm_object.
494 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
497 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
498 vm_pindex_t desired, int map_remove_only)
500 struct rb_vm_page_scan_info info;
505 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
509 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
511 if (lobject->type == OBJT_DEVICE ||
512 lobject->type == OBJT_MGTDEVICE ||
513 lobject->type == OBJT_PHYS)
515 if (lobject->paging_in_progress)
518 remove_mode = map_remove_only;
519 if (lobject->shadow_count > 1)
523 * scan the objects entire memory queue. We hold the
524 * object's token so the scan should not race anything.
526 info.limit = remove_mode;
528 info.desired = desired;
529 vm_page_rb_tree_RB_SCAN(&lobject->rb_memq, NULL,
530 vm_pageout_object_deactivate_pages_callback,
533 while ((tobject = lobject->backing_object) != NULL) {
534 KKASSERT(tobject != object);
535 vm_object_hold(tobject);
536 if (tobject == lobject->backing_object)
538 vm_object_drop(tobject);
540 if (lobject != object) {
542 vm_object_lock_swap();
543 vm_object_drop(lobject);
544 /* leaves tobject locked & at top */
548 if (lobject != object)
549 vm_object_drop(lobject); /* NULL ok */
553 * The caller must hold the vm_object.
556 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
558 struct rb_vm_page_scan_info *info = data;
561 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
564 mycpu->gd_cnt.v_pdpages++;
566 if (vm_page_busy_try(p, TRUE))
568 if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
572 if (!pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
577 actcount = pmap_ts_referenced(p);
579 vm_page_flag_set(p, PG_REFERENCED);
580 } else if (p->flags & PG_REFERENCED) {
584 vm_page_and_queue_spin_lock(p);
585 if (p->queue - p->pc != PQ_ACTIVE && (p->flags & PG_REFERENCED)) {
586 vm_page_and_queue_spin_unlock(p);
588 p->act_count += actcount;
589 vm_page_flag_clear(p, PG_REFERENCED);
590 } else if (p->queue - p->pc == PQ_ACTIVE) {
591 if ((p->flags & PG_REFERENCED) == 0) {
592 p->act_count -= min(p->act_count, ACT_DECLINE);
594 (vm_pageout_algorithm || (p->act_count == 0))) {
595 vm_page_and_queue_spin_unlock(p);
596 vm_page_protect(p, VM_PROT_NONE);
597 vm_page_deactivate(p);
599 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
601 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
603 vm_page_and_queue_spin_unlock(p);
606 vm_page_and_queue_spin_unlock(p);
608 vm_page_flag_clear(p, PG_REFERENCED);
610 vm_page_and_queue_spin_lock(p);
611 if (p->queue - p->pc == PQ_ACTIVE) {
612 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
613 p->act_count += ACT_ADVANCE;
614 TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
616 TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
619 vm_page_and_queue_spin_unlock(p);
621 } else if (p->queue - p->pc == PQ_INACTIVE) {
622 vm_page_and_queue_spin_unlock(p);
623 vm_page_protect(p, VM_PROT_NONE);
625 vm_page_and_queue_spin_unlock(p);
632 * Deactivate some number of pages in a map, try to do it fairly, but
633 * that is really hard to do.
636 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
639 vm_object_t obj, bigobj;
642 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
650 * first, search out the biggest object, and try to free pages from
653 tmpe = map->header.next;
654 while (tmpe != &map->header) {
655 switch(tmpe->maptype) {
656 case VM_MAPTYPE_NORMAL:
657 case VM_MAPTYPE_VPAGETABLE:
658 obj = tmpe->object.vm_object;
659 if ((obj != NULL) && (obj->shadow_count <= 1) &&
661 (bigobj->resident_page_count < obj->resident_page_count))) {
668 if (tmpe->wired_count > 0)
669 nothingwired = FALSE;
674 vm_object_hold(bigobj);
675 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
676 vm_object_drop(bigobj);
680 * Next, hunt around for other pages to deactivate. We actually
681 * do this search sort of wrong -- .text first is not the best idea.
683 tmpe = map->header.next;
684 while (tmpe != &map->header) {
685 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
687 switch(tmpe->maptype) {
688 case VM_MAPTYPE_NORMAL:
689 case VM_MAPTYPE_VPAGETABLE:
690 obj = tmpe->object.vm_object;
693 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
704 * Remove all mappings if a process is swapped out, this will free page
707 if (desired == 0 && nothingwired)
708 pmap_remove(vm_map_pmap(map),
709 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
715 * Called when the pageout scan wants to free a page. We no longer
716 * try to cycle the vm_object here with a reference & dealloc, which can
717 * cause a non-trivial object collapse in a critical path.
719 * It is unclear why we cycled the ref_count in the past, perhaps to try
720 * to optimize shadow chain collapses but I don't quite see why it would
721 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
722 * synchronously and not have to be kicked-start.
725 vm_pageout_page_free(vm_page_t m)
727 vm_page_protect(m, VM_PROT_NONE);
732 * vm_pageout_scan does the dirty work for the pageout daemon.
734 struct vm_pageout_scan_info {
735 struct proc *bigproc;
739 static int vm_pageout_scan_callback(struct proc *p, void *data);
742 vm_pageout_scan_inactive(int pass, int q, int avail_shortage,
743 int *vnodes_skippedp)
746 struct vm_page marker;
747 struct vnode *vpfailed; /* warning, allowed to be stale */
756 * Start scanning the inactive queue for pages we can move to the
757 * cache or free. The scan will stop when the target is reached or
758 * we have scanned the entire inactive queue. Note that m->act_count
759 * is not used to form decisions for the inactive queue, only for the
762 * maxlaunder limits the number of dirty pages we flush per scan.
763 * For most systems a smaller value (16 or 32) is more robust under
764 * extreme memory and disk pressure because any unnecessary writes
765 * to disk can result in extreme performance degredation. However,
766 * systems with excessive dirty pages (especially when MAP_NOSYNC is
767 * used) will die horribly with limited laundering. If the pageout
768 * daemon cannot clean enough pages in the first pass, we let it go
769 * all out in succeeding passes.
771 if ((maxlaunder = vm_max_launder) <= 1)
777 * Initialize our marker
779 bzero(&marker, sizeof(marker));
780 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
781 marker.queue = PQ_INACTIVE + q;
783 marker.wire_count = 1;
786 * Inactive queue scan.
788 * NOTE: The vm_page must be spinlocked before the queue to avoid
789 * deadlocks, so it is easiest to simply iterate the loop
790 * with the queue unlocked at the top.
794 vm_page_queues_spin_lock(PQ_INACTIVE + q);
795 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
796 maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt;
799 * Queue locked at top of loop to avoid stack marker issues.
801 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
802 maxscan-- > 0 && avail_shortage - delta > 0)
804 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
805 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
807 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
809 mycpu->gd_cnt.v_pdpages++;
812 * Skip marker pages (atomic against other markers to avoid
813 * infinite hop-over scans).
815 if (m->flags & PG_MARKER)
819 * Try to busy the page. Don't mess with pages which are
820 * already busy or reorder them in the queue.
822 if (vm_page_busy_try(m, TRUE))
826 * Remaining operations run with the page busy and neither
827 * the page or the queue will be spin-locked.
829 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
830 KKASSERT(m->queue - m->pc == PQ_INACTIVE);
834 * It is possible for a page to be busied ad-hoc (e.g. the
835 * pmap_collect() code) and wired and race against the
836 * allocation of a new page. vm_page_alloc() may be forced
837 * to deactivate the wired page in which case it winds up
838 * on the inactive queue and must be handled here. We
839 * correct the problem simply by unqueuing the page.
842 vm_page_unqueue_nowakeup(m);
844 kprintf("WARNING: pagedaemon: wired page on "
845 "inactive queue %p\n", m);
850 * A held page may be undergoing I/O, so skip it.
853 vm_page_and_queue_spin_lock(m);
854 if (m->queue - m->pc == PQ_INACTIVE) {
856 &vm_page_queues[PQ_INACTIVE + q].pl,
859 &vm_page_queues[PQ_INACTIVE + q].pl,
861 ++vm_swapcache_inactive_heuristic;
863 vm_page_and_queue_spin_unlock(m);
868 if (m->object == NULL || m->object->ref_count == 0) {
870 * If the object is not being used, we ignore previous
873 vm_page_flag_clear(m, PG_REFERENCED);
874 pmap_clear_reference(m);
875 /* fall through to end */
876 } else if (((m->flags & PG_REFERENCED) == 0) &&
877 (actcount = pmap_ts_referenced(m))) {
879 * Otherwise, if the page has been referenced while
880 * in the inactive queue, we bump the "activation
881 * count" upwards, making it less likely that the
882 * page will be added back to the inactive queue
883 * prematurely again. Here we check the page tables
884 * (or emulated bits, if any), given the upper level
885 * VM system not knowing anything about existing
889 m->act_count += (actcount + ACT_ADVANCE);
895 * (m) is still busied.
897 * If the upper level VM system knows about any page
898 * references, we activate the page. We also set the
899 * "activation count" higher than normal so that we will less
900 * likely place pages back onto the inactive queue again.
902 if ((m->flags & PG_REFERENCED) != 0) {
903 vm_page_flag_clear(m, PG_REFERENCED);
904 actcount = pmap_ts_referenced(m);
906 m->act_count += (actcount + ACT_ADVANCE + 1);
912 * If the upper level VM system doesn't know anything about
913 * the page being dirty, we have to check for it again. As
914 * far as the VM code knows, any partially dirty pages are
917 * Pages marked PG_WRITEABLE may be mapped into the user
918 * address space of a process running on another cpu. A
919 * user process (without holding the MP lock) running on
920 * another cpu may be able to touch the page while we are
921 * trying to remove it. vm_page_cache() will handle this
925 vm_page_test_dirty(m);
930 if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
932 * Invalid pages can be easily freed
934 vm_pageout_page_free(m);
935 mycpu->gd_cnt.v_dfree++;
937 } else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
939 * Clean pages can be placed onto the cache queue.
940 * This effectively frees them.
944 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
946 * Dirty pages need to be paged out, but flushing
947 * a page is extremely expensive verses freeing
948 * a clean page. Rather then artificially limiting
949 * the number of pages we can flush, we instead give
950 * dirty pages extra priority on the inactive queue
951 * by forcing them to be cycled through the queue
952 * twice before being flushed, after which the
953 * (now clean) page will cycle through once more
954 * before being freed. This significantly extends
955 * the thrash point for a heavily loaded machine.
957 vm_page_flag_set(m, PG_WINATCFLS);
958 vm_page_and_queue_spin_lock(m);
959 if (m->queue - m->pc == PQ_INACTIVE) {
961 &vm_page_queues[PQ_INACTIVE + q].pl,
964 &vm_page_queues[PQ_INACTIVE + q].pl,
966 ++vm_swapcache_inactive_heuristic;
968 vm_page_and_queue_spin_unlock(m);
970 } else if (maxlaunder > 0) {
972 * We always want to try to flush some dirty pages if
973 * we encounter them, to keep the system stable.
974 * Normally this number is small, but under extreme
975 * pressure where there are insufficient clean pages
976 * on the inactive queue, we may have to go all out.
978 int swap_pageouts_ok;
979 struct vnode *vp = NULL;
981 swap_pageouts_ok = 0;
984 (object->type != OBJT_SWAP) &&
985 (object->type != OBJT_DEFAULT)) {
986 swap_pageouts_ok = 1;
988 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
989 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
990 vm_page_count_min(0));
995 * We don't bother paging objects that are "dead".
996 * Those objects are in a "rundown" state.
998 if (!swap_pageouts_ok ||
1000 (object->flags & OBJ_DEAD)) {
1001 vm_page_and_queue_spin_lock(m);
1002 if (m->queue - m->pc == PQ_INACTIVE) {
1004 &vm_page_queues[PQ_INACTIVE + q].pl,
1007 &vm_page_queues[PQ_INACTIVE + q].pl,
1009 ++vm_swapcache_inactive_heuristic;
1011 vm_page_and_queue_spin_unlock(m);
1017 * (m) is still busied.
1019 * The object is already known NOT to be dead. It
1020 * is possible for the vget() to block the whole
1021 * pageout daemon, but the new low-memory handling
1022 * code should prevent it.
1024 * The previous code skipped locked vnodes and, worse,
1025 * reordered pages in the queue. This results in
1026 * completely non-deterministic operation because,
1027 * quite often, a vm_fault has initiated an I/O and
1028 * is holding a locked vnode at just the point where
1029 * the pageout daemon is woken up.
1031 * We can't wait forever for the vnode lock, we might
1032 * deadlock due to a vn_read() getting stuck in
1033 * vm_wait while holding this vnode. We skip the
1034 * vnode if we can't get it in a reasonable amount
1037 * vpfailed is used to (try to) avoid the case where
1038 * a large number of pages are associated with a
1039 * locked vnode, which could cause the pageout daemon
1040 * to stall for an excessive amount of time.
1042 if (object->type == OBJT_VNODE) {
1045 vp = object->handle;
1046 flags = LK_EXCLUSIVE;
1050 flags |= LK_TIMELOCK;
1055 * We have unbusied (m) temporarily so we can
1056 * acquire the vp lock without deadlocking.
1057 * (m) is held to prevent destruction.
1059 if (vget(vp, flags) != 0) {
1061 ++pageout_lock_miss;
1062 if (object->flags & OBJ_MIGHTBEDIRTY)
1069 * The page might have been moved to another
1070 * queue during potential blocking in vget()
1071 * above. The page might have been freed and
1072 * reused for another vnode. The object might
1073 * have been reused for another vnode.
1075 if (m->queue - m->pc != PQ_INACTIVE ||
1076 m->object != object ||
1077 object->handle != vp) {
1078 if (object->flags & OBJ_MIGHTBEDIRTY)
1086 * The page may have been busied during the
1087 * blocking in vput(); We don't move the
1088 * page back onto the end of the queue so that
1089 * statistics are more correct if we don't.
1091 if (vm_page_busy_try(m, TRUE)) {
1099 * (m) is busied again
1101 * We own the busy bit and remove our hold
1102 * bit. If the page is still held it
1103 * might be undergoing I/O, so skip it.
1105 if (m->hold_count) {
1106 vm_page_and_queue_spin_lock(m);
1107 if (m->queue - m->pc == PQ_INACTIVE) {
1108 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1109 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1110 ++vm_swapcache_inactive_heuristic;
1112 vm_page_and_queue_spin_unlock(m);
1113 if (object->flags & OBJ_MIGHTBEDIRTY)
1119 /* (m) is left busied as we fall through */
1123 * page is busy and not held here.
1125 * If a page is dirty, then it is either being washed
1126 * (but not yet cleaned) or it is still in the
1127 * laundry. If it is still in the laundry, then we
1128 * start the cleaning operation.
1130 * decrement inactive_shortage on success to account
1131 * for the (future) cleaned page. Otherwise we
1132 * could wind up laundering or cleaning too many
1135 count = vm_pageout_clean(m);
1137 maxlaunder -= count;
1140 * Clean ate busy, page no longer accessible
1150 * Systems with a ton of memory can wind up with huge
1151 * deactivation counts. Because the inactive scan is
1152 * doing a lot of flushing, the combination can result
1153 * in excessive paging even in situations where other
1154 * unrelated threads free up sufficient VM.
1156 * To deal with this we abort the nominal active->inactive
1157 * scan before we hit the inactive target when free+cache
1158 * levels have reached a reasonable target.
1160 * When deciding to stop early we need to add some slop to
1161 * the test and we need to return full completion to the caller
1162 * to prevent the caller from thinking there is something
1163 * wrong and issuing a low-memory+swap warning or pkill.
1165 vm_page_queues_spin_lock(PQ_INACTIVE + q);
1166 if (vm_paging_target() < -vm_max_launder) {
1168 * Stopping early, return full completion to caller.
1170 if (delta < avail_shortage)
1171 delta = avail_shortage;
1176 /* page queue still spin-locked */
1177 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
1178 vm_page_queues_spin_unlock(PQ_INACTIVE + q);
1184 vm_pageout_scan_active(int pass, int q,
1185 int avail_shortage, int inactive_shortage,
1186 int *recycle_countp)
1188 struct vm_page marker;
1195 * We want to move pages from the active queue to the inactive
1196 * queue to get the inactive queue to the inactive target. If
1197 * we still have a page shortage from above we try to directly free
1198 * clean pages instead of moving them.
1200 * If we do still have a shortage we keep track of the number of
1201 * pages we free or cache (recycle_count) as a measure of thrashing
1202 * between the active and inactive queues.
1204 * If we were able to completely satisfy the free+cache targets
1205 * from the inactive pool we limit the number of pages we move
1206 * from the active pool to the inactive pool to 2x the pages we
1207 * had removed from the inactive pool (with a minimum of 1/5 the
1208 * inactive target). If we were not able to completely satisfy
1209 * the free+cache targets we go for the whole target aggressively.
1211 * NOTE: Both variables can end up negative.
1212 * NOTE: We are still in a critical section.
1215 bzero(&marker, sizeof(marker));
1216 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1217 marker.queue = PQ_ACTIVE + q;
1219 marker.wire_count = 1;
1221 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1222 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1223 maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt;
1226 * Queue locked at top of loop to avoid stack marker issues.
1228 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1229 maxscan-- > 0 && (avail_shortage - delta > 0 ||
1230 inactive_shortage > 0))
1232 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1233 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1235 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1239 * Skip marker pages (atomic against other markers to avoid
1240 * infinite hop-over scans).
1242 if (m->flags & PG_MARKER)
1246 * Try to busy the page. Don't mess with pages which are
1247 * already busy or reorder them in the queue.
1249 if (vm_page_busy_try(m, TRUE))
1253 * Remaining operations run with the page busy and neither
1254 * the page or the queue will be spin-locked.
1256 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1257 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1261 * Don't deactivate pages that are held, even if we can
1262 * busy them. (XXX why not?)
1264 if (m->hold_count != 0) {
1265 vm_page_and_queue_spin_lock(m);
1266 if (m->queue - m->pc == PQ_ACTIVE) {
1268 &vm_page_queues[PQ_ACTIVE + q].pl,
1271 &vm_page_queues[PQ_ACTIVE + q].pl,
1274 vm_page_and_queue_spin_unlock(m);
1280 * The count for pagedaemon pages is done after checking the
1281 * page for eligibility...
1283 mycpu->gd_cnt.v_pdpages++;
1286 * Check to see "how much" the page has been used and clear
1287 * the tracking access bits. If the object has no references
1288 * don't bother paying the expense.
1291 if (m->object && m->object->ref_count != 0) {
1292 if (m->flags & PG_REFERENCED)
1294 actcount += pmap_ts_referenced(m);
1296 m->act_count += ACT_ADVANCE + actcount;
1297 if (m->act_count > ACT_MAX)
1298 m->act_count = ACT_MAX;
1301 vm_page_flag_clear(m, PG_REFERENCED);
1304 * actcount is only valid if the object ref_count is non-zero.
1305 * If the page does not have an object, actcount will be zero.
1307 if (actcount && m->object->ref_count != 0) {
1308 vm_page_and_queue_spin_lock(m);
1309 if (m->queue - m->pc == PQ_ACTIVE) {
1311 &vm_page_queues[PQ_ACTIVE + q].pl,
1314 &vm_page_queues[PQ_ACTIVE + q].pl,
1317 vm_page_and_queue_spin_unlock(m);
1320 switch(m->object->type) {
1323 m->act_count -= min(m->act_count,
1324 vm_anonmem_decline);
1327 m->act_count -= min(m->act_count,
1328 vm_filemem_decline);
1331 if (vm_pageout_algorithm ||
1332 (m->object == NULL) ||
1333 (m->object && (m->object->ref_count == 0)) ||
1334 m->act_count < pass + 1
1337 * Deactivate the page. If we had a
1338 * shortage from our inactive scan try to
1339 * free (cache) the page instead.
1341 * Don't just blindly cache the page if
1342 * we do not have a shortage from the
1343 * inactive scan, that could lead to
1344 * gigabytes being moved.
1346 --inactive_shortage;
1347 if (avail_shortage - delta > 0 ||
1348 (m->object && (m->object->ref_count == 0)))
1350 if (avail_shortage - delta > 0)
1352 vm_page_protect(m, VM_PROT_NONE);
1353 if (m->dirty == 0 &&
1354 (m->flags & PG_NEED_COMMIT) == 0 &&
1355 avail_shortage - delta > 0) {
1358 vm_page_deactivate(m);
1362 vm_page_deactivate(m);
1367 vm_page_and_queue_spin_lock(m);
1368 if (m->queue - m->pc == PQ_ACTIVE) {
1370 &vm_page_queues[PQ_ACTIVE + q].pl,
1373 &vm_page_queues[PQ_ACTIVE + q].pl,
1376 vm_page_and_queue_spin_unlock(m);
1381 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1385 * Clean out our local marker.
1387 * Page queue still spin-locked.
1389 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1390 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1396 * The number of actually free pages can drop down to v_free_reserved,
1397 * we try to build the free count back above v_free_min. Note that
1398 * vm_paging_needed() also returns TRUE if v_free_count is not at
1399 * least v_free_min so that is the minimum we must build the free
1402 * We use a slightly higher target to improve hysteresis,
1403 * ((v_free_target + v_free_min) / 2). Since v_free_target
1404 * is usually the same as v_cache_min this maintains about
1405 * half the pages in the free queue as are in the cache queue,
1406 * providing pretty good pipelining for pageout operation.
1408 * The system operator can manipulate vm.v_cache_min and
1409 * vm.v_free_target to tune the pageout demon. Be sure
1410 * to keep vm.v_free_min < vm.v_free_target.
1412 * Note that the original paging target is to get at least
1413 * (free_min + cache_min) into (free + cache). The slightly
1414 * higher target will shift additional pages from cache to free
1415 * without effecting the original paging target in order to
1416 * maintain better hysteresis and not have the free count always
1417 * be dead-on v_free_min.
1419 * NOTE: we are still in a critical section.
1421 * Pages moved from PQ_CACHE to totally free are not counted in the
1422 * pages_freed counter.
1425 vm_pageout_scan_cache(int avail_shortage, int pass,
1426 int vnodes_skipped, int recycle_count)
1428 static int lastkillticks;
1429 struct vm_pageout_scan_info info;
1432 while (vmstats.v_free_count <
1433 (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1435 * This steals some code from vm/vm_page.c
1437 static int cache_rover = 0;
1439 m = vm_page_list_find(PQ_CACHE,
1440 cache_rover & PQ_L2_MASK, FALSE);
1443 /* page is returned removed from its queue and spinlocked */
1444 if (vm_page_busy_try(m, TRUE)) {
1445 vm_page_deactivate_locked(m);
1446 vm_page_spin_unlock(m);
1449 vm_page_spin_unlock(m);
1450 pagedaemon_wakeup();
1454 * Remaining operations run with the page busy and neither
1455 * the page or the queue will be spin-locked.
1457 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1460 vm_page_deactivate(m);
1464 KKASSERT((m->flags & PG_MAPPED) == 0);
1465 KKASSERT(m->dirty == 0);
1466 cache_rover += PQ_PRIME2;
1467 vm_pageout_page_free(m);
1468 mycpu->gd_cnt.v_dfree++;
1471 #if !defined(NO_SWAPPING)
1473 * Idle process swapout -- run once per second.
1475 if (vm_swap_idle_enabled) {
1477 if (time_uptime != lsec) {
1478 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1486 * If we didn't get enough free pages, and we have skipped a vnode
1487 * in a writeable object, wakeup the sync daemon. And kick swapout
1488 * if we did not get enough free pages.
1490 if (vm_paging_target() > 0) {
1491 if (vnodes_skipped && vm_page_count_min(0))
1492 speedup_syncer(NULL);
1493 #if !defined(NO_SWAPPING)
1494 if (vm_swap_enabled && vm_page_count_target()) {
1496 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1502 * Handle catastrophic conditions. Under good conditions we should
1503 * be at the target, well beyond our minimum. If we could not even
1504 * reach our minimum the system is under heavy stress. But just being
1505 * under heavy stress does not trigger process killing.
1507 * We consider ourselves to have run out of memory if the swap pager
1508 * is full and avail_shortage is still positive. The secondary check
1509 * ensures that we do not kill processes if the instantanious
1510 * availability is good, even if the pageout demon pass says it
1511 * couldn't get to the target.
1513 if (swap_pager_almost_full &&
1515 (vm_page_count_min(recycle_count) || avail_shortage > 0)) {
1516 kprintf("Warning: system low on memory+swap "
1517 "shortage %d for %d ticks!\n",
1518 avail_shortage, ticks - swap_fail_ticks);
1520 if (swap_pager_full &&
1522 avail_shortage > 0 &&
1523 vm_paging_target() > 0 &&
1524 (unsigned int)(ticks - lastkillticks) >= hz) {
1526 * Kill something, maximum rate once per second to give
1527 * the process time to free up sufficient memory.
1529 lastkillticks = ticks;
1530 info.bigproc = NULL;
1532 allproc_scan(vm_pageout_scan_callback, &info);
1533 if (info.bigproc != NULL) {
1534 info.bigproc->p_nice = PRIO_MIN;
1535 info.bigproc->p_usched->resetpriority(
1536 FIRST_LWP_IN_PROC(info.bigproc));
1537 killproc(info.bigproc, "out of swap space");
1538 wakeup(&vmstats.v_free_count);
1539 PRELE(info.bigproc);
1545 vm_pageout_scan_callback(struct proc *p, void *data)
1547 struct vm_pageout_scan_info *info = data;
1551 * Never kill system processes or init. If we have configured swap
1552 * then try to avoid killing low-numbered pids.
1554 if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1555 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1559 lwkt_gettoken(&p->p_token);
1562 * if the process is in a non-running type state,
1565 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
1566 lwkt_reltoken(&p->p_token);
1571 * Get the approximate process size. Note that anonymous pages
1572 * with backing swap will be counted twice, but there should not
1573 * be too many such pages due to the stress the VM system is
1574 * under at this point.
1576 size = vmspace_anonymous_count(p->p_vmspace) +
1577 vmspace_swap_count(p->p_vmspace);
1580 * If the this process is bigger than the biggest one
1583 if (info->bigsize < size) {
1585 PRELE(info->bigproc);
1588 info->bigsize = size;
1590 lwkt_reltoken(&p->p_token);
1597 * This routine tries to maintain the pseudo LRU active queue,
1598 * so that during long periods of time where there is no paging,
1599 * that some statistic accumulation still occurs. This code
1600 * helps the situation where paging just starts to occur.
1603 vm_pageout_page_stats(int q)
1605 static int fullintervalcount = 0;
1606 struct vm_page marker;
1608 int pcount, tpcount; /* Number of pages to check */
1611 page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1612 vmstats.v_free_min) -
1613 (vmstats.v_free_count + vmstats.v_inactive_count +
1614 vmstats.v_cache_count);
1616 if (page_shortage <= 0)
1619 pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1620 fullintervalcount += vm_pageout_stats_interval;
1621 if (fullintervalcount < vm_pageout_full_stats_interval) {
1622 tpcount = (vm_pageout_stats_max * pcount) /
1623 vmstats.v_page_count + 1;
1624 if (pcount > tpcount)
1627 fullintervalcount = 0;
1630 bzero(&marker, sizeof(marker));
1631 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1632 marker.queue = PQ_ACTIVE + q;
1634 marker.wire_count = 1;
1636 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1637 TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1640 * Queue locked at top of loop to avoid stack marker issues.
1642 while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1647 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1648 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1649 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1653 * Skip marker pages (atomic against other markers to avoid
1654 * infinite hop-over scans).
1656 if (m->flags & PG_MARKER)
1660 * Ignore pages we can't busy
1662 if (vm_page_busy_try(m, TRUE))
1666 * Remaining operations run with the page busy and neither
1667 * the page or the queue will be spin-locked.
1669 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1670 KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1673 * We now have a safely busied page, the page and queue
1674 * spinlocks have been released.
1678 if (m->hold_count) {
1684 * Calculate activity
1687 if (m->flags & PG_REFERENCED) {
1688 vm_page_flag_clear(m, PG_REFERENCED);
1691 actcount += pmap_ts_referenced(m);
1694 * Update act_count and move page to end of queue.
1697 m->act_count += ACT_ADVANCE + actcount;
1698 if (m->act_count > ACT_MAX)
1699 m->act_count = ACT_MAX;
1700 vm_page_and_queue_spin_lock(m);
1701 if (m->queue - m->pc == PQ_ACTIVE) {
1703 &vm_page_queues[PQ_ACTIVE + q].pl,
1706 &vm_page_queues[PQ_ACTIVE + q].pl,
1709 vm_page_and_queue_spin_unlock(m);
1714 if (m->act_count == 0) {
1716 * We turn off page access, so that we have
1717 * more accurate RSS stats. We don't do this
1718 * in the normal page deactivation when the
1719 * system is loaded VM wise, because the
1720 * cost of the large number of page protect
1721 * operations would be higher than the value
1722 * of doing the operation.
1724 * We use the marker to save our place so
1725 * we can release the spin lock. both (m)
1726 * and (next) will be invalid.
1728 vm_page_protect(m, VM_PROT_NONE);
1729 vm_page_deactivate(m);
1731 m->act_count -= min(m->act_count, ACT_DECLINE);
1732 vm_page_and_queue_spin_lock(m);
1733 if (m->queue - m->pc == PQ_ACTIVE) {
1735 &vm_page_queues[PQ_ACTIVE + q].pl,
1738 &vm_page_queues[PQ_ACTIVE + q].pl,
1741 vm_page_and_queue_spin_unlock(m);
1745 vm_page_queues_spin_lock(PQ_ACTIVE + q);
1749 * Remove our local marker
1751 * Page queue still spin-locked.
1753 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1754 vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1758 vm_pageout_free_page_calc(vm_size_t count)
1760 if (count < vmstats.v_page_count)
1763 * free_reserved needs to include enough for the largest swap pager
1764 * structures plus enough for any pv_entry structs when paging.
1766 * v_free_min normal allocations
1767 * v_free_reserved system allocations
1768 * v_pageout_free_min allocations by pageout daemon
1769 * v_interrupt_free_min low level allocations (e.g swap structures)
1771 if (vmstats.v_page_count > 1024)
1772 vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1774 vmstats.v_free_min = 64;
1775 vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1776 vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1777 vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1778 vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1785 * vm_pageout is the high level pageout daemon.
1790 vm_pageout_thread(void)
1798 * Initialize some paging parameters.
1800 curthread->td_flags |= TDF_SYSTHREAD;
1802 vm_pageout_free_page_calc(vmstats.v_page_count);
1805 * v_free_target and v_cache_min control pageout hysteresis. Note
1806 * that these are more a measure of the VM cache queue hysteresis
1807 * then the VM free queue. Specifically, v_free_target is the
1808 * high water mark (free+cache pages).
1810 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1811 * low water mark, while v_free_min is the stop. v_cache_min must
1812 * be big enough to handle memory needs while the pageout daemon
1813 * is signalled and run to free more pages.
1815 if (vmstats.v_free_count > 6144)
1816 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1818 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1821 * NOTE: With the new buffer cache b_act_count we want the default
1822 * inactive target to be a percentage of available memory.
1824 * The inactive target essentially determines the minimum
1825 * number of 'temporary' pages capable of caching one-time-use
1826 * files when the VM system is otherwise full of pages
1827 * belonging to multi-time-use files or active program data.
1829 * NOTE: The inactive target is aggressively persued only if the
1830 * inactive queue becomes too small. If the inactive queue
1831 * is large enough to satisfy page movement to free+cache
1832 * then it is repopulated more slowly from the active queue.
1833 * This allows a general inactive_target default to be set.
1835 * There is an issue here for processes which sit mostly idle
1836 * 'overnight', such as sshd, tcsh, and X. Any movement from
1837 * the active queue will eventually cause such pages to
1838 * recycle eventually causing a lot of paging in the morning.
1839 * To reduce the incidence of this pages cycled out of the
1840 * buffer cache are moved directly to the inactive queue if
1841 * they were only used once or twice.
1843 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1844 * Increasing the value (up to 64) increases the number of
1845 * buffer recyclements which go directly to the inactive queue.
1847 if (vmstats.v_free_count > 2048) {
1848 vmstats.v_cache_min = vmstats.v_free_target;
1849 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1851 vmstats.v_cache_min = 0;
1852 vmstats.v_cache_max = 0;
1854 vmstats.v_inactive_target = vmstats.v_free_count / 4;
1856 /* XXX does not really belong here */
1857 if (vm_page_max_wired == 0)
1858 vm_page_max_wired = vmstats.v_free_count / 3;
1860 if (vm_pageout_stats_max == 0)
1861 vm_pageout_stats_max = vmstats.v_free_target;
1864 * Set interval in seconds for stats scan.
1866 if (vm_pageout_stats_interval == 0)
1867 vm_pageout_stats_interval = 5;
1868 if (vm_pageout_full_stats_interval == 0)
1869 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1873 * Set maximum free per pass
1875 if (vm_pageout_stats_free_max == 0)
1876 vm_pageout_stats_free_max = 5;
1878 swap_pager_swap_init();
1882 * The pageout daemon is never done, so loop forever.
1887 int inactive_shortage;
1888 int vnodes_skipped = 0;
1889 int recycle_count = 0;
1893 * Wait for an action request. If we timeout check to
1894 * see if paging is needed (in case the normal wakeup
1897 if (vm_pages_needed == 0) {
1898 error = tsleep(&vm_pages_needed,
1900 vm_pageout_stats_interval * hz);
1902 vm_paging_needed() == 0 &&
1903 vm_pages_needed == 0) {
1904 for (q = 0; q < PQ_L2_SIZE; ++q)
1905 vm_pageout_page_stats(q);
1908 vm_pages_needed = 1;
1911 mycpu->gd_cnt.v_pdwakeups++;
1914 * Do whatever cleanup that the pmap code can.
1919 * Scan for pageout. Try to avoid thrashing the system
1922 * Calculate our target for the number of free+cache pages we
1923 * want to get to. This is higher then the number that causes
1924 * allocations to stall (severe) in order to provide hysteresis,
1925 * and if we don't make it all the way but get to the minimum
1926 * we're happy. Goose it a bit if there are multiple requests
1929 * Don't reduce avail_shortage inside the loop or the
1930 * PQAVERAGE() calculation will break.
1932 avail_shortage = vm_paging_target() + vm_pageout_deficit;
1933 vm_pageout_deficit = 0;
1935 if (avail_shortage > 0) {
1938 for (q = 0; q < PQ_L2_SIZE; ++q) {
1939 delta += vm_pageout_scan_inactive(
1941 (q + q1iterator) & PQ_L2_MASK,
1942 PQAVERAGE(avail_shortage),
1944 if (avail_shortage - delta <= 0)
1947 avail_shortage -= delta;
1952 * Figure out how many active pages we must deactivate. If
1953 * we were able to reach our target with just the inactive
1954 * scan above we limit the number of active pages we
1955 * deactivate to reduce unnecessary work.
1957 inactive_shortage = vmstats.v_inactive_target -
1958 vmstats.v_inactive_count;
1961 * If we were unable to free sufficient inactive pages to
1962 * satisfy the free/cache queue requirements then simply
1963 * reaching the inactive target may not be good enough.
1964 * Try to deactivate pages in excess of the target based
1967 * However to prevent thrashing the VM system do not
1968 * deactivate more than an additional 1/10 the inactive
1969 * target's worth of active pages.
1971 if (avail_shortage > 0) {
1972 tmp = avail_shortage * 2;
1973 if (tmp > vmstats.v_inactive_target / 10)
1974 tmp = vmstats.v_inactive_target / 10;
1975 inactive_shortage += tmp;
1979 * Only trigger on inactive shortage. Triggering on
1980 * avail_shortage can starve the active queue with
1981 * unnecessary active->inactive transitions and destroy
1984 if (/*avail_shortage > 0 ||*/ inactive_shortage > 0) {
1987 for (q = 0; q < PQ_L2_SIZE; ++q) {
1988 delta += vm_pageout_scan_active(
1990 (q + q2iterator) & PQ_L2_MASK,
1991 PQAVERAGE(avail_shortage),
1992 PQAVERAGE(inactive_shortage),
1994 if (inactive_shortage - delta <= 0 &&
1995 avail_shortage - delta <= 0) {
1999 inactive_shortage -= delta;
2000 avail_shortage -= delta;
2005 * Finally free enough cache pages to meet our free page
2006 * requirement and take more drastic measures if we are
2009 vm_pageout_scan_cache(avail_shortage, pass,
2010 vnodes_skipped, recycle_count);
2013 * Wait for more work.
2015 if (avail_shortage > 0) {
2017 if (pass < 10 && vm_pages_needed > 1) {
2019 * Normal operation, additional processes
2020 * have already kicked us. Retry immediately
2021 * unless swap space is completely full in
2022 * which case delay a bit.
2024 if (swap_pager_full) {
2025 tsleep(&vm_pages_needed, 0, "pdelay",
2027 } /* else immediate retry */
2028 } else if (pass < 10) {
2030 * Normal operation, fewer processes. Delay
2031 * a bit but allow wakeups.
2033 vm_pages_needed = 0;
2034 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2035 vm_pages_needed = 1;
2036 } else if (swap_pager_full == 0) {
2038 * We've taken too many passes, forced delay.
2040 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2043 * Running out of memory, catastrophic
2044 * back-off to one-second intervals.
2046 tsleep(&vm_pages_needed, 0, "pdelay", hz);
2048 } else if (vm_pages_needed) {
2050 * Interlocked wakeup of waiters (non-optional).
2052 * Similar to vm_page_free_wakeup() in vm_page.c,
2056 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2057 !vm_page_count_target()) {
2058 vm_pages_needed = 0;
2059 wakeup(&vmstats.v_free_count);
2067 static struct kproc_desc page_kp = {
2072 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp);
2076 * Called after allocating a page out of the cache or free queue
2077 * to possibly wake the pagedaemon up to replentish our supply.
2079 * We try to generate some hysteresis by waking the pagedaemon up
2080 * when our free+cache pages go below the free_min+cache_min level.
2081 * The pagedaemon tries to get the count back up to at least the
2082 * minimum, and through to the target level if possible.
2084 * If the pagedaemon is already active bump vm_pages_needed as a hint
2085 * that there are even more requests pending.
2091 pagedaemon_wakeup(void)
2093 if (vm_paging_needed() && curthread != pagethread) {
2094 if (vm_pages_needed == 0) {
2095 vm_pages_needed = 1; /* SMP race ok */
2096 wakeup(&vm_pages_needed);
2097 } else if (vm_page_count_min(0)) {
2098 ++vm_pages_needed; /* SMP race ok */
2103 #if !defined(NO_SWAPPING)
2110 vm_req_vmdaemon(void)
2112 static int lastrun = 0;
2114 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2115 wakeup(&vm_daemon_needed);
2120 static int vm_daemon_callback(struct proc *p, void *data __unused);
2129 * XXX vm_daemon_needed specific token?
2132 tsleep(&vm_daemon_needed, 0, "psleep", 0);
2133 if (vm_pageout_req_swapout) {
2134 swapout_procs(vm_pageout_req_swapout);
2135 vm_pageout_req_swapout = 0;
2138 * scan the processes for exceeding their rlimits or if
2139 * process is swapped out -- deactivate pages
2141 allproc_scan(vm_daemon_callback, NULL);
2146 vm_daemon_callback(struct proc *p, void *data __unused)
2149 vm_pindex_t limit, size;
2152 * if this is a system process or if we have already
2153 * looked at this process, skip it.
2155 lwkt_gettoken(&p->p_token);
2157 if (p->p_flags & (P_SYSTEM | P_WEXIT)) {
2158 lwkt_reltoken(&p->p_token);
2163 * if the process is in a non-running type state,
2166 if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
2167 lwkt_reltoken(&p->p_token);
2174 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2175 p->p_rlimit[RLIMIT_RSS].rlim_max));
2178 * let processes that are swapped out really be
2179 * swapped out. Set the limit to nothing to get as
2180 * many pages out to swap as possible.
2182 if (p->p_flags & P_SWAPPEDOUT)
2187 size = vmspace_resident_count(vm);
2188 if (limit >= 0 && size >= limit) {
2189 vm_pageout_map_deactivate_pages(&vm->vm_map, limit);
2193 lwkt_reltoken(&p->p_token);