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 1.33 2008/03/20 06:02:50 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 and
396 * mark the pages read-only.
398 * We must make the pages read-only. This will also force the
399 * modified bit in the related pmaps to be cleared. The pager
400 * cannot clear the bit for us since the I/O completion code
401 * typically runs from an interrupt. The act of making the page
402 * read-only handles the case for us.
405 for (i
= 0; i
< count
; i
++) {
406 KASSERT(mc
[i
]->valid
== VM_PAGE_BITS_ALL
, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc
[i
], i
, count
));
407 vm_page_io_start(mc
[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 * If page couldn't be paged out, then reactivate the
441 * page so it doesn't clog the inactive list. (We
442 * will try paging out it again later).
444 vm_page_activate(mt
);
451 * If the operation is still going, leave the page busy to
452 * block all other accesses. Also, leave the paging in
453 * progress indicator set so that we don't attempt an object
456 if (pageout_status
[i
] != VM_PAGER_PEND
) {
457 vm_object_pip_wakeup(object
);
458 vm_page_io_finish(mt
);
459 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt
))
460 vm_page_protect(mt
, VM_PROT_READ
);
466 #if !defined(NO_SWAPPING)
468 * vm_pageout_object_deactivate_pages
470 * deactivate enough pages to satisfy the inactive target
471 * requirements or if vm_page_proc_limit is set, then
472 * deactivate all of the pages in the object and its
475 * The object and map must be locked.
477 static int vm_pageout_object_deactivate_pages_callback(vm_page_t
, void *);
480 vm_pageout_object_deactivate_pages(vm_map_t map
, vm_object_t object
,
481 vm_pindex_t desired
, int map_remove_only
)
483 struct rb_vm_page_scan_info info
;
486 if (object
->type
== OBJT_DEVICE
|| object
->type
== OBJT_PHYS
)
490 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
492 if (object
->paging_in_progress
)
495 remove_mode
= map_remove_only
;
496 if (object
->shadow_count
> 1)
500 * scan the objects entire memory queue. spl protection is
501 * required to avoid an interrupt unbusy/free race against
505 info
.limit
= remove_mode
;
507 info
.desired
= desired
;
508 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
509 vm_pageout_object_deactivate_pages_callback
,
513 object
= object
->backing_object
;
518 vm_pageout_object_deactivate_pages_callback(vm_page_t p
, void *data
)
520 struct rb_vm_page_scan_info
*info
= data
;
523 if (pmap_resident_count(vm_map_pmap(info
->map
)) <= info
->desired
) {
526 mycpu
->gd_cnt
.v_pdpages
++;
527 if (p
->wire_count
!= 0 || p
->hold_count
!= 0 || p
->busy
!= 0 ||
528 (p
->flags
& (PG_BUSY
|PG_UNMANAGED
)) ||
529 !pmap_page_exists_quick(vm_map_pmap(info
->map
), p
)) {
533 actcount
= pmap_ts_referenced(p
);
535 vm_page_flag_set(p
, PG_REFERENCED
);
536 } else if (p
->flags
& PG_REFERENCED
) {
540 if ((p
->queue
!= PQ_ACTIVE
) &&
541 (p
->flags
& PG_REFERENCED
)) {
543 p
->act_count
+= actcount
;
544 vm_page_flag_clear(p
, PG_REFERENCED
);
545 } else if (p
->queue
== PQ_ACTIVE
) {
546 if ((p
->flags
& PG_REFERENCED
) == 0) {
547 p
->act_count
-= min(p
->act_count
, ACT_DECLINE
);
548 if (!info
->limit
&& (vm_pageout_algorithm
|| (p
->act_count
== 0))) {
549 vm_page_protect(p
, VM_PROT_NONE
);
550 vm_page_deactivate(p
);
552 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
553 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
557 vm_page_flag_clear(p
, PG_REFERENCED
);
558 if (p
->act_count
< (ACT_MAX
- ACT_ADVANCE
))
559 p
->act_count
+= ACT_ADVANCE
;
560 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
561 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, p
, pageq
);
563 } else if (p
->queue
== PQ_INACTIVE
) {
564 vm_page_protect(p
, VM_PROT_NONE
);
570 * deactivate some number of pages in a map, try to do it fairly, but
571 * that is really hard to do.
574 vm_pageout_map_deactivate_pages(vm_map_t map
, vm_pindex_t desired
)
577 vm_object_t obj
, bigobj
;
580 if (lockmgr(&map
->lock
, LK_EXCLUSIVE
| LK_NOWAIT
)) {
588 * first, search out the biggest object, and try to free pages from
591 tmpe
= map
->header
.next
;
592 while (tmpe
!= &map
->header
) {
593 switch(tmpe
->maptype
) {
594 case VM_MAPTYPE_NORMAL
:
595 case VM_MAPTYPE_VPAGETABLE
:
596 obj
= tmpe
->object
.vm_object
;
597 if ((obj
!= NULL
) && (obj
->shadow_count
<= 1) &&
599 (bigobj
->resident_page_count
< obj
->resident_page_count
))) {
606 if (tmpe
->wired_count
> 0)
607 nothingwired
= FALSE
;
612 vm_pageout_object_deactivate_pages(map
, bigobj
, desired
, 0);
615 * Next, hunt around for other pages to deactivate. We actually
616 * do this search sort of wrong -- .text first is not the best idea.
618 tmpe
= map
->header
.next
;
619 while (tmpe
!= &map
->header
) {
620 if (pmap_resident_count(vm_map_pmap(map
)) <= desired
)
622 switch(tmpe
->maptype
) {
623 case VM_MAPTYPE_NORMAL
:
624 case VM_MAPTYPE_VPAGETABLE
:
625 obj
= tmpe
->object
.vm_object
;
627 vm_pageout_object_deactivate_pages(map
, obj
, desired
, 0);
636 * Remove all mappings if a process is swapped out, this will free page
639 if (desired
== 0 && nothingwired
)
640 pmap_remove(vm_map_pmap(map
),
641 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
647 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We
648 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
649 * be trivially freed.
652 vm_pageout_page_free(vm_page_t m
)
654 vm_object_t object
= m
->object
;
655 int type
= object
->type
;
657 if (type
== OBJT_SWAP
|| type
== OBJT_DEFAULT
)
658 vm_object_reference(object
);
660 vm_page_protect(m
, VM_PROT_NONE
);
662 if (type
== OBJT_SWAP
|| type
== OBJT_DEFAULT
)
663 vm_object_deallocate(object
);
667 * vm_pageout_scan does the dirty work for the pageout daemon.
670 struct vm_pageout_scan_info
{
671 struct proc
*bigproc
;
675 static int vm_pageout_scan_callback(struct proc
*p
, void *data
);
678 vm_pageout_scan(int pass
)
680 struct vm_pageout_scan_info info
;
682 struct vm_page marker
;
683 int page_shortage
, maxscan
, pcount
;
684 int addl_page_shortage
, addl_page_shortage_init
;
687 int vnodes_skipped
= 0;
691 * Do whatever cleanup that the pmap code can.
695 addl_page_shortage_init
= vm_pageout_deficit
;
696 vm_pageout_deficit
= 0;
699 * Calculate the number of pages we want to either free or move
702 page_shortage
= vm_paging_target() + addl_page_shortage_init
;
705 * Initialize our marker
707 bzero(&marker
, sizeof(marker
));
708 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
709 marker
.queue
= PQ_INACTIVE
;
710 marker
.wire_count
= 1;
713 * Start scanning the inactive queue for pages we can move to the
714 * cache or free. The scan will stop when the target is reached or
715 * we have scanned the entire inactive queue. Note that m->act_count
716 * is not used to form decisions for the inactive queue, only for the
719 * maxlaunder limits the number of dirty pages we flush per scan.
720 * For most systems a smaller value (16 or 32) is more robust under
721 * extreme memory and disk pressure because any unnecessary writes
722 * to disk can result in extreme performance degredation. However,
723 * systems with excessive dirty pages (especially when MAP_NOSYNC is
724 * used) will die horribly with limited laundering. If the pageout
725 * daemon cannot clean enough pages in the first pass, we let it go
726 * all out in succeeding passes.
728 if ((maxlaunder
= vm_max_launder
) <= 1)
734 * We will generally be in a critical section throughout the
735 * scan, but we can release it temporarily when we are sitting on a
736 * non-busy page without fear. this is required to prevent an
737 * interrupt from unbusying or freeing a page prior to our busy
738 * check, leaving us on the wrong queue or checking the wrong
743 addl_page_shortage
= addl_page_shortage_init
;
744 maxscan
= vmstats
.v_inactive_count
;
745 for (m
= TAILQ_FIRST(&vm_page_queues
[PQ_INACTIVE
].pl
);
746 m
!= NULL
&& maxscan
-- > 0 && page_shortage
> 0;
749 mycpu
->gd_cnt
.v_pdpages
++;
752 * Give interrupts a chance
758 * It's easier for some of the conditions below to just loop
759 * and catch queue changes here rather then check everywhere
762 if (m
->queue
!= PQ_INACTIVE
)
764 next
= TAILQ_NEXT(m
, pageq
);
769 if (m
->flags
& PG_MARKER
)
773 * A held page may be undergoing I/O, so skip it.
776 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
777 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
778 addl_page_shortage
++;
783 * Dont mess with busy pages, keep in the front of the
784 * queue, most likely are being paged out.
786 if (m
->busy
|| (m
->flags
& PG_BUSY
)) {
787 addl_page_shortage
++;
791 if (m
->object
->ref_count
== 0) {
793 * If the object is not being used, we ignore previous
796 vm_page_flag_clear(m
, PG_REFERENCED
);
797 pmap_clear_reference(m
);
799 } else if (((m
->flags
& PG_REFERENCED
) == 0) &&
800 (actcount
= pmap_ts_referenced(m
))) {
802 * Otherwise, if the page has been referenced while
803 * in the inactive queue, we bump the "activation
804 * count" upwards, making it less likely that the
805 * page will be added back to the inactive queue
806 * prematurely again. Here we check the page tables
807 * (or emulated bits, if any), given the upper level
808 * VM system not knowing anything about existing
812 m
->act_count
+= (actcount
+ ACT_ADVANCE
);
817 * If the upper level VM system knows about any page
818 * references, we activate the page. We also set the
819 * "activation count" higher than normal so that we will less
820 * likely place pages back onto the inactive queue again.
822 if ((m
->flags
& PG_REFERENCED
) != 0) {
823 vm_page_flag_clear(m
, PG_REFERENCED
);
824 actcount
= pmap_ts_referenced(m
);
826 m
->act_count
+= (actcount
+ ACT_ADVANCE
+ 1);
831 * If the upper level VM system doesn't know anything about
832 * the page being dirty, we have to check for it again. As
833 * far as the VM code knows, any partially dirty pages are
836 * Pages marked PG_WRITEABLE may be mapped into the user
837 * address space of a process running on another cpu. A
838 * user process (without holding the MP lock) running on
839 * another cpu may be able to touch the page while we are
840 * trying to remove it. To prevent this from occuring we
841 * must call pmap_remove_all() or otherwise make the page
842 * read-only. If the race occured pmap_remove_all() is
843 * responsible for setting m->dirty.
846 vm_page_test_dirty(m
);
848 if (m
->dirty
== 0 && (m
->flags
& PG_WRITEABLE
) != 0)
857 * Invalid pages can be easily freed
859 vm_pageout_page_free(m
);
860 mycpu
->gd_cnt
.v_dfree
++;
862 } else if (m
->dirty
== 0) {
864 * Clean pages can be placed onto the cache queue.
865 * This effectively frees them.
869 } else if ((m
->flags
& PG_WINATCFLS
) == 0 && pass
== 0) {
871 * Dirty pages need to be paged out, but flushing
872 * a page is extremely expensive verses freeing
873 * a clean page. Rather then artificially limiting
874 * the number of pages we can flush, we instead give
875 * dirty pages extra priority on the inactive queue
876 * by forcing them to be cycled through the queue
877 * twice before being flushed, after which the
878 * (now clean) page will cycle through once more
879 * before being freed. This significantly extends
880 * the thrash point for a heavily loaded machine.
882 vm_page_flag_set(m
, PG_WINATCFLS
);
883 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
884 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
885 } else if (maxlaunder
> 0) {
887 * We always want to try to flush some dirty pages if
888 * we encounter them, to keep the system stable.
889 * Normally this number is small, but under extreme
890 * pressure where there are insufficient clean pages
891 * on the inactive queue, we may have to go all out.
893 int swap_pageouts_ok
;
894 struct vnode
*vp
= NULL
;
898 if ((object
->type
!= OBJT_SWAP
) && (object
->type
!= OBJT_DEFAULT
)) {
899 swap_pageouts_ok
= 1;
901 swap_pageouts_ok
= !(defer_swap_pageouts
|| disable_swap_pageouts
);
902 swap_pageouts_ok
|= (!disable_swap_pageouts
&& defer_swap_pageouts
&&
903 vm_page_count_min());
908 * We don't bother paging objects that are "dead".
909 * Those objects are in a "rundown" state.
911 if (!swap_pageouts_ok
|| (object
->flags
& OBJ_DEAD
)) {
912 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
913 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
918 * The object is already known NOT to be dead. It
919 * is possible for the vget() to block the whole
920 * pageout daemon, but the new low-memory handling
921 * code should prevent it.
923 * The previous code skipped locked vnodes and, worse,
924 * reordered pages in the queue. This results in
925 * completely non-deterministic operation because,
926 * quite often, a vm_fault has initiated an I/O and
927 * is holding a locked vnode at just the point where
928 * the pageout daemon is woken up.
930 * We can't wait forever for the vnode lock, we might
931 * deadlock due to a vn_read() getting stuck in
932 * vm_wait while holding this vnode. We skip the
933 * vnode if we can't get it in a reasonable amount
937 if (object
->type
== OBJT_VNODE
) {
940 if (vget(vp
, LK_EXCLUSIVE
|LK_NOOBJ
|LK_TIMELOCK
)) {
942 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
948 * The page might have been moved to another
949 * queue during potential blocking in vget()
950 * above. The page might have been freed and
951 * reused for another vnode. The object might
952 * have been reused for another vnode.
954 if (m
->queue
!= PQ_INACTIVE
||
955 m
->object
!= object
||
956 object
->handle
!= vp
) {
957 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
964 * The page may have been busied during the
965 * blocking in vput(); We don't move the
966 * page back onto the end of the queue so that
967 * statistics are more correct if we don't.
969 if (m
->busy
|| (m
->flags
& PG_BUSY
)) {
975 * If the page has become held it might
976 * be undergoing I/O, so skip it
979 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
980 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, pageq
);
981 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
989 * If a page is dirty, then it is either being washed
990 * (but not yet cleaned) or it is still in the
991 * laundry. If it is still in the laundry, then we
992 * start the cleaning operation.
994 * This operation may cluster, invalidating the 'next'
995 * pointer. To prevent an inordinate number of
996 * restarts we use our marker to remember our place.
998 * decrement page_shortage on success to account for
999 * the (future) cleaned page. Otherwise we could wind
1000 * up laundering or cleaning too many pages.
1002 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_INACTIVE
].pl
, m
, &marker
, pageq
);
1003 if (vm_pageout_clean(m
) != 0) {
1007 next
= TAILQ_NEXT(&marker
, pageq
);
1008 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
].pl
, &marker
, pageq
);
1015 * Compute the number of pages we want to try to move from the
1016 * active queue to the inactive queue.
1018 page_shortage
= vm_paging_target() +
1019 vmstats
.v_inactive_target
- vmstats
.v_inactive_count
;
1020 page_shortage
+= addl_page_shortage
;
1023 * Scan the active queue for things we can deactivate. We nominally
1024 * track the per-page activity counter and use it to locate
1025 * deactivation candidates.
1027 * NOTE: we are still in a critical section.
1029 pcount
= vmstats
.v_active_count
;
1030 m
= TAILQ_FIRST(&vm_page_queues
[PQ_ACTIVE
].pl
);
1032 while ((m
!= NULL
) && (pcount
-- > 0) && (page_shortage
> 0)) {
1034 * Give interrupts a chance.
1040 * If the page was ripped out from under us, just stop.
1042 if (m
->queue
!= PQ_ACTIVE
)
1044 next
= TAILQ_NEXT(m
, pageq
);
1047 * Don't deactivate pages that are busy.
1049 if ((m
->busy
!= 0) ||
1050 (m
->flags
& PG_BUSY
) ||
1051 (m
->hold_count
!= 0)) {
1052 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1053 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1059 * The count for pagedaemon pages is done after checking the
1060 * page for eligibility...
1062 mycpu
->gd_cnt
.v_pdpages
++;
1065 * Check to see "how much" the page has been used.
1068 if (m
->object
->ref_count
!= 0) {
1069 if (m
->flags
& PG_REFERENCED
) {
1072 actcount
+= pmap_ts_referenced(m
);
1074 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1075 if (m
->act_count
> ACT_MAX
)
1076 m
->act_count
= ACT_MAX
;
1081 * Since we have "tested" this bit, we need to clear it now.
1083 vm_page_flag_clear(m
, PG_REFERENCED
);
1086 * Only if an object is currently being used, do we use the
1087 * page activation count stats.
1089 if (actcount
&& (m
->object
->ref_count
!= 0)) {
1090 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1091 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1093 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1094 if (vm_pageout_algorithm
||
1095 m
->object
->ref_count
== 0 ||
1096 m
->act_count
< pass
) {
1098 if (m
->object
->ref_count
== 0) {
1099 vm_page_protect(m
, VM_PROT_NONE
);
1103 vm_page_deactivate(m
);
1105 vm_page_deactivate(m
);
1108 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1109 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1116 * We try to maintain some *really* free pages, this allows interrupt
1117 * code to be guaranteed space. Since both cache and free queues
1118 * are considered basically 'free', moving pages from cache to free
1119 * does not effect other calculations.
1121 * NOTE: we are still in a critical section.
1124 while (vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
1125 static int cache_rover
= 0;
1126 m
= vm_page_list_find(PQ_CACHE
, cache_rover
, FALSE
);
1129 if ((m
->flags
& (PG_BUSY
|PG_UNMANAGED
)) ||
1134 kprintf("Warning: busy page %p found in cache\n", m
);
1136 vm_page_deactivate(m
);
1139 cache_rover
= (cache_rover
+ PQ_PRIME2
) & PQ_L2_MASK
;
1140 vm_pageout_page_free(m
);
1141 mycpu
->gd_cnt
.v_dfree
++;
1146 #if !defined(NO_SWAPPING)
1148 * Idle process swapout -- run once per second.
1150 if (vm_swap_idle_enabled
) {
1152 if (time_second
!= lsec
) {
1153 vm_pageout_req_swapout
|= VM_SWAP_IDLE
;
1161 * If we didn't get enough free pages, and we have skipped a vnode
1162 * in a writeable object, wakeup the sync daemon. And kick swapout
1163 * if we did not get enough free pages.
1165 if (vm_paging_target() > 0) {
1166 if (vnodes_skipped
&& vm_page_count_min())
1168 #if !defined(NO_SWAPPING)
1169 if (vm_swap_enabled
&& vm_page_count_target()) {
1171 vm_pageout_req_swapout
|= VM_SWAP_NORMAL
;
1177 * If we are out of swap and were not able to reach our paging
1178 * target, kill the largest process.
1180 if ((vm_swap_size
< 64 && vm_page_count_min()) ||
1181 (swap_pager_full
&& vm_paging_target() > 0)) {
1183 if ((vm_swap_size
< 64 || swap_pager_full
) && vm_page_count_min()) {
1185 info
.bigproc
= NULL
;
1187 allproc_scan(vm_pageout_scan_callback
, &info
);
1188 if (info
.bigproc
!= NULL
) {
1189 killproc(info
.bigproc
, "out of swap space");
1190 info
.bigproc
->p_nice
= PRIO_MIN
;
1191 info
.bigproc
->p_usched
->resetpriority(
1192 FIRST_LWP_IN_PROC(info
.bigproc
));
1193 wakeup(&vmstats
.v_free_count
);
1194 PRELE(info
.bigproc
);
1200 vm_pageout_scan_callback(struct proc
*p
, void *data
)
1202 struct vm_pageout_scan_info
*info
= data
;
1206 * if this is a system process, skip it
1208 if ((p
->p_flag
& P_SYSTEM
) || (p
->p_pid
== 1) ||
1209 ((p
->p_pid
< 48) && (vm_swap_size
!= 0))) {
1214 * if the process is in a non-running type state,
1217 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
) {
1222 * get the process size
1224 size
= vmspace_resident_count(p
->p_vmspace
) +
1225 vmspace_swap_count(p
->p_vmspace
);
1228 * If the this process is bigger than the biggest one
1231 if (size
> info
->bigsize
) {
1233 PRELE(info
->bigproc
);
1236 info
->bigsize
= size
;
1242 * This routine tries to maintain the pseudo LRU active queue,
1243 * so that during long periods of time where there is no paging,
1244 * that some statistic accumulation still occurs. This code
1245 * helps the situation where paging just starts to occur.
1248 vm_pageout_page_stats(void)
1251 int pcount
,tpcount
; /* Number of pages to check */
1252 static int fullintervalcount
= 0;
1256 (vmstats
.v_inactive_target
+ vmstats
.v_cache_max
+ vmstats
.v_free_min
) -
1257 (vmstats
.v_free_count
+ vmstats
.v_inactive_count
+ vmstats
.v_cache_count
);
1259 if (page_shortage
<= 0)
1264 pcount
= vmstats
.v_active_count
;
1265 fullintervalcount
+= vm_pageout_stats_interval
;
1266 if (fullintervalcount
< vm_pageout_full_stats_interval
) {
1267 tpcount
= (vm_pageout_stats_max
* vmstats
.v_active_count
) / vmstats
.v_page_count
;
1268 if (pcount
> tpcount
)
1271 fullintervalcount
= 0;
1274 m
= TAILQ_FIRST(&vm_page_queues
[PQ_ACTIVE
].pl
);
1275 while ((m
!= NULL
) && (pcount
-- > 0)) {
1278 if (m
->queue
!= PQ_ACTIVE
) {
1282 next
= TAILQ_NEXT(m
, pageq
);
1284 * Don't deactivate pages that are busy.
1286 if ((m
->busy
!= 0) ||
1287 (m
->flags
& PG_BUSY
) ||
1288 (m
->hold_count
!= 0)) {
1289 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1290 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1296 if (m
->flags
& PG_REFERENCED
) {
1297 vm_page_flag_clear(m
, PG_REFERENCED
);
1301 actcount
+= pmap_ts_referenced(m
);
1303 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1304 if (m
->act_count
> ACT_MAX
)
1305 m
->act_count
= ACT_MAX
;
1306 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1307 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1309 if (m
->act_count
== 0) {
1311 * We turn off page access, so that we have
1312 * more accurate RSS stats. We don't do this
1313 * in the normal page deactivation when the
1314 * system is loaded VM wise, because the
1315 * cost of the large number of page protect
1316 * operations would be higher than the value
1317 * of doing the operation.
1319 vm_page_protect(m
, VM_PROT_NONE
);
1320 vm_page_deactivate(m
);
1322 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1323 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1324 TAILQ_INSERT_TAIL(&vm_page_queues
[PQ_ACTIVE
].pl
, m
, pageq
);
1334 vm_pageout_free_page_calc(vm_size_t count
)
1336 if (count
< vmstats
.v_page_count
)
1339 * free_reserved needs to include enough for the largest swap pager
1340 * structures plus enough for any pv_entry structs when paging.
1342 if (vmstats
.v_page_count
> 1024)
1343 vmstats
.v_free_min
= 4 + (vmstats
.v_page_count
- 1024) / 200;
1345 vmstats
.v_free_min
= 4;
1346 vmstats
.v_pageout_free_min
= (2*MAXBSIZE
)/PAGE_SIZE
+
1347 vmstats
.v_interrupt_free_min
;
1348 vmstats
.v_free_reserved
= vm_pageout_page_count
+
1349 vmstats
.v_pageout_free_min
+ (count
/ 768) + PQ_L2_SIZE
;
1350 vmstats
.v_free_severe
= vmstats
.v_free_min
/ 2;
1351 vmstats
.v_free_min
+= vmstats
.v_free_reserved
;
1352 vmstats
.v_free_severe
+= vmstats
.v_free_reserved
;
1358 * vm_pageout is the high level pageout daemon.
1366 * Initialize some paging parameters.
1369 vmstats
.v_interrupt_free_min
= 2;
1370 if (vmstats
.v_page_count
< 2000)
1371 vm_pageout_page_count
= 8;
1373 vm_pageout_free_page_calc(vmstats
.v_page_count
);
1375 * v_free_target and v_cache_min control pageout hysteresis. Note
1376 * that these are more a measure of the VM cache queue hysteresis
1377 * then the VM free queue. Specifically, v_free_target is the
1378 * high water mark (free+cache pages).
1380 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1381 * low water mark, while v_free_min is the stop. v_cache_min must
1382 * be big enough to handle memory needs while the pageout daemon
1383 * is signalled and run to free more pages.
1385 if (vmstats
.v_free_count
> 6144)
1386 vmstats
.v_free_target
= 4 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1388 vmstats
.v_free_target
= 2 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1390 if (vmstats
.v_free_count
> 2048) {
1391 vmstats
.v_cache_min
= vmstats
.v_free_target
;
1392 vmstats
.v_cache_max
= 2 * vmstats
.v_cache_min
;
1393 vmstats
.v_inactive_target
= (3 * vmstats
.v_free_target
) / 2;
1395 vmstats
.v_cache_min
= 0;
1396 vmstats
.v_cache_max
= 0;
1397 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 4;
1399 if (vmstats
.v_inactive_target
> vmstats
.v_free_count
/ 3)
1400 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 3;
1402 /* XXX does not really belong here */
1403 if (vm_page_max_wired
== 0)
1404 vm_page_max_wired
= vmstats
.v_free_count
/ 3;
1406 if (vm_pageout_stats_max
== 0)
1407 vm_pageout_stats_max
= vmstats
.v_free_target
;
1410 * Set interval in seconds for stats scan.
1412 if (vm_pageout_stats_interval
== 0)
1413 vm_pageout_stats_interval
= 5;
1414 if (vm_pageout_full_stats_interval
== 0)
1415 vm_pageout_full_stats_interval
= vm_pageout_stats_interval
* 4;
1419 * Set maximum free per pass
1421 if (vm_pageout_stats_free_max
== 0)
1422 vm_pageout_stats_free_max
= 5;
1424 swap_pager_swap_init();
1427 * The pageout daemon is never done, so loop forever.
1433 * If we have enough free memory, wakeup waiters. Do
1434 * not clear vm_pages_needed until we reach our target,
1435 * otherwise we may be woken up over and over again and
1436 * waste a lot of cpu.
1439 if (vm_pages_needed
&& !vm_page_count_min()) {
1440 if (vm_paging_needed() <= 0)
1441 vm_pages_needed
= 0;
1442 wakeup(&vmstats
.v_free_count
);
1444 if (vm_pages_needed
) {
1446 * Still not done, take a second pass without waiting
1447 * (unlimited dirty cleaning), otherwise sleep a bit
1452 tsleep(&vm_pages_needed
, 0, "psleep", hz
/2);
1455 * Good enough, sleep & handle stats. Prime the pass
1462 error
= tsleep(&vm_pages_needed
,
1463 0, "psleep", vm_pageout_stats_interval
* hz
);
1464 if (error
&& !vm_pages_needed
) {
1467 vm_pageout_page_stats();
1472 if (vm_pages_needed
)
1473 mycpu
->gd_cnt
.v_pdwakeups
++;
1475 vm_pageout_scan(pass
);
1476 vm_pageout_deficit
= 0;
1481 pagedaemon_wakeup(void)
1483 if (!vm_pages_needed
&& curthread
!= pagethread
) {
1485 wakeup(&vm_pages_needed
);
1489 #if !defined(NO_SWAPPING)
1491 vm_req_vmdaemon(void)
1493 static int lastrun
= 0;
1495 if ((ticks
> (lastrun
+ hz
)) || (ticks
< lastrun
)) {
1496 wakeup(&vm_daemon_needed
);
1501 static int vm_daemon_callback(struct proc
*p
, void *data __unused
);
1507 tsleep(&vm_daemon_needed
, 0, "psleep", 0);
1508 if (vm_pageout_req_swapout
) {
1509 swapout_procs(vm_pageout_req_swapout
);
1510 vm_pageout_req_swapout
= 0;
1513 * scan the processes for exceeding their rlimits or if
1514 * process is swapped out -- deactivate pages
1516 allproc_scan(vm_daemon_callback
, NULL
);
1521 vm_daemon_callback(struct proc
*p
, void *data __unused
)
1523 vm_pindex_t limit
, size
;
1526 * if this is a system process or if we have already
1527 * looked at this process, skip it.
1529 if (p
->p_flag
& (P_SYSTEM
| P_WEXIT
))
1533 * if the process is in a non-running type state,
1536 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
)
1542 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
1543 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
1546 * let processes that are swapped out really be
1547 * swapped out. Set the limit to nothing to get as
1548 * many pages out to swap as possible.
1550 if (p
->p_flag
& P_SWAPPEDOUT
)
1553 size
= vmspace_resident_count(p
->p_vmspace
);
1554 if (limit
>= 0 && size
>= limit
) {
1555 vm_pageout_map_deactivate_pages(
1556 &p
->p_vmspace
->vm_map
, limit
);