2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
64 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
68 * The proverbial page-out daemon.
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/kernel.h>
76 #include <sys/kthread.h>
77 #include <sys/resourcevar.h>
78 #include <sys/signalvar.h>
79 #include <sys/vnode.h>
80 #include <sys/vmmeter.h>
81 #include <sys/sysctl.h>
84 #include <vm/vm_param.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_pager.h>
91 #include <vm/swap_pager.h>
92 #include <vm/vm_extern.h>
94 #include <sys/thread2.h>
95 #include <sys/spinlock2.h>
96 #include <vm/vm_page2.h>
99 * System initialization
102 /* the kernel process "vm_pageout"*/
103 static int vm_pageout_page(vm_page_t m
, int *max_launderp
,
104 int *vnodes_skippedp
, struct vnode
**vpfailedp
,
105 int pass
, int vmflush_flags
);
106 static int vm_pageout_clean_helper (vm_page_t
, int);
107 static int vm_pageout_free_page_calc (vm_size_t count
);
108 static void vm_pageout_page_free(vm_page_t m
) ;
109 struct thread
*pagethread
;
111 #if !defined(NO_SWAPPING)
112 /* the kernel process "vm_daemon"*/
113 static void vm_daemon (void);
114 static struct thread
*vmthread
;
116 static struct kproc_desc vm_kp
= {
121 SYSINIT(vmdaemon
, SI_SUB_KTHREAD_VM
, SI_ORDER_FIRST
, kproc_start
, &vm_kp
);
124 int vm_pages_needed
= 0; /* Event on which pageout daemon sleeps */
125 int vm_pageout_deficit
= 0; /* Estimated number of pages deficit */
126 int vm_pageout_pages_needed
= 0;/* pageout daemon needs pages */
127 int vm_page_free_hysteresis
= 16;
129 #if !defined(NO_SWAPPING)
130 static int vm_pageout_req_swapout
;
131 static int vm_daemon_needed
;
133 static int vm_max_launder
= 4096;
134 static int vm_pageout_stats_max
=0, vm_pageout_stats_interval
= 0;
135 static int vm_pageout_full_stats_interval
= 0;
136 static int vm_pageout_stats_free_max
=0, vm_pageout_algorithm
=0;
137 static int defer_swap_pageouts
=0;
138 static int disable_swap_pageouts
=0;
139 static u_int vm_anonmem_decline
= ACT_DECLINE
;
140 static u_int vm_filemem_decline
= ACT_DECLINE
* 2;
142 #if defined(NO_SWAPPING)
143 static int vm_swap_enabled
=0;
144 static int vm_swap_idle_enabled
=0;
146 static int vm_swap_enabled
=1;
147 static int vm_swap_idle_enabled
=0;
149 int vm_pageout_memuse_mode
=1; /* 0-disable, 1-passive, 2-active swp*/
151 SYSCTL_UINT(_vm
, VM_PAGEOUT_ALGORITHM
, anonmem_decline
,
152 CTLFLAG_RW
, &vm_anonmem_decline
, 0, "active->inactive anon memory");
154 SYSCTL_INT(_vm
, VM_PAGEOUT_ALGORITHM
, filemem_decline
,
155 CTLFLAG_RW
, &vm_filemem_decline
, 0, "active->inactive file cache");
157 SYSCTL_INT(_vm
, OID_AUTO
, page_free_hysteresis
,
158 CTLFLAG_RW
, &vm_page_free_hysteresis
, 0,
159 "Free more pages than the minimum required");
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");
175 SYSCTL_INT(_vm
, OID_AUTO
, pageout_memuse_mode
,
176 CTLFLAG_RW
, &vm_pageout_memuse_mode
, 0, "memoryuse resource mode");
178 #if defined(NO_SWAPPING)
179 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
180 CTLFLAG_RD
, &vm_swap_enabled
, 0, "");
181 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
182 CTLFLAG_RD
, &vm_swap_idle_enabled
, 0, "");
184 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
185 CTLFLAG_RW
, &vm_swap_enabled
, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
187 CTLFLAG_RW
, &vm_swap_idle_enabled
, 0, "Allow swapout on idle criteria");
190 SYSCTL_INT(_vm
, OID_AUTO
, defer_swapspace_pageouts
,
191 CTLFLAG_RW
, &defer_swap_pageouts
, 0, "Give preference to dirty pages in mem");
193 SYSCTL_INT(_vm
, OID_AUTO
, disable_swapspace_pageouts
,
194 CTLFLAG_RW
, &disable_swap_pageouts
, 0, "Disallow swapout of dirty pages");
196 static int pageout_lock_miss
;
197 SYSCTL_INT(_vm
, OID_AUTO
, pageout_lock_miss
,
198 CTLFLAG_RD
, &pageout_lock_miss
, 0, "vget() lock misses during pageout");
200 int vm_page_max_wired
; /* XXX max # of wired pages system-wide */
202 #if !defined(NO_SWAPPING)
203 static void vm_req_vmdaemon (void);
205 static void vm_pageout_page_stats(int q
);
208 * Calculate approximately how many pages on each queue to try to
209 * clean. An exact calculation creates an edge condition when the
210 * queues are unbalanced so add significant slop. The queue scans
211 * will stop early when targets are reached and will start where they
212 * left off on the next pass.
214 * We need to be generous here because there are all sorts of loading
215 * conditions that can cause edge cases if try to average over all queues.
216 * In particular, storage subsystems have become so fast that paging
217 * activity can become quite frantic. Eventually we will probably need
218 * two paging threads, one for dirty pages and one for clean, to deal
219 * with the bandwidth requirements.
221 * So what we do is calculate a value that can be satisfied nominally by
222 * only having to scan half the queues.
230 avg
= ((n
+ (PQ_L2_SIZE
- 1)) / (PQ_L2_SIZE
/ 2) + 1);
232 avg
= ((n
- (PQ_L2_SIZE
- 1)) / (PQ_L2_SIZE
/ 2) - 1);
238 * vm_pageout_clean_helper:
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.
248 vm_pageout_clean_helper(vm_page_t m
, int vmflush_flags
)
251 vm_page_t mc
[BLIST_MAX_ALLOC
];
253 int ib
, is
, page_base
;
254 vm_pindex_t pindex
= m
->pindex
;
259 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
260 * with the new swapper, but we could have serious problems paging
261 * out other object types if there is insufficient memory.
263 * Unfortunately, checking free memory here is far too late, so the
264 * check has been moved up a procedural level.
268 * Don't mess with the page if it's busy, held, or special
270 * XXX do we really need to check hold_count here? hold_count
271 * isn't supposed to mess with vm_page ops except prevent the
272 * page from being reused.
274 if (m
->hold_count
!= 0 || (m
->flags
& PG_UNMANAGED
)) {
280 * Place page in cluster. Align cluster for optimal swap space
281 * allocation (whether it is swap or not). This is typically ~16-32
282 * pages, which also tends to align the cluster to multiples of the
283 * filesystem block size if backed by a filesystem.
285 page_base
= pindex
% BLIST_MAX_ALLOC
;
291 * Scan object for clusterable pages.
293 * We can cluster ONLY if: ->> the page is NOT
294 * clean, wired, busy, held, or mapped into a
295 * buffer, and one of the following:
296 * 1) The page is inactive, or a seldom used
299 * 2) we force the issue.
301 * During heavy mmap/modification loads the pageout
302 * daemon can really fragment the underlying file
303 * due to flushing pages out of order and not trying
304 * align the clusters (which leave sporatic out-of-order
305 * holes). To solve this problem we do the reverse scan
306 * first and attempt to align our cluster, then do a
307 * forward scan if room remains.
309 vm_object_hold(object
);
314 p
= vm_page_lookup_busy_try(object
, pindex
- page_base
+ ib
,
316 if (error
|| p
== NULL
)
318 if ((p
->queue
- p
->pc
) == PQ_CACHE
||
319 (p
->flags
& PG_UNMANAGED
)) {
323 vm_page_test_dirty(p
);
324 if (((p
->dirty
& p
->valid
) == 0 &&
325 (p
->flags
& PG_NEED_COMMIT
) == 0) ||
326 p
->wire_count
!= 0 || /* may be held by buf cache */
327 p
->hold_count
!= 0) { /* may be undergoing I/O */
331 if (p
->queue
- p
->pc
!= PQ_INACTIVE
) {
332 if (p
->queue
- p
->pc
!= PQ_ACTIVE
||
333 (vmflush_flags
& VM_PAGER_ALLOW_ACTIVE
) == 0) {
340 * Try to maintain page groupings in the cluster.
342 if (m
->flags
& PG_WINATCFLS
)
343 vm_page_flag_set(p
, PG_WINATCFLS
);
345 vm_page_flag_clear(p
, PG_WINATCFLS
);
346 p
->act_count
= m
->act_count
;
353 while (is
< BLIST_MAX_ALLOC
&&
354 pindex
- page_base
+ is
< object
->size
) {
357 p
= vm_page_lookup_busy_try(object
, pindex
- page_base
+ is
,
359 if (error
|| p
== NULL
)
361 if (((p
->queue
- p
->pc
) == PQ_CACHE
) ||
362 (p
->flags
& PG_UNMANAGED
)) {
366 vm_page_test_dirty(p
);
367 if (((p
->dirty
& p
->valid
) == 0 &&
368 (p
->flags
& PG_NEED_COMMIT
) == 0) ||
369 p
->wire_count
!= 0 || /* may be held by buf cache */
370 p
->hold_count
!= 0) { /* may be undergoing I/O */
374 if (p
->queue
- p
->pc
!= PQ_INACTIVE
) {
375 if (p
->queue
- p
->pc
!= PQ_ACTIVE
||
376 (vmflush_flags
& VM_PAGER_ALLOW_ACTIVE
) == 0) {
383 * Try to maintain page groupings in the cluster.
385 if (m
->flags
& PG_WINATCFLS
)
386 vm_page_flag_set(p
, PG_WINATCFLS
);
388 vm_page_flag_clear(p
, PG_WINATCFLS
);
389 p
->act_count
= m
->act_count
;
395 vm_object_drop(object
);
398 * we allow reads during pageouts...
400 return vm_pageout_flush(&mc
[ib
], is
- ib
, vmflush_flags
);
404 * vm_pageout_flush() - launder the given pages
406 * The given pages are laundered. Note that we setup for the start of
407 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
408 * reference count all in here rather then in the parent. If we want
409 * the parent to do more sophisticated things we may have to change
412 * The pages in the array must be busied by the caller and will be
413 * unbusied by this function.
416 vm_pageout_flush(vm_page_t
*mc
, int count
, int vmflush_flags
)
419 int pageout_status
[count
];
424 * Initiate I/O. Bump the vm_page_t->busy counter.
426 for (i
= 0; i
< count
; i
++) {
427 KASSERT(mc
[i
]->valid
== VM_PAGE_BITS_ALL
,
428 ("vm_pageout_flush page %p index %d/%d: partially "
429 "invalid page", mc
[i
], i
, count
));
430 vm_page_io_start(mc
[i
]);
434 * We must make the pages read-only. This will also force the
435 * modified bit in the related pmaps to be cleared. The pager
436 * cannot clear the bit for us since the I/O completion code
437 * typically runs from an interrupt. The act of making the page
438 * read-only handles the case for us.
440 * Then we can unbusy the pages, we still hold a reference by virtue
443 for (i
= 0; i
< count
; i
++) {
444 if (vmflush_flags
& VM_PAGER_TRY_TO_CACHE
)
445 vm_page_protect(mc
[i
], VM_PROT_NONE
);
447 vm_page_protect(mc
[i
], VM_PROT_READ
);
448 vm_page_wakeup(mc
[i
]);
451 object
= mc
[0]->object
;
452 vm_object_pip_add(object
, count
);
454 vm_pager_put_pages(object
, mc
, count
,
456 ((object
== &kernel_object
) ? VM_PAGER_PUT_SYNC
: 0)),
459 for (i
= 0; i
< count
; i
++) {
460 vm_page_t mt
= mc
[i
];
462 switch (pageout_status
[i
]) {
471 * Page outside of range of object. Right now we
472 * essentially lose the changes by pretending it
475 vm_page_busy_wait(mt
, FALSE
, "pgbad");
476 pmap_clear_modify(mt
);
483 * A page typically cannot be paged out when we
484 * have run out of swap. We leave the page
485 * marked inactive and will try to page it out
488 * Starvation of the active page list is used to
489 * determine when the system is massively memory
498 * If not PENDing this was a synchronous operation and we
499 * clean up after the I/O. If it is PENDing the mess is
500 * cleaned up asynchronously.
502 * Also nominally act on the caller's wishes if the caller
503 * wants to try to really clean (cache or free) the page.
505 * Also nominally deactivate the page if the system is
508 if (pageout_status
[i
] != VM_PAGER_PEND
) {
509 vm_page_busy_wait(mt
, FALSE
, "pgouw");
510 vm_page_io_finish(mt
);
511 if (vmflush_flags
& VM_PAGER_TRY_TO_CACHE
) {
512 vm_page_try_to_cache(mt
);
513 } else if (vm_page_count_severe()) {
514 vm_page_deactivate(mt
);
519 vm_object_pip_wakeup(object
);
525 #if !defined(NO_SWAPPING)
528 * Callback function, page busied for us. We must dispose of the busy
529 * condition. Any related pmap pages may be held but will not be locked.
533 vm_pageout_mdp_callback(struct pmap_pgscan_info
*info
, vm_offset_t va
,
540 * Basic tests - There should never be a marker, and we can stop
541 * once the RSS is below the required level.
543 KKASSERT((p
->flags
& PG_MARKER
) == 0);
544 if (pmap_resident_tlnw_count(info
->pmap
) <= info
->limit
) {
549 mycpu
->gd_cnt
.v_pdpages
++;
551 if (p
->wire_count
|| p
->hold_count
|| (p
->flags
& PG_UNMANAGED
)) {
559 * Check if the page has been referened recently. If it has,
560 * activate it and skip.
562 actcount
= pmap_ts_referenced(p
);
564 vm_page_flag_set(p
, PG_REFERENCED
);
565 } else if (p
->flags
& PG_REFERENCED
) {
570 if (p
->queue
- p
->pc
!= PQ_ACTIVE
) {
571 vm_page_and_queue_spin_lock(p
);
572 if (p
->queue
- p
->pc
!= PQ_ACTIVE
) {
573 vm_page_and_queue_spin_unlock(p
);
576 vm_page_and_queue_spin_unlock(p
);
579 p
->act_count
+= actcount
;
580 if (p
->act_count
> ACT_MAX
)
581 p
->act_count
= ACT_MAX
;
583 vm_page_flag_clear(p
, PG_REFERENCED
);
589 * Remove the page from this particular pmap. Once we do this, our
590 * pmap scans will not see it again (unless it gets faulted in), so
591 * we must actively dispose of or deal with the page.
593 pmap_remove_specific(info
->pmap
, p
);
596 * If the page is not mapped to another process (i.e. as would be
597 * typical if this were a shared page from a library) then deactivate
598 * the page and clean it in two passes only.
600 * If the page hasn't been referenced since the last check, remove it
601 * from the pmap. If it is no longer mapped, deactivate it
602 * immediately, accelerating the normal decline.
604 * Once the page has been removed from the pmap the RSS code no
605 * longer tracks it so we have to make sure that it is staged for
606 * potential flush action.
608 if ((p
->flags
& PG_MAPPED
) == 0) {
609 if (p
->queue
- p
->pc
== PQ_ACTIVE
) {
610 vm_page_deactivate(p
);
612 if (p
->queue
- p
->pc
== PQ_INACTIVE
) {
618 * Ok, try to fully clean the page and any nearby pages such that at
619 * least the requested page is freed or moved to the cache queue.
621 * We usually do this synchronously to allow us to get the page into
622 * the CACHE queue quickly, which will prevent memory exhaustion if
623 * a process with a memoryuse limit is running away. However, the
624 * sysadmin may desire to set vm.swap_user_async which relaxes this
625 * and improves write performance.
628 int max_launder
= 0x7FFF;
629 int vnodes_skipped
= 0;
631 struct vnode
*vpfailed
= NULL
;
635 if (vm_pageout_memuse_mode
>= 2) {
636 vmflush_flags
= VM_PAGER_TRY_TO_CACHE
|
637 VM_PAGER_ALLOW_ACTIVE
;
638 if (swap_user_async
== 0)
639 vmflush_flags
|= VM_PAGER_PUT_SYNC
;
640 vm_page_flag_set(p
, PG_WINATCFLS
);
642 vm_pageout_page(p
, &max_launder
,
644 &vpfailed
, 1, vmflush_flags
);
658 * Deactivate some number of pages in a map due to set RLIMIT_RSS limits.
659 * that is relatively difficult to do. We try to keep track of where we
660 * left off last time to reduce scan overhead.
662 * Called when vm_pageout_memuse_mode is >= 1.
665 vm_pageout_map_deactivate_pages(vm_map_t map
, vm_pindex_t limit
)
667 vm_offset_t pgout_offset
;
668 struct pmap_pgscan_info info
;
671 pgout_offset
= map
->pgout_offset
;
674 kprintf("%016jx ", pgout_offset
);
676 if (pgout_offset
< VM_MIN_USER_ADDRESS
)
677 pgout_offset
= VM_MIN_USER_ADDRESS
;
678 if (pgout_offset
>= VM_MAX_USER_ADDRESS
)
680 info
.pmap
= vm_map_pmap(map
);
682 info
.beg_addr
= pgout_offset
;
683 info
.end_addr
= VM_MAX_USER_ADDRESS
;
684 info
.callback
= vm_pageout_mdp_callback
;
686 info
.actioncount
= 0;
690 pgout_offset
= info
.offset
;
692 kprintf("%016jx %08lx %08lx\n", pgout_offset
,
693 info
.cleancount
, info
.actioncount
);
696 if (pgout_offset
!= VM_MAX_USER_ADDRESS
&&
697 pmap_resident_tlnw_count(vm_map_pmap(map
)) > limit
) {
699 } else if (retries
&&
700 pmap_resident_tlnw_count(vm_map_pmap(map
)) > limit
) {
704 map
->pgout_offset
= pgout_offset
;
709 * Called when the pageout scan wants to free a page. We no longer
710 * try to cycle the vm_object here with a reference & dealloc, which can
711 * cause a non-trivial object collapse in a critical path.
713 * It is unclear why we cycled the ref_count in the past, perhaps to try
714 * to optimize shadow chain collapses but I don't quite see why it would
715 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
716 * synchronously and not have to be kicked-start.
719 vm_pageout_page_free(vm_page_t m
)
721 vm_page_protect(m
, VM_PROT_NONE
);
726 * vm_pageout_scan does the dirty work for the pageout daemon.
728 struct vm_pageout_scan_info
{
729 struct proc
*bigproc
;
733 static int vm_pageout_scan_callback(struct proc
*p
, void *data
);
736 vm_pageout_scan_inactive(int pass
, int q
, int avail_shortage
,
740 struct vm_page marker
;
741 struct vnode
*vpfailed
; /* warning, allowed to be stale */
747 * Start scanning the inactive queue for pages we can move to the
748 * cache or free. The scan will stop when the target is reached or
749 * we have scanned the entire inactive queue. Note that m->act_count
750 * is not used to form decisions for the inactive queue, only for the
753 * max_launder limits the number of dirty pages we flush per scan.
754 * For most systems a smaller value (16 or 32) is more robust under
755 * extreme memory and disk pressure because any unnecessary writes
756 * to disk can result in extreme performance degredation. However,
757 * systems with excessive dirty pages (especially when MAP_NOSYNC is
758 * used) will die horribly with limited laundering. If the pageout
759 * daemon cannot clean enough pages in the first pass, we let it go
760 * all out in succeeding passes.
762 if ((max_launder
= vm_max_launder
) <= 1)
768 * Initialize our marker
770 bzero(&marker
, sizeof(marker
));
771 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
772 marker
.queue
= PQ_INACTIVE
+ q
;
774 marker
.wire_count
= 1;
777 * Inactive queue scan.
779 * NOTE: The vm_page must be spinlocked before the queue to avoid
780 * deadlocks, so it is easiest to simply iterate the loop
781 * with the queue unlocked at the top.
785 vm_page_queues_spin_lock(PQ_INACTIVE
+ q
);
786 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, &marker
, pageq
);
787 maxscan
= vm_page_queues
[PQ_INACTIVE
+ q
].lcnt
;
790 * Queue locked at top of loop to avoid stack marker issues.
792 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
793 maxscan
-- > 0 && avail_shortage
- delta
> 0)
797 KKASSERT(m
->queue
== PQ_INACTIVE
+ q
);
798 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
800 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, m
,
802 mycpu
->gd_cnt
.v_pdpages
++;
805 * Skip marker pages (atomic against other markers to avoid
806 * infinite hop-over scans).
808 if (m
->flags
& PG_MARKER
)
812 * Try to busy the page. Don't mess with pages which are
813 * already busy or reorder them in the queue.
815 if (vm_page_busy_try(m
, TRUE
))
819 * Remaining operations run with the page busy and neither
820 * the page or the queue will be spin-locked.
822 vm_page_queues_spin_unlock(PQ_INACTIVE
+ q
);
823 KKASSERT(m
->queue
== PQ_INACTIVE
+ q
);
825 count
= vm_pageout_page(m
, &max_launder
, vnodes_skipped
,
830 * Systems with a ton of memory can wind up with huge
831 * deactivation counts. Because the inactive scan is
832 * doing a lot of flushing, the combination can result
833 * in excessive paging even in situations where other
834 * unrelated threads free up sufficient VM.
836 * To deal with this we abort the nominal active->inactive
837 * scan before we hit the inactive target when free+cache
838 * levels have reached a reasonable target.
840 * When deciding to stop early we need to add some slop to
841 * the test and we need to return full completion to the caller
842 * to prevent the caller from thinking there is something
843 * wrong and issuing a low-memory+swap warning or pkill.
845 * A deficit forces paging regardless of the state of the
846 * VM page queues (used for RSS enforcement).
849 vm_page_queues_spin_lock(PQ_INACTIVE
+ q
);
850 if (vm_paging_target() < -vm_max_launder
) {
852 * Stopping early, return full completion to caller.
854 if (delta
< avail_shortage
)
855 delta
= avail_shortage
;
860 /* page queue still spin-locked */
861 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, &marker
, pageq
);
862 vm_page_queues_spin_unlock(PQ_INACTIVE
+ q
);
868 * Pageout the specified page, return the total number of pages paged out
869 * (this routine may cluster).
871 * The page must be busied and soft-busied by the caller and will be disposed
872 * of by this function.
875 vm_pageout_page(vm_page_t m
, int *max_launderp
, int *vnodes_skippedp
,
876 struct vnode
**vpfailedp
, int pass
, int vmflush_flags
)
883 * It is possible for a page to be busied ad-hoc (e.g. the
884 * pmap_collect() code) and wired and race against the
885 * allocation of a new page. vm_page_alloc() may be forced
886 * to deactivate the wired page in which case it winds up
887 * on the inactive queue and must be handled here. We
888 * correct the problem simply by unqueuing the page.
891 vm_page_unqueue_nowakeup(m
);
893 kprintf("WARNING: pagedaemon: wired page on "
894 "inactive queue %p\n", m
);
899 * A held page may be undergoing I/O, so skip it.
902 vm_page_and_queue_spin_lock(m
);
903 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
905 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
907 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
908 ++vm_swapcache_inactive_heuristic
;
910 vm_page_and_queue_spin_unlock(m
);
915 if (m
->object
== NULL
|| m
->object
->ref_count
== 0) {
917 * If the object is not being used, we ignore previous
920 vm_page_flag_clear(m
, PG_REFERENCED
);
921 pmap_clear_reference(m
);
922 /* fall through to end */
923 } else if (((m
->flags
& PG_REFERENCED
) == 0) &&
924 (actcount
= pmap_ts_referenced(m
))) {
926 * Otherwise, if the page has been referenced while
927 * in the inactive queue, we bump the "activation
928 * count" upwards, making it less likely that the
929 * page will be added back to the inactive queue
930 * prematurely again. Here we check the page tables
931 * (or emulated bits, if any), given the upper level
932 * VM system not knowing anything about existing
936 m
->act_count
+= (actcount
+ ACT_ADVANCE
);
942 * (m) is still busied.
944 * If the upper level VM system knows about any page
945 * references, we activate the page. We also set the
946 * "activation count" higher than normal so that we will less
947 * likely place pages back onto the inactive queue again.
949 if ((m
->flags
& PG_REFERENCED
) != 0) {
950 vm_page_flag_clear(m
, PG_REFERENCED
);
951 actcount
= pmap_ts_referenced(m
);
953 m
->act_count
+= (actcount
+ ACT_ADVANCE
+ 1);
959 * If the upper level VM system doesn't know anything about
960 * the page being dirty, we have to check for it again. As
961 * far as the VM code knows, any partially dirty pages are
964 * Pages marked PG_WRITEABLE may be mapped into the user
965 * address space of a process running on another cpu. A
966 * user process (without holding the MP lock) running on
967 * another cpu may be able to touch the page while we are
968 * trying to remove it. vm_page_cache() will handle this
972 vm_page_test_dirty(m
);
977 if (m
->valid
== 0 && (m
->flags
& PG_NEED_COMMIT
) == 0) {
979 * Invalid pages can be easily freed
981 vm_pageout_page_free(m
);
982 mycpu
->gd_cnt
.v_dfree
++;
984 } else if (m
->dirty
== 0 && (m
->flags
& PG_NEED_COMMIT
) == 0) {
986 * Clean pages can be placed onto the cache queue.
987 * This effectively frees them.
991 } else if ((m
->flags
& PG_WINATCFLS
) == 0 && pass
== 0) {
993 * Dirty pages need to be paged out, but flushing
994 * a page is extremely expensive verses freeing
995 * a clean page. Rather then artificially limiting
996 * the number of pages we can flush, we instead give
997 * dirty pages extra priority on the inactive queue
998 * by forcing them to be cycled through the queue
999 * twice before being flushed, after which the
1000 * (now clean) page will cycle through once more
1001 * before being freed. This significantly extends
1002 * the thrash point for a heavily loaded machine.
1004 vm_page_flag_set(m
, PG_WINATCFLS
);
1005 vm_page_and_queue_spin_lock(m
);
1006 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
1008 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1010 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1011 ++vm_swapcache_inactive_heuristic
;
1013 vm_page_and_queue_spin_unlock(m
);
1015 } else if (*max_launderp
> 0) {
1017 * We always want to try to flush some dirty pages if
1018 * we encounter them, to keep the system stable.
1019 * Normally this number is small, but under extreme
1020 * pressure where there are insufficient clean pages
1021 * on the inactive queue, we may have to go all out.
1023 int swap_pageouts_ok
;
1024 struct vnode
*vp
= NULL
;
1026 swap_pageouts_ok
= 0;
1029 (object
->type
!= OBJT_SWAP
) &&
1030 (object
->type
!= OBJT_DEFAULT
)) {
1031 swap_pageouts_ok
= 1;
1033 swap_pageouts_ok
= !(defer_swap_pageouts
|| disable_swap_pageouts
);
1034 swap_pageouts_ok
|= (!disable_swap_pageouts
&& defer_swap_pageouts
&&
1035 vm_page_count_min(0));
1039 * We don't bother paging objects that are "dead".
1040 * Those objects are in a "rundown" state.
1042 if (!swap_pageouts_ok
||
1044 (object
->flags
& OBJ_DEAD
)) {
1045 vm_page_and_queue_spin_lock(m
);
1046 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
1048 &vm_page_queues
[m
->queue
].pl
,
1051 &vm_page_queues
[m
->queue
].pl
,
1053 ++vm_swapcache_inactive_heuristic
;
1055 vm_page_and_queue_spin_unlock(m
);
1061 * (m) is still busied.
1063 * The object is already known NOT to be dead. It
1064 * is possible for the vget() to block the whole
1065 * pageout daemon, but the new low-memory handling
1066 * code should prevent it.
1068 * The previous code skipped locked vnodes and, worse,
1069 * reordered pages in the queue. This results in
1070 * completely non-deterministic operation because,
1071 * quite often, a vm_fault has initiated an I/O and
1072 * is holding a locked vnode at just the point where
1073 * the pageout daemon is woken up.
1075 * We can't wait forever for the vnode lock, we might
1076 * deadlock due to a vn_read() getting stuck in
1077 * vm_wait while holding this vnode. We skip the
1078 * vnode if we can't get it in a reasonable amount
1081 * vpfailed is used to (try to) avoid the case where
1082 * a large number of pages are associated with a
1083 * locked vnode, which could cause the pageout daemon
1084 * to stall for an excessive amount of time.
1086 if (object
->type
== OBJT_VNODE
) {
1089 vp
= object
->handle
;
1090 flags
= LK_EXCLUSIVE
;
1091 if (vp
== *vpfailedp
)
1094 flags
|= LK_TIMELOCK
;
1099 * We have unbusied (m) temporarily so we can
1100 * acquire the vp lock without deadlocking.
1101 * (m) is held to prevent destruction.
1103 if (vget(vp
, flags
) != 0) {
1105 ++pageout_lock_miss
;
1106 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1113 * The page might have been moved to another
1114 * queue during potential blocking in vget()
1115 * above. The page might have been freed and
1116 * reused for another vnode. The object might
1117 * have been reused for another vnode.
1119 if (m
->queue
- m
->pc
!= PQ_INACTIVE
||
1120 m
->object
!= object
||
1121 object
->handle
!= vp
) {
1122 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1130 * The page may have been busied during the
1131 * blocking in vput(); We don't move the
1132 * page back onto the end of the queue so that
1133 * statistics are more correct if we don't.
1135 if (vm_page_busy_try(m
, TRUE
)) {
1143 * (m) is busied again
1145 * We own the busy bit and remove our hold
1146 * bit. If the page is still held it
1147 * might be undergoing I/O, so skip it.
1149 if (m
->hold_count
) {
1150 vm_page_and_queue_spin_lock(m
);
1151 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
1152 TAILQ_REMOVE(&vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1153 TAILQ_INSERT_TAIL(&vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1154 ++vm_swapcache_inactive_heuristic
;
1156 vm_page_and_queue_spin_unlock(m
);
1157 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1163 /* (m) is left busied as we fall through */
1167 * page is busy and not held here.
1169 * If a page is dirty, then it is either being washed
1170 * (but not yet cleaned) or it is still in the
1171 * laundry. If it is still in the laundry, then we
1172 * start the cleaning operation.
1174 * decrement inactive_shortage on success to account
1175 * for the (future) cleaned page. Otherwise we
1176 * could wind up laundering or cleaning too many
1179 * NOTE: Cleaning the page here does not cause
1180 * force_deficit to be adjusted, because the
1181 * page is not being freed or moved to the
1184 count
= vm_pageout_clean_helper(m
, vmflush_flags
);
1185 *max_launderp
-= count
;
1188 * Clean ate busy, page no longer accessible
1199 vm_pageout_scan_active(int pass
, int q
,
1200 int avail_shortage
, int inactive_shortage
,
1201 int *recycle_countp
)
1203 struct vm_page marker
;
1210 * We want to move pages from the active queue to the inactive
1211 * queue to get the inactive queue to the inactive target. If
1212 * we still have a page shortage from above we try to directly free
1213 * clean pages instead of moving them.
1215 * If we do still have a shortage we keep track of the number of
1216 * pages we free or cache (recycle_count) as a measure of thrashing
1217 * between the active and inactive queues.
1219 * If we were able to completely satisfy the free+cache targets
1220 * from the inactive pool we limit the number of pages we move
1221 * from the active pool to the inactive pool to 2x the pages we
1222 * had removed from the inactive pool (with a minimum of 1/5 the
1223 * inactive target). If we were not able to completely satisfy
1224 * the free+cache targets we go for the whole target aggressively.
1226 * NOTE: Both variables can end up negative.
1227 * NOTE: We are still in a critical section.
1230 bzero(&marker
, sizeof(marker
));
1231 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
1232 marker
.queue
= PQ_ACTIVE
+ q
;
1234 marker
.wire_count
= 1;
1236 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1237 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1238 maxscan
= vm_page_queues
[PQ_ACTIVE
+ q
].lcnt
;
1241 * Queue locked at top of loop to avoid stack marker issues.
1243 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
1244 maxscan
-- > 0 && (avail_shortage
- delta
> 0 ||
1245 inactive_shortage
> 0))
1247 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1248 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1250 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, m
,
1254 * Skip marker pages (atomic against other markers to avoid
1255 * infinite hop-over scans).
1257 if (m
->flags
& PG_MARKER
)
1261 * Try to busy the page. Don't mess with pages which are
1262 * already busy or reorder them in the queue.
1264 if (vm_page_busy_try(m
, TRUE
))
1268 * Remaining operations run with the page busy and neither
1269 * the page or the queue will be spin-locked.
1271 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1272 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1275 * Don't deactivate pages that are held, even if we can
1276 * busy them. (XXX why not?)
1278 if (m
->hold_count
!= 0) {
1279 vm_page_and_queue_spin_lock(m
);
1280 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1282 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1285 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1288 vm_page_and_queue_spin_unlock(m
);
1294 * The count for pagedaemon pages is done after checking the
1295 * page for eligibility...
1297 mycpu
->gd_cnt
.v_pdpages
++;
1300 * Check to see "how much" the page has been used and clear
1301 * the tracking access bits. If the object has no references
1302 * don't bother paying the expense.
1305 if (m
->object
&& m
->object
->ref_count
!= 0) {
1306 if (m
->flags
& PG_REFERENCED
)
1308 actcount
+= pmap_ts_referenced(m
);
1310 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1311 if (m
->act_count
> ACT_MAX
)
1312 m
->act_count
= ACT_MAX
;
1315 vm_page_flag_clear(m
, PG_REFERENCED
);
1318 * actcount is only valid if the object ref_count is non-zero.
1319 * If the page does not have an object, actcount will be zero.
1321 if (actcount
&& m
->object
->ref_count
!= 0) {
1322 vm_page_and_queue_spin_lock(m
);
1323 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1325 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1328 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1331 vm_page_and_queue_spin_unlock(m
);
1334 switch(m
->object
->type
) {
1337 m
->act_count
-= min(m
->act_count
,
1338 vm_anonmem_decline
);
1341 m
->act_count
-= min(m
->act_count
,
1342 vm_filemem_decline
);
1345 if (vm_pageout_algorithm
||
1346 (m
->object
== NULL
) ||
1347 (m
->object
&& (m
->object
->ref_count
== 0)) ||
1348 m
->act_count
< pass
+ 1
1351 * Deactivate the page. If we had a
1352 * shortage from our inactive scan try to
1353 * free (cache) the page instead.
1355 * Don't just blindly cache the page if
1356 * we do not have a shortage from the
1357 * inactive scan, that could lead to
1358 * gigabytes being moved.
1360 --inactive_shortage
;
1361 if (avail_shortage
- delta
> 0 ||
1362 (m
->object
&& (m
->object
->ref_count
== 0)))
1364 if (avail_shortage
- delta
> 0)
1366 vm_page_protect(m
, VM_PROT_NONE
);
1367 if (m
->dirty
== 0 &&
1368 (m
->flags
& PG_NEED_COMMIT
) == 0 &&
1369 avail_shortage
- delta
> 0) {
1372 vm_page_deactivate(m
);
1376 vm_page_deactivate(m
);
1381 vm_page_and_queue_spin_lock(m
);
1382 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1384 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1387 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1390 vm_page_and_queue_spin_unlock(m
);
1396 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1400 * Clean out our local marker.
1402 * Page queue still spin-locked.
1404 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1405 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1411 * The number of actually free pages can drop down to v_free_reserved,
1412 * we try to build the free count back above v_free_min. Note that
1413 * vm_paging_needed() also returns TRUE if v_free_count is not at
1414 * least v_free_min so that is the minimum we must build the free
1417 * We use a slightly higher target to improve hysteresis,
1418 * ((v_free_target + v_free_min) / 2). Since v_free_target
1419 * is usually the same as v_cache_min this maintains about
1420 * half the pages in the free queue as are in the cache queue,
1421 * providing pretty good pipelining for pageout operation.
1423 * The system operator can manipulate vm.v_cache_min and
1424 * vm.v_free_target to tune the pageout demon. Be sure
1425 * to keep vm.v_free_min < vm.v_free_target.
1427 * Note that the original paging target is to get at least
1428 * (free_min + cache_min) into (free + cache). The slightly
1429 * higher target will shift additional pages from cache to free
1430 * without effecting the original paging target in order to
1431 * maintain better hysteresis and not have the free count always
1432 * be dead-on v_free_min.
1434 * NOTE: we are still in a critical section.
1436 * Pages moved from PQ_CACHE to totally free are not counted in the
1437 * pages_freed counter.
1440 vm_pageout_scan_cache(int avail_shortage
, int pass
,
1441 int vnodes_skipped
, int recycle_count
)
1443 static int lastkillticks
;
1444 struct vm_pageout_scan_info info
;
1447 while (vmstats
.v_free_count
<
1448 (vmstats
.v_free_min
+ vmstats
.v_free_target
) / 2) {
1450 * This steals some code from vm/vm_page.c
1452 static int cache_rover
= 0;
1454 m
= vm_page_list_find(PQ_CACHE
,
1455 cache_rover
& PQ_L2_MASK
, FALSE
);
1458 /* page is returned removed from its queue and spinlocked */
1459 if (vm_page_busy_try(m
, TRUE
)) {
1460 vm_page_deactivate_locked(m
);
1461 vm_page_spin_unlock(m
);
1464 vm_page_spin_unlock(m
);
1465 pagedaemon_wakeup();
1469 * Remaining operations run with the page busy and neither
1470 * the page or the queue will be spin-locked.
1472 if ((m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) ||
1475 vm_page_deactivate(m
);
1479 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
1480 KKASSERT(m
->dirty
== 0);
1481 cache_rover
+= PQ_PRIME2
;
1482 vm_pageout_page_free(m
);
1483 mycpu
->gd_cnt
.v_dfree
++;
1486 #if !defined(NO_SWAPPING)
1488 * Idle process swapout -- run once per second.
1490 if (vm_swap_idle_enabled
) {
1492 if (time_uptime
!= lsec
) {
1493 atomic_set_int(&vm_pageout_req_swapout
, VM_SWAP_IDLE
);
1501 * If we didn't get enough free pages, and we have skipped a vnode
1502 * in a writeable object, wakeup the sync daemon. And kick swapout
1503 * if we did not get enough free pages.
1505 if (vm_paging_target() > 0) {
1506 if (vnodes_skipped
&& vm_page_count_min(0))
1507 speedup_syncer(NULL
);
1508 #if !defined(NO_SWAPPING)
1509 if (vm_swap_enabled
&& vm_page_count_target()) {
1510 atomic_set_int(&vm_pageout_req_swapout
, VM_SWAP_NORMAL
);
1517 * Handle catastrophic conditions. Under good conditions we should
1518 * be at the target, well beyond our minimum. If we could not even
1519 * reach our minimum the system is under heavy stress. But just being
1520 * under heavy stress does not trigger process killing.
1522 * We consider ourselves to have run out of memory if the swap pager
1523 * is full and avail_shortage is still positive. The secondary check
1524 * ensures that we do not kill processes if the instantanious
1525 * availability is good, even if the pageout demon pass says it
1526 * couldn't get to the target.
1528 if (swap_pager_almost_full
&&
1530 (vm_page_count_min(recycle_count
) || avail_shortage
> 0)) {
1531 kprintf("Warning: system low on memory+swap "
1532 "shortage %d for %d ticks!\n",
1533 avail_shortage
, ticks
- swap_fail_ticks
);
1535 if (swap_pager_full
&&
1537 avail_shortage
> 0 &&
1538 vm_paging_target() > 0 &&
1539 (unsigned int)(ticks
- lastkillticks
) >= hz
) {
1541 * Kill something, maximum rate once per second to give
1542 * the process time to free up sufficient memory.
1544 lastkillticks
= ticks
;
1545 info
.bigproc
= NULL
;
1547 allproc_scan(vm_pageout_scan_callback
, &info
);
1548 if (info
.bigproc
!= NULL
) {
1549 info
.bigproc
->p_nice
= PRIO_MIN
;
1550 info
.bigproc
->p_usched
->resetpriority(
1551 FIRST_LWP_IN_PROC(info
.bigproc
));
1552 atomic_set_int(&info
.bigproc
->p_flags
, P_LOWMEMKILL
);
1553 killproc(info
.bigproc
, "out of swap space");
1554 wakeup(&vmstats
.v_free_count
);
1555 PRELE(info
.bigproc
);
1561 vm_pageout_scan_callback(struct proc
*p
, void *data
)
1563 struct vm_pageout_scan_info
*info
= data
;
1567 * Never kill system processes or init. If we have configured swap
1568 * then try to avoid killing low-numbered pids.
1570 if ((p
->p_flags
& P_SYSTEM
) || (p
->p_pid
== 1) ||
1571 ((p
->p_pid
< 48) && (vm_swap_size
!= 0))) {
1575 lwkt_gettoken(&p
->p_token
);
1578 * if the process is in a non-running type state,
1581 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
&& p
->p_stat
!= SCORE
) {
1582 lwkt_reltoken(&p
->p_token
);
1587 * Get the approximate process size. Note that anonymous pages
1588 * with backing swap will be counted twice, but there should not
1589 * be too many such pages due to the stress the VM system is
1590 * under at this point.
1592 size
= vmspace_anonymous_count(p
->p_vmspace
) +
1593 vmspace_swap_count(p
->p_vmspace
);
1596 * If the this process is bigger than the biggest one
1599 if (info
->bigsize
< size
) {
1601 PRELE(info
->bigproc
);
1604 info
->bigsize
= size
;
1606 lwkt_reltoken(&p
->p_token
);
1613 * This routine tries to maintain the pseudo LRU active queue,
1614 * so that during long periods of time where there is no paging,
1615 * that some statistic accumulation still occurs. This code
1616 * helps the situation where paging just starts to occur.
1619 vm_pageout_page_stats(int q
)
1621 static int fullintervalcount
= 0;
1622 struct vm_page marker
;
1624 int pcount
, tpcount
; /* Number of pages to check */
1627 page_shortage
= (vmstats
.v_inactive_target
+ vmstats
.v_cache_max
+
1628 vmstats
.v_free_min
) -
1629 (vmstats
.v_free_count
+ vmstats
.v_inactive_count
+
1630 vmstats
.v_cache_count
);
1632 if (page_shortage
<= 0)
1635 pcount
= vm_page_queues
[PQ_ACTIVE
+ q
].lcnt
;
1636 fullintervalcount
+= vm_pageout_stats_interval
;
1637 if (fullintervalcount
< vm_pageout_full_stats_interval
) {
1638 tpcount
= (vm_pageout_stats_max
* pcount
) /
1639 vmstats
.v_page_count
+ 1;
1640 if (pcount
> tpcount
)
1643 fullintervalcount
= 0;
1646 bzero(&marker
, sizeof(marker
));
1647 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
1648 marker
.queue
= PQ_ACTIVE
+ q
;
1650 marker
.wire_count
= 1;
1652 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1653 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1656 * Queue locked at top of loop to avoid stack marker issues.
1658 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
1663 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1664 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1665 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, m
,
1669 * Skip marker pages (atomic against other markers to avoid
1670 * infinite hop-over scans).
1672 if (m
->flags
& PG_MARKER
)
1676 * Ignore pages we can't busy
1678 if (vm_page_busy_try(m
, TRUE
))
1682 * Remaining operations run with the page busy and neither
1683 * the page or the queue will be spin-locked.
1685 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1686 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1689 * We now have a safely busied page, the page and queue
1690 * spinlocks have been released.
1694 if (m
->hold_count
) {
1700 * Calculate activity
1703 if (m
->flags
& PG_REFERENCED
) {
1704 vm_page_flag_clear(m
, PG_REFERENCED
);
1707 actcount
+= pmap_ts_referenced(m
);
1710 * Update act_count and move page to end of queue.
1713 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1714 if (m
->act_count
> ACT_MAX
)
1715 m
->act_count
= ACT_MAX
;
1716 vm_page_and_queue_spin_lock(m
);
1717 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1719 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1722 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1725 vm_page_and_queue_spin_unlock(m
);
1730 if (m
->act_count
== 0) {
1732 * We turn off page access, so that we have
1733 * more accurate RSS stats. We don't do this
1734 * in the normal page deactivation when the
1735 * system is loaded VM wise, because the
1736 * cost of the large number of page protect
1737 * operations would be higher than the value
1738 * of doing the operation.
1740 * We use the marker to save our place so
1741 * we can release the spin lock. both (m)
1742 * and (next) will be invalid.
1744 vm_page_protect(m
, VM_PROT_NONE
);
1745 vm_page_deactivate(m
);
1747 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1748 vm_page_and_queue_spin_lock(m
);
1749 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1751 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1754 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1757 vm_page_and_queue_spin_unlock(m
);
1761 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1765 * Remove our local marker
1767 * Page queue still spin-locked.
1769 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1770 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1774 vm_pageout_free_page_calc(vm_size_t count
)
1776 if (count
< vmstats
.v_page_count
)
1779 * free_reserved needs to include enough for the largest swap pager
1780 * structures plus enough for any pv_entry structs when paging.
1782 * v_free_min normal allocations
1783 * v_free_reserved system allocations
1784 * v_pageout_free_min allocations by pageout daemon
1785 * v_interrupt_free_min low level allocations (e.g swap structures)
1787 if (vmstats
.v_page_count
> 1024)
1788 vmstats
.v_free_min
= 64 + (vmstats
.v_page_count
- 1024) / 200;
1790 vmstats
.v_free_min
= 64;
1793 * Make sure the vmmeter slop can't blow out our global minimums.
1795 if (vmstats
.v_free_min
< VMMETER_SLOP_COUNT
* ncpus
* 10)
1796 vmstats
.v_free_min
= VMMETER_SLOP_COUNT
* ncpus
* 10;
1798 vmstats
.v_free_reserved
= vmstats
.v_free_min
* 4 / 8 + 7;
1799 vmstats
.v_free_severe
= vmstats
.v_free_min
* 4 / 8 + 0;
1800 vmstats
.v_pageout_free_min
= vmstats
.v_free_min
* 2 / 8 + 7;
1801 vmstats
.v_interrupt_free_min
= vmstats
.v_free_min
* 1 / 8 + 7;
1808 * vm_pageout is the high level pageout daemon.
1813 vm_pageout_thread(void)
1821 * Initialize some paging parameters.
1823 curthread
->td_flags
|= TDF_SYSTHREAD
;
1825 vm_pageout_free_page_calc(vmstats
.v_page_count
);
1828 * v_free_target and v_cache_min control pageout hysteresis. Note
1829 * that these are more a measure of the VM cache queue hysteresis
1830 * then the VM free queue. Specifically, v_free_target is the
1831 * high water mark (free+cache pages).
1833 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1834 * low water mark, while v_free_min is the stop. v_cache_min must
1835 * be big enough to handle memory needs while the pageout daemon
1836 * is signalled and run to free more pages.
1838 if (vmstats
.v_free_count
> 6144)
1839 vmstats
.v_free_target
= 4 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1841 vmstats
.v_free_target
= 2 * vmstats
.v_free_min
+ vmstats
.v_free_reserved
;
1844 * NOTE: With the new buffer cache b_act_count we want the default
1845 * inactive target to be a percentage of available memory.
1847 * The inactive target essentially determines the minimum
1848 * number of 'temporary' pages capable of caching one-time-use
1849 * files when the VM system is otherwise full of pages
1850 * belonging to multi-time-use files or active program data.
1852 * NOTE: The inactive target is aggressively persued only if the
1853 * inactive queue becomes too small. If the inactive queue
1854 * is large enough to satisfy page movement to free+cache
1855 * then it is repopulated more slowly from the active queue.
1856 * This allows a general inactive_target default to be set.
1858 * There is an issue here for processes which sit mostly idle
1859 * 'overnight', such as sshd, tcsh, and X. Any movement from
1860 * the active queue will eventually cause such pages to
1861 * recycle eventually causing a lot of paging in the morning.
1862 * To reduce the incidence of this pages cycled out of the
1863 * buffer cache are moved directly to the inactive queue if
1864 * they were only used once or twice.
1866 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1867 * Increasing the value (up to 64) increases the number of
1868 * buffer recyclements which go directly to the inactive queue.
1870 if (vmstats
.v_free_count
> 2048) {
1871 vmstats
.v_cache_min
= vmstats
.v_free_target
;
1872 vmstats
.v_cache_max
= 2 * vmstats
.v_cache_min
;
1874 vmstats
.v_cache_min
= 0;
1875 vmstats
.v_cache_max
= 0;
1877 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 4;
1879 /* XXX does not really belong here */
1880 if (vm_page_max_wired
== 0)
1881 vm_page_max_wired
= vmstats
.v_free_count
/ 3;
1883 if (vm_pageout_stats_max
== 0)
1884 vm_pageout_stats_max
= vmstats
.v_free_target
;
1887 * Set interval in seconds for stats scan.
1889 if (vm_pageout_stats_interval
== 0)
1890 vm_pageout_stats_interval
= 5;
1891 if (vm_pageout_full_stats_interval
== 0)
1892 vm_pageout_full_stats_interval
= vm_pageout_stats_interval
* 4;
1896 * Set maximum free per pass
1898 if (vm_pageout_stats_free_max
== 0)
1899 vm_pageout_stats_free_max
= 5;
1901 swap_pager_swap_init();
1905 * The pageout daemon is never done, so loop forever.
1910 int inactive_shortage
;
1911 int vnodes_skipped
= 0;
1912 int recycle_count
= 0;
1916 * Wait for an action request. If we timeout check to
1917 * see if paging is needed (in case the normal wakeup
1920 if (vm_pages_needed
== 0) {
1921 error
= tsleep(&vm_pages_needed
,
1923 vm_pageout_stats_interval
* hz
);
1925 vm_paging_needed() == 0 &&
1926 vm_pages_needed
== 0) {
1927 for (q
= 0; q
< PQ_L2_SIZE
; ++q
)
1928 vm_pageout_page_stats(q
);
1931 vm_pages_needed
= 1;
1934 mycpu
->gd_cnt
.v_pdwakeups
++;
1937 * Scan for INACTIVE->CLEAN/PAGEOUT
1939 * This routine tries to avoid thrashing the system with
1940 * unnecessary activity.
1942 * Calculate our target for the number of free+cache pages we
1943 * want to get to. This is higher then the number that causes
1944 * allocations to stall (severe) in order to provide hysteresis,
1945 * and if we don't make it all the way but get to the minimum
1946 * we're happy. Goose it a bit if there are multiple requests
1949 * Don't reduce avail_shortage inside the loop or the
1950 * PQAVERAGE() calculation will break.
1952 * NOTE! deficit is differentiated from avail_shortage as
1953 * REQUIRING at least (deficit) pages to be cleaned,
1954 * even if the page queues are in good shape. This
1955 * is used primarily for handling per-process
1956 * RLIMIT_RSS and may also see small values when
1957 * processes block due to low memory.
1960 avail_shortage
= vm_paging_target() + vm_pageout_deficit
;
1961 vm_pageout_deficit
= 0;
1963 if (avail_shortage
> 0) {
1966 for (q
= 0; q
< PQ_L2_SIZE
; ++q
) {
1967 delta
+= vm_pageout_scan_inactive(
1969 (q
+ q1iterator
) & PQ_L2_MASK
,
1970 PQAVERAGE(avail_shortage
),
1972 if (avail_shortage
- delta
<= 0)
1975 avail_shortage
-= delta
;
1980 * Figure out how many active pages we must deactivate. If
1981 * we were able to reach our target with just the inactive
1982 * scan above we limit the number of active pages we
1983 * deactivate to reduce unnecessary work.
1986 inactive_shortage
= vmstats
.v_inactive_target
-
1987 vmstats
.v_inactive_count
;
1990 * If we were unable to free sufficient inactive pages to
1991 * satisfy the free/cache queue requirements then simply
1992 * reaching the inactive target may not be good enough.
1993 * Try to deactivate pages in excess of the target based
1996 * However to prevent thrashing the VM system do not
1997 * deactivate more than an additional 1/10 the inactive
1998 * target's worth of active pages.
2000 if (avail_shortage
> 0) {
2001 tmp
= avail_shortage
* 2;
2002 if (tmp
> vmstats
.v_inactive_target
/ 10)
2003 tmp
= vmstats
.v_inactive_target
/ 10;
2004 inactive_shortage
+= tmp
;
2008 * Only trigger a pmap cleanup on inactive shortage.
2010 if (inactive_shortage
> 0) {
2015 * Scan for ACTIVE->INACTIVE
2017 * Only trigger on inactive shortage. Triggering on
2018 * avail_shortage can starve the active queue with
2019 * unnecessary active->inactive transitions and destroy
2022 if (/*avail_shortage > 0 ||*/ inactive_shortage
> 0) {
2025 for (q
= 0; q
< PQ_L2_SIZE
; ++q
) {
2026 delta
+= vm_pageout_scan_active(
2028 (q
+ q2iterator
) & PQ_L2_MASK
,
2029 PQAVERAGE(avail_shortage
),
2030 PQAVERAGE(inactive_shortage
),
2032 if (inactive_shortage
- delta
<= 0 &&
2033 avail_shortage
- delta
<= 0) {
2037 inactive_shortage
-= delta
;
2038 avail_shortage
-= delta
;
2043 * Scan for CACHE->FREE
2045 * Finally free enough cache pages to meet our free page
2046 * requirement and take more drastic measures if we are
2050 vm_pageout_scan_cache(avail_shortage
, pass
,
2051 vnodes_skipped
, recycle_count
);
2054 * Wait for more work.
2056 if (avail_shortage
> 0) {
2058 if (pass
< 10 && vm_pages_needed
> 1) {
2060 * Normal operation, additional processes
2061 * have already kicked us. Retry immediately
2062 * unless swap space is completely full in
2063 * which case delay a bit.
2065 if (swap_pager_full
) {
2066 tsleep(&vm_pages_needed
, 0, "pdelay",
2068 } /* else immediate retry */
2069 } else if (pass
< 10) {
2071 * Normal operation, fewer processes. Delay
2072 * a bit but allow wakeups.
2074 vm_pages_needed
= 0;
2075 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
2076 vm_pages_needed
= 1;
2077 } else if (swap_pager_full
== 0) {
2079 * We've taken too many passes, forced delay.
2081 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
2084 * Running out of memory, catastrophic
2085 * back-off to one-second intervals.
2087 tsleep(&vm_pages_needed
, 0, "pdelay", hz
);
2089 } else if (vm_pages_needed
) {
2091 * Interlocked wakeup of waiters (non-optional).
2093 * Similar to vm_page_free_wakeup() in vm_page.c,
2097 if (!vm_page_count_min(vm_page_free_hysteresis
) ||
2098 !vm_page_count_target()) {
2099 vm_pages_needed
= 0;
2100 wakeup(&vmstats
.v_free_count
);
2108 static struct kproc_desc page_kp
= {
2113 SYSINIT(pagedaemon
, SI_SUB_KTHREAD_PAGE
, SI_ORDER_FIRST
, kproc_start
, &page_kp
);
2117 * Called after allocating a page out of the cache or free queue
2118 * to possibly wake the pagedaemon up to replentish our supply.
2120 * We try to generate some hysteresis by waking the pagedaemon up
2121 * when our free+cache pages go below the free_min+cache_min level.
2122 * The pagedaemon tries to get the count back up to at least the
2123 * minimum, and through to the target level if possible.
2125 * If the pagedaemon is already active bump vm_pages_needed as a hint
2126 * that there are even more requests pending.
2132 pagedaemon_wakeup(void)
2134 if (vm_paging_needed() && curthread
!= pagethread
) {
2135 if (vm_pages_needed
== 0) {
2136 vm_pages_needed
= 1; /* SMP race ok */
2137 wakeup(&vm_pages_needed
);
2138 } else if (vm_page_count_min(0)) {
2139 ++vm_pages_needed
; /* SMP race ok */
2144 #if !defined(NO_SWAPPING)
2151 vm_req_vmdaemon(void)
2153 static int lastrun
= 0;
2155 if ((ticks
> (lastrun
+ hz
)) || (ticks
< lastrun
)) {
2156 wakeup(&vm_daemon_needed
);
2161 static int vm_daemon_callback(struct proc
*p
, void *data __unused
);
2172 tsleep(&vm_daemon_needed
, 0, "psleep", 0);
2173 req_swapout
= atomic_swap_int(&vm_pageout_req_swapout
, 0);
2179 swapout_procs(vm_pageout_req_swapout
);
2182 * scan the processes for exceeding their rlimits or if
2183 * process is swapped out -- deactivate pages
2185 allproc_scan(vm_daemon_callback
, NULL
);
2190 vm_daemon_callback(struct proc
*p
, void *data __unused
)
2193 vm_pindex_t limit
, size
;
2196 * if this is a system process or if we have already
2197 * looked at this process, skip it.
2199 lwkt_gettoken(&p
->p_token
);
2201 if (p
->p_flags
& (P_SYSTEM
| P_WEXIT
)) {
2202 lwkt_reltoken(&p
->p_token
);
2207 * if the process is in a non-running type state,
2210 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
&& p
->p_stat
!= SCORE
) {
2211 lwkt_reltoken(&p
->p_token
);
2218 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
2219 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
2222 * let processes that are swapped out really be
2223 * swapped out. Set the limit to nothing to get as
2224 * many pages out to swap as possible.
2226 if (p
->p_flags
& P_SWAPPEDOUT
)
2231 size
= pmap_resident_tlnw_count(&vm
->vm_pmap
);
2232 if (limit
>= 0 && size
> 4096 &&
2233 size
- 4096 >= limit
&& vm_pageout_memuse_mode
>= 1) {
2234 vm_pageout_map_deactivate_pages(&vm
->vm_map
, limit
);
2238 lwkt_reltoken(&p
->p_token
);