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 1.35 2008/05/09 07:24:48 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 void 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 extern int vm_swap_size
;
143 static int vm_max_launder
= 32;
144 static int vm_pageout_stats_max
=0, vm_pageout_stats_interval
= 0;
145 static int vm_pageout_full_stats_interval
= 0;
146 static int vm_pageout_stats_free_max
=0, vm_pageout_algorithm
=0;
147 static int defer_swap_pageouts
=0;
148 static int disable_swap_pageouts
=0;
150 #if defined(NO_SWAPPING)
151 static int vm_swap_enabled
=0;
152 static int vm_swap_idle_enabled
=0;
154 static int vm_swap_enabled
=1;
155 static int vm_swap_idle_enabled
=0;
158 SYSCTL_INT(_vm
, VM_PAGEOUT_ALGORITHM
, pageout_algorithm
,
159 CTLFLAG_RW
, &vm_pageout_algorithm
, 0, "LRU page mgmt");
161 SYSCTL_INT(_vm
, OID_AUTO
, max_launder
,
162 CTLFLAG_RW
, &vm_max_launder
, 0, "Limit dirty flushes in pageout");
164 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_max
,
165 CTLFLAG_RW
, &vm_pageout_stats_max
, 0, "Max pageout stats scan length");
167 SYSCTL_INT(_vm
, OID_AUTO
, pageout_full_stats_interval
,
168 CTLFLAG_RW
, &vm_pageout_full_stats_interval
, 0, "Interval for full stats scan");
170 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_interval
,
171 CTLFLAG_RW
, &vm_pageout_stats_interval
, 0, "Interval for partial stats scan");
173 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_free_max
,
174 CTLFLAG_RW
, &vm_pageout_stats_free_max
, 0, "Not implemented");
176 #if defined(NO_SWAPPING)
177 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
178 CTLFLAG_RD
, &vm_swap_enabled
, 0, "");
179 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
180 CTLFLAG_RD
, &vm_swap_idle_enabled
, 0, "");
182 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
183 CTLFLAG_RW
, &vm_swap_enabled
, 0, "Enable entire process swapout");
184 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
185 CTLFLAG_RW
, &vm_swap_idle_enabled
, 0, "Allow swapout on idle criteria");
188 SYSCTL_INT(_vm
, OID_AUTO
, defer_swapspace_pageouts
,
189 CTLFLAG_RW
, &defer_swap_pageouts
, 0, "Give preference to dirty pages in mem");
191 SYSCTL_INT(_vm
, OID_AUTO
, disable_swapspace_pageouts
,
192 CTLFLAG_RW
, &disable_swap_pageouts
, 0, "Disallow swapout of dirty pages");
194 static int pageout_lock_miss
;
195 SYSCTL_INT(_vm
, OID_AUTO
, pageout_lock_miss
,
196 CTLFLAG_RD
, &pageout_lock_miss
, 0, "vget() lock misses during pageout");
199 SYSCTL_INT(_vm
, OID_AUTO
, vm_load
,
200 CTLFLAG_RD
, &vm_load
, 0, "load on the VM system");
201 int vm_load_enable
= 1;
202 SYSCTL_INT(_vm
, OID_AUTO
, vm_load_enable
,
203 CTLFLAG_RW
, &vm_load_enable
, 0, "enable vm_load rate limiting");
206 SYSCTL_INT(_vm
, OID_AUTO
, vm_load_debug
,
207 CTLFLAG_RW
, &vm_load_debug
, 0, "debug vm_load");
210 #define VM_PAGEOUT_PAGE_COUNT 16
211 int vm_pageout_page_count
= VM_PAGEOUT_PAGE_COUNT
;
213 int vm_page_max_wired
; /* XXX max # of wired pages system-wide */
215 #if !defined(NO_SWAPPING)
216 typedef void freeer_fcn_t (vm_map_t
, vm_object_t
, vm_pindex_t
, int);
217 static void vm_pageout_map_deactivate_pages (vm_map_t
, vm_pindex_t
);
218 static freeer_fcn_t vm_pageout_object_deactivate_pages
;
219 static void vm_req_vmdaemon (void);
221 static void vm_pageout_page_stats(void);
227 vm_fault_ratecheck(void)
229 if (vm_pages_needed
) {
241 * Clean the page and remove it from the laundry. The page must not be
244 * We set the busy bit to cause potential page faults on this page to
245 * block. Note the careful timing, however, the busy bit isn't set till
246 * late and we cannot do anything that will mess with the page.
250 vm_pageout_clean(vm_page_t m
)
253 vm_page_t mc
[2*vm_pageout_page_count
];
255 int ib
, is
, page_base
;
256 vm_pindex_t pindex
= m
->pindex
;
261 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
262 * with the new swapper, but we could have serious problems paging
263 * out other object types if there is insufficient memory.
265 * Unfortunately, checking free memory here is far too late, so the
266 * check has been moved up a procedural level.
270 * Don't mess with the page if it's busy, held, or special
272 if ((m
->hold_count
!= 0) ||
273 ((m
->busy
!= 0) || (m
->flags
& (PG_BUSY
|PG_UNMANAGED
)))) {
277 mc
[vm_pageout_page_count
] = m
;
279 page_base
= vm_pageout_page_count
;
284 * Scan object for clusterable pages.
286 * We can cluster ONLY if: ->> the page is NOT
287 * clean, wired, busy, held, or mapped into a
288 * buffer, and one of the following:
289 * 1) The page is inactive, or a seldom used
292 * 2) we force the issue.
294 * During heavy mmap/modification loads the pageout
295 * daemon can really fragment the underlying file
296 * due to flushing pages out of order and not trying
297 * align the clusters (which leave sporatic out-of-order
298 * holes). To solve this problem we do the reverse scan
299 * first and attempt to align our cluster, then do a
300 * forward scan if room remains.
304 while (ib
&& pageout_count
< vm_pageout_page_count
) {
312 if ((p
= vm_page_lookup(object
, pindex
- ib
)) == NULL
) {
316 if (((p
->queue
- p
->pc
) == PQ_CACHE
) ||
317 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) || p
->busy
) {
321 vm_page_test_dirty(p
);
322 if ((p
->dirty
& p
->valid
) == 0 ||
323 p
->queue
!= PQ_INACTIVE
||
324 p
->wire_count
!= 0 || /* may be held by buf cache */
325 p
->hold_count
!= 0) { /* may be undergoing I/O */
333 * alignment boundry, stop here and switch directions. Do
336 if ((pindex
- (ib
- 1)) % vm_pageout_page_count
== 0)
340 while (pageout_count
< vm_pageout_page_count
&&
341 pindex
+ is
< object
->size
) {
344 if ((p
= vm_page_lookup(object
, pindex
+ is
)) == NULL
)
346 if (((p
->queue
- p
->pc
) == PQ_CACHE
) ||
347 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) || p
->busy
) {
350 vm_page_test_dirty(p
);
351 if ((p
->dirty
& p
->valid
) == 0 ||
352 p
->queue
!= PQ_INACTIVE
||
353 p
->wire_count
!= 0 || /* may be held by buf cache */
354 p
->hold_count
!= 0) { /* may be undergoing I/O */
357 mc
[page_base
+ pageout_count
] = p
;
363 * If we exhausted our forward scan, continue with the reverse scan
364 * when possible, even past a page boundry. This catches boundry
367 if (ib
&& pageout_count
< vm_pageout_page_count
)
371 * we allow reads during pageouts...
373 return vm_pageout_flush(&mc
[page_base
], pageout_count
, 0);
377 * vm_pageout_flush() - launder the given pages
379 * The given pages are laundered. Note that we setup for the start of
380 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
381 * reference count all in here rather then in the parent. If we want
382 * the parent to do more sophisticated things we may have to change
387 vm_pageout_flush(vm_page_t
*mc
, int count
, int flags
)
390 int pageout_status
[count
];
395 * Initiate I/O. Bump the vm_page_t->busy counter.
397 for (i
= 0; i
< count
; i
++) {
398 KASSERT(mc
[i
]->valid
== VM_PAGE_BITS_ALL
, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc
[i
], i
, count
));
399 vm_page_io_start(mc
[i
]);
403 * We must make the pages read-only. This will also force the
404 * modified bit in the related pmaps to be cleared. The pager
405 * cannot clear the bit for us since the I/O completion code
406 * typically runs from an interrupt. The act of making the page
407 * read-only handles the case for us.
409 for (i
= 0; i
< count
; i
++) {
410 vm_page_protect(mc
[i
], VM_PROT_READ
);
413 object
= mc
[0]->object
;
414 vm_object_pip_add(object
, count
);
416 vm_pager_put_pages(object
, mc
, count
,
417 (flags
| ((object
== &kernel_object
) ? VM_PAGER_PUT_SYNC
: 0)),
420 for (i
= 0; i
< count
; i
++) {
421 vm_page_t mt
= mc
[i
];
423 switch (pageout_status
[i
]) {
432 * Page outside of range of object. Right now we
433 * essentially lose the changes by pretending it
436 pmap_clear_modify(mt
);
442 * If page couldn't be paged out, then reactivate the
443 * page so it doesn't clog the inactive list. (We
444 * will try paging out it again later).
446 vm_page_activate(mt
);
453 * If the operation is still going, leave the page busy to
454 * block all other accesses. Also, leave the paging in
455 * progress indicator set so that we don't attempt an object
458 * For any pages which have completed synchronously,
459 * deactivate the page if we are under a severe deficit.
460 * Do not try to enter them into the cache, though, they
461 * might still be read-heavy.
463 if (pageout_status
[i
] != VM_PAGER_PEND
) {
464 vm_object_pip_wakeup(object
);
465 vm_page_io_finish(mt
);
466 if (vm_page_count_severe())
467 vm_page_deactivate(mt
);
469 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt
))
470 vm_page_protect(mt
, VM_PROT_READ
);
477 #if !defined(NO_SWAPPING)
479 * vm_pageout_object_deactivate_pages
481 * deactivate enough pages to satisfy the inactive target
482 * requirements or if vm_page_proc_limit is set, then
483 * deactivate all of the pages in the object and its
486 * The object and map must be locked.
488 static int vm_pageout_object_deactivate_pages_callback(vm_page_t
, void *);
491 vm_pageout_object_deactivate_pages(vm_map_t map
, vm_object_t object
,
492 vm_pindex_t desired
, int map_remove_only
)
494 struct rb_vm_page_scan_info info
;
497 if (object
->type
== OBJT_DEVICE
|| object
->type
== OBJT_PHYS
)
501 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
503 if (object
->paging_in_progress
)
506 remove_mode
= map_remove_only
;
507 if (object
->shadow_count
> 1)
511 * scan the objects entire memory queue. spl protection is
512 * required to avoid an interrupt unbusy/free race against
516 info
.limit
= remove_mode
;
518 info
.desired
= desired
;
519 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
520 vm_pageout_object_deactivate_pages_callback
,
524 object
= object
->backing_object
;
529 vm_pageout_object_deactivate_pages_callback(vm_page_t p
, void *data
)
531 struct rb_vm_page_scan_info
*info
= data
;
534 if (pmap_resident_count(vm_map_pmap(info
->map
)) <= info
->desired
) {
537 mycpu
->gd_cnt
.v_pdpages
++;
538 if (p
->wire_count
!= 0 || p
->hold_count
!= 0 || p
->busy
!= 0 ||
539 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) ||
540 !pmap_page_exists_quick(vm_map_pmap(info
->map
), p
)) {
544 actcount
= pmap_ts_referenced(p
);
546 vm_page_flag_set(p
, PG_REFERENCED
);
547 } else if (p
->flags
& PG_REFERENCED
) {
551 if ((p
->queue
!= PQ_ACTIVE
) &&
552 (p
->flags
& PG_REFERENCED
)) {
554 p
->act_count
+= actcount
;
555 vm_page_flag_clear(p
, PG_REFERENCED
);
556 } else if (p
->queue
== PQ_ACTIVE
) {
557 if ((p
->flags
& PG_REFERENCED
) == 0) {
558 p
->act_count
-= min(p
->act_count
, ACT_DECLINE
);
559 if (!info
->limit
&& (vm_pageout_algorithm
|| (p
->act_count
== 0))) {
561 vm_page_protect(p
, VM_PROT_NONE
);
563 vm_page_deactivate(p
);
565 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
566 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
570 vm_page_flag_clear(p
, PG_REFERENCED
);
571 if (p
->act_count
< (ACT_MAX
- ACT_ADVANCE
))
572 p
->act_count
+= ACT_ADVANCE
;
573 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
574 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
576 } else if (p
->queue
== PQ_INACTIVE
) {
578 vm_page_protect(p
, VM_PROT_NONE
);
585 * deactivate some number of pages in a map, try to do it fairly, but
586 * that is really hard to do.
589 vm_pageout_map_deactivate_pages(vm_map_t map
, vm_pindex_t desired
)
592 vm_object_t obj
, bigobj
;
595 if (lockmgr(&map
->lock
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
603 * first, search out the biggest object, and try to free pages from
606 tmpe
= map
->header
.next
;
607 while (tmpe
!= &map
->header
) {
608 switch(tmpe
->maptype
) {
609 case VM_MAPTYPE_NORMAL
:
610 case VM_MAPTYPE_VPAGETABLE
:
611 obj
= tmpe
->object
.vm_object
;
612 if ((obj
!= NULL
) && (obj
->shadow_count
<= 1) &&
614 (bigobj
->resident_page_count
< obj
->resident_page_count
))) {
621 if (tmpe
->wired_count
> 0)
622 nothingwired
= FALSE
;
627 vm_pageout_object_deactivate_pages(map
, bigobj
, desired
, 0);
630 * Next, hunt around for other pages to deactivate. We actually
631 * do this search sort of wrong -- .text first is not the best idea.
633 tmpe
= map
->header
.next
;
634 while (tmpe
!= &map
->header
) {
635 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
637 switch(tmpe
->maptype
) {
638 case VM_MAPTYPE_NORMAL
:
639 case VM_MAPTYPE_VPAGETABLE
:
640 obj
= tmpe
->object
.vm_object
;
642 vm_pageout_object_deactivate_pages(map
, obj
, desired
, 0);
651 * Remove all mappings if a process is swapped out, this will free page
654 if (desired
== 0 && nothingwired
)
655 pmap_remove(vm_map_pmap(map
),
656 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
662 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We
663 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
664 * be trivially freed.
667 vm_pageout_page_free(vm_page_t m
)
669 vm_object_t object
= m
->object
;
670 int type
= object
->type
;
672 if (type
== OBJT_SWAP
|| type
== OBJT_DEFAULT
)
673 vm_object_reference(object
);
675 vm_page_protect(m
, VM_PROT_NONE
);
677 if (type
== OBJT_SWAP
|| type
== OBJT_DEFAULT
)
678 vm_object_deallocate(object
);
682 * vm_pageout_scan does the dirty work for the pageout daemon.
685 struct vm_pageout_scan_info
{
686 struct proc
*bigproc
;
690 static int vm_pageout_scan_callback(struct proc
*p
, void *data
);
693 vm_pageout_scan(int pass
)
695 struct vm_pageout_scan_info info
;
697 struct vm_page marker
;
698 int page_shortage
, maxscan
, pcount
;
699 int addl_page_shortage
, addl_page_shortage_init
;
702 int vnodes_skipped
= 0;
706 * Do whatever cleanup that the pmap code can.
710 addl_page_shortage_init
= vm_pageout_deficit
;
711 vm_pageout_deficit
= 0;
714 * Calculate the number of pages we want to either free or move
717 page_shortage
= vm_paging_target() + addl_page_shortage_init
;
720 * Initialize our marker
722 bzero(&marker
, sizeof(marker
));
723 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
724 marker
.queue
= PQ_INACTIVE
;
725 marker
.wire_count
= 1;
728 * Start scanning the inactive queue for pages we can move to the
729 * cache or free. The scan will stop when the target is reached or
730 * we have scanned the entire inactive queue. Note that m->act_count
731 * is not used to form decisions for the inactive queue, only for the
734 * maxlaunder limits the number of dirty pages we flush per scan.
735 * For most systems a smaller value (16 or 32) is more robust under
736 * extreme memory and disk pressure because any unnecessary writes
737 * to disk can result in extreme performance degredation. However,
738 * systems with excessive dirty pages (especially when MAP_NOSYNC is
739 * used) will die horribly with limited laundering. If the pageout
740 * daemon cannot clean enough pages in the first pass, we let it go
741 * all out in succeeding passes.
743 if ((maxlaunder
= vm_max_launder
) <= 1)
749 * We will generally be in a critical section throughout the
750 * scan, but we can release it temporarily when we are sitting on a
751 * non-busy page without fear. this is required to prevent an
752 * interrupt from unbusying or freeing a page prior to our busy
753 * check, leaving us on the wrong queue or checking the wrong
758 addl_page_shortage
= addl_page_shortage_init
;
759 maxscan
= vmstats
.v_inactive_count
;
760 for (m
= TAILQ_FIRST(&vm_page_queues
[PQ_INACTIVE
].pl
);
761 m
!= NULL
&& maxscan
-- > 0 && page_shortage
> 0;
764 mycpu
->gd_cnt
.v_pdpages
++;
767 * Give interrupts a chance
773 * It's easier for some of the conditions below to just loop
774 * and catch queue changes here rather then check everywhere
777 if (m
->queue
!= PQ_INACTIVE
)
779 next
= TAILQ_NEXT(m
, pageq
);
784 if (m
->flags
& PG_MARKER
)
788 * A held page may be undergoing I/O, so skip it.
791 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
792 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
793 addl_page_shortage
++;
798 * Dont mess with busy pages, keep in the front of the
799 * queue, most likely are being paged out.
801 if (m
->busy
|| (m
->flags
& PG_BUSY
)) {
802 addl_page_shortage
++;
806 if (m
->object
->ref_count
== 0) {
808 * If the object is not being used, we ignore previous
811 vm_page_flag_clear(m
, PG_REFERENCED
);
812 pmap_clear_reference(m
);
814 } else if (((m
->flags
& PG_REFERENCED
) == 0) &&
815 (actcount
= pmap_ts_referenced(m
))) {
817 * Otherwise, if the page has been referenced while
818 * in the inactive queue, we bump the "activation
819 * count" upwards, making it less likely that the
820 * page will be added back to the inactive queue
821 * prematurely again. Here we check the page tables
822 * (or emulated bits, if any), given the upper level
823 * VM system not knowing anything about existing
827 m
->act_count
+= (actcount
+ ACT_ADVANCE
);
832 * If the upper level VM system knows about any page
833 * references, we activate the page. We also set the
834 * "activation count" higher than normal so that we will less
835 * likely place pages back onto the inactive queue again.
837 if ((m
->flags
& PG_REFERENCED
) != 0) {
838 vm_page_flag_clear(m
, PG_REFERENCED
);
839 actcount
= pmap_ts_referenced(m
);
841 m
->act_count
+= (actcount
+ ACT_ADVANCE
+ 1);
846 * If the upper level VM system doesn't know anything about
847 * the page being dirty, we have to check for it again. As
848 * far as the VM code knows, any partially dirty pages are
851 * Pages marked PG_WRITEABLE may be mapped into the user
852 * address space of a process running on another cpu. A
853 * user process (without holding the MP lock) running on
854 * another cpu may be able to touch the page while we are
855 * trying to remove it. vm_page_cache() will handle this
859 vm_page_test_dirty(m
);
861 if (m
->dirty
== 0 && (m
->flags
& PG_WRITEABLE
) != 0)
870 * Invalid pages can be easily freed
872 vm_pageout_page_free(m
);
873 mycpu
->gd_cnt
.v_dfree
++;
875 } else if (m
->dirty
== 0) {
877 * Clean pages can be placed onto the cache queue.
878 * This effectively frees them.
882 } else if ((m
->flags
& PG_WINATCFLS
) == 0 && pass
== 0) {
884 * Dirty pages need to be paged out, but flushing
885 * a page is extremely expensive verses freeing
886 * a clean page. Rather then artificially limiting
887 * the number of pages we can flush, we instead give
888 * dirty pages extra priority on the inactive queue
889 * by forcing them to be cycled through the queue
890 * twice before being flushed, after which the
891 * (now clean) page will cycle through once more
892 * before being freed. This significantly extends
893 * the thrash point for a heavily loaded machine.
895 vm_page_flag_set(m
, PG_WINATCFLS
);
896 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
897 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
898 } else if (maxlaunder
> 0) {
900 * We always want to try to flush some dirty pages if
901 * we encounter them, to keep the system stable.
902 * Normally this number is small, but under extreme
903 * pressure where there are insufficient clean pages
904 * on the inactive queue, we may have to go all out.
906 int swap_pageouts_ok
;
907 struct vnode
*vp
= NULL
;
911 if ((object
->type
!= OBJT_SWAP
) && (object
->type
!= OBJT_DEFAULT
)) {
912 swap_pageouts_ok
= 1;
914 swap_pageouts_ok
= !(defer_swap_pageouts
|| disable_swap_pageouts
);
915 swap_pageouts_ok
|= (!disable_swap_pageouts
&& defer_swap_pageouts
&&
916 vm_page_count_min());
921 * We don't bother paging objects that are "dead".
922 * Those objects are in a "rundown" state.
924 if (!swap_pageouts_ok
|| (object
->flags
& OBJ_DEAD
)) {
925 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
926 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
931 * The object is already known NOT to be dead. It
932 * is possible for the vget() to block the whole
933 * pageout daemon, but the new low-memory handling
934 * code should prevent it.
936 * The previous code skipped locked vnodes and, worse,
937 * reordered pages in the queue. This results in
938 * completely non-deterministic operation because,
939 * quite often, a vm_fault has initiated an I/O and
940 * is holding a locked vnode at just the point where
941 * the pageout daemon is woken up.
943 * We can't wait forever for the vnode lock, we might
944 * deadlock due to a vn_read() getting stuck in
945 * vm_wait while holding this vnode. We skip the
946 * vnode if we can't get it in a reasonable amount
950 if (object
->type
== OBJT_VNODE
) {
953 if (vget(vp
, LK_EXCLUSIVE
|LK_NOOBJ
|LK_TIMELOCK
)) {
955 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
961 * The page might have been moved to another
962 * queue during potential blocking in vget()
963 * above. The page might have been freed and
964 * reused for another vnode. The object might
965 * have been reused for another vnode.
967 if (m
->queue
!= PQ_INACTIVE
||
968 m
->object
!= object
||
969 object
->handle
!= vp
) {
970 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
977 * The page may have been busied during the
978 * blocking in vput(); We don't move the
979 * page back onto the end of the queue so that
980 * statistics are more correct if we don't.
982 if (m
->busy
|| (m
->flags
& PG_BUSY
)) {
988 * If the page has become held it might
989 * be undergoing I/O, so skip it
992 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
993 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
994 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1002 * If a page is dirty, then it is either being washed
1003 * (but not yet cleaned) or it is still in the
1004 * laundry. If it is still in the laundry, then we
1005 * start the cleaning operation.
1007 * This operation may cluster, invalidating the 'next'
1008 * pointer. To prevent an inordinate number of
1009 * restarts we use our marker to remember our place.
1011 * decrement page_shortage on success to account for
1012 * the (future) cleaned page. Otherwise we could wind
1013 * up laundering or cleaning too many pages.
1015 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, &marker
, pageq
);
1016 if (vm_pageout_clean(m
) != 0) {
1020 next
= TAILQ_NEXT(&marker
, pageq
);
1021 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, &marker
, pageq
);
1028 * Compute the number of pages we want to try to move from the
1029 * active queue to the inactive queue.
1031 page_shortage
= vm_paging_target() +
1032 vmstats
.v_inactive_target
- vmstats
.v_inactive_count
;
1033 page_shortage
+= addl_page_shortage
;
1036 * Scan the active queue for things we can deactivate. We nominally
1037 * track the per-page activity counter and use it to locate
1038 * deactivation candidates.
1040 * NOTE: we are still in a critical section.
1042 pcount
= vmstats
.v_active_count
;
1043 m
= TAILQ_FIRST(&vm_page_queues
[PQ_ACTIVE
].pl
);
1045 while ((m
!= NULL
) && (pcount
-- > 0) && (page_shortage
> 0)) {
1047 * Give interrupts a chance.
1053 * If the page was ripped out from under us, just stop.
1055 if (m
->queue
!= PQ_ACTIVE
)
1057 next
= TAILQ_NEXT(m
, pageq
);
1060 * Don't deactivate pages that are busy.
1062 if ((m
->busy
!= 0) ||
1063 (m
->flags
& PG_BUSY
) ||
1064 (m
->hold_count
!= 0)) {
1065 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1066 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1072 * The count for pagedaemon pages is done after checking the
1073 * page for eligibility...
1075 mycpu
->gd_cnt
.v_pdpages
++;
1078 * Check to see "how much" the page has been used.
1081 if (m
->object
->ref_count
!= 0) {
1082 if (m
->flags
& PG_REFERENCED
) {
1085 actcount
+= pmap_ts_referenced(m
);
1087 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1088 if (m
->act_count
> ACT_MAX
)
1089 m
->act_count
= ACT_MAX
;
1094 * Since we have "tested" this bit, we need to clear it now.
1096 vm_page_flag_clear(m
, PG_REFERENCED
);
1099 * Only if an object is currently being used, do we use the
1100 * page activation count stats.
1102 if (actcount
&& (m
->object
->ref_count
!= 0)) {
1103 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1104 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1106 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1107 if (vm_pageout_algorithm
||
1108 m
->object
->ref_count
== 0 ||
1109 m
->act_count
< pass
) {
1111 if (m
->object
->ref_count
== 0) {
1113 vm_page_protect(m
, VM_PROT_NONE
);
1118 vm_page_deactivate(m
);
1120 vm_page_deactivate(m
);
1123 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1124 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1131 * We try to maintain some *really* free pages, this allows interrupt
1132 * code to be guaranteed space. Since both cache and free queues
1133 * are considered basically 'free', moving pages from cache to free
1134 * does not effect other calculations.
1136 * NOTE: we are still in a critical section.
1139 while (vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
1140 static int cache_rover
= 0;
1141 m
= vm_page_list_find(PQ_CACHE
, cache_rover
, FALSE
);
1144 if ((m
->flags
& (PG_BUSY
|PG_UNMANAGED
)) ||
1149 kprintf("Warning: busy page %p found in cache\n", m
);
1151 vm_page_deactivate(m
);
1154 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
1155 KKASSERT(m
->dirty
== 0);
1156 cache_rover
= (cache_rover
+ PQ_PRIME2
) & PQ_L2_MASK
;
1157 vm_pageout_page_free(m
);
1158 mycpu
->gd_cnt
.v_dfree
++;
1163 #if !defined(NO_SWAPPING)
1165 * Idle process swapout -- run once per second.
1167 if (vm_swap_idle_enabled
) {
1169 if (time_second
!= lsec
) {
1170 vm_pageout_req_swapout
|= VM_SWAP_IDLE
;
1178 * If we didn't get enough free pages, and we have skipped a vnode
1179 * in a writeable object, wakeup the sync daemon. And kick swapout
1180 * if we did not get enough free pages.
1182 if (vm_paging_target() > 0) {
1183 if (vnodes_skipped
&& vm_page_count_min())
1185 #if !defined(NO_SWAPPING)
1186 if (vm_swap_enabled
&& vm_page_count_target()) {
1188 vm_pageout_req_swapout
|= VM_SWAP_NORMAL
;
1194 * If we are out of swap and were not able to reach our paging
1195 * target, kill the largest process.
1197 if ((vm_swap_size
< 64 && vm_page_count_min()) ||
1198 (swap_pager_full
&& vm_paging_target() > 0)) {
1200 if ((vm_swap_size
< 64 || swap_pager_full
) && vm_page_count_min()) {
1202 info
.bigproc
= NULL
;
1204 allproc_scan(vm_pageout_scan_callback
, &info
);
1205 if (info
.bigproc
!= NULL
) {
1206 killproc(info
.bigproc
, "out of swap space");
1207 info
.bigproc
->p_nice
= PRIO_MIN
;
1208 info
.bigproc
->p_usched
->resetpriority(
1209 FIRST_LWP_IN_PROC(info
.bigproc
));
1210 wakeup(&vmstats
.v_free_count
);
1211 PRELE(info
.bigproc
);
1217 vm_pageout_scan_callback(struct proc
*p
, void *data
)
1219 struct vm_pageout_scan_info
*info
= data
;
1223 * if this is a system process, skip it
1225 if ((p
->p_flag
& P_SYSTEM
) || (p
->p_pid
== 1) ||
1226 ((p
->p_pid
< 48) && (vm_swap_size
!= 0))) {
1231 * if the process is in a non-running type state,
1234 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
) {
1239 * get the process size
1241 size
= vmspace_resident_count(p
->p_vmspace
) +
1242 vmspace_swap_count(p
->p_vmspace
);
1245 * If the this process is bigger than the biggest one
1248 if (size
> info
->bigsize
) {
1250 PRELE(info
->bigproc
);
1253 info
->bigsize
= size
;
1259 * This routine tries to maintain the pseudo LRU active queue,
1260 * so that during long periods of time where there is no paging,
1261 * that some statistic accumulation still occurs. This code
1262 * helps the situation where paging just starts to occur.
1265 vm_pageout_page_stats(void)
1268 int pcount
,tpcount
; /* Number of pages to check */
1269 static int fullintervalcount
= 0;
1273 (vmstats
.v_inactive_target
+ vmstats
.v_cache_max
+ vmstats
.v_free_min
) -
1274 (vmstats
.v_free_count
+ vmstats
.v_inactive_count
+ vmstats
.v_cache_count
);
1276 if (page_shortage
<= 0)
1281 pcount
= vmstats
.v_active_count
;
1282 fullintervalcount
+= vm_pageout_stats_interval
;
1283 if (fullintervalcount
< vm_pageout_full_stats_interval
) {
1284 tpcount
= (vm_pageout_stats_max
* vmstats
.v_active_count
) / vmstats
.v_page_count
;
1285 if (pcount
> tpcount
)
1288 fullintervalcount
= 0;
1291 m
= TAILQ_FIRST(&vm_page_queues
[PQ_ACTIVE
].pl
);
1292 while ((m
!= NULL
) && (pcount
-- > 0)) {
1295 if (m
->queue
!= PQ_ACTIVE
) {
1299 next
= TAILQ_NEXT(m
, pageq
);
1301 * Don't deactivate pages that are busy.
1303 if ((m
->busy
!= 0) ||
1304 (m
->flags
& PG_BUSY
) ||
1305 (m
->hold_count
!= 0)) {
1306 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1307 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1313 if (m
->flags
& PG_REFERENCED
) {
1314 vm_page_flag_clear(m
, PG_REFERENCED
);
1318 actcount
+= pmap_ts_referenced(m
);
1320 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1321 if (m
->act_count
> ACT_MAX
)
1322 m
->act_count
= ACT_MAX
;
1323 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1324 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1326 if (m
->act_count
== 0) {
1328 * We turn off page access, so that we have
1329 * more accurate RSS stats. We don't do this
1330 * in the normal page deactivation when the
1331 * system is loaded VM wise, because the
1332 * cost of the large number of page protect
1333 * operations would be higher than the value
1334 * of doing the operation.
1337 vm_page_protect(m
, VM_PROT_NONE
);
1339 vm_page_deactivate(m
);
1341 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1342 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1343 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1353 vm_pageout_free_page_calc(vm_size_t count
)
1355 if (count
< vmstats
.v_page_count
)
1358 * free_reserved needs to include enough for the largest swap pager
1359 * structures plus enough for any pv_entry structs when paging.
1361 if (vmstats
.v_page_count
> 1024)
1362 vmstats
.v_free_min
= 4 + (vmstats
.v_page_count
- 1024) / 200;
1364 vmstats
.v_free_min
= 4;
1365 vmstats
.v_pageout_free_min
= (2*MAXBSIZE
)/PAGE_SIZE
+
1366 vmstats
.v_interrupt_free_min
;
1367 vmstats
.v_free_reserved
= vm_pageout_page_count
+
1368 vmstats
.v_pageout_free_min
+ (count
/ 768) + PQ_L2_SIZE
;
1369 vmstats
.v_free_severe
= vmstats
.v_free_min
/ 2;
1370 vmstats
.v_free_min
+= vmstats
.v_free_reserved
;
1371 vmstats
.v_free_severe
+= vmstats
.v_free_reserved
;
1377 * vm_pageout is the high level pageout daemon.
1385 * Initialize some paging parameters.
1388 vmstats
.v_interrupt_free_min
= 2;
1389 if (vmstats
.v_page_count
< 2000)
1390 vm_pageout_page_count
= 8;
1392 vm_pageout_free_page_calc(vmstats
.v_page_count
);
1394 * v_free_target and v_cache_min control pageout hysteresis. Note
1395 * that these are more a measure of the VM cache queue hysteresis
1396 * then the VM free queue. Specifically, v_free_target is the
1397 * high water mark (free+cache pages).
1399 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1400 * low water mark, while v_free_min is the stop. v_cache_min must
1401 * be big enough to handle memory needs while the pageout daemon
1402 * is signalled and run to free more pages.
1404 if (vmstats
.v_free_count
> 6144)
1405 vmstats
.v_free_target
= 4 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1407 vmstats
.v_free_target
= 2 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1409 if (vmstats
.v_free_count
> 2048) {
1410 vmstats
.v_cache_min
= vmstats
.v_free_target
;
1411 vmstats
.v_cache_max
= 2 * vmstats
.v_cache_min
;
1412 vmstats
.v_inactive_target
= (3 * vmstats
.v_free_target
) / 2;
1414 vmstats
.v_cache_min
= 0;
1415 vmstats
.v_cache_max
= 0;
1416 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 4;
1418 if (vmstats
.v_inactive_target
> vmstats
.v_free_count
/ 3)
1419 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 3;
1421 /* XXX does not really belong here */
1422 if (vm_page_max_wired
== 0)
1423 vm_page_max_wired
= vmstats
.v_free_count
/ 3;
1425 if (vm_pageout_stats_max
== 0)
1426 vm_pageout_stats_max
= vmstats
.v_free_target
;
1429 * Set interval in seconds for stats scan.
1431 if (vm_pageout_stats_interval
== 0)
1432 vm_pageout_stats_interval
= 5;
1433 if (vm_pageout_full_stats_interval
== 0)
1434 vm_pageout_full_stats_interval
= vm_pageout_stats_interval
* 4;
1438 * Set maximum free per pass
1440 if (vm_pageout_stats_free_max
== 0)
1441 vm_pageout_stats_free_max
= 5;
1443 swap_pager_swap_init();
1446 * The pageout daemon is never done, so loop forever.
1452 * If we have enough free memory, wakeup waiters. Do
1453 * not clear vm_pages_needed until we reach our target,
1454 * otherwise we may be woken up over and over again and
1455 * waste a lot of cpu.
1458 if (vm_pages_needed
&& !vm_page_count_min()) {
1459 if (vm_paging_needed() <= 0)
1460 vm_pages_needed
= 0;
1461 wakeup(&vmstats
.v_free_count
);
1463 if (vm_pages_needed
) {
1465 * Still not done, take a second pass without waiting
1466 * (unlimited dirty cleaning), otherwise sleep a bit
1471 tsleep(&vm_pages_needed
, 0, "psleep", hz
/2);
1474 * Good enough, sleep & handle stats. Prime the pass
1481 error
= tsleep(&vm_pages_needed
,
1482 0, "psleep", vm_pageout_stats_interval
* hz
);
1483 if (error
&& !vm_pages_needed
) {
1486 vm_pageout_page_stats();
1491 if (vm_pages_needed
)
1492 mycpu
->gd_cnt
.v_pdwakeups
++;
1494 vm_pageout_scan(pass
);
1495 vm_pageout_deficit
= 0;
1500 pagedaemon_wakeup(void)
1502 if (!vm_pages_needed
&& curthread
!= pagethread
) {
1504 wakeup(&vm_pages_needed
);
1508 #if !defined(NO_SWAPPING)
1510 vm_req_vmdaemon(void)
1512 static int lastrun
= 0;
1514 if ((ticks
> (lastrun
+ hz
)) || (ticks
< lastrun
)) {
1515 wakeup(&vm_daemon_needed
);
1520 static int vm_daemon_callback(struct proc
*p
, void *data __unused
);
1526 tsleep(&vm_daemon_needed
, 0, "psleep", 0);
1527 if (vm_pageout_req_swapout
) {
1528 swapout_procs(vm_pageout_req_swapout
);
1529 vm_pageout_req_swapout
= 0;
1532 * scan the processes for exceeding their rlimits or if
1533 * process is swapped out -- deactivate pages
1535 allproc_scan(vm_daemon_callback
, NULL
);
1540 vm_daemon_callback(struct proc
*p
, void *data __unused
)
1542 vm_pindex_t limit
, size
;
1545 * if this is a system process or if we have already
1546 * looked at this process, skip it.
1548 if (p
->p_flag
& (P_SYSTEM
| P_WEXIT
))
1552 * if the process is in a non-running type state,
1555 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
)
1561 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
1562 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
1565 * let processes that are swapped out really be
1566 * swapped out. Set the limit to nothing to get as
1567 * many pages out to swap as possible.
1569 if (p
->p_flag
& P_SWAPPEDOUT
)
1572 size
= vmspace_resident_count(p
->p_vmspace
);
1573 if (limit
>= 0 && size
>= limit
) {
1574 vm_pageout_map_deactivate_pages(
1575 &p
->p_vmspace
->vm_map
, limit
);