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 static struct thread
*emergpager
;
110 struct thread
*pagethread
;
111 static int sequence_emerg_pager
;
113 #if !defined(NO_SWAPPING)
114 /* the kernel process "vm_daemon"*/
115 static void vm_daemon (void);
116 static struct thread
*vmthread
;
118 static struct kproc_desc vm_kp
= {
123 SYSINIT(vmdaemon
, SI_SUB_KTHREAD_VM
, SI_ORDER_FIRST
, kproc_start
, &vm_kp
);
126 int vm_pages_needed
= 0; /* Event on which pageout daemon sleeps */
127 int vm_pageout_deficit
= 0; /* Estimated number of pages deficit */
128 int vm_pageout_pages_needed
= 0;/* pageout daemon needs pages */
129 int vm_page_free_hysteresis
= 16;
130 static int vm_pagedaemon_time
;
132 #if !defined(NO_SWAPPING)
133 static int vm_pageout_req_swapout
;
134 static int vm_daemon_needed
;
136 static int vm_max_launder
= 4096;
137 static int vm_emerg_launder
= 100;
138 static int vm_pageout_stats_max
=0, vm_pageout_stats_interval
= 0;
139 static int vm_pageout_full_stats_interval
= 0;
140 static int vm_pageout_stats_free_max
=0, vm_pageout_algorithm
=0;
141 static int defer_swap_pageouts
=0;
142 static int disable_swap_pageouts
=0;
143 static u_int vm_anonmem_decline
= ACT_DECLINE
;
144 static u_int vm_filemem_decline
= ACT_DECLINE
* 2;
146 #if defined(NO_SWAPPING)
147 static int vm_swap_enabled
=0;
148 static int vm_swap_idle_enabled
=0;
150 static int vm_swap_enabled
=1;
151 static int vm_swap_idle_enabled
=0;
153 int vm_pageout_memuse_mode
=1; /* 0-disable, 1-passive, 2-active swp*/
155 SYSCTL_UINT(_vm
, VM_PAGEOUT_ALGORITHM
, anonmem_decline
,
156 CTLFLAG_RW
, &vm_anonmem_decline
, 0, "active->inactive anon memory");
158 SYSCTL_INT(_vm
, VM_PAGEOUT_ALGORITHM
, filemem_decline
,
159 CTLFLAG_RW
, &vm_filemem_decline
, 0, "active->inactive file cache");
161 SYSCTL_INT(_vm
, OID_AUTO
, page_free_hysteresis
,
162 CTLFLAG_RW
, &vm_page_free_hysteresis
, 0,
163 "Free more pages than the minimum required");
165 SYSCTL_INT(_vm
, OID_AUTO
, max_launder
,
166 CTLFLAG_RW
, &vm_max_launder
, 0, "Limit dirty flushes in pageout");
167 SYSCTL_INT(_vm
, OID_AUTO
, emerg_launder
,
168 CTLFLAG_RW
, &vm_emerg_launder
, 0, "Emergency pager minimum");
170 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_max
,
171 CTLFLAG_RW
, &vm_pageout_stats_max
, 0, "Max pageout stats scan length");
173 SYSCTL_INT(_vm
, OID_AUTO
, pageout_full_stats_interval
,
174 CTLFLAG_RW
, &vm_pageout_full_stats_interval
, 0, "Interval for full stats scan");
176 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_interval
,
177 CTLFLAG_RW
, &vm_pageout_stats_interval
, 0, "Interval for partial stats scan");
179 SYSCTL_INT(_vm
, OID_AUTO
, pageout_stats_free_max
,
180 CTLFLAG_RW
, &vm_pageout_stats_free_max
, 0, "Not implemented");
181 SYSCTL_INT(_vm
, OID_AUTO
, pageout_memuse_mode
,
182 CTLFLAG_RW
, &vm_pageout_memuse_mode
, 0, "memoryuse resource mode");
184 #if defined(NO_SWAPPING)
185 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
186 CTLFLAG_RD
, &vm_swap_enabled
, 0, "");
187 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
188 CTLFLAG_RD
, &vm_swap_idle_enabled
, 0, "");
190 SYSCTL_INT(_vm
, VM_SWAPPING_ENABLED
, swap_enabled
,
191 CTLFLAG_RW
, &vm_swap_enabled
, 0, "Enable entire process swapout");
192 SYSCTL_INT(_vm
, OID_AUTO
, swap_idle_enabled
,
193 CTLFLAG_RW
, &vm_swap_idle_enabled
, 0, "Allow swapout on idle criteria");
196 SYSCTL_INT(_vm
, OID_AUTO
, defer_swapspace_pageouts
,
197 CTLFLAG_RW
, &defer_swap_pageouts
, 0, "Give preference to dirty pages in mem");
199 SYSCTL_INT(_vm
, OID_AUTO
, disable_swapspace_pageouts
,
200 CTLFLAG_RW
, &disable_swap_pageouts
, 0, "Disallow swapout of dirty pages");
202 static int pageout_lock_miss
;
203 SYSCTL_INT(_vm
, OID_AUTO
, pageout_lock_miss
,
204 CTLFLAG_RD
, &pageout_lock_miss
, 0, "vget() lock misses during pageout");
206 int vm_page_max_wired
; /* XXX max # of wired pages system-wide */
208 #if !defined(NO_SWAPPING)
209 static void vm_req_vmdaemon (void);
211 static void vm_pageout_page_stats(int q
);
214 * Calculate approximately how many pages on each queue to try to
215 * clean. An exact calculation creates an edge condition when the
216 * queues are unbalanced so add significant slop. The queue scans
217 * will stop early when targets are reached and will start where they
218 * left off on the next pass.
220 * We need to be generous here because there are all sorts of loading
221 * conditions that can cause edge cases if try to average over all queues.
222 * In particular, storage subsystems have become so fast that paging
223 * activity can become quite frantic. Eventually we will probably need
224 * two paging threads, one for dirty pages and one for clean, to deal
225 * with the bandwidth requirements.
227 * So what we do is calculate a value that can be satisfied nominally by
228 * only having to scan half the queues.
236 avg
= ((n
+ (PQ_L2_SIZE
- 1)) / (PQ_L2_SIZE
/ 2) + 1);
238 avg
= ((n
- (PQ_L2_SIZE
- 1)) / (PQ_L2_SIZE
/ 2) - 1);
244 * vm_pageout_clean_helper:
246 * Clean the page and remove it from the laundry. The page must be busied
247 * by the caller and will be disposed of (put away, flushed) by this routine.
250 vm_pageout_clean_helper(vm_page_t m
, int vmflush_flags
)
253 vm_page_t mc
[BLIST_MAX_ALLOC
];
255 int ib
, is
, page_base
;
256 vm_pindex_t pindex
= m
->pindex
;
261 * Don't mess with the page if it's held or special.
263 * XXX do we really need to check hold_count here? hold_count
264 * isn't supposed to mess with vm_page ops except prevent the
265 * page from being reused.
267 if (m
->hold_count
!= 0 || (m
->flags
& PG_UNMANAGED
)) {
273 * Place page in cluster. Align cluster for optimal swap space
274 * allocation (whether it is swap or not). This is typically ~16-32
275 * pages, which also tends to align the cluster to multiples of the
276 * filesystem block size if backed by a filesystem.
278 page_base
= pindex
% BLIST_MAX_ALLOC
;
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.
302 vm_object_hold(object
);
307 p
= vm_page_lookup_busy_try(object
, pindex
- page_base
+ ib
,
309 if (error
|| p
== NULL
)
311 if ((p
->queue
- p
->pc
) == PQ_CACHE
||
312 (p
->flags
& PG_UNMANAGED
)) {
316 vm_page_test_dirty(p
);
317 if (((p
->dirty
& p
->valid
) == 0 &&
318 (p
->flags
& PG_NEED_COMMIT
) == 0) ||
319 p
->wire_count
!= 0 || /* may be held by buf cache */
320 p
->hold_count
!= 0) { /* may be undergoing I/O */
324 if (p
->queue
- p
->pc
!= PQ_INACTIVE
) {
325 if (p
->queue
- p
->pc
!= PQ_ACTIVE
||
326 (vmflush_flags
& VM_PAGER_ALLOW_ACTIVE
) == 0) {
333 * Try to maintain page groupings in the cluster.
335 if (m
->flags
& PG_WINATCFLS
)
336 vm_page_flag_set(p
, PG_WINATCFLS
);
338 vm_page_flag_clear(p
, PG_WINATCFLS
);
339 p
->act_count
= m
->act_count
;
346 while (is
< BLIST_MAX_ALLOC
&&
347 pindex
- page_base
+ is
< object
->size
) {
350 p
= vm_page_lookup_busy_try(object
, pindex
- page_base
+ is
,
352 if (error
|| p
== NULL
)
354 if (((p
->queue
- p
->pc
) == PQ_CACHE
) ||
355 (p
->flags
& PG_UNMANAGED
)) {
359 vm_page_test_dirty(p
);
360 if (((p
->dirty
& p
->valid
) == 0 &&
361 (p
->flags
& PG_NEED_COMMIT
) == 0) ||
362 p
->wire_count
!= 0 || /* may be held by buf cache */
363 p
->hold_count
!= 0) { /* may be undergoing I/O */
367 if (p
->queue
- p
->pc
!= PQ_INACTIVE
) {
368 if (p
->queue
- p
->pc
!= PQ_ACTIVE
||
369 (vmflush_flags
& VM_PAGER_ALLOW_ACTIVE
) == 0) {
376 * Try to maintain page groupings in the cluster.
378 if (m
->flags
& PG_WINATCFLS
)
379 vm_page_flag_set(p
, PG_WINATCFLS
);
381 vm_page_flag_clear(p
, PG_WINATCFLS
);
382 p
->act_count
= m
->act_count
;
388 vm_object_drop(object
);
391 * we allow reads during pageouts...
393 return vm_pageout_flush(&mc
[ib
], is
- ib
, vmflush_flags
);
397 * vm_pageout_flush() - launder the given pages
399 * The given pages are laundered. Note that we setup for the start of
400 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
401 * reference count all in here rather then in the parent. If we want
402 * the parent to do more sophisticated things we may have to change
405 * The pages in the array must be busied by the caller and will be
406 * unbusied by this function.
409 vm_pageout_flush(vm_page_t
*mc
, int count
, int vmflush_flags
)
412 int pageout_status
[count
];
417 * Initiate I/O. Bump the vm_page_t->busy counter.
419 for (i
= 0; i
< count
; i
++) {
420 KASSERT(mc
[i
]->valid
== VM_PAGE_BITS_ALL
,
421 ("vm_pageout_flush page %p index %d/%d: partially "
422 "invalid page", mc
[i
], i
, count
));
423 vm_page_io_start(mc
[i
]);
427 * We must make the pages read-only. This will also force the
428 * modified bit in the related pmaps to be cleared. The pager
429 * cannot clear the bit for us since the I/O completion code
430 * typically runs from an interrupt. The act of making the page
431 * read-only handles the case for us.
433 * Then we can unbusy the pages, we still hold a reference by virtue
436 for (i
= 0; i
< count
; i
++) {
437 if (vmflush_flags
& VM_PAGER_TRY_TO_CACHE
)
438 vm_page_protect(mc
[i
], VM_PROT_NONE
);
440 vm_page_protect(mc
[i
], VM_PROT_READ
);
441 vm_page_wakeup(mc
[i
]);
444 object
= mc
[0]->object
;
445 vm_object_pip_add(object
, count
);
447 vm_pager_put_pages(object
, mc
, count
,
449 ((object
== &kernel_object
) ?
450 VM_PAGER_PUT_SYNC
: 0)),
453 for (i
= 0; i
< count
; i
++) {
454 vm_page_t mt
= mc
[i
];
456 switch (pageout_status
[i
]) {
465 * Page outside of range of object. Right now we
466 * essentially lose the changes by pretending it
469 vm_page_busy_wait(mt
, FALSE
, "pgbad");
470 pmap_clear_modify(mt
);
477 * A page typically cannot be paged out when we
478 * have run out of swap. We leave the page
479 * marked inactive and will try to page it out
482 * Starvation of the active page list is used to
483 * determine when the system is massively memory
492 * If not PENDing this was a synchronous operation and we
493 * clean up after the I/O. If it is PENDing the mess is
494 * cleaned up asynchronously.
496 * Also nominally act on the caller's wishes if the caller
497 * wants to try to really clean (cache or free) the page.
499 * Also nominally deactivate the page if the system is
502 if (pageout_status
[i
] != VM_PAGER_PEND
) {
503 vm_page_busy_wait(mt
, FALSE
, "pgouw");
504 vm_page_io_finish(mt
);
505 if (vmflush_flags
& VM_PAGER_TRY_TO_CACHE
) {
506 vm_page_try_to_cache(mt
);
507 } else if (vm_page_count_severe()) {
508 vm_page_deactivate(mt
);
513 vm_object_pip_wakeup(object
);
519 #if !defined(NO_SWAPPING)
522 * Callback function, page busied for us. We must dispose of the busy
523 * condition. Any related pmap pages may be held but will not be locked.
527 vm_pageout_mdp_callback(struct pmap_pgscan_info
*info
, vm_offset_t va
,
534 * Basic tests - There should never be a marker, and we can stop
535 * once the RSS is below the required level.
537 KKASSERT((p
->flags
& PG_MARKER
) == 0);
538 if (pmap_resident_tlnw_count(info
->pmap
) <= info
->limit
) {
543 mycpu
->gd_cnt
.v_pdpages
++;
545 if (p
->wire_count
|| p
->hold_count
|| (p
->flags
& PG_UNMANAGED
)) {
553 * Check if the page has been referened recently. If it has,
554 * activate it and skip.
556 actcount
= pmap_ts_referenced(p
);
558 vm_page_flag_set(p
, PG_REFERENCED
);
559 } else if (p
->flags
& PG_REFERENCED
) {
564 if (p
->queue
- p
->pc
!= PQ_ACTIVE
) {
565 vm_page_and_queue_spin_lock(p
);
566 if (p
->queue
- p
->pc
!= PQ_ACTIVE
) {
567 vm_page_and_queue_spin_unlock(p
);
570 vm_page_and_queue_spin_unlock(p
);
573 p
->act_count
+= actcount
;
574 if (p
->act_count
> ACT_MAX
)
575 p
->act_count
= ACT_MAX
;
577 vm_page_flag_clear(p
, PG_REFERENCED
);
583 * Remove the page from this particular pmap. Once we do this, our
584 * pmap scans will not see it again (unless it gets faulted in), so
585 * we must actively dispose of or deal with the page.
587 pmap_remove_specific(info
->pmap
, p
);
590 * If the page is not mapped to another process (i.e. as would be
591 * typical if this were a shared page from a library) then deactivate
592 * the page and clean it in two passes only.
594 * If the page hasn't been referenced since the last check, remove it
595 * from the pmap. If it is no longer mapped, deactivate it
596 * immediately, accelerating the normal decline.
598 * Once the page has been removed from the pmap the RSS code no
599 * longer tracks it so we have to make sure that it is staged for
600 * potential flush action.
602 if ((p
->flags
& PG_MAPPED
) == 0) {
603 if (p
->queue
- p
->pc
== PQ_ACTIVE
) {
604 vm_page_deactivate(p
);
606 if (p
->queue
- p
->pc
== PQ_INACTIVE
) {
612 * Ok, try to fully clean the page and any nearby pages such that at
613 * least the requested page is freed or moved to the cache queue.
615 * We usually do this synchronously to allow us to get the page into
616 * the CACHE queue quickly, which will prevent memory exhaustion if
617 * a process with a memoryuse limit is running away. However, the
618 * sysadmin may desire to set vm.swap_user_async which relaxes this
619 * and improves write performance.
622 int max_launder
= 0x7FFF;
623 int vnodes_skipped
= 0;
625 struct vnode
*vpfailed
= NULL
;
629 if (vm_pageout_memuse_mode
>= 2) {
630 vmflush_flags
= VM_PAGER_TRY_TO_CACHE
|
631 VM_PAGER_ALLOW_ACTIVE
;
632 if (swap_user_async
== 0)
633 vmflush_flags
|= VM_PAGER_PUT_SYNC
;
634 vm_page_flag_set(p
, PG_WINATCFLS
);
636 vm_pageout_page(p
, &max_launder
,
638 &vpfailed
, 1, vmflush_flags
);
648 * Must be at end to avoid SMP races.
656 * Deactivate some number of pages in a map due to set RLIMIT_RSS limits.
657 * that is relatively difficult to do. We try to keep track of where we
658 * left off last time to reduce scan overhead.
660 * Called when vm_pageout_memuse_mode is >= 1.
663 vm_pageout_map_deactivate_pages(vm_map_t map
, vm_pindex_t limit
)
665 vm_offset_t pgout_offset
;
666 struct pmap_pgscan_info info
;
669 pgout_offset
= map
->pgout_offset
;
672 kprintf("%016jx ", pgout_offset
);
674 if (pgout_offset
< VM_MIN_USER_ADDRESS
)
675 pgout_offset
= VM_MIN_USER_ADDRESS
;
676 if (pgout_offset
>= VM_MAX_USER_ADDRESS
)
678 info
.pmap
= vm_map_pmap(map
);
680 info
.beg_addr
= pgout_offset
;
681 info
.end_addr
= VM_MAX_USER_ADDRESS
;
682 info
.callback
= vm_pageout_mdp_callback
;
684 info
.actioncount
= 0;
688 pgout_offset
= info
.offset
;
690 kprintf("%016jx %08lx %08lx\n", pgout_offset
,
691 info
.cleancount
, info
.actioncount
);
694 if (pgout_offset
!= VM_MAX_USER_ADDRESS
&&
695 pmap_resident_tlnw_count(vm_map_pmap(map
)) > limit
) {
697 } else if (retries
&&
698 pmap_resident_tlnw_count(vm_map_pmap(map
)) > limit
) {
702 map
->pgout_offset
= pgout_offset
;
707 * Called when the pageout scan wants to free a page. We no longer
708 * try to cycle the vm_object here with a reference & dealloc, which can
709 * cause a non-trivial object collapse in a critical path.
711 * It is unclear why we cycled the ref_count in the past, perhaps to try
712 * to optimize shadow chain collapses but I don't quite see why it would
713 * be necessary. An OBJ_DEAD object should terminate any and all vm_pages
714 * synchronously and not have to be kicked-start.
717 vm_pageout_page_free(vm_page_t m
)
719 vm_page_protect(m
, VM_PROT_NONE
);
724 * vm_pageout_scan does the dirty work for the pageout daemon.
726 struct vm_pageout_scan_info
{
727 struct proc
*bigproc
;
731 static int vm_pageout_scan_callback(struct proc
*p
, void *data
);
734 * Scan inactive queue
736 * WARNING! Can be called from two pagedaemon threads simultaneously.
739 vm_pageout_scan_inactive(int pass
, int q
, int avail_shortage
,
743 struct vm_page marker
;
744 struct vnode
*vpfailed
; /* warning, allowed to be stale */
750 isep
= (curthread
== emergpager
);
753 * Start scanning the inactive queue for pages we can move to the
754 * cache or free. The scan will stop when the target is reached or
755 * we have scanned the entire inactive queue. Note that m->act_count
756 * is not used to form decisions for the inactive queue, only for the
759 * max_launder limits the number of dirty pages we flush per scan.
760 * For most systems a smaller value (16 or 32) is more robust under
761 * extreme memory and disk pressure because any unnecessary writes
762 * to disk can result in extreme performance degredation. However,
763 * systems with excessive dirty pages (especially when MAP_NOSYNC is
764 * used) will die horribly with limited laundering. If the pageout
765 * daemon cannot clean enough pages in the first pass, we let it go
766 * all out in succeeding passes.
768 * NOTE! THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
771 if ((max_launder
= vm_max_launder
) <= 1)
777 * Initialize our marker
779 bzero(&marker
, sizeof(marker
));
780 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
781 marker
.queue
= PQ_INACTIVE
+ q
;
783 marker
.wire_count
= 1;
786 * Inactive queue scan.
788 * NOTE: The vm_page must be spinlocked before the queue to avoid
789 * deadlocks, so it is easiest to simply iterate the loop
790 * with the queue unlocked at the top.
794 vm_page_queues_spin_lock(PQ_INACTIVE
+ q
);
795 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, &marker
, pageq
);
796 maxscan
= vm_page_queues
[PQ_INACTIVE
+ q
].lcnt
;
799 * Queue locked at top of loop to avoid stack marker issues.
801 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
802 maxscan
-- > 0 && avail_shortage
- delta
> 0)
806 KKASSERT(m
->queue
== PQ_INACTIVE
+ q
);
807 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
,
809 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, m
,
811 mycpu
->gd_cnt
.v_pdpages
++;
814 * Skip marker pages (atomic against other markers to avoid
815 * infinite hop-over scans).
817 if (m
->flags
& PG_MARKER
)
821 * Try to busy the page. Don't mess with pages which are
822 * already busy or reorder them in the queue.
824 if (vm_page_busy_try(m
, TRUE
))
828 * Remaining operations run with the page busy and neither
829 * the page or the queue will be spin-locked.
831 vm_page_queues_spin_unlock(PQ_INACTIVE
+ q
);
832 KKASSERT(m
->queue
== PQ_INACTIVE
+ q
);
835 * The emergency pager runs when the primary pager gets
836 * stuck, which typically means the primary pager deadlocked
837 * on a vnode-backed page. Therefore, the emergency pager
838 * must skip vnode-backed pages.
841 if (m
->object
&& m
->object
->type
== OBJT_VNODE
) {
843 vm_page_queues_spin_lock(PQ_INACTIVE
+ q
);
850 * Try to pageout the page and perhaps other nearby pages.
852 count
= vm_pageout_page(m
, &max_launder
, vnodes_skipped
,
857 * Systems with a ton of memory can wind up with huge
858 * deactivation counts. Because the inactive scan is
859 * doing a lot of flushing, the combination can result
860 * in excessive paging even in situations where other
861 * unrelated threads free up sufficient VM.
863 * To deal with this we abort the nominal active->inactive
864 * scan before we hit the inactive target when free+cache
865 * levels have reached a reasonable target.
867 * When deciding to stop early we need to add some slop to
868 * the test and we need to return full completion to the caller
869 * to prevent the caller from thinking there is something
870 * wrong and issuing a low-memory+swap warning or pkill.
872 * A deficit forces paging regardless of the state of the
873 * VM page queues (used for RSS enforcement).
876 vm_page_queues_spin_lock(PQ_INACTIVE
+ q
);
877 if (vm_paging_target() < -vm_max_launder
) {
879 * Stopping early, return full completion to caller.
881 if (delta
< avail_shortage
)
882 delta
= avail_shortage
;
887 /* page queue still spin-locked */
888 TAILQ_REMOVE(&vm_page_queues
[PQ_INACTIVE
+ q
].pl
, &marker
, pageq
);
889 vm_page_queues_spin_unlock(PQ_INACTIVE
+ q
);
895 * Pageout the specified page, return the total number of pages paged out
896 * (this routine may cluster).
898 * The page must be busied and soft-busied by the caller and will be disposed
899 * of by this function.
902 vm_pageout_page(vm_page_t m
, int *max_launderp
, int *vnodes_skippedp
,
903 struct vnode
**vpfailedp
, int pass
, int vmflush_flags
)
910 * It is possible for a page to be busied ad-hoc (e.g. the
911 * pmap_collect() code) and wired and race against the
912 * allocation of a new page. vm_page_alloc() may be forced
913 * to deactivate the wired page in which case it winds up
914 * on the inactive queue and must be handled here. We
915 * correct the problem simply by unqueuing the page.
918 vm_page_unqueue_nowakeup(m
);
920 kprintf("WARNING: pagedaemon: wired page on "
921 "inactive queue %p\n", m
);
926 * A held page may be undergoing I/O, so skip it.
929 vm_page_and_queue_spin_lock(m
);
930 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
932 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
934 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
935 ++vm_swapcache_inactive_heuristic
;
937 vm_page_and_queue_spin_unlock(m
);
942 if (m
->object
== NULL
|| m
->object
->ref_count
== 0) {
944 * If the object is not being used, we ignore previous
947 vm_page_flag_clear(m
, PG_REFERENCED
);
948 pmap_clear_reference(m
);
949 /* fall through to end */
950 } else if (((m
->flags
& PG_REFERENCED
) == 0) &&
951 (actcount
= pmap_ts_referenced(m
))) {
953 * Otherwise, if the page has been referenced while
954 * in the inactive queue, we bump the "activation
955 * count" upwards, making it less likely that the
956 * page will be added back to the inactive queue
957 * prematurely again. Here we check the page tables
958 * (or emulated bits, if any), given the upper level
959 * VM system not knowing anything about existing
963 m
->act_count
+= (actcount
+ ACT_ADVANCE
);
969 * (m) is still busied.
971 * If the upper level VM system knows about any page
972 * references, we activate the page. We also set the
973 * "activation count" higher than normal so that we will less
974 * likely place pages back onto the inactive queue again.
976 if ((m
->flags
& PG_REFERENCED
) != 0) {
977 vm_page_flag_clear(m
, PG_REFERENCED
);
978 actcount
= pmap_ts_referenced(m
);
980 m
->act_count
+= (actcount
+ ACT_ADVANCE
+ 1);
986 * If the upper level VM system doesn't know anything about
987 * the page being dirty, we have to check for it again. As
988 * far as the VM code knows, any partially dirty pages are
991 * Pages marked PG_WRITEABLE may be mapped into the user
992 * address space of a process running on another cpu. A
993 * user process (without holding the MP lock) running on
994 * another cpu may be able to touch the page while we are
995 * trying to remove it. vm_page_cache() will handle this
999 vm_page_test_dirty(m
);
1004 if (m
->valid
== 0 && (m
->flags
& PG_NEED_COMMIT
) == 0) {
1006 * Invalid pages can be easily freed
1008 vm_pageout_page_free(m
);
1009 mycpu
->gd_cnt
.v_dfree
++;
1011 } else if (m
->dirty
== 0 && (m
->flags
& PG_NEED_COMMIT
) == 0) {
1013 * Clean pages can be placed onto the cache queue.
1014 * This effectively frees them.
1018 } else if ((m
->flags
& PG_WINATCFLS
) == 0 && pass
== 0) {
1020 * Dirty pages need to be paged out, but flushing
1021 * a page is extremely expensive verses freeing
1022 * a clean page. Rather then artificially limiting
1023 * the number of pages we can flush, we instead give
1024 * dirty pages extra priority on the inactive queue
1025 * by forcing them to be cycled through the queue
1026 * twice before being flushed, after which the
1027 * (now clean) page will cycle through once more
1028 * before being freed. This significantly extends
1029 * the thrash point for a heavily loaded machine.
1031 vm_page_flag_set(m
, PG_WINATCFLS
);
1032 vm_page_and_queue_spin_lock(m
);
1033 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
1035 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1037 &vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1038 ++vm_swapcache_inactive_heuristic
;
1040 vm_page_and_queue_spin_unlock(m
);
1042 } else if (*max_launderp
> 0) {
1044 * We always want to try to flush some dirty pages if
1045 * we encounter them, to keep the system stable.
1046 * Normally this number is small, but under extreme
1047 * pressure where there are insufficient clean pages
1048 * on the inactive queue, we may have to go all out.
1050 int swap_pageouts_ok
;
1051 struct vnode
*vp
= NULL
;
1053 swap_pageouts_ok
= 0;
1056 (object
->type
!= OBJT_SWAP
) &&
1057 (object
->type
!= OBJT_DEFAULT
)) {
1058 swap_pageouts_ok
= 1;
1060 swap_pageouts_ok
= !(defer_swap_pageouts
||
1061 disable_swap_pageouts
);
1062 swap_pageouts_ok
|= (!disable_swap_pageouts
&&
1063 defer_swap_pageouts
&&
1064 vm_page_count_min(0));
1068 * We don't bother paging objects that are "dead".
1069 * Those objects are in a "rundown" state.
1071 if (!swap_pageouts_ok
||
1073 (object
->flags
& OBJ_DEAD
)) {
1074 vm_page_and_queue_spin_lock(m
);
1075 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
1077 &vm_page_queues
[m
->queue
].pl
,
1080 &vm_page_queues
[m
->queue
].pl
,
1082 ++vm_swapcache_inactive_heuristic
;
1084 vm_page_and_queue_spin_unlock(m
);
1090 * (m) is still busied.
1092 * The object is already known NOT to be dead. It
1093 * is possible for the vget() to block the whole
1094 * pageout daemon, but the new low-memory handling
1095 * code should prevent it.
1097 * The previous code skipped locked vnodes and, worse,
1098 * reordered pages in the queue. This results in
1099 * completely non-deterministic operation because,
1100 * quite often, a vm_fault has initiated an I/O and
1101 * is holding a locked vnode at just the point where
1102 * the pageout daemon is woken up.
1104 * We can't wait forever for the vnode lock, we might
1105 * deadlock due to a vn_read() getting stuck in
1106 * vm_wait while holding this vnode. We skip the
1107 * vnode if we can't get it in a reasonable amount
1110 * vpfailed is used to (try to) avoid the case where
1111 * a large number of pages are associated with a
1112 * locked vnode, which could cause the pageout daemon
1113 * to stall for an excessive amount of time.
1115 if (object
->type
== OBJT_VNODE
) {
1118 vp
= object
->handle
;
1119 flags
= LK_EXCLUSIVE
;
1120 if (vp
== *vpfailedp
)
1123 flags
|= LK_TIMELOCK
;
1128 * We have unbusied (m) temporarily so we can
1129 * acquire the vp lock without deadlocking.
1130 * (m) is held to prevent destruction.
1132 if (vget(vp
, flags
) != 0) {
1134 ++pageout_lock_miss
;
1135 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1142 * The page might have been moved to another
1143 * queue during potential blocking in vget()
1144 * above. The page might have been freed and
1145 * reused for another vnode. The object might
1146 * have been reused for another vnode.
1148 if (m
->queue
- m
->pc
!= PQ_INACTIVE
||
1149 m
->object
!= object
||
1150 object
->handle
!= vp
) {
1151 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1159 * The page may have been busied during the
1160 * blocking in vput(); We don't move the
1161 * page back onto the end of the queue so that
1162 * statistics are more correct if we don't.
1164 if (vm_page_busy_try(m
, TRUE
)) {
1172 * (m) is busied again
1174 * We own the busy bit and remove our hold
1175 * bit. If the page is still held it
1176 * might be undergoing I/O, so skip it.
1178 if (m
->hold_count
) {
1179 vm_page_and_queue_spin_lock(m
);
1180 if (m
->queue
- m
->pc
== PQ_INACTIVE
) {
1181 TAILQ_REMOVE(&vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1182 TAILQ_INSERT_TAIL(&vm_page_queues
[m
->queue
].pl
, m
, pageq
);
1183 ++vm_swapcache_inactive_heuristic
;
1185 vm_page_and_queue_spin_unlock(m
);
1186 if (object
->flags
& OBJ_MIGHTBEDIRTY
)
1192 /* (m) is left busied as we fall through */
1196 * page is busy and not held here.
1198 * If a page is dirty, then it is either being washed
1199 * (but not yet cleaned) or it is still in the
1200 * laundry. If it is still in the laundry, then we
1201 * start the cleaning operation.
1203 * decrement inactive_shortage on success to account
1204 * for the (future) cleaned page. Otherwise we
1205 * could wind up laundering or cleaning too many
1208 * NOTE: Cleaning the page here does not cause
1209 * force_deficit to be adjusted, because the
1210 * page is not being freed or moved to the
1213 count
= vm_pageout_clean_helper(m
, vmflush_flags
);
1214 *max_launderp
-= count
;
1217 * Clean ate busy, page no longer accessible
1230 * WARNING! Can be called from two pagedaemon threads simultaneously.
1233 vm_pageout_scan_active(int pass
, int q
,
1234 int avail_shortage
, int inactive_shortage
,
1235 int *recycle_countp
)
1237 struct vm_page marker
;
1244 isep
= (curthread
== emergpager
);
1247 * We want to move pages from the active queue to the inactive
1248 * queue to get the inactive queue to the inactive target. If
1249 * we still have a page shortage from above we try to directly free
1250 * clean pages instead of moving them.
1252 * If we do still have a shortage we keep track of the number of
1253 * pages we free or cache (recycle_count) as a measure of thrashing
1254 * between the active and inactive queues.
1256 * If we were able to completely satisfy the free+cache targets
1257 * from the inactive pool we limit the number of pages we move
1258 * from the active pool to the inactive pool to 2x the pages we
1259 * had removed from the inactive pool (with a minimum of 1/5 the
1260 * inactive target). If we were not able to completely satisfy
1261 * the free+cache targets we go for the whole target aggressively.
1263 * NOTE: Both variables can end up negative.
1264 * NOTE: We are still in a critical section.
1266 * NOTE! THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
1270 bzero(&marker
, sizeof(marker
));
1271 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
1272 marker
.queue
= PQ_ACTIVE
+ q
;
1274 marker
.wire_count
= 1;
1276 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1277 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1278 maxscan
= vm_page_queues
[PQ_ACTIVE
+ q
].lcnt
;
1281 * Queue locked at top of loop to avoid stack marker issues.
1283 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
1284 maxscan
-- > 0 && (avail_shortage
- delta
> 0 ||
1285 inactive_shortage
> 0))
1287 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1288 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1290 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, m
,
1294 * Skip marker pages (atomic against other markers to avoid
1295 * infinite hop-over scans).
1297 if (m
->flags
& PG_MARKER
)
1301 * Try to busy the page. Don't mess with pages which are
1302 * already busy or reorder them in the queue.
1304 if (vm_page_busy_try(m
, TRUE
))
1308 * Remaining operations run with the page busy and neither
1309 * the page or the queue will be spin-locked.
1311 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1312 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1315 * Don't deactivate pages that are held, even if we can
1316 * busy them. (XXX why not?)
1318 if (m
->hold_count
!= 0) {
1319 vm_page_and_queue_spin_lock(m
);
1320 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1322 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1325 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1328 vm_page_and_queue_spin_unlock(m
);
1334 * The emergency pager ignores vnode-backed pages as these
1335 * are the pages that probably bricked the main pager.
1337 if (isep
&& m
->object
&& m
->object
->type
== OBJT_VNODE
) {
1338 vm_page_and_queue_spin_lock(m
);
1339 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1341 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1344 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1347 vm_page_and_queue_spin_unlock(m
);
1353 * The count for pagedaemon pages is done after checking the
1354 * page for eligibility...
1356 mycpu
->gd_cnt
.v_pdpages
++;
1359 * Check to see "how much" the page has been used and clear
1360 * the tracking access bits. If the object has no references
1361 * don't bother paying the expense.
1364 if (m
->object
&& m
->object
->ref_count
!= 0) {
1365 if (m
->flags
& PG_REFERENCED
)
1367 actcount
+= pmap_ts_referenced(m
);
1369 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1370 if (m
->act_count
> ACT_MAX
)
1371 m
->act_count
= ACT_MAX
;
1374 vm_page_flag_clear(m
, PG_REFERENCED
);
1377 * actcount is only valid if the object ref_count is non-zero.
1378 * If the page does not have an object, actcount will be zero.
1380 if (actcount
&& m
->object
->ref_count
!= 0) {
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
);
1393 switch(m
->object
->type
) {
1396 m
->act_count
-= min(m
->act_count
,
1397 vm_anonmem_decline
);
1400 m
->act_count
-= min(m
->act_count
,
1401 vm_filemem_decline
);
1404 if (vm_pageout_algorithm
||
1405 (m
->object
== NULL
) ||
1406 (m
->object
&& (m
->object
->ref_count
== 0)) ||
1407 m
->act_count
< pass
+ 1
1410 * Deactivate the page. If we had a
1411 * shortage from our inactive scan try to
1412 * free (cache) the page instead.
1414 * Don't just blindly cache the page if
1415 * we do not have a shortage from the
1416 * inactive scan, that could lead to
1417 * gigabytes being moved.
1419 --inactive_shortage
;
1420 if (avail_shortage
- delta
> 0 ||
1421 (m
->object
&& (m
->object
->ref_count
== 0)))
1423 if (avail_shortage
- delta
> 0)
1425 vm_page_protect(m
, VM_PROT_NONE
);
1426 if (m
->dirty
== 0 &&
1427 (m
->flags
& PG_NEED_COMMIT
) == 0 &&
1428 avail_shortage
- delta
> 0) {
1431 vm_page_deactivate(m
);
1435 vm_page_deactivate(m
);
1440 vm_page_and_queue_spin_lock(m
);
1441 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1443 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1446 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1449 vm_page_and_queue_spin_unlock(m
);
1455 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1459 * Clean out our local marker.
1461 * Page queue still spin-locked.
1463 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1464 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1470 * The number of actually free pages can drop down to v_free_reserved,
1471 * we try to build the free count back above v_free_min. Note that
1472 * vm_paging_needed() also returns TRUE if v_free_count is not at
1473 * least v_free_min so that is the minimum we must build the free
1476 * We use a slightly higher target to improve hysteresis,
1477 * ((v_free_target + v_free_min) / 2). Since v_free_target
1478 * is usually the same as v_cache_min this maintains about
1479 * half the pages in the free queue as are in the cache queue,
1480 * providing pretty good pipelining for pageout operation.
1482 * The system operator can manipulate vm.v_cache_min and
1483 * vm.v_free_target to tune the pageout demon. Be sure
1484 * to keep vm.v_free_min < vm.v_free_target.
1486 * Note that the original paging target is to get at least
1487 * (free_min + cache_min) into (free + cache). The slightly
1488 * higher target will shift additional pages from cache to free
1489 * without effecting the original paging target in order to
1490 * maintain better hysteresis and not have the free count always
1491 * be dead-on v_free_min.
1493 * NOTE: we are still in a critical section.
1495 * Pages moved from PQ_CACHE to totally free are not counted in the
1496 * pages_freed counter.
1498 * WARNING! Can be called from two pagedaemon threads simultaneously.
1501 vm_pageout_scan_cache(int avail_shortage
, int pass
,
1502 int vnodes_skipped
, int recycle_count
)
1504 static int lastkillticks
;
1505 struct vm_pageout_scan_info info
;
1509 isep
= (curthread
== emergpager
);
1511 while (vmstats
.v_free_count
<
1512 (vmstats
.v_free_min
+ vmstats
.v_free_target
) / 2) {
1514 * This steals some code from vm/vm_page.c
1516 static int cache_rover
= 0;
1518 m
= vm_page_list_find(PQ_CACHE
, cache_rover
& PQ_L2_MASK
);
1521 /* page is returned removed from its queue and spinlocked */
1522 if (vm_page_busy_try(m
, TRUE
)) {
1523 vm_page_deactivate_locked(m
);
1524 vm_page_spin_unlock(m
);
1527 vm_page_spin_unlock(m
);
1528 pagedaemon_wakeup();
1532 * Remaining operations run with the page busy and neither
1533 * the page or the queue will be spin-locked.
1535 if ((m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) ||
1538 vm_page_deactivate(m
);
1542 KKASSERT((m
->flags
& PG_MAPPED
) == 0);
1543 KKASSERT(m
->dirty
== 0);
1544 cache_rover
+= PQ_PRIME2
;
1545 vm_pageout_page_free(m
);
1546 mycpu
->gd_cnt
.v_dfree
++;
1549 #if !defined(NO_SWAPPING)
1551 * Idle process swapout -- run once per second.
1553 if (vm_swap_idle_enabled
) {
1555 if (time_uptime
!= lsec
) {
1556 atomic_set_int(&vm_pageout_req_swapout
, VM_SWAP_IDLE
);
1564 * If we didn't get enough free pages, and we have skipped a vnode
1565 * in a writeable object, wakeup the sync daemon. And kick swapout
1566 * if we did not get enough free pages.
1568 if (vm_paging_target() > 0) {
1569 if (vnodes_skipped
&& vm_page_count_min(0))
1570 speedup_syncer(NULL
);
1571 #if !defined(NO_SWAPPING)
1572 if (vm_swap_enabled
&& vm_page_count_target()) {
1573 atomic_set_int(&vm_pageout_req_swapout
, VM_SWAP_NORMAL
);
1580 * Handle catastrophic conditions. Under good conditions we should
1581 * be at the target, well beyond our minimum. If we could not even
1582 * reach our minimum the system is under heavy stress. But just being
1583 * under heavy stress does not trigger process killing.
1585 * We consider ourselves to have run out of memory if the swap pager
1586 * is full and avail_shortage is still positive. The secondary check
1587 * ensures that we do not kill processes if the instantanious
1588 * availability is good, even if the pageout demon pass says it
1589 * couldn't get to the target.
1591 * NOTE! THE EMERGENCY PAGER (isep) DOES NOT HANDLE SWAP FULL
1594 if (swap_pager_almost_full
&&
1597 (vm_page_count_min(recycle_count
) || avail_shortage
> 0)) {
1598 kprintf("Warning: system low on memory+swap "
1599 "shortage %d for %d ticks!\n",
1600 avail_shortage
, ticks
- swap_fail_ticks
);
1602 kprintf("Metrics: spaf=%d spf=%d pass=%d avail=%d target=%d last=%u\n",
1603 swap_pager_almost_full
,
1608 (unsigned int)(ticks
- lastkillticks
));
1610 if (swap_pager_full
&&
1613 avail_shortage
> 0 &&
1614 vm_paging_target() > 0 &&
1615 (unsigned int)(ticks
- lastkillticks
) >= hz
) {
1617 * Kill something, maximum rate once per second to give
1618 * the process time to free up sufficient memory.
1620 lastkillticks
= ticks
;
1621 info
.bigproc
= NULL
;
1623 allproc_scan(vm_pageout_scan_callback
, &info
, 0);
1624 if (info
.bigproc
!= NULL
) {
1625 kprintf("Try to kill process %d %s\n",
1626 info
.bigproc
->p_pid
, info
.bigproc
->p_comm
);
1627 info
.bigproc
->p_nice
= PRIO_MIN
;
1628 info
.bigproc
->p_usched
->resetpriority(
1629 FIRST_LWP_IN_PROC(info
.bigproc
));
1630 atomic_set_int(&info
.bigproc
->p_flags
, P_LOWMEMKILL
);
1631 killproc(info
.bigproc
, "out of swap space");
1632 wakeup(&vmstats
.v_free_count
);
1633 PRELE(info
.bigproc
);
1639 vm_pageout_scan_callback(struct proc
*p
, void *data
)
1641 struct vm_pageout_scan_info
*info
= data
;
1645 * Never kill system processes or init. If we have configured swap
1646 * then try to avoid killing low-numbered pids.
1648 if ((p
->p_flags
& P_SYSTEM
) || (p
->p_pid
== 1) ||
1649 ((p
->p_pid
< 48) && (vm_swap_size
!= 0))) {
1653 lwkt_gettoken(&p
->p_token
);
1656 * if the process is in a non-running type state,
1659 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
&& p
->p_stat
!= SCORE
) {
1660 lwkt_reltoken(&p
->p_token
);
1665 * Get the approximate process size. Note that anonymous pages
1666 * with backing swap will be counted twice, but there should not
1667 * be too many such pages due to the stress the VM system is
1668 * under at this point.
1670 size
= vmspace_anonymous_count(p
->p_vmspace
) +
1671 vmspace_swap_count(p
->p_vmspace
);
1674 * If the this process is bigger than the biggest one
1677 if (info
->bigsize
< size
) {
1679 PRELE(info
->bigproc
);
1682 info
->bigsize
= size
;
1684 lwkt_reltoken(&p
->p_token
);
1691 * This routine tries to maintain the pseudo LRU active queue,
1692 * so that during long periods of time where there is no paging,
1693 * that some statistic accumulation still occurs. This code
1694 * helps the situation where paging just starts to occur.
1697 vm_pageout_page_stats(int q
)
1699 static int fullintervalcount
= 0;
1700 struct vm_page marker
;
1702 int pcount
, tpcount
; /* Number of pages to check */
1705 page_shortage
= (vmstats
.v_inactive_target
+ vmstats
.v_cache_max
+
1706 vmstats
.v_free_min
) -
1707 (vmstats
.v_free_count
+ vmstats
.v_inactive_count
+
1708 vmstats
.v_cache_count
);
1710 if (page_shortage
<= 0)
1713 pcount
= vm_page_queues
[PQ_ACTIVE
+ q
].lcnt
;
1714 fullintervalcount
+= vm_pageout_stats_interval
;
1715 if (fullintervalcount
< vm_pageout_full_stats_interval
) {
1716 tpcount
= (vm_pageout_stats_max
* pcount
) /
1717 vmstats
.v_page_count
+ 1;
1718 if (pcount
> tpcount
)
1721 fullintervalcount
= 0;
1724 bzero(&marker
, sizeof(marker
));
1725 marker
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_MARKER
;
1726 marker
.queue
= PQ_ACTIVE
+ q
;
1728 marker
.wire_count
= 1;
1730 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1731 TAILQ_INSERT_HEAD(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1734 * Queue locked at top of loop to avoid stack marker issues.
1736 while ((m
= TAILQ_NEXT(&marker
, pageq
)) != NULL
&&
1741 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1742 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1743 TAILQ_INSERT_AFTER(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, m
,
1747 * Skip marker pages (atomic against other markers to avoid
1748 * infinite hop-over scans).
1750 if (m
->flags
& PG_MARKER
)
1754 * Ignore pages we can't busy
1756 if (vm_page_busy_try(m
, TRUE
))
1760 * Remaining operations run with the page busy and neither
1761 * the page or the queue will be spin-locked.
1763 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1764 KKASSERT(m
->queue
== PQ_ACTIVE
+ q
);
1767 * We now have a safely busied page, the page and queue
1768 * spinlocks have been released.
1772 if (m
->hold_count
) {
1778 * Calculate activity
1781 if (m
->flags
& PG_REFERENCED
) {
1782 vm_page_flag_clear(m
, PG_REFERENCED
);
1785 actcount
+= pmap_ts_referenced(m
);
1788 * Update act_count and move page to end of queue.
1791 m
->act_count
+= ACT_ADVANCE
+ actcount
;
1792 if (m
->act_count
> ACT_MAX
)
1793 m
->act_count
= ACT_MAX
;
1794 vm_page_and_queue_spin_lock(m
);
1795 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1797 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1800 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1803 vm_page_and_queue_spin_unlock(m
);
1808 if (m
->act_count
== 0) {
1810 * We turn off page access, so that we have
1811 * more accurate RSS stats. We don't do this
1812 * in the normal page deactivation when the
1813 * system is loaded VM wise, because the
1814 * cost of the large number of page protect
1815 * operations would be higher than the value
1816 * of doing the operation.
1818 * We use the marker to save our place so
1819 * we can release the spin lock. both (m)
1820 * and (next) will be invalid.
1822 vm_page_protect(m
, VM_PROT_NONE
);
1823 vm_page_deactivate(m
);
1825 m
->act_count
-= min(m
->act_count
, ACT_DECLINE
);
1826 vm_page_and_queue_spin_lock(m
);
1827 if (m
->queue
- m
->pc
== PQ_ACTIVE
) {
1829 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1832 &vm_page_queues
[PQ_ACTIVE
+ q
].pl
,
1835 vm_page_and_queue_spin_unlock(m
);
1839 vm_page_queues_spin_lock(PQ_ACTIVE
+ q
);
1843 * Remove our local marker
1845 * Page queue still spin-locked.
1847 TAILQ_REMOVE(&vm_page_queues
[PQ_ACTIVE
+ q
].pl
, &marker
, pageq
);
1848 vm_page_queues_spin_unlock(PQ_ACTIVE
+ q
);
1852 vm_pageout_free_page_calc(vm_size_t count
)
1854 if (count
< vmstats
.v_page_count
)
1857 * free_reserved needs to include enough for the largest swap pager
1858 * structures plus enough for any pv_entry structs when paging.
1860 * v_free_min normal allocations
1861 * v_free_reserved system allocations
1862 * v_pageout_free_min allocations by pageout daemon
1863 * v_interrupt_free_min low level allocations (e.g swap structures)
1865 if (vmstats
.v_page_count
> 1024)
1866 vmstats
.v_free_min
= 64 + (vmstats
.v_page_count
- 1024) / 200;
1868 vmstats
.v_free_min
= 64;
1871 * Make sure the vmmeter slop can't blow out our global minimums.
1873 * However, to accomodate weird configurations (vkernels with many
1874 * cpus and little memory, or artifically reduced hw.physmem), do
1875 * not allow v_free_min to exceed 1/20 of ram or the pageout demon
1876 * will go out of control.
1878 if (vmstats
.v_free_min
< VMMETER_SLOP_COUNT
* ncpus
* 10)
1879 vmstats
.v_free_min
= VMMETER_SLOP_COUNT
* ncpus
* 10;
1880 if (vmstats
.v_free_min
> vmstats
.v_page_count
/ 20)
1881 vmstats
.v_free_min
= vmstats
.v_page_count
/ 20;
1883 vmstats
.v_free_reserved
= vmstats
.v_free_min
* 4 / 8 + 7;
1884 vmstats
.v_free_severe
= vmstats
.v_free_min
* 4 / 8 + 0;
1885 vmstats
.v_pageout_free_min
= vmstats
.v_free_min
* 2 / 8 + 7;
1886 vmstats
.v_interrupt_free_min
= vmstats
.v_free_min
* 1 / 8 + 7;
1893 * vm_pageout is the high level pageout daemon. TWO kernel threads run
1894 * this daemon, the primary pageout daemon and the emergency pageout daemon.
1896 * The emergency pageout daemon takes over when the primary pageout daemon
1897 * deadlocks. The emergency pageout daemon ONLY pages out to swap, thus
1898 * avoiding the many low-memory deadlocks which can occur when paging out
1902 vm_pageout_thread(void)
1911 curthread
->td_flags
|= TDF_SYSTHREAD
;
1914 * We only need to setup once.
1918 if (curthread
== emergpager
) {
1924 * Initialize some paging parameters.
1926 vm_pageout_free_page_calc(vmstats
.v_page_count
);
1929 * v_free_target and v_cache_min control pageout hysteresis. Note
1930 * that these are more a measure of the VM cache queue hysteresis
1931 * then the VM free queue. Specifically, v_free_target is the
1932 * high water mark (free+cache pages).
1934 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1935 * low water mark, while v_free_min is the stop. v_cache_min must
1936 * be big enough to handle memory needs while the pageout daemon
1937 * is signalled and run to free more pages.
1939 if (vmstats
.v_free_count
> 6144)
1940 vmstats
.v_free_target
= 4 * vmstats
.v_free_min
+
1941 vmstats
.v_free_reserved
;
1943 vmstats
.v_free_target
= 2 * vmstats
.v_free_min
+
1944 vmstats
.v_free_reserved
;
1947 * NOTE: With the new buffer cache b_act_count we want the default
1948 * inactive target to be a percentage of available memory.
1950 * The inactive target essentially determines the minimum
1951 * number of 'temporary' pages capable of caching one-time-use
1952 * files when the VM system is otherwise full of pages
1953 * belonging to multi-time-use files or active program data.
1955 * NOTE: The inactive target is aggressively persued only if the
1956 * inactive queue becomes too small. If the inactive queue
1957 * is large enough to satisfy page movement to free+cache
1958 * then it is repopulated more slowly from the active queue.
1959 * This allows a general inactive_target default to be set.
1961 * There is an issue here for processes which sit mostly idle
1962 * 'overnight', such as sshd, tcsh, and X. Any movement from
1963 * the active queue will eventually cause such pages to
1964 * recycle eventually causing a lot of paging in the morning.
1965 * To reduce the incidence of this pages cycled out of the
1966 * buffer cache are moved directly to the inactive queue if
1967 * they were only used once or twice.
1969 * The vfs.vm_cycle_point sysctl can be used to adjust this.
1970 * Increasing the value (up to 64) increases the number of
1971 * buffer recyclements which go directly to the inactive queue.
1973 if (vmstats
.v_free_count
> 2048) {
1974 vmstats
.v_cache_min
= vmstats
.v_free_target
;
1975 vmstats
.v_cache_max
= 2 * vmstats
.v_cache_min
;
1977 vmstats
.v_cache_min
= 0;
1978 vmstats
.v_cache_max
= 0;
1980 vmstats
.v_inactive_target
= vmstats
.v_free_count
/ 4;
1982 /* XXX does not really belong here */
1983 if (vm_page_max_wired
== 0)
1984 vm_page_max_wired
= vmstats
.v_free_count
/ 3;
1986 if (vm_pageout_stats_max
== 0)
1987 vm_pageout_stats_max
= vmstats
.v_free_target
;
1990 * Set interval in seconds for stats scan.
1992 if (vm_pageout_stats_interval
== 0)
1993 vm_pageout_stats_interval
= 5;
1994 if (vm_pageout_full_stats_interval
== 0)
1995 vm_pageout_full_stats_interval
= vm_pageout_stats_interval
* 4;
1999 * Set maximum free per pass
2001 if (vm_pageout_stats_free_max
== 0)
2002 vm_pageout_stats_free_max
= 5;
2004 swap_pager_swap_init();
2007 atomic_swap_int(&sequence_emerg_pager
, 1);
2008 wakeup(&sequence_emerg_pager
);
2012 * Sequence emergency pager startup
2015 while (sequence_emerg_pager
== 0)
2016 tsleep(&sequence_emerg_pager
, 0, "pstartup", hz
);
2020 * The pageout daemon is never done, so loop forever.
2022 * WARNING! This code is being executed by two kernel threads
2023 * potentially simultaneously.
2028 int inactive_shortage
;
2029 int vnodes_skipped
= 0;
2030 int recycle_count
= 0;
2034 * Wait for an action request. If we timeout check to
2035 * see if paging is needed (in case the normal wakeup
2040 * Emergency pagedaemon monitors the primary
2041 * pagedaemon while vm_pages_needed != 0.
2043 * The emergency pagedaemon only runs if VM paging
2044 * is needed and the primary pagedaemon has not
2045 * updated vm_pagedaemon_time for more than 2 seconds.
2047 if (vm_pages_needed
)
2048 tsleep(&vm_pagedaemon_time
, 0, "psleep", hz
);
2050 tsleep(&vm_pagedaemon_time
, 0, "psleep", hz
*10);
2051 if (vm_pages_needed
== 0) {
2055 if ((int)(ticks
- vm_pagedaemon_time
) < hz
* 2) {
2058 kprintf("Emergency pager finished\n");
2063 if (emrunning
== 0) {
2065 kprintf("Emergency pager running\n");
2069 * Primary pagedaemon
2071 if (vm_pages_needed
== 0) {
2072 error
= tsleep(&vm_pages_needed
,
2074 vm_pageout_stats_interval
* hz
);
2076 vm_paging_needed() == 0 &&
2077 vm_pages_needed
== 0) {
2078 for (q
= 0; q
< PQ_L2_SIZE
; ++q
)
2079 vm_pageout_page_stats(q
);
2082 vm_pagedaemon_time
= ticks
;
2083 vm_pages_needed
= 1;
2086 * Wake the emergency pagedaemon up so it
2087 * can monitor us. It will automatically
2088 * go back into a long sleep when
2089 * vm_pages_needed returns to 0.
2091 wakeup(&vm_pagedaemon_time
);
2095 mycpu
->gd_cnt
.v_pdwakeups
++;
2098 * Scan for INACTIVE->CLEAN/PAGEOUT
2100 * This routine tries to avoid thrashing the system with
2101 * unnecessary activity.
2103 * Calculate our target for the number of free+cache pages we
2104 * want to get to. This is higher then the number that causes
2105 * allocations to stall (severe) in order to provide hysteresis,
2106 * and if we don't make it all the way but get to the minimum
2107 * we're happy. Goose it a bit if there are multiple requests
2110 * Don't reduce avail_shortage inside the loop or the
2111 * PQAVERAGE() calculation will break.
2113 * NOTE! deficit is differentiated from avail_shortage as
2114 * REQUIRING at least (deficit) pages to be cleaned,
2115 * even if the page queues are in good shape. This
2116 * is used primarily for handling per-process
2117 * RLIMIT_RSS and may also see small values when
2118 * processes block due to low memory.
2122 vm_pagedaemon_time
= ticks
;
2123 avail_shortage
= vm_paging_target() + vm_pageout_deficit
;
2124 vm_pageout_deficit
= 0;
2126 if (avail_shortage
> 0) {
2129 for (q
= 0; q
< PQ_L2_SIZE
; ++q
) {
2130 delta
+= vm_pageout_scan_inactive(
2132 (q
+ q1iterator
) & PQ_L2_MASK
,
2133 PQAVERAGE(avail_shortage
),
2135 if (avail_shortage
- delta
<= 0)
2138 avail_shortage
-= delta
;
2143 * Figure out how many active pages we must deactivate. If
2144 * we were able to reach our target with just the inactive
2145 * scan above we limit the number of active pages we
2146 * deactivate to reduce unnecessary work.
2150 vm_pagedaemon_time
= ticks
;
2151 inactive_shortage
= vmstats
.v_inactive_target
-
2152 vmstats
.v_inactive_count
;
2155 * If we were unable to free sufficient inactive pages to
2156 * satisfy the free/cache queue requirements then simply
2157 * reaching the inactive target may not be good enough.
2158 * Try to deactivate pages in excess of the target based
2161 * However to prevent thrashing the VM system do not
2162 * deactivate more than an additional 1/10 the inactive
2163 * target's worth of active pages.
2165 if (avail_shortage
> 0) {
2166 tmp
= avail_shortage
* 2;
2167 if (tmp
> vmstats
.v_inactive_target
/ 10)
2168 tmp
= vmstats
.v_inactive_target
/ 10;
2169 inactive_shortage
+= tmp
;
2173 * Only trigger a pmap cleanup on inactive shortage.
2175 if (isep
== 0 && inactive_shortage
> 0) {
2180 * Scan for ACTIVE->INACTIVE
2182 * Only trigger on inactive shortage. Triggering on
2183 * avail_shortage can starve the active queue with
2184 * unnecessary active->inactive transitions and destroy
2187 * If this is the emergency pager, always try to move
2188 * a few pages from active to inactive because the inactive
2189 * queue might have enough pages, but not enough anonymous
2192 if (isep
&& inactive_shortage
< vm_emerg_launder
)
2193 inactive_shortage
= vm_emerg_launder
;
2195 if (/*avail_shortage > 0 ||*/ inactive_shortage
> 0) {
2198 for (q
= 0; q
< PQ_L2_SIZE
; ++q
) {
2199 delta
+= vm_pageout_scan_active(
2201 (q
+ q2iterator
) & PQ_L2_MASK
,
2202 PQAVERAGE(avail_shortage
),
2203 PQAVERAGE(inactive_shortage
),
2205 if (inactive_shortage
- delta
<= 0 &&
2206 avail_shortage
- delta
<= 0) {
2210 inactive_shortage
-= delta
;
2211 avail_shortage
-= delta
;
2216 * Scan for CACHE->FREE
2218 * Finally free enough cache pages to meet our free page
2219 * requirement and take more drastic measures if we are
2224 vm_pagedaemon_time
= ticks
;
2225 vm_pageout_scan_cache(avail_shortage
, pass
,
2226 vnodes_skipped
, recycle_count
);
2229 * Wait for more work.
2231 if (avail_shortage
> 0) {
2233 if (pass
< 10 && vm_pages_needed
> 1) {
2235 * Normal operation, additional processes
2236 * have already kicked us. Retry immediately
2237 * unless swap space is completely full in
2238 * which case delay a bit.
2240 if (swap_pager_full
) {
2241 tsleep(&vm_pages_needed
, 0, "pdelay",
2243 } /* else immediate retry */
2244 } else if (pass
< 10) {
2246 * Normal operation, fewer processes. Delay
2247 * a bit but allow wakeups. vm_pages_needed
2248 * is only adjusted against the primary
2252 vm_pages_needed
= 0;
2253 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
2255 vm_pages_needed
= 1;
2256 } else if (swap_pager_full
== 0) {
2258 * We've taken too many passes, forced delay.
2260 tsleep(&vm_pages_needed
, 0, "pdelay", hz
/ 10);
2263 * Running out of memory, catastrophic
2264 * back-off to one-second intervals.
2266 tsleep(&vm_pages_needed
, 0, "pdelay", hz
);
2268 } else if (vm_pages_needed
) {
2270 * Interlocked wakeup of waiters (non-optional).
2272 * Similar to vm_page_free_wakeup() in vm_page.c,
2276 if (!vm_page_count_min(vm_page_free_hysteresis
) ||
2277 !vm_page_count_target()) {
2278 vm_pages_needed
= 0;
2279 wakeup(&vmstats
.v_free_count
);
2287 static struct kproc_desc pg1_kp
= {
2292 SYSINIT(pagedaemon
, SI_SUB_KTHREAD_PAGE
, SI_ORDER_FIRST
, kproc_start
, &pg1_kp
);
2294 static struct kproc_desc pg2_kp
= {
2299 SYSINIT(emergpager
, SI_SUB_KTHREAD_PAGE
, SI_ORDER_ANY
, kproc_start
, &pg2_kp
);
2303 * Called after allocating a page out of the cache or free queue
2304 * to possibly wake the pagedaemon up to replentish our supply.
2306 * We try to generate some hysteresis by waking the pagedaemon up
2307 * when our free+cache pages go below the free_min+cache_min level.
2308 * The pagedaemon tries to get the count back up to at least the
2309 * minimum, and through to the target level if possible.
2311 * If the pagedaemon is already active bump vm_pages_needed as a hint
2312 * that there are even more requests pending.
2318 pagedaemon_wakeup(void)
2320 if (vm_paging_needed() && curthread
!= pagethread
) {
2321 if (vm_pages_needed
== 0) {
2322 vm_pages_needed
= 1; /* SMP race ok */
2323 wakeup(&vm_pages_needed
);
2324 } else if (vm_page_count_min(0)) {
2325 ++vm_pages_needed
; /* SMP race ok */
2330 #if !defined(NO_SWAPPING)
2337 vm_req_vmdaemon(void)
2339 static int lastrun
= 0;
2341 if ((ticks
> (lastrun
+ hz
)) || (ticks
< lastrun
)) {
2342 wakeup(&vm_daemon_needed
);
2347 static int vm_daemon_callback(struct proc
*p
, void *data __unused
);
2358 tsleep(&vm_daemon_needed
, 0, "psleep", 0);
2359 req_swapout
= atomic_swap_int(&vm_pageout_req_swapout
, 0);
2365 swapout_procs(vm_pageout_req_swapout
);
2368 * scan the processes for exceeding their rlimits or if
2369 * process is swapped out -- deactivate pages
2371 allproc_scan(vm_daemon_callback
, NULL
, 0);
2376 vm_daemon_callback(struct proc
*p
, void *data __unused
)
2379 vm_pindex_t limit
, size
;
2382 * if this is a system process or if we have already
2383 * looked at this process, skip it.
2385 lwkt_gettoken(&p
->p_token
);
2387 if (p
->p_flags
& (P_SYSTEM
| P_WEXIT
)) {
2388 lwkt_reltoken(&p
->p_token
);
2393 * if the process is in a non-running type state,
2396 if (p
->p_stat
!= SACTIVE
&& p
->p_stat
!= SSTOP
&& p
->p_stat
!= SCORE
) {
2397 lwkt_reltoken(&p
->p_token
);
2404 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
2405 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
2408 * let processes that are swapped out really be
2409 * swapped out. Set the limit to nothing to get as
2410 * many pages out to swap as possible.
2412 if (p
->p_flags
& P_SWAPPEDOUT
)
2417 size
= pmap_resident_tlnw_count(&vm
->vm_pmap
);
2418 if (limit
>= 0 && size
> 4096 &&
2419 size
- 4096 >= limit
&& vm_pageout_memuse_mode
>= 1) {
2420 vm_pageout_map_deactivate_pages(&vm
->vm_map
, limit
);
2424 lwkt_reltoken(&p
->p_token
);