2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
43 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44 * All rights reserved.
46 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48 * Permission to use, copy, modify and distribute this software and
49 * its documentation is hereby granted, provided that both the copyright
50 * notice and this permission notice appear in all copies of the
51 * software, derivative works or modified versions, and any portions
52 * thereof, and that both notices appear in supporting documentation.
54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58 * Carnegie Mellon requests users of this software to return to
60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
61 * School of Computer Science
62 * Carnegie Mellon University
63 * Pittsburgh PA 15213-3890
65 * any improvements or extensions that they make and grant Carnegie the
66 * rights to redistribute these changes.
68 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
69 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.36 2008/07/01 02:02:56 dillon Exp $
73 * The proverbial page-out daemon.
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
81 #include <sys/kthread.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/vnode.h>
85 #include <sys/vmmeter.h>
86 #include <sys/sysctl.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pager.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_extern.h>
99 #include <sys/thread2.h>
100 #include <vm/vm_page2.h>
103 * System initialization
106 /* the kernel process "vm_pageout"*/
107 static void vm_pageout (void);
108 static int vm_pageout_clean (vm_page_t
);
109 static int vm_pageout_scan (int pass
);
110 static int vm_pageout_free_page_calc (vm_size_t count
);
111 struct thread
*pagethread
;
113 static struct kproc_desc page_kp
= {
118 SYSINIT(pagedaemon
, SI_SUB_KTHREAD_PAGE
, SI_ORDER_FIRST
, kproc_start
, &page_kp
)
120 #if !defined(NO_SWAPPING)
121 /* the kernel process "vm_daemon"*/
122 static void vm_daemon (void);
123 static struct thread
*vmthread
;
125 static struct kproc_desc vm_kp
= {
130 SYSINIT(vmdaemon
, SI_SUB_KTHREAD_VM
, SI_ORDER_FIRST
, kproc_start
, &vm_kp
)
134 int vm_pages_needed
=0; /* Event on which pageout daemon sleeps */
135 int vm_pageout_deficit
=0; /* Estimated number of pages deficit */
136 int vm_pageout_pages_needed
=0; /* flag saying that the pageout daemon needs pages */
138 #if !defined(NO_SWAPPING)
139 static int vm_pageout_req_swapout
; /* XXX */
140 static int vm_daemon_needed
;
142 static int vm_max_launder
= 32;
143 static int vm_pageout_stats_max
=0, vm_pageout_stats_interval
= 0;
144 static int vm_pageout_full_stats_interval
= 0;
145 static int vm_pageout_stats_free_max
=0, vm_pageout_algorithm
=0;
146 static int defer_swap_pageouts
=0;
147 static int disable_swap_pageouts
=0;
149 #if defined(NO_SWAPPING)
150 static int vm_swap_enabled
=0;
151 static int vm_swap_idle_enabled
=0;
153 static int vm_swap_enabled
=1;
154 static int vm_swap_idle_enabled
=0;
157 SYSCTL_INT(_vm
, VM_PAGEOUT_ALGORITHM
, pageout_algorithm
,
158 CTLFLAG_RW
, &vm_pageout_algorithm
, 0, "LRU page mgmt");
160 SYSCTL_INT(_vm
, OID_AUTO
, max_launder
,
161 CTLFLAG_RW
, &vm_max_launder
, 0, "Limit dirty flushes in pageout");
163 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_max
,
164 CTLFLAG_RW
, &vm_pageout_stats_max
, 0, "Max pageout stats scan length");
166 SYSCTL_INT(_vm
, OID_AUTO
, pageout_full_stats_interval
,
167 CTLFLAG_RW
, &vm_pageout_full_stats_interval
, 0, "Interval for full stats scan");
169 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_interval
,
170 CTLFLAG_RW
, &vm_pageout_stats_interval
, 0, "Interval for partial stats scan");
172 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_free_max
,
173 CTLFLAG_RW
, &vm_pageout_stats_free_max
, 0, "Not implemented");
175 #if defined(NO_SWAPPING)
176 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
177 CTLFLAG_RD
, &vm_swap_enabled
, 0, "");
178 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
179 CTLFLAG_RD
, &vm_swap_idle_enabled
, 0, "");
181 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
182 CTLFLAG_RW
, &vm_swap_enabled
, 0, "Enable entire process swapout");
183 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
184 CTLFLAG_RW
, &vm_swap_idle_enabled
, 0, "Allow swapout on idle criteria");
187 SYSCTL_INT(_vm
, OID_AUTO
, defer_swapspace_pageouts
,
188 CTLFLAG_RW
, &defer_swap_pageouts
, 0, "Give preference to dirty pages in mem");
190 SYSCTL_INT(_vm
, OID_AUTO
, disable_swapspace_pageouts
,
191 CTLFLAG_RW
, &disable_swap_pageouts
, 0, "Disallow swapout of dirty pages");
193 static int pageout_lock_miss
;
194 SYSCTL_INT(_vm
, OID_AUTO
, pageout_lock_miss
,
195 CTLFLAG_RD
, &pageout_lock_miss
, 0, "vget() lock misses during pageout");
198 SYSCTL_INT(_vm
, OID_AUTO
, vm_load
,
199 CTLFLAG_RD
, &vm_load
, 0, "load on the VM system");
200 int vm_load_enable
= 1;
201 SYSCTL_INT(_vm
, OID_AUTO
, vm_load_enable
,
202 CTLFLAG_RW
, &vm_load_enable
, 0, "enable vm_load rate limiting");
205 SYSCTL_INT(_vm
, OID_AUTO
, vm_load_debug
,
206 CTLFLAG_RW
, &vm_load_debug
, 0, "debug vm_load");
209 #define VM_PAGEOUT_PAGE_COUNT 16
210 int vm_pageout_page_count
= VM_PAGEOUT_PAGE_COUNT
;
212 int vm_page_max_wired
; /* XXX max # of wired pages system-wide */
214 #if !defined(NO_SWAPPING)
215 typedef void freeer_fcn_t (vm_map_t
, vm_object_t
, vm_pindex_t
, int);
216 static void vm_pageout_map_deactivate_pages (vm_map_t
, vm_pindex_t
);
217 static freeer_fcn_t vm_pageout_object_deactivate_pages
;
218 static void vm_req_vmdaemon (void);
220 static void vm_pageout_page_stats(void);
223 * Update vm_load to slow down faulting processes.
226 vm_fault_ratecheck(void)
228 if (vm_pages_needed
) {
240 * Clean the page and remove it from the laundry. The page must not be
243 * We set the busy bit to cause potential page faults on this page to
244 * block. Note the careful timing, however, the busy bit isn't set till
245 * late and we cannot do anything that will mess with the page.
249 vm_pageout_clean(vm_page_t m
)
252 vm_page_t mc
[2*vm_pageout_page_count
];
254 int ib
, is
, page_base
;
255 vm_pindex_t pindex
= m
->pindex
;
260 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
261 * with the new swapper, but we could have serious problems paging
262 * out other object types if there is insufficient memory.
264 * Unfortunately, checking free memory here is far too late, so the
265 * check has been moved up a procedural level.
269 * Don't mess with the page if it's busy, held, or special
271 if ((m
->hold_count
!= 0) ||
272 ((m
->busy
!= 0) || (m
->flags
& (PG_BUSY
|PG_UNMANAGED
)))) {
276 mc
[vm_pageout_page_count
] = m
;
278 page_base
= vm_pageout_page_count
;
283 * Scan object for clusterable pages.
285 * We can cluster ONLY if: ->> the page is NOT
286 * clean, wired, busy, held, or mapped into a
287 * buffer, and one of the following:
288 * 1) The page is inactive, or a seldom used
291 * 2) we force the issue.
293 * During heavy mmap/modification loads the pageout
294 * daemon can really fragment the underlying file
295 * due to flushing pages out of order and not trying
296 * align the clusters (which leave sporatic out-of-order
297 * holes). To solve this problem we do the reverse scan
298 * first and attempt to align our cluster, then do a
299 * forward scan if room remains.
303 while (ib
&& pageout_count
< vm_pageout_page_count
) {
311 if ((p
= vm_page_lookup(object
, pindex
- ib
)) == NULL
) {
315 if (((p
->queue
- p
->pc
) == PQ_CACHE
) ||
316 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) || p
->busy
) {
320 vm_page_test_dirty(p
);
321 if ((p
->dirty
& p
->valid
) == 0 ||
322 p
->queue
!= PQ_INACTIVE
||
323 p
->wire_count
!= 0 || /* may be held by buf cache */
324 p
->hold_count
!= 0) { /* may be undergoing I/O */
332 * alignment boundry, stop here and switch directions. Do
335 if ((pindex
- (ib
- 1)) % vm_pageout_page_count
== 0)
339 while (pageout_count
< vm_pageout_page_count
&&
340 pindex
+ is
< object
->size
) {
343 if ((p
= vm_page_lookup(object
, pindex
+ is
)) == NULL
)
345 if (((p
->queue
- p
->pc
) == PQ_CACHE
) ||
346 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) || p
->busy
) {
349 vm_page_test_dirty(p
);
350 if ((p
->dirty
& p
->valid
) == 0 ||
351 p
->queue
!= PQ_INACTIVE
||
352 p
->wire_count
!= 0 || /* may be held by buf cache */
353 p
->hold_count
!= 0) { /* may be undergoing I/O */
356 mc
[page_base
+ pageout_count
] = p
;
362 * If we exhausted our forward scan, continue with the reverse scan
363 * when possible, even past a page boundry. This catches boundry
366 if (ib
&& pageout_count
< vm_pageout_page_count
)
370 * we allow reads during pageouts...
372 return vm_pageout_flush(&mc
[page_base
], pageout_count
, 0);
376 * vm_pageout_flush() - launder the given pages
378 * The given pages are laundered. Note that we setup for the start of
379 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
380 * reference count all in here rather then in the parent. If we want
381 * the parent to do more sophisticated things we may have to change
385 vm_pageout_flush(vm_page_t
*mc
, int count
, int flags
)
388 int pageout_status
[count
];
393 * Initiate I/O. Bump the vm_page_t->busy counter.
395 for (i
= 0; i
< count
; i
++) {
396 KASSERT(mc
[i
]->valid
== VM_PAGE_BITS_ALL
, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc
[i
], i
, count
));
397 vm_page_io_start(mc
[i
]);
401 * We must make the pages read-only. This will also force the
402 * modified bit in the related pmaps to be cleared. The pager
403 * cannot clear the bit for us since the I/O completion code
404 * typically runs from an interrupt. The act of making the page
405 * read-only handles the case for us.
407 for (i
= 0; i
< count
; i
++) {
408 vm_page_protect(mc
[i
], VM_PROT_READ
);
411 object
= mc
[0]->object
;
412 vm_object_pip_add(object
, count
);
414 vm_pager_put_pages(object
, mc
, count
,
415 (flags
| ((object
== &kernel_object
) ? VM_PAGER_PUT_SYNC
: 0)),
418 for (i
= 0; i
< count
; i
++) {
419 vm_page_t mt
= mc
[i
];
421 switch (pageout_status
[i
]) {
430 * Page outside of range of object. Right now we
431 * essentially lose the changes by pretending it
434 pmap_clear_modify(mt
);
440 * A page typically cannot be paged out when we
441 * have run out of swap. We leave the page
442 * marked inactive and will try to page it out
445 * Starvation of the active page list is used to
446 * determine when the system is massively memory
455 * If the operation is still going, leave the page busy to
456 * block all other accesses. Also, leave the paging in
457 * progress indicator set so that we don't attempt an object
460 * For any pages which have completed synchronously,
461 * deactivate the page if we are under a severe deficit.
462 * Do not try to enter them into the cache, though, they
463 * might still be read-heavy.
465 if (pageout_status
[i
] != VM_PAGER_PEND
) {
466 vm_object_pip_wakeup(object
);
467 vm_page_io_finish(mt
);
468 if (vm_page_count_severe())
469 vm_page_deactivate(mt
);
471 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt
))
472 vm_page_protect(mt
, VM_PROT_READ
);
479 #if !defined(NO_SWAPPING)
481 * vm_pageout_object_deactivate_pages
483 * deactivate enough pages to satisfy the inactive target
484 * requirements or if vm_page_proc_limit is set, then
485 * deactivate all of the pages in the object and its
488 * The object and map must be locked.
490 static int vm_pageout_object_deactivate_pages_callback(vm_page_t
, void *);
493 vm_pageout_object_deactivate_pages(vm_map_t map
, vm_object_t object
,
494 vm_pindex_t desired
, int map_remove_only
)
496 struct rb_vm_page_scan_info info
;
499 if (object
->type
== OBJT_DEVICE
|| object
->type
== OBJT_PHYS
)
503 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
505 if (object
->paging_in_progress
)
508 remove_mode
= map_remove_only
;
509 if (object
->shadow_count
> 1)
513 * scan the objects entire memory queue. spl protection is
514 * required to avoid an interrupt unbusy/free race against
518 info
.limit
= remove_mode
;
520 info
.desired
= desired
;
521 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
522 vm_pageout_object_deactivate_pages_callback
,
526 object
= object
->backing_object
;
531 vm_pageout_object_deactivate_pages_callback(vm_page_t p
, void *data
)
533 struct rb_vm_page_scan_info
*info
= data
;
536 if (pmap_resident_count(vm_map_pmap(info
->map
)) <= info
->desired
) {
539 mycpu
->gd_cnt
.v_pdpages
++;
540 if (p
->wire_count
!= 0 || p
->hold_count
!= 0 || p
->busy
!= 0 ||
541 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) ||
542 !pmap_page_exists_quick(vm_map_pmap(info
->map
), p
)) {
546 actcount
= pmap_ts_referenced(p
);
548 vm_page_flag_set(p
, PG_REFERENCED
);
549 } else if (p
->flags
& PG_REFERENCED
) {
553 if ((p
->queue
!= PQ_ACTIVE
) &&
554 (p
->flags
& PG_REFERENCED
)) {
556 p
->act_count
+= actcount
;
557 vm_page_flag_clear(p
, PG_REFERENCED
);
558 } else if (p
->queue
== PQ_ACTIVE
) {
559 if ((p
->flags
& PG_REFERENCED
) == 0) {
560 p
->act_count
-= min(p
->act_count
, ACT_DECLINE
);
561 if (!info
->limit
&& (vm_pageout_algorithm
|| (p
->act_count
== 0))) {
563 vm_page_protect(p
, VM_PROT_NONE
);
565 vm_page_deactivate(p
);
567 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
568 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
572 vm_page_flag_clear(p
, PG_REFERENCED
);
573 if (p
->act_count
< (ACT_MAX
- ACT_ADVANCE
))
574 p
->act_count
+= ACT_ADVANCE
;
575 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
576 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
578 } else if (p
->queue
== PQ_INACTIVE
) {
580 vm_page_protect(p
, VM_PROT_NONE
);
587 * deactivate some number of pages in a map, try to do it fairly, but
588 * that is really hard to do.
591 vm_pageout_map_deactivate_pages(vm_map_t map
, vm_pindex_t desired
)
594 vm_object_t obj
, bigobj
;
597 if (lockmgr(&map
->lock
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
605 * first, search out the biggest object, and try to free pages from
608 tmpe
= map
->header
.next
;
609 while (tmpe
!= &map
->header
) {
610 switch(tmpe
->maptype
) {
611 case VM_MAPTYPE_NORMAL
:
612 case VM_MAPTYPE_VPAGETABLE
:
613 obj
= tmpe
->object
.vm_object
;
614 if ((obj
!= NULL
) && (obj
->shadow_count
<= 1) &&
616 (bigobj
->resident_page_count
< obj
->resident_page_count
))) {
623 if (tmpe
->wired_count
> 0)
624 nothingwired
= FALSE
;
629 vm_pageout_object_deactivate_pages(map
, bigobj
, desired
, 0);
632 * Next, hunt around for other pages to deactivate. We actually
633 * do this search sort of wrong -- .text first is not the best idea.
635 tmpe
= map
->header
.next
;
636 while (tmpe
!= &map
->header
) {
637 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
639 switch(tmpe
->maptype
) {
640 case VM_MAPTYPE_NORMAL
:
641 case VM_MAPTYPE_VPAGETABLE
:
642 obj
= tmpe
->object
.vm_object
;
644 vm_pageout_object_deactivate_pages(map
, obj
, desired
, 0);
653 * Remove all mappings if a process is swapped out, this will free page
656 if (desired
== 0 && nothingwired
)
657 pmap_remove(vm_map_pmap(map
),
658 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
664 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We
665 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
666 * be trivially freed.
669 vm_pageout_page_free(vm_page_t m
)
671 vm_object_t object
= m
->object
;
672 int type
= object
->type
;
674 if (type
== OBJT_SWAP
|| type
== OBJT_DEFAULT
)
675 vm_object_reference(object
);
677 vm_page_protect(m
, VM_PROT_NONE
);
679 if (type
== OBJT_SWAP
|| type
== OBJT_DEFAULT
)
680 vm_object_deallocate(object
);
684 * vm_pageout_scan does the dirty work for the pageout daemon.
686 struct vm_pageout_scan_info
{
687 struct proc
*bigproc
;
691 static int vm_pageout_scan_callback(struct proc
*p
, void *data
);
694 vm_pageout_scan(int pass
)
696 struct vm_pageout_scan_info info
;
698 struct vm_page marker
;
701 int inactive_shortage
, active_shortage
;
702 int inactive_original_shortage
;
705 int vnodes_skipped
= 0;
709 * Do whatever cleanup that the pmap code can.
714 * Calculate our target for the number of free+cache pages we
715 * want to get to. This is higher then the number that causes
716 * allocations to stall (severe) in order to provide hysteresis,
717 * and if we don't make it all the way but get to the minimum
720 inactive_shortage
= vm_paging_target() + vm_pageout_deficit
;
721 inactive_original_shortage
= inactive_shortage
;
722 vm_pageout_deficit
= 0;
725 * Initialize our marker
727 bzero(&marker
, sizeof(marker
));
728 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
729 marker
.queue
= PQ_INACTIVE
;
730 marker
.wire_count
= 1;
733 * Start scanning the inactive queue for pages we can move to the
734 * cache or free. The scan will stop when the target is reached or
735 * we have scanned the entire inactive queue. Note that m->act_count
736 * is not used to form decisions for the inactive queue, only for the
739 * maxlaunder limits the number of dirty pages we flush per scan.
740 * For most systems a smaller value (16 or 32) is more robust under
741 * extreme memory and disk pressure because any unnecessary writes
742 * to disk can result in extreme performance degredation. However,
743 * systems with excessive dirty pages (especially when MAP_NOSYNC is
744 * used) will die horribly with limited laundering. If the pageout
745 * daemon cannot clean enough pages in the first pass, we let it go
746 * all out in succeeding passes.
748 if ((maxlaunder
= vm_max_launder
) <= 1)
754 * We will generally be in a critical section throughout the
755 * scan, but we can release it temporarily when we are sitting on a
756 * non-busy page without fear. this is required to prevent an
757 * interrupt from unbusying or freeing a page prior to our busy
758 * check, leaving us on the wrong queue or checking the wrong
763 maxscan
= vmstats
.v_inactive_count
;
764 for (m
= TAILQ_FIRST(&vm_page_queues
[PQ_INACTIVE
].pl
);
765 m
!= NULL
&& maxscan
-- > 0 && inactive_shortage
> 0;
768 mycpu
->gd_cnt
.v_pdpages
++;
771 * Give interrupts a chance
777 * It's easier for some of the conditions below to just loop
778 * and catch queue changes here rather then check everywhere
781 if (m
->queue
!= PQ_INACTIVE
)
783 next
= TAILQ_NEXT(m
, pageq
);
788 if (m
->flags
& PG_MARKER
)
792 * A held page may be undergoing I/O, so skip it.
795 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
796 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
801 * Dont mess with busy pages, keep in the front of the
802 * queue, most likely are being paged out.
804 if (m
->busy
|| (m
->flags
& PG_BUSY
)) {
808 if (m
->object
->ref_count
== 0) {
810 * If the object is not being used, we ignore previous
813 vm_page_flag_clear(m
, PG_REFERENCED
);
814 pmap_clear_reference(m
);
816 } else if (((m
->flags
& PG_REFERENCED
) == 0) &&
817 (actcount
= pmap_ts_referenced(m
))) {
819 * Otherwise, if the page has been referenced while
820 * in the inactive queue, we bump the "activation
821 * count" upwards, making it less likely that the
822 * page will be added back to the inactive queue
823 * prematurely again. Here we check the page tables
824 * (or emulated bits, if any), given the upper level
825 * VM system not knowing anything about existing
829 m
->act_count
+= (actcount
+ ACT_ADVANCE
);
834 * If the upper level VM system knows about any page
835 * references, we activate the page. We also set the
836 * "activation count" higher than normal so that we will less
837 * likely place pages back onto the inactive queue again.
839 if ((m
->flags
& PG_REFERENCED
) != 0) {
840 vm_page_flag_clear(m
, PG_REFERENCED
);
841 actcount
= pmap_ts_referenced(m
);
843 m
->act_count
+= (actcount
+ ACT_ADVANCE
+ 1);
848 * If the upper level VM system doesn't know anything about
849 * the page being dirty, we have to check for it again. As
850 * far as the VM code knows, any partially dirty pages are
853 * Pages marked PG_WRITEABLE may be mapped into the user
854 * address space of a process running on another cpu. A
855 * user process (without holding the MP lock) running on
856 * another cpu may be able to touch the page while we are
857 * trying to remove it. vm_page_cache() will handle this
861 vm_page_test_dirty(m
);
868 * Invalid pages can be easily freed
870 vm_pageout_page_free(m
);
871 mycpu
->gd_cnt
.v_dfree
++;
873 } else if (m
->dirty
== 0) {
875 * Clean pages can be placed onto the cache queue.
876 * This effectively frees them.
880 } else if ((m
->flags
& PG_WINATCFLS
) == 0 && pass
== 0) {
882 * Dirty pages need to be paged out, but flushing
883 * a page is extremely expensive verses freeing
884 * a clean page. Rather then artificially limiting
885 * the number of pages we can flush, we instead give
886 * dirty pages extra priority on the inactive queue
887 * by forcing them to be cycled through the queue
888 * twice before being flushed, after which the
889 * (now clean) page will cycle through once more
890 * before being freed. This significantly extends
891 * the thrash point for a heavily loaded machine.
893 vm_page_flag_set(m
, PG_WINATCFLS
);
894 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
895 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
896 } else if (maxlaunder
> 0) {
898 * We always want to try to flush some dirty pages if
899 * we encounter them, to keep the system stable.
900 * Normally this number is small, but under extreme
901 * pressure where there are insufficient clean pages
902 * on the inactive queue, we may have to go all out.
904 int swap_pageouts_ok
;
905 struct vnode
*vp
= NULL
;
909 if ((object
->type
!= OBJT_SWAP
) && (object
->type
!= OBJT_DEFAULT
)) {
910 swap_pageouts_ok
= 1;
912 swap_pageouts_ok
= !(defer_swap_pageouts
|| disable_swap_pageouts
);
913 swap_pageouts_ok
|= (!disable_swap_pageouts
&& defer_swap_pageouts
&&
914 vm_page_count_min(0));
919 * We don't bother paging objects that are "dead".
920 * Those objects are in a "rundown" state.
922 if (!swap_pageouts_ok
|| (object
->flags
& OBJ_DEAD
)) {
923 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
924 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
929 * The object is already known NOT to be dead. It
930 * is possible for the vget() to block the whole
931 * pageout daemon, but the new low-memory handling
932 * code should prevent it.
934 * The previous code skipped locked vnodes and, worse,
935 * reordered pages in the queue. This results in
936 * completely non-deterministic operation because,
937 * quite often, a vm_fault has initiated an I/O and
938 * is holding a locked vnode at just the point where
939 * the pageout daemon is woken up.
941 * We can't wait forever for the vnode lock, we might
942 * deadlock due to a vn_read() getting stuck in
943 * vm_wait while holding this vnode. We skip the
944 * vnode if we can't get it in a reasonable amount
948 if (object
->type
== OBJT_VNODE
) {
951 if (vget(vp
, LK_EXCLUSIVE
|LK_NOOBJ
|LK_TIMELOCK
)) {
953 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
959 * The page might have been moved to another
960 * queue during potential blocking in vget()
961 * above. The page might have been freed and
962 * reused for another vnode. The object might
963 * have been reused for another vnode.
965 if (m
->queue
!= PQ_INACTIVE
||
966 m
->object
!= object
||
967 object
->handle
!= vp
) {
968 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
975 * The page may have been busied during the
976 * blocking in vput(); We don't move the
977 * page back onto the end of the queue so that
978 * statistics are more correct if we don't.
980 if (m
->busy
|| (m
->flags
& PG_BUSY
)) {
986 * If the page has become held it might
987 * be undergoing I/O, so skip it
990 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
991 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
992 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1000 * If a page is dirty, then it is either being washed
1001 * (but not yet cleaned) or it is still in the
1002 * laundry. If it is still in the laundry, then we
1003 * start the cleaning operation.
1005 * This operation may cluster, invalidating the 'next'
1006 * pointer. To prevent an inordinate number of
1007 * restarts we use our marker to remember our place.
1009 * decrement inactive_shortage on success to account
1010 * for the (future) cleaned page. Otherwise we
1011 * could wind up laundering or cleaning too many
1014 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, &marker
, pageq
);
1015 if (vm_pageout_clean(m
) != 0) {
1016 --inactive_shortage
;
1019 next
= TAILQ_NEXT(&marker
, pageq
);
1020 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, &marker
, pageq
);
1027 * We want to move pages from the active queue to the inactive
1028 * queue to get the inactive queue to the inactive target. If
1029 * we still have a page shortage from above we try to directly free
1030 * clean pages instead of moving them.
1032 * If we do still have a shortage we keep track of the number of
1033 * pages we free or cache (recycle_count) as a measure of thrashing
1034 * between the active and inactive queues.
1036 * If we were able to completely satisfy the free+cache targets
1037 * from the inactive pool we limit the number of pages we move
1038 * from the active pool to the inactive pool to 2x the pages we
1039 * had removed from the inactive pool (with a minimum of 1/5 the
1040 * inactive target). If we were not able to completely satisfy
1041 * the free+cache targets we go for the whole target aggressively.
1043 * NOTE: Both variables can end up negative.
1044 * NOTE: We are still in a critical section.
1046 active_shortage
= vmstats
.v_inactive_target
- vmstats
.v_inactive_count
;
1047 if (inactive_original_shortage
< vmstats
.v_inactive_target
/ 10)
1048 inactive_original_shortage
= vmstats
.v_inactive_target
/ 10;
1049 if (inactive_shortage
<= 0 &&
1050 active_shortage
> inactive_original_shortage
* 2) {
1051 active_shortage
= inactive_original_shortage
* 2;
1054 pcount
= vmstats
.v_active_count
;
1056 m
= TAILQ_FIRST(&vm_page_queues
[PQ_ACTIVE
].pl
);
1058 while ((m
!= NULL
) && (pcount
-- > 0) &&
1059 (inactive_shortage
> 0 || active_shortage
> 0)
1062 * Give interrupts a chance.
1068 * If the page was ripped out from under us, just stop.
1070 if (m
->queue
!= PQ_ACTIVE
)
1072 next
= TAILQ_NEXT(m
, pageq
);
1075 * Don't deactivate pages that are busy.
1077 if ((m
->busy
!= 0) ||
1078 (m
->flags
& PG_BUSY
) ||
1079 (m
->hold_count
!= 0)) {
1080 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1081 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1087 * The count for pagedaemon pages is done after checking the
1088 * page for eligibility...
1090 mycpu
->gd_cnt
.v_pdpages
++;
1093 * Check to see "how much" the page has been used and clear
1094 * the tracking access bits. If the object has no references
1095 * don't bother paying the expense.
1098 if (m
->object
->ref_count
!= 0) {
1099 if (m
->flags
& PG_REFERENCED
)
1101 actcount
+= pmap_ts_referenced(m
);
1103 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1104 if (m
->act_count
> ACT_MAX
)
1105 m
->act_count
= ACT_MAX
;
1108 vm_page_flag_clear(m
, PG_REFERENCED
);
1111 * actcount is only valid if the object ref_count is non-zero.
1113 if (actcount
&& m
->object
->ref_count
!= 0) {
1114 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1115 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1117 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1118 if (vm_pageout_algorithm
||
1119 m
->object
->ref_count
== 0 ||
1120 m
->act_count
< pass
+ 1
1123 * Deactivate the page. If we had a
1124 * shortage from our inactive scan try to
1125 * free (cache) the page instead.
1127 * Don't just blindly cache the page if
1128 * we do not have a shortage from the
1129 * inactive scan, that could lead to
1130 * gigabytes being moved.
1133 if (inactive_shortage
> 0 ||
1134 m
->object
->ref_count
== 0) {
1135 if (inactive_shortage
> 0)
1138 vm_page_protect(m
, VM_PROT_NONE
);
1140 if (m
->dirty
== 0 &&
1141 inactive_shortage
> 0) {
1142 --inactive_shortage
;
1145 vm_page_deactivate(m
);
1148 vm_page_deactivate(m
);
1151 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1152 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1159 * We try to maintain some *really* free pages, this allows interrupt
1160 * code to be guaranteed space. Since both cache and free queues
1161 * are considered basically 'free', moving pages from cache to free
1162 * does not effect other calculations.
1164 * NOTE: we are still in a critical section.
1166 * Pages moved from PQ_CACHE to totally free are not counted in the
1167 * pages_freed counter.
1169 while (vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
1170 static int cache_rover
= 0;
1171 m
= vm_page_list_find(PQ_CACHE
, cache_rover
, FALSE
);
1174 if ((m
->flags
& (PG_BUSY
|PG_UNMANAGED
)) ||
1179 kprintf("Warning: busy page %p found in cache\n", m
);
1181 vm_page_deactivate(m
);
1184 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
1185 KKASSERT(m
->dirty
== 0);
1186 cache_rover
= (cache_rover
+ PQ_PRIME2
) & PQ_L2_MASK
;
1187 vm_pageout_page_free(m
);
1188 mycpu
->gd_cnt
.v_dfree
++;
1193 #if !defined(NO_SWAPPING)
1195 * Idle process swapout -- run once per second.
1197 if (vm_swap_idle_enabled
) {
1199 if (time_second
!= lsec
) {
1200 vm_pageout_req_swapout
|= VM_SWAP_IDLE
;
1208 * If we didn't get enough free pages, and we have skipped a vnode
1209 * in a writeable object, wakeup the sync daemon. And kick swapout
1210 * if we did not get enough free pages.
1212 if (vm_paging_target() > 0) {
1213 if (vnodes_skipped
&& vm_page_count_min(0))
1215 #if !defined(NO_SWAPPING)
1216 if (vm_swap_enabled
&& vm_page_count_target()) {
1218 vm_pageout_req_swapout
|= VM_SWAP_NORMAL
;
1224 * Handle catastrophic conditions. Under good conditions we should
1225 * be at the target, well beyond our minimum. If we could not even
1226 * reach our minimum the system is under heavy stress.
1228 * Determine whether we have run out of memory. This occurs when
1229 * swap_pager_full is TRUE and the only pages left in the page
1230 * queues are dirty. We will still likely have page shortages.
1232 * - swap_pager_full is set if insufficient swap was
1233 * available to satisfy a requested pageout.
1235 * - the inactive queue is bloated (4 x size of active queue),
1236 * meaning it is unable to get rid of dirty pages and.
1238 * - vm_page_count_min() without counting pages recycled from the
1239 * active queue (recycle_count) means we could not recover
1240 * enough pages to meet bare minimum needs. This test only
1241 * works if the inactive queue is bloated.
1243 * - due to a positive inactive_shortage we shifted the remaining
1244 * dirty pages from the active queue to the inactive queue
1245 * trying to find clean ones to free.
1247 if (swap_pager_full
&& vm_page_count_min(recycle_count
))
1248 kprintf("Warning: system low on memory+swap!\n");
1249 if (swap_pager_full
&& vm_page_count_min(recycle_count
) &&
1250 vmstats
.v_inactive_count
> vmstats
.v_active_count
* 4 &&
1251 inactive_shortage
> 0) {
1255 info
.bigproc
= NULL
;
1257 allproc_scan(vm_pageout_scan_callback
, &info
);
1258 if (info
.bigproc
!= NULL
) {
1259 killproc(info
.bigproc
, "out of swap space");
1260 info
.bigproc
->p_nice
= PRIO_MIN
;
1261 info
.bigproc
->p_usched
->resetpriority(
1262 FIRST_LWP_IN_PROC(info
.bigproc
));
1263 wakeup(&vmstats
.v_free_count
);
1264 PRELE(info
.bigproc
);
1267 return(inactive_shortage
);
1271 vm_pageout_scan_callback(struct proc
*p
, void *data
)
1273 struct vm_pageout_scan_info
*info
= data
;
1277 * Never kill system processes or init. If we have configured swap
1278 * then try to avoid killing low-numbered pids.
1280 if ((p
->p_flag
& P_SYSTEM
) || (p
->p_pid
== 1) ||
1281 ((p
->p_pid
< 48) && (vm_swap_size
!= 0))) {
1286 * if the process is in a non-running type state,
1289 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
)
1293 * Get the approximate process size. Note that anonymous pages
1294 * with backing swap will be counted twice, but there should not
1295 * be too many such pages due to the stress the VM system is
1296 * under at this point.
1298 size
= vmspace_anonymous_count(p
->p_vmspace
) +
1299 vmspace_swap_count(p
->p_vmspace
);
1302 * If the this process is bigger than the biggest one
1305 if (info
->bigsize
< size
) {
1307 PRELE(info
->bigproc
);
1310 info
->bigsize
= size
;
1316 * This routine tries to maintain the pseudo LRU active queue,
1317 * so that during long periods of time where there is no paging,
1318 * that some statistic accumulation still occurs. This code
1319 * helps the situation where paging just starts to occur.
1322 vm_pageout_page_stats(void)
1325 int pcount
,tpcount
; /* Number of pages to check */
1326 static int fullintervalcount
= 0;
1330 (vmstats
.v_inactive_target
+ vmstats
.v_cache_max
+ vmstats
.v_free_min
) -
1331 (vmstats
.v_free_count
+ vmstats
.v_inactive_count
+ vmstats
.v_cache_count
);
1333 if (page_shortage
<= 0)
1338 pcount
= vmstats
.v_active_count
;
1339 fullintervalcount
+= vm_pageout_stats_interval
;
1340 if (fullintervalcount
< vm_pageout_full_stats_interval
) {
1341 tpcount
= (vm_pageout_stats_max
* vmstats
.v_active_count
) / vmstats
.v_page_count
;
1342 if (pcount
> tpcount
)
1345 fullintervalcount
= 0;
1348 m
= TAILQ_FIRST(&vm_page_queues
[PQ_ACTIVE
].pl
);
1349 while ((m
!= NULL
) && (pcount
-- > 0)) {
1352 if (m
->queue
!= PQ_ACTIVE
) {
1356 next
= TAILQ_NEXT(m
, pageq
);
1358 * Don't deactivate pages that are busy.
1360 if ((m
->busy
!= 0) ||
1361 (m
->flags
& PG_BUSY
) ||
1362 (m
->hold_count
!= 0)) {
1363 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1364 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1370 if (m
->flags
& PG_REFERENCED
) {
1371 vm_page_flag_clear(m
, PG_REFERENCED
);
1375 actcount
+= pmap_ts_referenced(m
);
1377 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1378 if (m
->act_count
> ACT_MAX
)
1379 m
->act_count
= ACT_MAX
;
1380 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1381 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1383 if (m
->act_count
== 0) {
1385 * We turn off page access, so that we have
1386 * more accurate RSS stats. We don't do this
1387 * in the normal page deactivation when the
1388 * system is loaded VM wise, because the
1389 * cost of the large number of page protect
1390 * operations would be higher than the value
1391 * of doing the operation.
1394 vm_page_protect(m
, VM_PROT_NONE
);
1396 vm_page_deactivate(m
);
1398 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1399 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1400 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1410 vm_pageout_free_page_calc(vm_size_t count
)
1412 if (count
< vmstats
.v_page_count
)
1415 * free_reserved needs to include enough for the largest swap pager
1416 * structures plus enough for any pv_entry structs when paging.
1418 if (vmstats
.v_page_count
> 1024)
1419 vmstats
.v_free_min
= 4 + (vmstats
.v_page_count
- 1024) / 200;
1421 vmstats
.v_free_min
= 4;
1422 vmstats
.v_pageout_free_min
= (2*MAXBSIZE
)/PAGE_SIZE
+
1423 vmstats
.v_interrupt_free_min
;
1424 vmstats
.v_free_reserved
= vm_pageout_page_count
+
1425 vmstats
.v_pageout_free_min
+ (count
/ 768) + PQ_L2_SIZE
;
1426 vmstats
.v_free_severe
= vmstats
.v_free_min
/ 2;
1427 vmstats
.v_free_min
+= vmstats
.v_free_reserved
;
1428 vmstats
.v_free_severe
+= vmstats
.v_free_reserved
;
1434 * vm_pageout is the high level pageout daemon.
1440 int inactive_shortage
;
1443 * Initialize some paging parameters.
1445 curthread
->td_flags
|= TDF_SYSTHREAD
;
1447 vmstats
.v_interrupt_free_min
= 2;
1448 if (vmstats
.v_page_count
< 2000)
1449 vm_pageout_page_count
= 8;
1451 vm_pageout_free_page_calc(vmstats
.v_page_count
);
1454 * v_free_target and v_cache_min control pageout hysteresis. Note
1455 * that these are more a measure of the VM cache queue hysteresis
1456 * then the VM free queue. Specifically, v_free_target is the
1457 * high water mark (free+cache pages).
1459 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1460 * low water mark, while v_free_min is the stop. v_cache_min must
1461 * be big enough to handle memory needs while the pageout daemon
1462 * is signalled and run to free more pages.
1464 if (vmstats
.v_free_count
> 6144)
1465 vmstats
.v_free_target
= 4 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1467 vmstats
.v_free_target
= 2 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1470 * NOTE: With the new buffer cache b_act_count we want the default
1471 * inactive target to be a percentage of available memory.
1473 * The inactive target essentially determines the minimum
1474 * number of 'temporary' pages capable of caching one-time-use
1475 * files when the VM system is otherwise full of pages
1476 * belonging to multi-time-use files or active program data.
1478 * NOTE: The inactive target is aggressively persued only if the
1479 * inactive queue becomes too small. If the inactive queue
1480 * is large enough to satisfy page movement to free+cache
1481 * then it is repopulated more slowly from the active queue.
1482 * This allows a generate inactive_target default to be set.
1484 * There is an issue here for processes which sit mostly idle
1485 * 'overnight', such as sshd, tcsh, and X. Any movement from
1486 * the active queue will eventually cause such pages to
1487 * recycle eventually causing a lot of paging in the morning.
1488 * To reduce the incidence of this pages cycled out of the
1489 * buffer cache are moved directly to the inactive queue if
1490 * they were only used once or twice. The vfs.vm_cycle_point
1491 * sysctl can be used to adjust this.
1493 if (vmstats
.v_free_count
> 2048) {
1494 vmstats
.v_cache_min
= vmstats
.v_free_target
;
1495 vmstats
.v_cache_max
= 2 * vmstats
.v_cache_min
;
1497 vmstats
.v_cache_min
= 0;
1498 vmstats
.v_cache_max
= 0;
1500 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 2;
1502 /* XXX does not really belong here */
1503 if (vm_page_max_wired
== 0)
1504 vm_page_max_wired
= vmstats
.v_free_count
/ 3;
1506 if (vm_pageout_stats_max
== 0)
1507 vm_pageout_stats_max
= vmstats
.v_free_target
;
1510 * Set interval in seconds for stats scan.
1512 if (vm_pageout_stats_interval
== 0)
1513 vm_pageout_stats_interval
= 5;
1514 if (vm_pageout_full_stats_interval
== 0)
1515 vm_pageout_full_stats_interval
= vm_pageout_stats_interval
* 4;
1519 * Set maximum free per pass
1521 if (vm_pageout_stats_free_max
== 0)
1522 vm_pageout_stats_free_max
= 5;
1524 swap_pager_swap_init();
1528 * The pageout daemon is never done, so loop forever.
1534 * Wait for an action request
1537 if (vm_pages_needed
== 0) {
1538 error
= tsleep(&vm_pages_needed
,
1540 vm_pageout_stats_interval
* hz
);
1541 if (error
&& vm_pages_needed
== 0) {
1542 vm_pageout_page_stats();
1545 vm_pages_needed
= 1;
1550 * If we have enough free memory, wakeup waiters.
1551 * (This is optional here)
1554 if (!vm_page_count_min(0))
1555 wakeup(&vmstats
.v_free_count
);
1556 mycpu
->gd_cnt
.v_pdwakeups
++;
1560 * Scan for pageout. Try to avoid thrashing the system
1563 inactive_shortage
= vm_pageout_scan(pass
);
1564 if (inactive_shortage
> 0) {
1566 if (swap_pager_full
) {
1568 * Running out of memory, catastrophic back-off
1569 * to one-second intervals.
1571 tsleep(&vm_pages_needed
, 0, "pdelay", hz
);
1572 } else if (pass
< 10 && vm_pages_needed
> 1) {
1574 * Normal operation, additional processes
1575 * have already kicked us. Retry immediately.
1577 } else if (pass
< 10) {
1579 * Normal operation, fewer processes. Delay
1580 * a bit but allow wakeups.
1582 vm_pages_needed
= 0;
1583 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
1584 vm_pages_needed
= 1;
1587 * We've taken too many passes, forced delay.
1589 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
1593 * Interlocked wakeup of waiters (non-optional)
1596 if (vm_pages_needed
&& !vm_page_count_min(0)) {
1597 wakeup(&vmstats
.v_free_count
);
1598 vm_pages_needed
= 0;
1605 * Called after allocating a page out of the cache or free queue
1606 * to possibly wake the pagedaemon up to replentish our supply.
1608 * We try to generate some hysteresis by waking the pagedaemon up
1609 * when our free+cache pages go below the severe level. The pagedaemon
1610 * tries to get the count back up to at least the minimum, and through
1611 * to the target level if possible.
1613 * If the pagedaemon is already active bump vm_pages_needed as a hint
1614 * that there are even more requests pending.
1617 pagedaemon_wakeup(void)
1619 if (vm_page_count_severe() && curthread
!= pagethread
) {
1620 if (vm_pages_needed
== 0) {
1621 vm_pages_needed
= 1;
1622 wakeup(&vm_pages_needed
);
1623 } else if (vm_page_count_min(0)) {
1629 #if !defined(NO_SWAPPING)
1631 vm_req_vmdaemon(void)
1633 static int lastrun
= 0;
1635 if ((ticks
> (lastrun
+ hz
)) || (ticks
< lastrun
)) {
1636 wakeup(&vm_daemon_needed
);
1641 static int vm_daemon_callback(struct proc
*p
, void *data __unused
);
1647 tsleep(&vm_daemon_needed
, 0, "psleep", 0);
1648 if (vm_pageout_req_swapout
) {
1649 swapout_procs(vm_pageout_req_swapout
);
1650 vm_pageout_req_swapout
= 0;
1653 * scan the processes for exceeding their rlimits or if
1654 * process is swapped out -- deactivate pages
1656 allproc_scan(vm_daemon_callback
, NULL
);
1661 vm_daemon_callback(struct proc
*p
, void *data __unused
)
1663 vm_pindex_t limit
, size
;
1666 * if this is a system process or if we have already
1667 * looked at this process, skip it.
1669 if (p
->p_flag
& (P_SYSTEM
| P_WEXIT
))
1673 * if the process is in a non-running type state,
1676 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
)
1682 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
1683 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
1686 * let processes that are swapped out really be
1687 * swapped out. Set the limit to nothing to get as
1688 * many pages out to swap as possible.
1690 if (p
->p_flag
& P_SWAPPEDOUT
)
1693 size
= vmspace_resident_count(p
->p_vmspace
);
1694 if (limit
>= 0 && size
>= limit
) {
1695 vm_pageout_map_deactivate_pages(
1696 &p
->p_vmspace
->vm_map
, limit
);