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 * 4. 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; /* flag saying that the pageout daemon needs pages */
124 #if !defined(NO_SWAPPING)
125 static int vm_pageout_req_swapout
; /* XXX */
126 static int vm_daemon_needed
;
128 static int vm_max_launder
= 32;
129 static int vm_pageout_stats_max
=0, vm_pageout_stats_interval
= 0;
130 static int vm_pageout_full_stats_interval
= 0;
131 static int vm_pageout_stats_free_max
=0, vm_pageout_algorithm
=0;
132 static int defer_swap_pageouts
=0;
133 static int disable_swap_pageouts
=0;
135 #if defined(NO_SWAPPING)
136 static int vm_swap_enabled
=0;
137 static int vm_swap_idle_enabled
=0;
139 static int vm_swap_enabled
=1;
140 static int vm_swap_idle_enabled
=0;
143 SYSCTL_INT(_vm
, VM_PAGEOUT_ALGORITHM
, pageout_algorithm
,
144 CTLFLAG_RW
, &vm_pageout_algorithm
, 0, "LRU page mgmt");
146 SYSCTL_INT(_vm
, OID_AUTO
, max_launder
,
147 CTLFLAG_RW
, &vm_max_launder
, 0, "Limit dirty flushes in pageout");
149 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_max
,
150 CTLFLAG_RW
, &vm_pageout_stats_max
, 0, "Max pageout stats scan length");
152 SYSCTL_INT(_vm
, OID_AUTO
, pageout_full_stats_interval
,
153 CTLFLAG_RW
, &vm_pageout_full_stats_interval
, 0, "Interval for full stats scan");
155 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_interval
,
156 CTLFLAG_RW
, &vm_pageout_stats_interval
, 0, "Interval for partial stats scan");
158 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_free_max
,
159 CTLFLAG_RW
, &vm_pageout_stats_free_max
, 0, "Not implemented");
161 #if defined(NO_SWAPPING)
162 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
163 CTLFLAG_RD
, &vm_swap_enabled
, 0, "");
164 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
165 CTLFLAG_RD
, &vm_swap_idle_enabled
, 0, "");
167 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
168 CTLFLAG_RW
, &vm_swap_enabled
, 0, "Enable entire process swapout");
169 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
170 CTLFLAG_RW
, &vm_swap_idle_enabled
, 0, "Allow swapout on idle criteria");
173 SYSCTL_INT(_vm
, OID_AUTO
, defer_swapspace_pageouts
,
174 CTLFLAG_RW
, &defer_swap_pageouts
, 0, "Give preference to dirty pages in mem");
176 SYSCTL_INT(_vm
, OID_AUTO
, disable_swapspace_pageouts
,
177 CTLFLAG_RW
, &disable_swap_pageouts
, 0, "Disallow swapout of dirty pages");
179 static int pageout_lock_miss
;
180 SYSCTL_INT(_vm
, OID_AUTO
, pageout_lock_miss
,
181 CTLFLAG_RD
, &pageout_lock_miss
, 0, "vget() lock misses during pageout");
183 #define VM_PAGEOUT_PAGE_COUNT 16
184 int vm_pageout_page_count
= VM_PAGEOUT_PAGE_COUNT
;
186 int vm_page_max_wired
; /* XXX max # of wired pages system-wide */
188 #if !defined(NO_SWAPPING)
189 typedef void freeer_fcn_t (vm_map_t
, vm_object_t
, vm_pindex_t
, int);
190 static void vm_pageout_map_deactivate_pages (vm_map_t
, vm_pindex_t
);
191 static freeer_fcn_t vm_pageout_object_deactivate_pages
;
192 static void vm_req_vmdaemon (void);
194 static void vm_pageout_page_stats(int q
);
200 return((n
+ (PQ_L2_SIZE
- 1)) / PQ_L2_SIZE
+ 1);
202 return((n
- (PQ_L2_SIZE
- 1)) / PQ_L2_SIZE
- 1);
208 * Clean the page and remove it from the laundry. The page must not be
211 * We set the busy bit to cause potential page faults on this page to
212 * block. Note the careful timing, however, the busy bit isn't set till
213 * late and we cannot do anything that will mess with the page.
216 vm_pageout_clean(vm_page_t m
)
219 vm_page_t mc
[2*vm_pageout_page_count
];
222 int ib
, is
, page_base
;
223 vm_pindex_t pindex
= m
->pindex
;
228 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
229 * with the new swapper, but we could have serious problems paging
230 * out other object types if there is insufficient memory.
232 * Unfortunately, checking free memory here is far too late, so the
233 * check has been moved up a procedural level.
237 * Don't mess with the page if it's busy, held, or special
239 * XXX do we really need to check hold_count here? hold_count
240 * isn't supposed to mess with vm_page ops except prevent the
241 * page from being reused.
243 if (m
->hold_count
!= 0 || (m
->flags
& PG_UNMANAGED
)) {
248 mc
[vm_pageout_page_count
] = m
;
250 page_base
= vm_pageout_page_count
;
255 * Scan object for clusterable pages.
257 * We can cluster ONLY if: ->> the page is NOT
258 * clean, wired, busy, held, or mapped into a
259 * buffer, and one of the following:
260 * 1) The page is inactive, or a seldom used
263 * 2) we force the issue.
265 * During heavy mmap/modification loads the pageout
266 * daemon can really fragment the underlying file
267 * due to flushing pages out of order and not trying
268 * align the clusters (which leave sporatic out-of-order
269 * holes). To solve this problem we do the reverse scan
270 * first and attempt to align our cluster, then do a
271 * forward scan if room remains.
274 vm_object_hold(object
);
276 while (ib
&& pageout_count
< vm_pageout_page_count
) {
284 p
= vm_page_lookup_busy_try(object
, pindex
- ib
, TRUE
, &error
);
285 if (error
|| p
== NULL
) {
289 if ((p
->queue
- p
->pc
) == PQ_CACHE
||
290 (p
->flags
& PG_UNMANAGED
)) {
295 vm_page_test_dirty(p
);
296 if (((p
->dirty
& p
->valid
) == 0 &&
297 (p
->flags
& PG_NEED_COMMIT
) == 0) ||
298 p
->queue
- p
->pc
!= PQ_INACTIVE
||
299 p
->wire_count
!= 0 || /* may be held by buf cache */
300 p
->hold_count
!= 0) { /* may be undergoing I/O */
309 * alignment boundry, stop here and switch directions. Do
312 if ((pindex
- (ib
- 1)) % vm_pageout_page_count
== 0)
316 while (pageout_count
< vm_pageout_page_count
&&
317 pindex
+ is
< object
->size
) {
320 p
= vm_page_lookup_busy_try(object
, pindex
+ is
, TRUE
, &error
);
321 if (error
|| p
== NULL
)
323 if (((p
->queue
- p
->pc
) == PQ_CACHE
) ||
324 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) || p
->busy
) {
328 vm_page_test_dirty(p
);
329 if (((p
->dirty
& p
->valid
) == 0 &&
330 (p
->flags
& PG_NEED_COMMIT
) == 0) ||
331 p
->queue
- p
->pc
!= PQ_INACTIVE
||
332 p
->wire_count
!= 0 || /* may be held by buf cache */
333 p
->hold_count
!= 0) { /* may be undergoing I/O */
337 mc
[page_base
+ pageout_count
] = p
;
343 * If we exhausted our forward scan, continue with the reverse scan
344 * when possible, even past a page boundry. This catches boundry
347 if (ib
&& pageout_count
< vm_pageout_page_count
)
350 vm_object_drop(object
);
353 * we allow reads during pageouts...
355 return vm_pageout_flush(&mc
[page_base
], pageout_count
, 0);
359 * vm_pageout_flush() - launder the given pages
361 * The given pages are laundered. Note that we setup for the start of
362 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
363 * reference count all in here rather then in the parent. If we want
364 * the parent to do more sophisticated things we may have to change
367 * The pages in the array must be busied by the caller and will be
368 * unbusied by this function.
371 vm_pageout_flush(vm_page_t
*mc
, int count
, int flags
)
374 int pageout_status
[count
];
379 * Initiate I/O. Bump the vm_page_t->busy counter.
381 for (i
= 0; i
< count
; i
++) {
382 KASSERT(mc
[i
]->valid
== VM_PAGE_BITS_ALL
,
383 ("vm_pageout_flush page %p index %d/%d: partially "
384 "invalid page", mc
[i
], i
, count
));
385 vm_page_io_start(mc
[i
]);
389 * We must make the pages read-only. This will also force the
390 * modified bit in the related pmaps to be cleared. The pager
391 * cannot clear the bit for us since the I/O completion code
392 * typically runs from an interrupt. The act of making the page
393 * read-only handles the case for us.
395 * Then we can unbusy the pages, we still hold a reference by virtue
398 for (i
= 0; i
< count
; i
++) {
399 vm_page_protect(mc
[i
], VM_PROT_READ
);
400 vm_page_wakeup(mc
[i
]);
403 object
= mc
[0]->object
;
404 vm_object_pip_add(object
, count
);
406 vm_pager_put_pages(object
, mc
, count
,
407 (flags
| ((object
== &kernel_object
) ? VM_PAGER_PUT_SYNC
: 0)),
410 for (i
= 0; i
< count
; i
++) {
411 vm_page_t mt
= mc
[i
];
413 switch (pageout_status
[i
]) {
422 * Page outside of range of object. Right now we
423 * essentially lose the changes by pretending it
426 vm_page_busy_wait(mt
, FALSE
, "pgbad");
427 pmap_clear_modify(mt
);
434 * A page typically cannot be paged out when we
435 * have run out of swap. We leave the page
436 * marked inactive and will try to page it out
439 * Starvation of the active page list is used to
440 * determine when the system is massively memory
449 * If the operation is still going, leave the page busy to
450 * block all other accesses. Also, leave the paging in
451 * progress indicator set so that we don't attempt an object
454 * For any pages which have completed synchronously,
455 * deactivate the page if we are under a severe deficit.
456 * Do not try to enter them into the cache, though, they
457 * might still be read-heavy.
459 if (pageout_status
[i
] != VM_PAGER_PEND
) {
460 vm_page_busy_wait(mt
, FALSE
, "pgouw");
461 if (vm_page_count_severe())
462 vm_page_deactivate(mt
);
464 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt
))
465 vm_page_protect(mt
, VM_PROT_READ
);
467 vm_page_io_finish(mt
);
469 vm_object_pip_wakeup(object
);
475 #if !defined(NO_SWAPPING)
477 * deactivate enough pages to satisfy the inactive target
478 * requirements or if vm_page_proc_limit is set, then
479 * deactivate all of the pages in the object and its
482 * The map must be locked.
483 * The caller must hold the vm_object.
485 static int vm_pageout_object_deactivate_pages_callback(vm_page_t
, void *);
488 vm_pageout_object_deactivate_pages(vm_map_t map
, vm_object_t object
,
489 vm_pindex_t desired
, int map_remove_only
)
491 struct rb_vm_page_scan_info info
;
496 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
500 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
502 if (lobject
->type
== OBJT_DEVICE
|| lobject
->type
== OBJT_PHYS
)
504 if (lobject
->paging_in_progress
)
507 remove_mode
= map_remove_only
;
508 if (lobject
->shadow_count
> 1)
512 * scan the objects entire memory queue. We hold the
513 * object's token so the scan should not race anything.
515 info
.limit
= remove_mode
;
517 info
.desired
= desired
;
518 vm_page_rb_tree_RB_SCAN(&lobject
->rb_memq
, NULL
,
519 vm_pageout_object_deactivate_pages_callback
,
522 while ((tobject
= lobject
->backing_object
) != NULL
) {
523 KKASSERT(tobject
!= object
);
524 vm_object_hold(tobject
);
525 if (tobject
== lobject
->backing_object
)
527 vm_object_drop(tobject
);
529 if (lobject
!= object
) {
530 vm_object_lock_swap();
531 vm_object_drop(lobject
);
535 if (lobject
!= object
)
536 vm_object_drop(lobject
);
540 * The caller must hold the vm_object.
543 vm_pageout_object_deactivate_pages_callback(vm_page_t p
, void *data
)
545 struct rb_vm_page_scan_info
*info
= data
;
548 if (pmap_resident_count(vm_map_pmap(info
->map
)) <= info
->desired
) {
551 mycpu
->gd_cnt
.v_pdpages
++;
553 if (vm_page_busy_try(p
, TRUE
))
555 if (p
->wire_count
|| p
->hold_count
|| (p
->flags
& PG_UNMANAGED
)) {
559 if (!pmap_page_exists_quick(vm_map_pmap(info
->map
), p
)) {
564 actcount
= pmap_ts_referenced(p
);
566 vm_page_flag_set(p
, PG_REFERENCED
);
567 } else if (p
->flags
& PG_REFERENCED
) {
571 vm_page_and_queue_spin_lock(p
);
572 if (p
->queue
- p
->pc
!= PQ_ACTIVE
&& (p
->flags
& PG_REFERENCED
)) {
573 vm_page_and_queue_spin_unlock(p
);
575 p
->act_count
+= actcount
;
576 vm_page_flag_clear(p
, PG_REFERENCED
);
577 } else if (p
->queue
- p
->pc
== PQ_ACTIVE
) {
578 if ((p
->flags
& PG_REFERENCED
) == 0) {
579 p
->act_count
-= min(p
->act_count
, ACT_DECLINE
);
581 (vm_pageout_algorithm
|| (p
->act_count
== 0))) {
582 vm_page_and_queue_spin_unlock(p
);
583 vm_page_protect(p
, VM_PROT_NONE
);
584 vm_page_deactivate(p
);
586 TAILQ_REMOVE(&vm_page_queues
[p
->queue
].pl
,
588 TAILQ_INSERT_TAIL(&vm_page_queues
[p
->queue
].pl
,
590 vm_page_and_queue_spin_unlock(p
);
593 vm_page_and_queue_spin_unlock(p
);
595 vm_page_flag_clear(p
, PG_REFERENCED
);
597 vm_page_and_queue_spin_lock(p
);
598 if (p
->queue
- p
->pc
== PQ_ACTIVE
) {
599 if (p
->act_count
< (ACT_MAX
- ACT_ADVANCE
))
600 p
->act_count
+= ACT_ADVANCE
;
601 TAILQ_REMOVE(&vm_page_queues
[p
->queue
].pl
,
603 TAILQ_INSERT_TAIL(&vm_page_queues
[p
->queue
].pl
,
606 vm_page_and_queue_spin_unlock(p
);
608 } else if (p
->queue
- p
->pc
== PQ_INACTIVE
) {
609 vm_page_and_queue_spin_unlock(p
);
610 vm_page_protect(p
, VM_PROT_NONE
);
612 vm_page_and_queue_spin_unlock(p
);
619 * Deactivate some number of pages in a map, try to do it fairly, but
620 * that is really hard to do.
623 vm_pageout_map_deactivate_pages(vm_map_t map
, vm_pindex_t desired
)
626 vm_object_t obj
, bigobj
;
629 if (lockmgr(&map
->lock
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
637 * first, search out the biggest object, and try to free pages from
640 tmpe
= map
->header
.next
;
641 while (tmpe
!= &map
->header
) {
642 switch(tmpe
->maptype
) {
643 case VM_MAPTYPE_NORMAL
:
644 case VM_MAPTYPE_VPAGETABLE
:
645 obj
= tmpe
->object
.vm_object
;
646 if ((obj
!= NULL
) && (obj
->shadow_count
<= 1) &&
648 (bigobj
->resident_page_count
< obj
->resident_page_count
))) {
655 if (tmpe
->wired_count
> 0)
656 nothingwired
= FALSE
;
661 vm_object_hold(bigobj
);
662 vm_pageout_object_deactivate_pages(map
, bigobj
, desired
, 0);
663 vm_object_drop(bigobj
);
667 * Next, hunt around for other pages to deactivate. We actually
668 * do this search sort of wrong -- .text first is not the best idea.
670 tmpe
= map
->header
.next
;
671 while (tmpe
!= &map
->header
) {
672 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
674 switch(tmpe
->maptype
) {
675 case VM_MAPTYPE_NORMAL
:
676 case VM_MAPTYPE_VPAGETABLE
:
677 obj
= tmpe
->object
.vm_object
;
680 vm_pageout_object_deactivate_pages(map
, obj
, desired
, 0);
691 * Remove all mappings if a process is swapped out, this will free page
694 if (desired
== 0 && nothingwired
)
695 pmap_remove(vm_map_pmap(map
),
696 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
702 * Called when the pageout scan wants to free a page. We no longer
703 * try to cycle the vm_object here with a reference & dealloc, which can
704 * cause a non-trivial object collapse in a critical path.
706 * It is unclear why we cycled the ref_count in the past, perhaps to try
707 * to optimize shadow chain collapses but I don't quite see why it would
708 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
709 * synchronously and not have to be kicked-start.
712 vm_pageout_page_free(vm_page_t m
)
714 vm_page_protect(m
, VM_PROT_NONE
);
719 * vm_pageout_scan does the dirty work for the pageout daemon.
721 struct vm_pageout_scan_info
{
722 struct proc
*bigproc
;
726 static int vm_pageout_scan_callback(struct proc
*p
, void *data
);
729 vm_pageout_scan_inactive(int pass
, int q
, int avail_shortage
,
730 int *vnodes_skippedp
)
733 struct vm_page marker
;
734 struct vnode
*vpfailed
; /* warning, allowed to be stale */
742 * Start scanning the inactive queue for pages we can move to the
743 * cache or free. The scan will stop when the target is reached or
744 * we have scanned the entire inactive queue. Note that m->act_count
745 * is not used to form decisions for the inactive queue, only for the
748 * maxlaunder limits the number of dirty pages we flush per scan.
749 * For most systems a smaller value (16 or 32) is more robust under
750 * extreme memory and disk pressure because any unnecessary writes
751 * to disk can result in extreme performance degredation. However,
752 * systems with excessive dirty pages (especially when MAP_NOSYNC is
753 * used) will die horribly with limited laundering. If the pageout
754 * daemon cannot clean enough pages in the first pass, we let it go
755 * all out in succeeding passes.
757 if ((maxlaunder
= vm_max_launder
) <= 1)
763 * Initialize our marker
765 bzero(&marker
, sizeof(marker
));
766 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
767 marker
.queue
= PQ_INACTIVE
+ q
;
769 marker
.wire_count
= 1;
772 * Inactive queue scan.
774 * NOTE: The vm_page must be spinlocked before the queue to avoid
775 * deadlocks, so it is easiest to simply iterate the loop
776 * with the queue unlocked at the top.
780 vm_page_queues_spin_lock(PQ_INACTIVE
+ q
);
781 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, &marker
, pageq
);
782 maxscan
= vm_page_queues
[PQ_INACTIVE
+ q
].lcnt
;
783 vm_page_queues_spin_unlock(PQ_INACTIVE
+ q
);
785 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
786 maxscan
-- > 0 && avail_shortage
- delta
> 0)
788 vm_page_and_queue_spin_lock(m
);
789 if (m
!= TAILQ_NEXT(&marker
, pageq
)) {
790 vm_page_and_queue_spin_unlock(m
);
794 KKASSERT(m
->queue
- m
->pc
== PQ_INACTIVE
);
795 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
797 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, m
,
799 mycpu
->gd_cnt
.v_pdpages
++;
804 if (m
->flags
& PG_MARKER
) {
805 vm_page_and_queue_spin_unlock(m
);
810 * Try to busy the page. Don't mess with pages which are
811 * already busy or reorder them in the queue.
813 if (vm_page_busy_try(m
, TRUE
)) {
814 vm_page_and_queue_spin_unlock(m
);
817 vm_page_and_queue_spin_unlock(m
);
818 KKASSERT(m
->queue
- m
->pc
== PQ_INACTIVE
);
823 * The page has been successfully busied and is now no
824 * longer spinlocked. The queue is no longer spinlocked
829 * It is possible for a page to be busied ad-hoc (e.g. the
830 * pmap_collect() code) and wired and race against the
831 * allocation of a new page. vm_page_alloc() may be forced
832 * to deactivate the wired page in which case it winds up
833 * on the inactive queue and must be handled here. We
834 * correct the problem simply by unqueuing the page.
837 vm_page_unqueue_nowakeup(m
);
839 kprintf("WARNING: pagedaemon: wired page on "
840 "inactive queue %p\n", m
);
845 * A held page may be undergoing I/O, so skip it.
848 vm_page_and_queue_spin_lock(m
);
849 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
851 &vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
854 &vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
856 ++vm_swapcache_inactive_heuristic
;
858 vm_page_and_queue_spin_unlock(m
);
863 if (m
->object
== NULL
|| m
->object
->ref_count
== 0) {
865 * If the object is not being used, we ignore previous
868 vm_page_flag_clear(m
, PG_REFERENCED
);
869 pmap_clear_reference(m
);
870 /* fall through to end */
871 } else if (((m
->flags
& PG_REFERENCED
) == 0) &&
872 (actcount
= pmap_ts_referenced(m
))) {
874 * Otherwise, if the page has been referenced while
875 * in the inactive queue, we bump the "activation
876 * count" upwards, making it less likely that the
877 * page will be added back to the inactive queue
878 * prematurely again. Here we check the page tables
879 * (or emulated bits, if any), given the upper level
880 * VM system not knowing anything about existing
884 m
->act_count
+= (actcount
+ ACT_ADVANCE
);
890 * (m) is still busied.
892 * If the upper level VM system knows about any page
893 * references, we activate the page. We also set the
894 * "activation count" higher than normal so that we will less
895 * likely place pages back onto the inactive queue again.
897 if ((m
->flags
& PG_REFERENCED
) != 0) {
898 vm_page_flag_clear(m
, PG_REFERENCED
);
899 actcount
= pmap_ts_referenced(m
);
901 m
->act_count
+= (actcount
+ ACT_ADVANCE
+ 1);
907 * If the upper level VM system doesn't know anything about
908 * the page being dirty, we have to check for it again. As
909 * far as the VM code knows, any partially dirty pages are
912 * Pages marked PG_WRITEABLE may be mapped into the user
913 * address space of a process running on another cpu. A
914 * user process (without holding the MP lock) running on
915 * another cpu may be able to touch the page while we are
916 * trying to remove it. vm_page_cache() will handle this
920 vm_page_test_dirty(m
);
925 if (m
->valid
== 0 && (m
->flags
& PG_NEED_COMMIT
) == 0) {
927 * Invalid pages can be easily freed
929 vm_pageout_page_free(m
);
930 mycpu
->gd_cnt
.v_dfree
++;
932 } else if (m
->dirty
== 0 && (m
->flags
& PG_NEED_COMMIT
) == 0) {
934 * Clean pages can be placed onto the cache queue.
935 * This effectively frees them.
939 } else if ((m
->flags
& PG_WINATCFLS
) == 0 && pass
== 0) {
941 * Dirty pages need to be paged out, but flushing
942 * a page is extremely expensive verses freeing
943 * a clean page. Rather then artificially limiting
944 * the number of pages we can flush, we instead give
945 * dirty pages extra priority on the inactive queue
946 * by forcing them to be cycled through the queue
947 * twice before being flushed, after which the
948 * (now clean) page will cycle through once more
949 * before being freed. This significantly extends
950 * the thrash point for a heavily loaded machine.
952 vm_page_flag_set(m
, PG_WINATCFLS
);
953 vm_page_and_queue_spin_lock(m
);
954 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
956 &vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
959 &vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
961 ++vm_swapcache_inactive_heuristic
;
963 vm_page_and_queue_spin_unlock(m
);
965 } else if (maxlaunder
> 0) {
967 * We always want to try to flush some dirty pages if
968 * we encounter them, to keep the system stable.
969 * Normally this number is small, but under extreme
970 * pressure where there are insufficient clean pages
971 * on the inactive queue, we may have to go all out.
973 int swap_pageouts_ok
;
974 struct vnode
*vp
= NULL
;
976 swap_pageouts_ok
= 0;
979 (object
->type
!= OBJT_SWAP
) &&
980 (object
->type
!= OBJT_DEFAULT
)) {
981 swap_pageouts_ok
= 1;
983 swap_pageouts_ok
= !(defer_swap_pageouts
|| disable_swap_pageouts
);
984 swap_pageouts_ok
|= (!disable_swap_pageouts
&& defer_swap_pageouts
&&
985 vm_page_count_min(0));
990 * We don't bother paging objects that are "dead".
991 * Those objects are in a "rundown" state.
993 if (!swap_pageouts_ok
||
995 (object
->flags
& OBJ_DEAD
)) {
996 vm_page_and_queue_spin_lock(m
);
997 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
999 &vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
1002 &vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
1004 ++vm_swapcache_inactive_heuristic
;
1006 vm_page_and_queue_spin_unlock(m
);
1012 * (m) is still busied.
1014 * The object is already known NOT to be dead. It
1015 * is possible for the vget() to block the whole
1016 * pageout daemon, but the new low-memory handling
1017 * code should prevent it.
1019 * The previous code skipped locked vnodes and, worse,
1020 * reordered pages in the queue. This results in
1021 * completely non-deterministic operation because,
1022 * quite often, a vm_fault has initiated an I/O and
1023 * is holding a locked vnode at just the point where
1024 * the pageout daemon is woken up.
1026 * We can't wait forever for the vnode lock, we might
1027 * deadlock due to a vn_read() getting stuck in
1028 * vm_wait while holding this vnode. We skip the
1029 * vnode if we can't get it in a reasonable amount
1032 * vpfailed is used to (try to) avoid the case where
1033 * a large number of pages are associated with a
1034 * locked vnode, which could cause the pageout daemon
1035 * to stall for an excessive amount of time.
1037 if (object
->type
== OBJT_VNODE
) {
1040 vp
= object
->handle
;
1041 flags
= LK_EXCLUSIVE
| LK_NOOBJ
;
1045 flags
|= LK_TIMELOCK
;
1050 * We have unbusied (m) temporarily so we can
1051 * acquire the vp lock without deadlocking.
1052 * (m) is held to prevent destruction.
1054 if (vget(vp
, flags
) != 0) {
1056 ++pageout_lock_miss
;
1057 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1064 * The page might have been moved to another
1065 * queue during potential blocking in vget()
1066 * above. The page might have been freed and
1067 * reused for another vnode. The object might
1068 * have been reused for another vnode.
1070 if (m
->queue
- m
->pc
!= PQ_INACTIVE
||
1071 m
->object
!= object
||
1072 object
->handle
!= vp
) {
1073 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1081 * The page may have been busied during the
1082 * blocking in vput(); We don't move the
1083 * page back onto the end of the queue so that
1084 * statistics are more correct if we don't.
1086 if (vm_page_busy_try(m
, TRUE
)) {
1094 * (m) is busied again
1096 * We own the busy bit and remove our hold
1097 * bit. If the page is still held it
1098 * might be undergoing I/O, so skip it.
1100 if (m
->hold_count
) {
1101 vm_page_and_queue_spin_lock(m
);
1102 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
1103 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, m
, pageq
);
1104 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, m
, pageq
);
1105 ++vm_swapcache_inactive_heuristic
;
1107 vm_page_and_queue_spin_unlock(m
);
1108 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1114 /* (m) is left busied as we fall through */
1118 * page is busy and not held here.
1120 * If a page is dirty, then it is either being washed
1121 * (but not yet cleaned) or it is still in the
1122 * laundry. If it is still in the laundry, then we
1123 * start the cleaning operation.
1125 * decrement inactive_shortage on success to account
1126 * for the (future) cleaned page. Otherwise we
1127 * could wind up laundering or cleaning too many
1130 if (vm_pageout_clean(m
) != 0) {
1134 /* clean ate busy, page no longer accessible */
1141 vm_page_queues_spin_lock(PQ_INACTIVE
+ q
);
1142 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, &marker
, pageq
);
1143 vm_page_queues_spin_unlock(PQ_INACTIVE
+ q
);
1148 vm_pageout_scan_active(int pass
, int q
,
1149 int avail_shortage
, int inactive_shortage
,
1150 int *recycle_countp
)
1152 struct vm_page marker
;
1159 * We want to move pages from the active queue to the inactive
1160 * queue to get the inactive queue to the inactive target. If
1161 * we still have a page shortage from above we try to directly free
1162 * clean pages instead of moving them.
1164 * If we do still have a shortage we keep track of the number of
1165 * pages we free or cache (recycle_count) as a measure of thrashing
1166 * between the active and inactive queues.
1168 * If we were able to completely satisfy the free+cache targets
1169 * from the inactive pool we limit the number of pages we move
1170 * from the active pool to the inactive pool to 2x the pages we
1171 * had removed from the inactive pool (with a minimum of 1/5 the
1172 * inactive target). If we were not able to completely satisfy
1173 * the free+cache targets we go for the whole target aggressively.
1175 * NOTE: Both variables can end up negative.
1176 * NOTE: We are still in a critical section.
1179 bzero(&marker
, sizeof(marker
));
1180 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
1181 marker
.queue
= PQ_ACTIVE
+ q
;
1183 marker
.wire_count
= 1;
1185 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1186 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1187 maxscan
= vm_page_queues
[PQ_ACTIVE
+ q
].lcnt
;
1188 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1190 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
1191 maxscan
-- > 0 && (avail_shortage
- delta
> 0 ||
1192 inactive_shortage
> 0))
1194 vm_page_and_queue_spin_lock(m
);
1195 if (m
!= TAILQ_NEXT(&marker
, pageq
)) {
1196 vm_page_and_queue_spin_unlock(m
);
1200 KKASSERT(m
->queue
- m
->pc
== PQ_ACTIVE
);
1201 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1203 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, m
,
1209 if (m
->flags
& PG_MARKER
) {
1210 vm_page_and_queue_spin_unlock(m
);
1215 * Try to busy the page. Don't mess with pages which are
1216 * already busy or reorder them in the queue.
1218 if (vm_page_busy_try(m
, TRUE
)) {
1219 vm_page_and_queue_spin_unlock(m
);
1224 * Don't deactivate pages that are held, even if we can
1225 * busy them. (XXX why not?)
1227 if (m
->hold_count
!= 0) {
1228 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1230 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1232 vm_page_and_queue_spin_unlock(m
);
1236 vm_page_and_queue_spin_unlock(m
);
1240 * The page has been successfully busied and the page and
1241 * queue are no longer locked.
1245 * The count for pagedaemon pages is done after checking the
1246 * page for eligibility...
1248 mycpu
->gd_cnt
.v_pdpages
++;
1251 * Check to see "how much" the page has been used and clear
1252 * the tracking access bits. If the object has no references
1253 * don't bother paying the expense.
1256 if (m
->object
&& m
->object
->ref_count
!= 0) {
1257 if (m
->flags
& PG_REFERENCED
)
1259 actcount
+= pmap_ts_referenced(m
);
1261 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1262 if (m
->act_count
> ACT_MAX
)
1263 m
->act_count
= ACT_MAX
;
1266 vm_page_flag_clear(m
, PG_REFERENCED
);
1269 * actcount is only valid if the object ref_count is non-zero.
1270 * If the page does not have an object, actcount will be zero.
1272 if (actcount
&& m
->object
->ref_count
!= 0) {
1273 vm_page_and_queue_spin_lock(m
);
1274 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1276 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1279 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1282 vm_page_and_queue_spin_unlock(m
);
1285 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1286 if (vm_pageout_algorithm
||
1287 (m
->object
== NULL
) ||
1288 (m
->object
&& (m
->object
->ref_count
== 0)) ||
1289 m
->act_count
< pass
+ 1
1292 * Deactivate the page. If we had a
1293 * shortage from our inactive scan try to
1294 * free (cache) the page instead.
1296 * Don't just blindly cache the page if
1297 * we do not have a shortage from the
1298 * inactive scan, that could lead to
1299 * gigabytes being moved.
1301 --inactive_shortage
;
1302 if (avail_shortage
- delta
> 0 ||
1303 (m
->object
&& (m
->object
->ref_count
== 0)))
1305 if (avail_shortage
- delta
> 0)
1307 vm_page_protect(m
, VM_PROT_NONE
);
1308 if (m
->dirty
== 0 &&
1309 (m
->flags
& PG_NEED_COMMIT
) == 0 &&
1310 avail_shortage
- delta
> 0) {
1313 vm_page_deactivate(m
);
1317 vm_page_deactivate(m
);
1322 vm_page_and_queue_spin_lock(m
);
1323 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1325 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1328 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1331 vm_page_and_queue_spin_unlock(m
);
1338 * Clean out our local marker.
1340 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1341 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1342 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1348 * The number of actually free pages can drop down to v_free_reserved,
1349 * we try to build the free count back above v_free_min. Note that
1350 * vm_paging_needed() also returns TRUE if v_free_count is not at
1351 * least v_free_min so that is the minimum we must build the free
1354 * We use a slightly higher target to improve hysteresis,
1355 * ((v_free_target + v_free_min) / 2). Since v_free_target
1356 * is usually the same as v_cache_min this maintains about
1357 * half the pages in the free queue as are in the cache queue,
1358 * providing pretty good pipelining for pageout operation.
1360 * The system operator can manipulate vm.v_cache_min and
1361 * vm.v_free_target to tune the pageout demon. Be sure
1362 * to keep vm.v_free_min < vm.v_free_target.
1364 * Note that the original paging target is to get at least
1365 * (free_min + cache_min) into (free + cache). The slightly
1366 * higher target will shift additional pages from cache to free
1367 * without effecting the original paging target in order to
1368 * maintain better hysteresis and not have the free count always
1369 * be dead-on v_free_min.
1371 * NOTE: we are still in a critical section.
1373 * Pages moved from PQ_CACHE to totally free are not counted in the
1374 * pages_freed counter.
1377 vm_pageout_scan_cache(int avail_shortage
, int vnodes_skipped
, int recycle_count
)
1379 struct vm_pageout_scan_info info
;
1382 while (vmstats
.v_free_count
<
1383 (vmstats
.v_free_min
+ vmstats
.v_free_target
) / 2) {
1385 * This steals some code from vm/vm_page.c
1387 static int cache_rover
= 0;
1389 m
= vm_page_list_find(PQ_CACHE
, cache_rover
& PQ_L2_MASK
, FALSE
);
1392 /* page is returned removed from its queue and spinlocked */
1393 if (vm_page_busy_try(m
, TRUE
)) {
1394 vm_page_deactivate_locked(m
);
1395 vm_page_spin_unlock(m
);
1397 kprintf("Warning: busy page %p found in cache\n", m
);
1401 vm_page_spin_unlock(m
);
1402 pagedaemon_wakeup();
1406 * Page has been successfully busied and it and its queue
1407 * is no longer spinlocked.
1409 if ((m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) ||
1412 vm_page_deactivate(m
);
1416 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
1417 KKASSERT(m
->dirty
== 0);
1418 cache_rover
+= PQ_PRIME2
;
1419 vm_pageout_page_free(m
);
1420 mycpu
->gd_cnt
.v_dfree
++;
1423 #if !defined(NO_SWAPPING)
1425 * Idle process swapout -- run once per second.
1427 if (vm_swap_idle_enabled
) {
1429 if (time_second
!= lsec
) {
1430 vm_pageout_req_swapout
|= VM_SWAP_IDLE
;
1438 * If we didn't get enough free pages, and we have skipped a vnode
1439 * in a writeable object, wakeup the sync daemon. And kick swapout
1440 * if we did not get enough free pages.
1442 if (vm_paging_target() > 0) {
1443 if (vnodes_skipped
&& vm_page_count_min(0))
1445 #if !defined(NO_SWAPPING)
1446 if (vm_swap_enabled
&& vm_page_count_target()) {
1448 vm_pageout_req_swapout
|= VM_SWAP_NORMAL
;
1454 * Handle catastrophic conditions. Under good conditions we should
1455 * be at the target, well beyond our minimum. If we could not even
1456 * reach our minimum the system is under heavy stress.
1458 * Determine whether we have run out of memory. This occurs when
1459 * swap_pager_full is TRUE and the only pages left in the page
1460 * queues are dirty. We will still likely have page shortages.
1462 * - swap_pager_full is set if insufficient swap was
1463 * available to satisfy a requested pageout.
1465 * - the inactive queue is bloated (4 x size of active queue),
1466 * meaning it is unable to get rid of dirty pages and.
1468 * - vm_page_count_min() without counting pages recycled from the
1469 * active queue (recycle_count) means we could not recover
1470 * enough pages to meet bare minimum needs. This test only
1471 * works if the inactive queue is bloated.
1473 * - due to a positive avail_shortage we shifted the remaining
1474 * dirty pages from the active queue to the inactive queue
1475 * trying to find clean ones to free.
1477 if (swap_pager_full
&& vm_page_count_min(recycle_count
))
1478 kprintf("Warning: system low on memory+swap!\n");
1479 if (swap_pager_full
&& vm_page_count_min(recycle_count
) &&
1480 vmstats
.v_inactive_count
> vmstats
.v_active_count
* 4 &&
1481 avail_shortage
> 0) {
1485 info
.bigproc
= NULL
;
1487 allproc_scan(vm_pageout_scan_callback
, &info
);
1488 if (info
.bigproc
!= NULL
) {
1489 killproc(info
.bigproc
, "out of swap space");
1490 info
.bigproc
->p_nice
= PRIO_MIN
;
1491 info
.bigproc
->p_usched
->resetpriority(
1492 FIRST_LWP_IN_PROC(info
.bigproc
));
1493 wakeup(&vmstats
.v_free_count
);
1494 PRELE(info
.bigproc
);
1500 * The caller must hold proc_token.
1503 vm_pageout_scan_callback(struct proc
*p
, void *data
)
1505 struct vm_pageout_scan_info
*info
= data
;
1509 * Never kill system processes or init. If we have configured swap
1510 * then try to avoid killing low-numbered pids.
1512 if ((p
->p_flags
& P_SYSTEM
) || (p
->p_pid
== 1) ||
1513 ((p
->p_pid
< 48) && (vm_swap_size
!= 0))) {
1518 * if the process is in a non-running type state,
1521 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
)
1525 * Get the approximate process size. Note that anonymous pages
1526 * with backing swap will be counted twice, but there should not
1527 * be too many such pages due to the stress the VM system is
1528 * under at this point.
1530 size
= vmspace_anonymous_count(p
->p_vmspace
) +
1531 vmspace_swap_count(p
->p_vmspace
);
1534 * If the this process is bigger than the biggest one
1537 if (info
->bigsize
< size
) {
1539 PRELE(info
->bigproc
);
1542 info
->bigsize
= size
;
1549 * This routine tries to maintain the pseudo LRU active queue,
1550 * so that during long periods of time where there is no paging,
1551 * that some statistic accumulation still occurs. This code
1552 * helps the situation where paging just starts to occur.
1555 vm_pageout_page_stats(int q
)
1557 static int fullintervalcount
= 0;
1558 struct vm_page marker
;
1560 int pcount
, tpcount
; /* Number of pages to check */
1563 page_shortage
= (vmstats
.v_inactive_target
+ vmstats
.v_cache_max
+
1564 vmstats
.v_free_min
) -
1565 (vmstats
.v_free_count
+ vmstats
.v_inactive_count
+
1566 vmstats
.v_cache_count
);
1568 if (page_shortage
<= 0)
1571 pcount
= vm_page_queues
[PQ_ACTIVE
+ q
].lcnt
;
1572 fullintervalcount
+= vm_pageout_stats_interval
;
1573 if (fullintervalcount
< vm_pageout_full_stats_interval
) {
1574 tpcount
= (vm_pageout_stats_max
* pcount
) /
1575 vmstats
.v_page_count
+ 1;
1576 if (pcount
> tpcount
)
1579 fullintervalcount
= 0;
1582 bzero(&marker
, sizeof(marker
));
1583 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
1584 marker
.queue
= PQ_ACTIVE
+ q
;
1586 marker
.wire_count
= 1;
1588 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1589 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1590 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1592 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
1597 vm_page_and_queue_spin_lock(m
);
1598 if (m
!= TAILQ_NEXT(&marker
, pageq
)) {
1599 vm_page_and_queue_spin_unlock(m
);
1603 KKASSERT(m
->queue
- m
->pc
== PQ_ACTIVE
);
1604 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1605 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, m
,
1611 if (m
->flags
& PG_MARKER
) {
1612 vm_page_and_queue_spin_unlock(m
);
1617 * Ignore pages we can't busy
1619 if (vm_page_busy_try(m
, TRUE
)) {
1620 vm_page_and_queue_spin_unlock(m
);
1623 vm_page_and_queue_spin_unlock(m
);
1624 KKASSERT(m
->queue
- m
->pc
== PQ_ACTIVE
);
1627 * We now have a safely busied page, the page and queue
1628 * spinlocks have been released.
1632 if (m
->hold_count
) {
1638 * Calculate activity
1641 if (m
->flags
& PG_REFERENCED
) {
1642 vm_page_flag_clear(m
, PG_REFERENCED
);
1645 actcount
+= pmap_ts_referenced(m
);
1648 * Update act_count and move page to end of queue.
1651 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1652 if (m
->act_count
> ACT_MAX
)
1653 m
->act_count
= ACT_MAX
;
1654 vm_page_and_queue_spin_lock(m
);
1655 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1657 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1660 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1663 vm_page_and_queue_spin_unlock(m
);
1668 if (m
->act_count
== 0) {
1670 * We turn off page access, so that we have
1671 * more accurate RSS stats. We don't do this
1672 * in the normal page deactivation when the
1673 * system is loaded VM wise, because the
1674 * cost of the large number of page protect
1675 * operations would be higher than the value
1676 * of doing the operation.
1678 * We use the marker to save our place so
1679 * we can release the spin lock. both (m)
1680 * and (next) will be invalid.
1682 vm_page_protect(m
, VM_PROT_NONE
);
1683 vm_page_deactivate(m
);
1685 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1686 vm_page_and_queue_spin_lock(m
);
1687 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1689 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1692 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1695 vm_page_and_queue_spin_unlock(m
);
1701 * Remove our local marker
1703 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1704 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1705 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1709 vm_pageout_free_page_calc(vm_size_t count
)
1711 if (count
< vmstats
.v_page_count
)
1714 * free_reserved needs to include enough for the largest swap pager
1715 * structures plus enough for any pv_entry structs when paging.
1717 * v_free_min normal allocations
1718 * v_free_reserved system allocations
1719 * v_pageout_free_min allocations by pageout daemon
1720 * v_interrupt_free_min low level allocations (e.g swap structures)
1722 if (vmstats
.v_page_count
> 1024)
1723 vmstats
.v_free_min
= 64 + (vmstats
.v_page_count
- 1024) / 200;
1725 vmstats
.v_free_min
= 64;
1726 vmstats
.v_free_reserved
= vmstats
.v_free_min
* 4 / 8 + 7;
1727 vmstats
.v_free_severe
= vmstats
.v_free_min
* 4 / 8 + 0;
1728 vmstats
.v_pageout_free_min
= vmstats
.v_free_min
* 2 / 8 + 7;
1729 vmstats
.v_interrupt_free_min
= vmstats
.v_free_min
* 1 / 8 + 7;
1736 * vm_pageout is the high level pageout daemon.
1741 vm_pageout_thread(void)
1747 * Initialize some paging parameters.
1749 curthread
->td_flags
|= TDF_SYSTHREAD
;
1751 if (vmstats
.v_page_count
< 2000)
1752 vm_pageout_page_count
= 8;
1754 vm_pageout_free_page_calc(vmstats
.v_page_count
);
1757 * v_free_target and v_cache_min control pageout hysteresis. Note
1758 * that these are more a measure of the VM cache queue hysteresis
1759 * then the VM free queue. Specifically, v_free_target is the
1760 * high water mark (free+cache pages).
1762 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1763 * low water mark, while v_free_min is the stop. v_cache_min must
1764 * be big enough to handle memory needs while the pageout daemon
1765 * is signalled and run to free more pages.
1767 if (vmstats
.v_free_count
> 6144)
1768 vmstats
.v_free_target
= 4 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1770 vmstats
.v_free_target
= 2 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1773 * NOTE: With the new buffer cache b_act_count we want the default
1774 * inactive target to be a percentage of available memory.
1776 * The inactive target essentially determines the minimum
1777 * number of 'temporary' pages capable of caching one-time-use
1778 * files when the VM system is otherwise full of pages
1779 * belonging to multi-time-use files or active program data.
1781 * NOTE: The inactive target is aggressively persued only if the
1782 * inactive queue becomes too small. If the inactive queue
1783 * is large enough to satisfy page movement to free+cache
1784 * then it is repopulated more slowly from the active queue.
1785 * This allows a general inactive_target default to be set.
1787 * There is an issue here for processes which sit mostly idle
1788 * 'overnight', such as sshd, tcsh, and X. Any movement from
1789 * the active queue will eventually cause such pages to
1790 * recycle eventually causing a lot of paging in the morning.
1791 * To reduce the incidence of this pages cycled out of the
1792 * buffer cache are moved directly to the inactive queue if
1793 * they were only used once or twice.
1795 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1796 * Increasing the value (up to 64) increases the number of
1797 * buffer recyclements which go directly to the inactive queue.
1799 if (vmstats
.v_free_count
> 2048) {
1800 vmstats
.v_cache_min
= vmstats
.v_free_target
;
1801 vmstats
.v_cache_max
= 2 * vmstats
.v_cache_min
;
1803 vmstats
.v_cache_min
= 0;
1804 vmstats
.v_cache_max
= 0;
1806 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 4;
1808 /* XXX does not really belong here */
1809 if (vm_page_max_wired
== 0)
1810 vm_page_max_wired
= vmstats
.v_free_count
/ 3;
1812 if (vm_pageout_stats_max
== 0)
1813 vm_pageout_stats_max
= vmstats
.v_free_target
;
1816 * Set interval in seconds for stats scan.
1818 if (vm_pageout_stats_interval
== 0)
1819 vm_pageout_stats_interval
= 5;
1820 if (vm_pageout_full_stats_interval
== 0)
1821 vm_pageout_full_stats_interval
= vm_pageout_stats_interval
* 4;
1825 * Set maximum free per pass
1827 if (vm_pageout_stats_free_max
== 0)
1828 vm_pageout_stats_free_max
= 5;
1830 swap_pager_swap_init();
1834 * The pageout daemon is never done, so loop forever.
1841 int inactive_shortage
;
1842 int vnodes_skipped
= 0;
1843 int recycle_count
= 0;
1847 * Wait for an action request. If we timeout check to
1848 * see if paging is needed (in case the normal wakeup
1851 if (vm_pages_needed
== 0) {
1852 error
= tsleep(&vm_pages_needed
,
1854 vm_pageout_stats_interval
* hz
);
1856 vm_paging_needed() == 0 &&
1857 vm_pages_needed
== 0) {
1858 for (q
= 0; q
< PQ_L2_SIZE
; ++q
)
1859 vm_pageout_page_stats(q
);
1862 vm_pages_needed
= 1;
1865 mycpu
->gd_cnt
.v_pdwakeups
++;
1868 * Do whatever cleanup that the pmap code can.
1873 * Scan for pageout. Try to avoid thrashing the system
1876 * Calculate our target for the number of free+cache pages we
1877 * want to get to. This is higher then the number that causes
1878 * allocations to stall (severe) in order to provide hysteresis,
1879 * and if we don't make it all the way but get to the minimum
1880 * we're happy. Goose it a bit if there are multipler
1881 * requests for memory.
1883 avail_shortage
= vm_paging_target() + vm_pageout_deficit
;
1884 vm_pageout_deficit
= 0;
1886 if (avail_shortage
> 0) {
1887 for (q
= 0; q
< PQ_L2_SIZE
; ++q
) {
1888 delta1
+= vm_pageout_scan_inactive(
1890 PQAVERAGE(avail_shortage
),
1893 avail_shortage
-= delta1
;
1897 * Figure out how many active pages we must deactivate. If
1898 * we were able to reach our target with just the inactive
1899 * scan above we limit the number of active pages we
1900 * deactivate to reduce unnecessary work.
1902 inactive_shortage
= vmstats
.v_inactive_target
-
1903 vmstats
.v_inactive_count
;
1906 * If we were unable to free sufficient inactive pages to
1907 * satisfy the free/cache queue requirements then simply
1908 * reaching the inactive target may not be good enough.
1909 * Try to deactivate pages in excess of the target based
1912 * However to prevent thrashing the VM system do not
1913 * deactivate more than an additional 1/10 the inactive
1914 * target's worth of active pages.
1916 if (avail_shortage
> 0) {
1917 tmp
= avail_shortage
* 2;
1918 if (tmp
> vmstats
.v_inactive_target
/ 10)
1919 tmp
= vmstats
.v_inactive_target
/ 10;
1920 inactive_shortage
+= tmp
;
1923 if (avail_shortage
> 0 || inactive_shortage
> 0) {
1925 for (q
= 0; q
< PQ_L2_SIZE
; ++q
) {
1926 delta2
+= vm_pageout_scan_active(
1928 PQAVERAGE(avail_shortage
),
1929 PQAVERAGE(inactive_shortage
),
1932 inactive_shortage
-= delta2
;
1933 avail_shortage
-= delta2
;
1937 * Finally free enough cache pages to meet our free page
1938 * requirement and take more drastic measures if we are
1941 vm_pageout_scan_cache(avail_shortage
, vnodes_skipped
,
1945 * Wait for more work.
1947 if (avail_shortage
> 0) {
1949 if (swap_pager_full
) {
1951 * Running out of memory, catastrophic back-off
1952 * to one-second intervals.
1954 tsleep(&vm_pages_needed
, 0, "pdelay", hz
);
1955 } else if (pass
< 10 && vm_pages_needed
> 1) {
1957 * Normal operation, additional processes
1958 * have already kicked us. Retry immediately.
1960 } else if (pass
< 10) {
1962 * Normal operation, fewer processes. Delay
1963 * a bit but allow wakeups.
1965 vm_pages_needed
= 0;
1966 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
1967 vm_pages_needed
= 1;
1970 * We've taken too many passes, forced delay.
1972 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
1976 * Interlocked wakeup of waiters (non-optional)
1979 if (vm_pages_needed
&& !vm_page_count_min(0)) {
1980 wakeup(&vmstats
.v_free_count
);
1981 vm_pages_needed
= 0;
1987 static struct kproc_desc page_kp
= {
1992 SYSINIT(pagedaemon
, SI_SUB_KTHREAD_PAGE
, SI_ORDER_FIRST
, kproc_start
, &page_kp
)
1996 * Called after allocating a page out of the cache or free queue
1997 * to possibly wake the pagedaemon up to replentish our supply.
1999 * We try to generate some hysteresis by waking the pagedaemon up
2000 * when our free+cache pages go below the free_min+cache_min level.
2001 * The pagedaemon tries to get the count back up to at least the
2002 * minimum, and through to the target level if possible.
2004 * If the pagedaemon is already active bump vm_pages_needed as a hint
2005 * that there are even more requests pending.
2011 pagedaemon_wakeup(void)
2013 if (vm_paging_needed() && curthread
!= pagethread
) {
2014 if (vm_pages_needed
== 0) {
2015 vm_pages_needed
= 1; /* SMP race ok */
2016 wakeup(&vm_pages_needed
);
2017 } else if (vm_page_count_min(0)) {
2018 ++vm_pages_needed
; /* SMP race ok */
2023 #if !defined(NO_SWAPPING)
2030 vm_req_vmdaemon(void)
2032 static int lastrun
= 0;
2034 if ((ticks
> (lastrun
+ hz
)) || (ticks
< lastrun
)) {
2035 wakeup(&vm_daemon_needed
);
2040 static int vm_daemon_callback(struct proc
*p
, void *data __unused
);
2049 * XXX vm_daemon_needed specific token?
2052 tsleep(&vm_daemon_needed
, 0, "psleep", 0);
2053 if (vm_pageout_req_swapout
) {
2054 swapout_procs(vm_pageout_req_swapout
);
2055 vm_pageout_req_swapout
= 0;
2058 * scan the processes for exceeding their rlimits or if
2059 * process is swapped out -- deactivate pages
2061 allproc_scan(vm_daemon_callback
, NULL
);
2066 * Caller must hold proc_token.
2069 vm_daemon_callback(struct proc
*p
, void *data __unused
)
2071 vm_pindex_t limit
, size
;
2074 * if this is a system process or if we have already
2075 * looked at this process, skip it.
2077 if (p
->p_flags
& (P_SYSTEM
| P_WEXIT
))
2081 * if the process is in a non-running type state,
2084 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
)
2090 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
2091 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
2094 * let processes that are swapped out really be
2095 * swapped out. Set the limit to nothing to get as
2096 * many pages out to swap as possible.
2098 if (p
->p_flags
& P_SWAPPEDOUT
)
2101 lwkt_gettoken(&p
->p_vmspace
->vm_map
.token
);
2102 size
= vmspace_resident_count(p
->p_vmspace
);
2103 if (limit
>= 0 && size
>= limit
) {
2104 vm_pageout_map_deactivate_pages(&p
->p_vmspace
->vm_map
, limit
);
2106 lwkt_reltoken(&p
->p_vmspace
->vm_map
.token
);