FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_pageout.c
1 /*-
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 * Copyright (c) 2005 Yahoo! Technologies Norway AS
9 * All rights reserved.
10 *
11 * This code is derived from software contributed to Berkeley by
12 * The Mach Operating System project at Carnegie-Mellon University.
13 *
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions
16 * are met:
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in the
21 * documentation and/or other materials provided with the distribution.
22 * 3. All advertising materials mentioning features or use of this software
23 * must display the following acknowledgement:
24 * This product includes software developed by the University of
25 * California, Berkeley and its contributors.
26 * 4. Neither the name of the University nor the names of its contributors
27 * may be used to endorse or promote products derived from this software
28 * without specific prior written permission.
29 *
30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40 * SUCH DAMAGE.
41 *
42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
43 *
44 *
45 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46 * All rights reserved.
47 *
48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49 *
50 * Permission to use, copy, modify and distribute this software and
51 * its documentation is hereby granted, provided that both the copyright
52 * notice and this permission notice appear in all copies of the
53 * software, derivative works or modified versions, and any portions
54 * thereof, and that both notices appear in supporting documentation.
55 *
56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59 *
60 * Carnegie Mellon requests users of this software to return to
61 *
62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
63 * School of Computer Science
64 * Carnegie Mellon University
65 * Pittsburgh PA 15213-3890
66 *
67 * any improvements or extensions that they make and grant Carnegie the
68 * rights to redistribute these changes.
69 */
70
71 /*
72 * The proverbial page-out daemon.
73 */
74
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD: releng/7.3/sys/vm/vm_pageout.c 197197 2009-09-14 17:34:49Z jhb $");
77
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
94 #include <sys/sx.h>
95 #include <sys/sysctl.h>
96
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/uma.h>
107
108 #include <machine/mutex.h>
109
110 /*
111 * System initialization
112 */
113
114 /* the kernel process "vm_pageout"*/
115 static void vm_pageout(void);
116 static int vm_pageout_clean(vm_page_t);
117 static void vm_pageout_scan(int pass);
118
119 struct proc *pageproc;
120
121 static struct kproc_desc page_kp = {
122 "pagedaemon",
123 vm_pageout,
124 &pageproc
125 };
126 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
127 &page_kp);
128
129 #if !defined(NO_SWAPPING)
130 /* the kernel process "vm_daemon"*/
131 static void vm_daemon(void);
132 static struct proc *vmproc;
133
134 static struct kproc_desc vm_kp = {
135 "vmdaemon",
136 vm_daemon,
137 &vmproc
138 };
139 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
140 #endif
141
142
143 int vm_pages_needed; /* Event on which pageout daemon sleeps */
144 int vm_pageout_deficit; /* Estimated number of pages deficit */
145 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
146
147 #if !defined(NO_SWAPPING)
148 static int vm_pageout_req_swapout; /* XXX */
149 static int vm_daemon_needed;
150 static struct mtx vm_daemon_mtx;
151 /* Allow for use by vm_pageout before vm_daemon is initialized. */
152 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
153 #endif
154 static int vm_max_launder = 32;
155 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
156 static int vm_pageout_full_stats_interval = 0;
157 static int vm_pageout_algorithm=0;
158 static int defer_swap_pageouts=0;
159 static int disable_swap_pageouts=0;
160
161 #if defined(NO_SWAPPING)
162 static int vm_swap_enabled=0;
163 static int vm_swap_idle_enabled=0;
164 #else
165 static int vm_swap_enabled=1;
166 static int vm_swap_idle_enabled=0;
167 #endif
168
169 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
170 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
171
172 SYSCTL_INT(_vm, OID_AUTO, max_launder,
173 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
174
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
176 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
177
178 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
179 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
180
181 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
182 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
183
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, "");
189 #else
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");
194 #endif
195
196 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
197 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
198
199 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
200 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
201
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");
205
206 #define VM_PAGEOUT_PAGE_COUNT 16
207 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
208
209 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
210 SYSCTL_INT(_vm, OID_AUTO, max_wired,
211 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
212
213 #if !defined(NO_SWAPPING)
214 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
215 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
216 static void vm_req_vmdaemon(int req);
217 #endif
218 static void vm_pageout_page_stats(void);
219
220 /*
221 * vm_pageout_fallback_object_lock:
222 *
223 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
224 * known to have failed and page queue must be either PQ_ACTIVE or
225 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
226 * while locking the vm object. Use marker page to detect page queue
227 * changes and maintain notion of next page on page queue. Return
228 * TRUE if no changes were detected, FALSE otherwise. vm object is
229 * locked on return.
230 *
231 * This function depends on both the lock portion of struct vm_object
232 * and normal struct vm_page being type stable.
233 */
234 boolean_t
235 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
236 {
237 struct vm_page marker;
238 boolean_t unchanged;
239 u_short queue;
240 vm_object_t object;
241
242 /*
243 * Initialize our marker
244 */
245 bzero(&marker, sizeof(marker));
246 marker.flags = PG_FICTITIOUS | PG_MARKER;
247 marker.oflags = VPO_BUSY;
248 marker.queue = m->queue;
249 marker.wire_count = 1;
250
251 queue = m->queue;
252 object = m->object;
253
254 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
255 m, &marker, pageq);
256 vm_page_unlock_queues();
257 VM_OBJECT_LOCK(object);
258 vm_page_lock_queues();
259
260 /* Page queue might have changed. */
261 *next = TAILQ_NEXT(&marker, pageq);
262 unchanged = (m->queue == queue &&
263 m->object == object &&
264 &marker == TAILQ_NEXT(m, pageq));
265 TAILQ_REMOVE(&vm_page_queues[queue].pl,
266 &marker, pageq);
267 return (unchanged);
268 }
269
270 /*
271 * vm_pageout_clean:
272 *
273 * Clean the page and remove it from the laundry.
274 *
275 * We set the busy bit to cause potential page faults on this page to
276 * block. Note the careful timing, however, the busy bit isn't set till
277 * late and we cannot do anything that will mess with the page.
278 */
279 static int
280 vm_pageout_clean(m)
281 vm_page_t m;
282 {
283 vm_object_t object;
284 vm_page_t mc[2*vm_pageout_page_count];
285 int pageout_count;
286 int ib, is, page_base;
287 vm_pindex_t pindex = m->pindex;
288
289 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
290 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
291
292 /*
293 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
294 * with the new swapper, but we could have serious problems paging
295 * out other object types if there is insufficient memory.
296 *
297 * Unfortunately, checking free memory here is far too late, so the
298 * check has been moved up a procedural level.
299 */
300
301 /*
302 * Can't clean the page if it's busy or held.
303 */
304 if ((m->hold_count != 0) ||
305 ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
306 return 0;
307 }
308
309 mc[vm_pageout_page_count] = m;
310 pageout_count = 1;
311 page_base = vm_pageout_page_count;
312 ib = 1;
313 is = 1;
314
315 /*
316 * Scan object for clusterable pages.
317 *
318 * We can cluster ONLY if: ->> the page is NOT
319 * clean, wired, busy, held, or mapped into a
320 * buffer, and one of the following:
321 * 1) The page is inactive, or a seldom used
322 * active page.
323 * -or-
324 * 2) we force the issue.
325 *
326 * During heavy mmap/modification loads the pageout
327 * daemon can really fragment the underlying file
328 * due to flushing pages out of order and not trying
329 * align the clusters (which leave sporatic out-of-order
330 * holes). To solve this problem we do the reverse scan
331 * first and attempt to align our cluster, then do a
332 * forward scan if room remains.
333 */
334 object = m->object;
335 more:
336 while (ib && pageout_count < vm_pageout_page_count) {
337 vm_page_t p;
338
339 if (ib > pindex) {
340 ib = 0;
341 break;
342 }
343
344 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
345 ib = 0;
346 break;
347 }
348 if ((p->oflags & VPO_BUSY) || p->busy) {
349 ib = 0;
350 break;
351 }
352 vm_page_test_dirty(p);
353 if ((p->dirty & p->valid) == 0 ||
354 p->queue != PQ_INACTIVE ||
355 p->wire_count != 0 || /* may be held by buf cache */
356 p->hold_count != 0) { /* may be undergoing I/O */
357 ib = 0;
358 break;
359 }
360 mc[--page_base] = p;
361 ++pageout_count;
362 ++ib;
363 /*
364 * alignment boundry, stop here and switch directions. Do
365 * not clear ib.
366 */
367 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
368 break;
369 }
370
371 while (pageout_count < vm_pageout_page_count &&
372 pindex + is < object->size) {
373 vm_page_t p;
374
375 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
376 break;
377 if ((p->oflags & VPO_BUSY) || p->busy) {
378 break;
379 }
380 vm_page_test_dirty(p);
381 if ((p->dirty & p->valid) == 0 ||
382 p->queue != PQ_INACTIVE ||
383 p->wire_count != 0 || /* may be held by buf cache */
384 p->hold_count != 0) { /* may be undergoing I/O */
385 break;
386 }
387 mc[page_base + pageout_count] = p;
388 ++pageout_count;
389 ++is;
390 }
391
392 /*
393 * If we exhausted our forward scan, continue with the reverse scan
394 * when possible, even past a page boundry. This catches boundry
395 * conditions.
396 */
397 if (ib && pageout_count < vm_pageout_page_count)
398 goto more;
399
400 /*
401 * we allow reads during pageouts...
402 */
403 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
404 }
405
406 /*
407 * vm_pageout_flush() - launder the given pages
408 *
409 * The given pages are laundered. Note that we setup for the start of
410 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
411 * reference count all in here rather then in the parent. If we want
412 * the parent to do more sophisticated things we may have to change
413 * the ordering.
414 */
415 int
416 vm_pageout_flush(vm_page_t *mc, int count, int flags)
417 {
418 vm_object_t object = mc[0]->object;
419 int pageout_status[count];
420 int numpagedout = 0;
421 int i;
422
423 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
424 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
425 /*
426 * Initiate I/O. Bump the vm_page_t->busy counter and
427 * mark the pages read-only.
428 *
429 * We do not have to fixup the clean/dirty bits here... we can
430 * allow the pager to do it after the I/O completes.
431 *
432 * NOTE! mc[i]->dirty may be partial or fragmented due to an
433 * edge case with file fragments.
434 */
435 for (i = 0; i < count; i++) {
436 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
437 ("vm_pageout_flush: partially invalid page %p index %d/%d",
438 mc[i], i, count));
439 vm_page_io_start(mc[i]);
440 pmap_remove_write(mc[i]);
441 }
442 vm_page_unlock_queues();
443 vm_object_pip_add(object, count);
444
445 vm_pager_put_pages(object, mc, count, flags, pageout_status);
446
447 vm_page_lock_queues();
448 for (i = 0; i < count; i++) {
449 vm_page_t mt = mc[i];
450
451 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
452 (mt->flags & PG_WRITEABLE) == 0,
453 ("vm_pageout_flush: page %p is not write protected", mt));
454 switch (pageout_status[i]) {
455 case VM_PAGER_OK:
456 case VM_PAGER_PEND:
457 numpagedout++;
458 break;
459 case VM_PAGER_BAD:
460 /*
461 * Page outside of range of object. Right now we
462 * essentially lose the changes by pretending it
463 * worked.
464 */
465 pmap_clear_modify(mt);
466 vm_page_undirty(mt);
467 break;
468 case VM_PAGER_ERROR:
469 case VM_PAGER_FAIL:
470 /*
471 * If page couldn't be paged out, then reactivate the
472 * page so it doesn't clog the inactive list. (We
473 * will try paging out it again later).
474 */
475 vm_page_activate(mt);
476 break;
477 case VM_PAGER_AGAIN:
478 break;
479 }
480
481 /*
482 * If the operation is still going, leave the page busy to
483 * block all other accesses. Also, leave the paging in
484 * progress indicator set so that we don't attempt an object
485 * collapse.
486 */
487 if (pageout_status[i] != VM_PAGER_PEND) {
488 vm_object_pip_wakeup(object);
489 vm_page_io_finish(mt);
490 if (vm_page_count_severe())
491 vm_page_try_to_cache(mt);
492 }
493 }
494 return numpagedout;
495 }
496
497 #if !defined(NO_SWAPPING)
498 /*
499 * vm_pageout_object_deactivate_pages
500 *
501 * deactivate enough pages to satisfy the inactive target
502 * requirements or if vm_page_proc_limit is set, then
503 * deactivate all of the pages in the object and its
504 * backing_objects.
505 *
506 * The object and map must be locked.
507 */
508 static void
509 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
510 pmap_t pmap;
511 vm_object_t first_object;
512 long desired;
513 {
514 vm_object_t backing_object, object;
515 vm_page_t p, next;
516 int actcount, rcount, remove_mode;
517
518 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
519 if (first_object->type == OBJT_DEVICE ||
520 first_object->type == OBJT_SG ||
521 first_object->type == OBJT_PHYS)
522 return;
523 for (object = first_object;; object = backing_object) {
524 if (pmap_resident_count(pmap) <= desired)
525 goto unlock_return;
526 if (object->paging_in_progress)
527 goto unlock_return;
528
529 remove_mode = 0;
530 if (object->shadow_count > 1)
531 remove_mode = 1;
532 /*
533 * scan the objects entire memory queue
534 */
535 rcount = object->resident_page_count;
536 p = TAILQ_FIRST(&object->memq);
537 vm_page_lock_queues();
538 while (p && (rcount-- > 0)) {
539 if (pmap_resident_count(pmap) <= desired) {
540 vm_page_unlock_queues();
541 goto unlock_return;
542 }
543 next = TAILQ_NEXT(p, listq);
544 cnt.v_pdpages++;
545 if (p->wire_count != 0 ||
546 p->hold_count != 0 ||
547 p->busy != 0 ||
548 (p->oflags & VPO_BUSY) ||
549 (p->flags & PG_UNMANAGED) ||
550 !pmap_page_exists_quick(pmap, p)) {
551 p = next;
552 continue;
553 }
554 actcount = pmap_ts_referenced(p);
555 if (actcount) {
556 vm_page_flag_set(p, PG_REFERENCED);
557 } else if (p->flags & PG_REFERENCED) {
558 actcount = 1;
559 }
560 if ((p->queue != PQ_ACTIVE) &&
561 (p->flags & PG_REFERENCED)) {
562 vm_page_activate(p);
563 p->act_count += actcount;
564 vm_page_flag_clear(p, PG_REFERENCED);
565 } else if (p->queue == PQ_ACTIVE) {
566 if ((p->flags & PG_REFERENCED) == 0) {
567 p->act_count -= min(p->act_count, ACT_DECLINE);
568 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
569 pmap_remove_all(p);
570 vm_page_deactivate(p);
571 } else {
572 vm_page_requeue(p);
573 }
574 } else {
575 vm_page_activate(p);
576 vm_page_flag_clear(p, PG_REFERENCED);
577 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
578 p->act_count += ACT_ADVANCE;
579 vm_page_requeue(p);
580 }
581 } else if (p->queue == PQ_INACTIVE) {
582 pmap_remove_all(p);
583 }
584 p = next;
585 }
586 vm_page_unlock_queues();
587 if ((backing_object = object->backing_object) == NULL)
588 goto unlock_return;
589 VM_OBJECT_LOCK(backing_object);
590 if (object != first_object)
591 VM_OBJECT_UNLOCK(object);
592 }
593 unlock_return:
594 if (object != first_object)
595 VM_OBJECT_UNLOCK(object);
596 }
597
598 /*
599 * deactivate some number of pages in a map, try to do it fairly, but
600 * that is really hard to do.
601 */
602 static void
603 vm_pageout_map_deactivate_pages(map, desired)
604 vm_map_t map;
605 long desired;
606 {
607 vm_map_entry_t tmpe;
608 vm_object_t obj, bigobj;
609 int nothingwired;
610
611 if (!vm_map_trylock(map))
612 return;
613
614 bigobj = NULL;
615 nothingwired = TRUE;
616
617 /*
618 * first, search out the biggest object, and try to free pages from
619 * that.
620 */
621 tmpe = map->header.next;
622 while (tmpe != &map->header) {
623 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
624 obj = tmpe->object.vm_object;
625 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
626 if (obj->shadow_count <= 1 &&
627 (bigobj == NULL ||
628 bigobj->resident_page_count < obj->resident_page_count)) {
629 if (bigobj != NULL)
630 VM_OBJECT_UNLOCK(bigobj);
631 bigobj = obj;
632 } else
633 VM_OBJECT_UNLOCK(obj);
634 }
635 }
636 if (tmpe->wired_count > 0)
637 nothingwired = FALSE;
638 tmpe = tmpe->next;
639 }
640
641 if (bigobj != NULL) {
642 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
643 VM_OBJECT_UNLOCK(bigobj);
644 }
645 /*
646 * Next, hunt around for other pages to deactivate. We actually
647 * do this search sort of wrong -- .text first is not the best idea.
648 */
649 tmpe = map->header.next;
650 while (tmpe != &map->header) {
651 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
652 break;
653 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
654 obj = tmpe->object.vm_object;
655 if (obj != NULL) {
656 VM_OBJECT_LOCK(obj);
657 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
658 VM_OBJECT_UNLOCK(obj);
659 }
660 }
661 tmpe = tmpe->next;
662 }
663
664 /*
665 * Remove all mappings if a process is swapped out, this will free page
666 * table pages.
667 */
668 if (desired == 0 && nothingwired) {
669 pmap_remove(vm_map_pmap(map), vm_map_min(map),
670 vm_map_max(map));
671 }
672 vm_map_unlock(map);
673 }
674 #endif /* !defined(NO_SWAPPING) */
675
676 /*
677 * vm_pageout_scan does the dirty work for the pageout daemon.
678 */
679 static void
680 vm_pageout_scan(int pass)
681 {
682 vm_page_t m, next;
683 struct vm_page marker;
684 int page_shortage, maxscan, pcount;
685 int addl_page_shortage, addl_page_shortage_init;
686 vm_object_t object;
687 int actcount;
688 int vnodes_skipped = 0;
689 int maxlaunder;
690
691 /*
692 * Decrease registered cache sizes.
693 */
694 EVENTHANDLER_INVOKE(vm_lowmem, 0);
695 /*
696 * We do this explicitly after the caches have been drained above.
697 */
698 uma_reclaim();
699
700 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
701
702 /*
703 * Calculate the number of pages we want to either free or move
704 * to the cache.
705 */
706 page_shortage = vm_paging_target() + addl_page_shortage_init;
707
708 /*
709 * Initialize our marker
710 */
711 bzero(&marker, sizeof(marker));
712 marker.flags = PG_FICTITIOUS | PG_MARKER;
713 marker.oflags = VPO_BUSY;
714 marker.queue = PQ_INACTIVE;
715 marker.wire_count = 1;
716
717 /*
718 * Start scanning the inactive queue for pages we can move to the
719 * cache or free. The scan will stop when the target is reached or
720 * we have scanned the entire inactive queue. Note that m->act_count
721 * is not used to form decisions for the inactive queue, only for the
722 * active queue.
723 *
724 * maxlaunder limits the number of dirty pages we flush per scan.
725 * For most systems a smaller value (16 or 32) is more robust under
726 * extreme memory and disk pressure because any unnecessary writes
727 * to disk can result in extreme performance degredation. However,
728 * systems with excessive dirty pages (especially when MAP_NOSYNC is
729 * used) will die horribly with limited laundering. If the pageout
730 * daemon cannot clean enough pages in the first pass, we let it go
731 * all out in succeeding passes.
732 */
733 if ((maxlaunder = vm_max_launder) <= 1)
734 maxlaunder = 1;
735 if (pass)
736 maxlaunder = 10000;
737 vm_page_lock_queues();
738 rescan0:
739 addl_page_shortage = addl_page_shortage_init;
740 maxscan = cnt.v_inactive_count;
741
742 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
743 m != NULL && maxscan-- > 0 && page_shortage > 0;
744 m = next) {
745
746 cnt.v_pdpages++;
747
748 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
749 goto rescan0;
750 }
751
752 next = TAILQ_NEXT(m, pageq);
753 object = m->object;
754
755 /*
756 * skip marker pages
757 */
758 if (m->flags & PG_MARKER)
759 continue;
760
761 /*
762 * A held page may be undergoing I/O, so skip it.
763 */
764 if (m->hold_count) {
765 vm_page_requeue(m);
766 addl_page_shortage++;
767 continue;
768 }
769 /*
770 * Don't mess with busy pages, keep in the front of the
771 * queue, most likely are being paged out.
772 */
773 if (!VM_OBJECT_TRYLOCK(object) &&
774 (!vm_pageout_fallback_object_lock(m, &next) ||
775 m->hold_count != 0)) {
776 VM_OBJECT_UNLOCK(object);
777 addl_page_shortage++;
778 continue;
779 }
780 if (m->busy || (m->oflags & VPO_BUSY)) {
781 VM_OBJECT_UNLOCK(object);
782 addl_page_shortage++;
783 continue;
784 }
785
786 /*
787 * If the object is not being used, we ignore previous
788 * references.
789 */
790 if (object->ref_count == 0) {
791 vm_page_flag_clear(m, PG_REFERENCED);
792 pmap_clear_reference(m);
793
794 /*
795 * Otherwise, if the page has been referenced while in the
796 * inactive queue, we bump the "activation count" upwards,
797 * making it less likely that the page will be added back to
798 * the inactive queue prematurely again. Here we check the
799 * page tables (or emulated bits, if any), given the upper
800 * level VM system not knowing anything about existing
801 * references.
802 */
803 } else if (((m->flags & PG_REFERENCED) == 0) &&
804 (actcount = pmap_ts_referenced(m))) {
805 vm_page_activate(m);
806 VM_OBJECT_UNLOCK(object);
807 m->act_count += (actcount + ACT_ADVANCE);
808 continue;
809 }
810
811 /*
812 * If the upper level VM system knows about any page
813 * references, we activate the page. We also set the
814 * "activation count" higher than normal so that we will less
815 * likely place pages back onto the inactive queue again.
816 */
817 if ((m->flags & PG_REFERENCED) != 0) {
818 vm_page_flag_clear(m, PG_REFERENCED);
819 actcount = pmap_ts_referenced(m);
820 vm_page_activate(m);
821 VM_OBJECT_UNLOCK(object);
822 m->act_count += (actcount + ACT_ADVANCE + 1);
823 continue;
824 }
825
826 /*
827 * If the upper level VM system doesn't know anything about
828 * the page being dirty, we have to check for it again. As
829 * far as the VM code knows, any partially dirty pages are
830 * fully dirty.
831 */
832 if (m->dirty == 0 && !pmap_is_modified(m)) {
833 /*
834 * Avoid a race condition: Unless write access is
835 * removed from the page, another processor could
836 * modify it before all access is removed by the call
837 * to vm_page_cache() below. If vm_page_cache() finds
838 * that the page has been modified when it removes all
839 * access, it panics because it cannot cache dirty
840 * pages. In principle, we could eliminate just write
841 * access here rather than all access. In the expected
842 * case, when there are no last instant modifications
843 * to the page, removing all access will be cheaper
844 * overall.
845 */
846 if ((m->flags & PG_WRITEABLE) != 0)
847 pmap_remove_all(m);
848 } else {
849 vm_page_dirty(m);
850 }
851
852 if (m->valid == 0) {
853 /*
854 * Invalid pages can be easily freed
855 */
856 vm_page_free(m);
857 cnt.v_dfree++;
858 --page_shortage;
859 } else if (m->dirty == 0) {
860 /*
861 * Clean pages can be placed onto the cache queue.
862 * This effectively frees them.
863 */
864 vm_page_cache(m);
865 --page_shortage;
866 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
867 /*
868 * Dirty pages need to be paged out, but flushing
869 * a page is extremely expensive verses freeing
870 * a clean page. Rather then artificially limiting
871 * the number of pages we can flush, we instead give
872 * dirty pages extra priority on the inactive queue
873 * by forcing them to be cycled through the queue
874 * twice before being flushed, after which the
875 * (now clean) page will cycle through once more
876 * before being freed. This significantly extends
877 * the thrash point for a heavily loaded machine.
878 */
879 vm_page_flag_set(m, PG_WINATCFLS);
880 vm_page_requeue(m);
881 } else if (maxlaunder > 0) {
882 /*
883 * We always want to try to flush some dirty pages if
884 * we encounter them, to keep the system stable.
885 * Normally this number is small, but under extreme
886 * pressure where there are insufficient clean pages
887 * on the inactive queue, we may have to go all out.
888 */
889 int swap_pageouts_ok, vfslocked = 0;
890 struct vnode *vp = NULL;
891 struct mount *mp = NULL;
892
893 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
894 swap_pageouts_ok = 1;
895 } else {
896 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
897 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
898 vm_page_count_min());
899
900 }
901
902 /*
903 * We don't bother paging objects that are "dead".
904 * Those objects are in a "rundown" state.
905 */
906 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
907 VM_OBJECT_UNLOCK(object);
908 vm_page_requeue(m);
909 continue;
910 }
911
912 /*
913 * Following operations may unlock
914 * vm_page_queue_mtx, invalidating the 'next'
915 * pointer. To prevent an inordinate number
916 * of restarts we use our marker to remember
917 * our place.
918 *
919 */
920 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
921 m, &marker, pageq);
922 /*
923 * The object is already known NOT to be dead. It
924 * is possible for the vget() to block the whole
925 * pageout daemon, but the new low-memory handling
926 * code should prevent it.
927 *
928 * The previous code skipped locked vnodes and, worse,
929 * reordered pages in the queue. This results in
930 * completely non-deterministic operation and, on a
931 * busy system, can lead to extremely non-optimal
932 * pageouts. For example, it can cause clean pages
933 * to be freed and dirty pages to be moved to the end
934 * of the queue. Since dirty pages are also moved to
935 * the end of the queue once-cleaned, this gives
936 * way too large a weighting to defering the freeing
937 * of dirty pages.
938 *
939 * We can't wait forever for the vnode lock, we might
940 * deadlock due to a vn_read() getting stuck in
941 * vm_wait while holding this vnode. We skip the
942 * vnode if we can't get it in a reasonable amount
943 * of time.
944 */
945 if (object->type == OBJT_VNODE) {
946 vp = object->handle;
947 if (vp->v_type == VREG &&
948 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
949 mp = NULL;
950 ++pageout_lock_miss;
951 if (object->flags & OBJ_MIGHTBEDIRTY)
952 vnodes_skipped++;
953 goto unlock_and_continue;
954 }
955 vm_page_unlock_queues();
956 vm_object_reference_locked(object);
957 VM_OBJECT_UNLOCK(object);
958 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
959 if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
960 curthread)) {
961 VM_OBJECT_LOCK(object);
962 vm_page_lock_queues();
963 ++pageout_lock_miss;
964 if (object->flags & OBJ_MIGHTBEDIRTY)
965 vnodes_skipped++;
966 vp = NULL;
967 goto unlock_and_continue;
968 }
969 VM_OBJECT_LOCK(object);
970 vm_page_lock_queues();
971 /*
972 * The page might have been moved to another
973 * queue during potential blocking in vget()
974 * above. The page might have been freed and
975 * reused for another vnode.
976 */
977 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
978 m->object != object ||
979 TAILQ_NEXT(m, pageq) != &marker) {
980 if (object->flags & OBJ_MIGHTBEDIRTY)
981 vnodes_skipped++;
982 goto unlock_and_continue;
983 }
984
985 /*
986 * The page may have been busied during the
987 * blocking in vget(). We don't move the
988 * page back onto the end of the queue so that
989 * statistics are more correct if we don't.
990 */
991 if (m->busy || (m->oflags & VPO_BUSY)) {
992 goto unlock_and_continue;
993 }
994
995 /*
996 * If the page has become held it might
997 * be undergoing I/O, so skip it
998 */
999 if (m->hold_count) {
1000 vm_page_requeue(m);
1001 if (object->flags & OBJ_MIGHTBEDIRTY)
1002 vnodes_skipped++;
1003 goto unlock_and_continue;
1004 }
1005 }
1006
1007 /*
1008 * If a page is dirty, then it is either being washed
1009 * (but not yet cleaned) or it is still in the
1010 * laundry. If it is still in the laundry, then we
1011 * start the cleaning operation.
1012 *
1013 * decrement page_shortage on success to account for
1014 * the (future) cleaned page. Otherwise we could wind
1015 * up laundering or cleaning too many pages.
1016 */
1017 if (vm_pageout_clean(m) != 0) {
1018 --page_shortage;
1019 --maxlaunder;
1020 }
1021 unlock_and_continue:
1022 VM_OBJECT_UNLOCK(object);
1023 if (mp != NULL) {
1024 vm_page_unlock_queues();
1025 if (vp != NULL)
1026 vput(vp);
1027 VFS_UNLOCK_GIANT(vfslocked);
1028 vm_object_deallocate(object);
1029 vn_finished_write(mp);
1030 vm_page_lock_queues();
1031 }
1032 next = TAILQ_NEXT(&marker, pageq);
1033 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1034 &marker, pageq);
1035 continue;
1036 }
1037 VM_OBJECT_UNLOCK(object);
1038 }
1039
1040 /*
1041 * Compute the number of pages we want to try to move from the
1042 * active queue to the inactive queue.
1043 */
1044 page_shortage = vm_paging_target() +
1045 cnt.v_inactive_target - cnt.v_inactive_count;
1046 page_shortage += addl_page_shortage;
1047
1048 /*
1049 * Scan the active queue for things we can deactivate. We nominally
1050 * track the per-page activity counter and use it to locate
1051 * deactivation candidates.
1052 */
1053 pcount = cnt.v_active_count;
1054 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1055
1056 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1057
1058 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1059 ("vm_pageout_scan: page %p isn't active", m));
1060
1061 next = TAILQ_NEXT(m, pageq);
1062 object = m->object;
1063 if ((m->flags & PG_MARKER) != 0) {
1064 m = next;
1065 continue;
1066 }
1067 if (!VM_OBJECT_TRYLOCK(object) &&
1068 !vm_pageout_fallback_object_lock(m, &next)) {
1069 VM_OBJECT_UNLOCK(object);
1070 m = next;
1071 continue;
1072 }
1073
1074 /*
1075 * Don't deactivate pages that are busy.
1076 */
1077 if ((m->busy != 0) ||
1078 (m->oflags & VPO_BUSY) ||
1079 (m->hold_count != 0)) {
1080 VM_OBJECT_UNLOCK(object);
1081 vm_page_requeue(m);
1082 m = next;
1083 continue;
1084 }
1085
1086 /*
1087 * The count for pagedaemon pages is done after checking the
1088 * page for eligibility...
1089 */
1090 cnt.v_pdpages++;
1091
1092 /*
1093 * Check to see "how much" the page has been used.
1094 */
1095 actcount = 0;
1096 if (object->ref_count != 0) {
1097 if (m->flags & PG_REFERENCED) {
1098 actcount += 1;
1099 }
1100 actcount += pmap_ts_referenced(m);
1101 if (actcount) {
1102 m->act_count += ACT_ADVANCE + actcount;
1103 if (m->act_count > ACT_MAX)
1104 m->act_count = ACT_MAX;
1105 }
1106 }
1107
1108 /*
1109 * Since we have "tested" this bit, we need to clear it now.
1110 */
1111 vm_page_flag_clear(m, PG_REFERENCED);
1112
1113 /*
1114 * Only if an object is currently being used, do we use the
1115 * page activation count stats.
1116 */
1117 if (actcount && (object->ref_count != 0)) {
1118 vm_page_requeue(m);
1119 } else {
1120 m->act_count -= min(m->act_count, ACT_DECLINE);
1121 if (vm_pageout_algorithm ||
1122 object->ref_count == 0 ||
1123 m->act_count == 0) {
1124 page_shortage--;
1125 if (object->ref_count == 0) {
1126 pmap_remove_all(m);
1127 if (m->dirty == 0)
1128 vm_page_cache(m);
1129 else
1130 vm_page_deactivate(m);
1131 } else {
1132 vm_page_deactivate(m);
1133 }
1134 } else {
1135 vm_page_requeue(m);
1136 }
1137 }
1138 VM_OBJECT_UNLOCK(object);
1139 m = next;
1140 }
1141 vm_page_unlock_queues();
1142 #if !defined(NO_SWAPPING)
1143 /*
1144 * Idle process swapout -- run once per second.
1145 */
1146 if (vm_swap_idle_enabled) {
1147 static long lsec;
1148 if (time_second != lsec) {
1149 vm_req_vmdaemon(VM_SWAP_IDLE);
1150 lsec = time_second;
1151 }
1152 }
1153 #endif
1154
1155 /*
1156 * If we didn't get enough free pages, and we have skipped a vnode
1157 * in a writeable object, wakeup the sync daemon. And kick swapout
1158 * if we did not get enough free pages.
1159 */
1160 if (vm_paging_target() > 0) {
1161 if (vnodes_skipped && vm_page_count_min())
1162 (void) speedup_syncer();
1163 #if !defined(NO_SWAPPING)
1164 if (vm_swap_enabled && vm_page_count_target())
1165 vm_req_vmdaemon(VM_SWAP_NORMAL);
1166 #endif
1167 }
1168
1169 /*
1170 * If we are critically low on one of RAM or swap and low on
1171 * the other, kill the largest process. However, we avoid
1172 * doing this on the first pass in order to give ourselves a
1173 * chance to flush out dirty vnode-backed pages and to allow
1174 * active pages to be moved to the inactive queue and reclaimed.
1175 */
1176 if (pass != 0 &&
1177 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1178 (swap_pager_full && vm_paging_target() > 0)))
1179 vm_pageout_oom(VM_OOM_MEM);
1180 }
1181
1182
1183 void
1184 vm_pageout_oom(int shortage)
1185 {
1186 struct proc *p, *bigproc;
1187 vm_offset_t size, bigsize;
1188 struct thread *td;
1189 struct vmspace *vm;
1190
1191 /*
1192 * We keep the process bigproc locked once we find it to keep anyone
1193 * from messing with it; however, there is a possibility of
1194 * deadlock if process B is bigproc and one of it's child processes
1195 * attempts to propagate a signal to B while we are waiting for A's
1196 * lock while walking this list. To avoid this, we don't block on
1197 * the process lock but just skip a process if it is already locked.
1198 */
1199 bigproc = NULL;
1200 bigsize = 0;
1201 sx_slock(&allproc_lock);
1202 FOREACH_PROC_IN_SYSTEM(p) {
1203 int breakout;
1204
1205 if (PROC_TRYLOCK(p) == 0)
1206 continue;
1207 /*
1208 * If this is a system or protected process, skip it.
1209 */
1210 if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1211 (p->p_pid == 1) ||
1212 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1213 PROC_UNLOCK(p);
1214 continue;
1215 }
1216 /*
1217 * If the process is in a non-running type state,
1218 * don't touch it. Check all the threads individually.
1219 */
1220 PROC_SLOCK(p);
1221 breakout = 0;
1222 FOREACH_THREAD_IN_PROC(p, td) {
1223 thread_lock(td);
1224 if (!TD_ON_RUNQ(td) &&
1225 !TD_IS_RUNNING(td) &&
1226 !TD_IS_SLEEPING(td)) {
1227 thread_unlock(td);
1228 breakout = 1;
1229 break;
1230 }
1231 thread_unlock(td);
1232 }
1233 PROC_SUNLOCK(p);
1234 if (breakout) {
1235 PROC_UNLOCK(p);
1236 continue;
1237 }
1238 /*
1239 * get the process size
1240 */
1241 vm = vmspace_acquire_ref(p);
1242 if (vm == NULL) {
1243 PROC_UNLOCK(p);
1244 continue;
1245 }
1246 if (!vm_map_trylock_read(&vm->vm_map)) {
1247 vmspace_free(vm);
1248 PROC_UNLOCK(p);
1249 continue;
1250 }
1251 size = vmspace_swap_count(vm);
1252 vm_map_unlock_read(&vm->vm_map);
1253 if (shortage == VM_OOM_MEM)
1254 size += vmspace_resident_count(vm);
1255 vmspace_free(vm);
1256 /*
1257 * if the this process is bigger than the biggest one
1258 * remember it.
1259 */
1260 if (size > bigsize) {
1261 if (bigproc != NULL)
1262 PROC_UNLOCK(bigproc);
1263 bigproc = p;
1264 bigsize = size;
1265 } else
1266 PROC_UNLOCK(p);
1267 }
1268 sx_sunlock(&allproc_lock);
1269 if (bigproc != NULL) {
1270 killproc(bigproc, "out of swap space");
1271 PROC_SLOCK(bigproc);
1272 sched_nice(bigproc, PRIO_MIN);
1273 PROC_SUNLOCK(bigproc);
1274 PROC_UNLOCK(bigproc);
1275 wakeup(&cnt.v_free_count);
1276 }
1277 }
1278
1279 /*
1280 * This routine tries to maintain the pseudo LRU active queue,
1281 * so that during long periods of time where there is no paging,
1282 * that some statistic accumulation still occurs. This code
1283 * helps the situation where paging just starts to occur.
1284 */
1285 static void
1286 vm_pageout_page_stats()
1287 {
1288 vm_object_t object;
1289 vm_page_t m,next;
1290 int pcount,tpcount; /* Number of pages to check */
1291 static int fullintervalcount = 0;
1292 int page_shortage;
1293
1294 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1295 page_shortage =
1296 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1297 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1298
1299 if (page_shortage <= 0)
1300 return;
1301
1302 pcount = cnt.v_active_count;
1303 fullintervalcount += vm_pageout_stats_interval;
1304 if (fullintervalcount < vm_pageout_full_stats_interval) {
1305 tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1306 cnt.v_page_count;
1307 if (pcount > tpcount)
1308 pcount = tpcount;
1309 } else {
1310 fullintervalcount = 0;
1311 }
1312
1313 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1314 while ((m != NULL) && (pcount-- > 0)) {
1315 int actcount;
1316
1317 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1318 ("vm_pageout_page_stats: page %p isn't active", m));
1319
1320 next = TAILQ_NEXT(m, pageq);
1321 object = m->object;
1322
1323 if ((m->flags & PG_MARKER) != 0) {
1324 m = next;
1325 continue;
1326 }
1327 if (!VM_OBJECT_TRYLOCK(object) &&
1328 !vm_pageout_fallback_object_lock(m, &next)) {
1329 VM_OBJECT_UNLOCK(object);
1330 m = next;
1331 continue;
1332 }
1333
1334 /*
1335 * Don't deactivate pages that are busy.
1336 */
1337 if ((m->busy != 0) ||
1338 (m->oflags & VPO_BUSY) ||
1339 (m->hold_count != 0)) {
1340 VM_OBJECT_UNLOCK(object);
1341 vm_page_requeue(m);
1342 m = next;
1343 continue;
1344 }
1345
1346 actcount = 0;
1347 if (m->flags & PG_REFERENCED) {
1348 vm_page_flag_clear(m, PG_REFERENCED);
1349 actcount += 1;
1350 }
1351
1352 actcount += pmap_ts_referenced(m);
1353 if (actcount) {
1354 m->act_count += ACT_ADVANCE + actcount;
1355 if (m->act_count > ACT_MAX)
1356 m->act_count = ACT_MAX;
1357 vm_page_requeue(m);
1358 } else {
1359 if (m->act_count == 0) {
1360 /*
1361 * We turn off page access, so that we have
1362 * more accurate RSS stats. We don't do this
1363 * in the normal page deactivation when the
1364 * system is loaded VM wise, because the
1365 * cost of the large number of page protect
1366 * operations would be higher than the value
1367 * of doing the operation.
1368 */
1369 pmap_remove_all(m);
1370 vm_page_deactivate(m);
1371 } else {
1372 m->act_count -= min(m->act_count, ACT_DECLINE);
1373 vm_page_requeue(m);
1374 }
1375 }
1376 VM_OBJECT_UNLOCK(object);
1377 m = next;
1378 }
1379 }
1380
1381 /*
1382 * vm_pageout is the high level pageout daemon.
1383 */
1384 static void
1385 vm_pageout()
1386 {
1387 int error, pass;
1388
1389 /*
1390 * Initialize some paging parameters.
1391 */
1392 cnt.v_interrupt_free_min = 2;
1393 if (cnt.v_page_count < 2000)
1394 vm_pageout_page_count = 8;
1395
1396 /*
1397 * v_free_reserved needs to include enough for the largest
1398 * swap pager structures plus enough for any pv_entry structs
1399 * when paging.
1400 */
1401 if (cnt.v_page_count > 1024)
1402 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1403 else
1404 cnt.v_free_min = 4;
1405 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1406 cnt.v_interrupt_free_min;
1407 cnt.v_free_reserved = vm_pageout_page_count +
1408 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1409 cnt.v_free_severe = cnt.v_free_min / 2;
1410 cnt.v_free_min += cnt.v_free_reserved;
1411 cnt.v_free_severe += cnt.v_free_reserved;
1412
1413 /*
1414 * v_free_target and v_cache_min control pageout hysteresis. Note
1415 * that these are more a measure of the VM cache queue hysteresis
1416 * then the VM free queue. Specifically, v_free_target is the
1417 * high water mark (free+cache pages).
1418 *
1419 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1420 * low water mark, while v_free_min is the stop. v_cache_min must
1421 * be big enough to handle memory needs while the pageout daemon
1422 * is signalled and run to free more pages.
1423 */
1424 if (cnt.v_free_count > 6144)
1425 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1426 else
1427 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1428
1429 if (cnt.v_free_count > 2048) {
1430 cnt.v_cache_min = cnt.v_free_target;
1431 cnt.v_cache_max = 2 * cnt.v_cache_min;
1432 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1433 } else {
1434 cnt.v_cache_min = 0;
1435 cnt.v_cache_max = 0;
1436 cnt.v_inactive_target = cnt.v_free_count / 4;
1437 }
1438 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1439 cnt.v_inactive_target = cnt.v_free_count / 3;
1440
1441 /* XXX does not really belong here */
1442 if (vm_page_max_wired == 0)
1443 vm_page_max_wired = cnt.v_free_count / 3;
1444
1445 if (vm_pageout_stats_max == 0)
1446 vm_pageout_stats_max = cnt.v_free_target;
1447
1448 /*
1449 * Set interval in seconds for stats scan.
1450 */
1451 if (vm_pageout_stats_interval == 0)
1452 vm_pageout_stats_interval = 5;
1453 if (vm_pageout_full_stats_interval == 0)
1454 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1455
1456 swap_pager_swap_init();
1457 pass = 0;
1458 /*
1459 * The pageout daemon is never done, so loop forever.
1460 */
1461 while (TRUE) {
1462 /*
1463 * If we have enough free memory, wakeup waiters. Do
1464 * not clear vm_pages_needed until we reach our target,
1465 * otherwise we may be woken up over and over again and
1466 * waste a lot of cpu.
1467 */
1468 mtx_lock(&vm_page_queue_free_mtx);
1469 if (vm_pages_needed && !vm_page_count_min()) {
1470 if (!vm_paging_needed())
1471 vm_pages_needed = 0;
1472 wakeup(&cnt.v_free_count);
1473 }
1474 if (vm_pages_needed) {
1475 /*
1476 * Still not done, take a second pass without waiting
1477 * (unlimited dirty cleaning), otherwise sleep a bit
1478 * and try again.
1479 */
1480 ++pass;
1481 if (pass > 1)
1482 msleep(&vm_pages_needed,
1483 &vm_page_queue_free_mtx, PVM, "psleep",
1484 hz / 2);
1485 } else {
1486 /*
1487 * Good enough, sleep & handle stats. Prime the pass
1488 * for the next run.
1489 */
1490 if (pass > 1)
1491 pass = 1;
1492 else
1493 pass = 0;
1494 error = msleep(&vm_pages_needed,
1495 &vm_page_queue_free_mtx, PVM, "psleep",
1496 vm_pageout_stats_interval * hz);
1497 if (error && !vm_pages_needed) {
1498 mtx_unlock(&vm_page_queue_free_mtx);
1499 pass = 0;
1500 vm_page_lock_queues();
1501 vm_pageout_page_stats();
1502 vm_page_unlock_queues();
1503 continue;
1504 }
1505 }
1506 if (vm_pages_needed)
1507 cnt.v_pdwakeups++;
1508 mtx_unlock(&vm_page_queue_free_mtx);
1509 vm_pageout_scan(pass);
1510 }
1511 }
1512
1513 /*
1514 * Unless the free page queue lock is held by the caller, this function
1515 * should be regarded as advisory. Specifically, the caller should
1516 * not msleep() on &cnt.v_free_count following this function unless
1517 * the free page queue lock is held until the msleep() is performed.
1518 */
1519 void
1520 pagedaemon_wakeup()
1521 {
1522
1523 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1524 vm_pages_needed = 1;
1525 wakeup(&vm_pages_needed);
1526 }
1527 }
1528
1529 #if !defined(NO_SWAPPING)
1530 static void
1531 vm_req_vmdaemon(int req)
1532 {
1533 static int lastrun = 0;
1534
1535 mtx_lock(&vm_daemon_mtx);
1536 vm_pageout_req_swapout |= req;
1537 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1538 wakeup(&vm_daemon_needed);
1539 lastrun = ticks;
1540 }
1541 mtx_unlock(&vm_daemon_mtx);
1542 }
1543
1544 static void
1545 vm_daemon()
1546 {
1547 struct rlimit rsslim;
1548 struct proc *p;
1549 struct thread *td;
1550 struct vmspace *vm;
1551 int breakout, swapout_flags;
1552
1553 while (TRUE) {
1554 mtx_lock(&vm_daemon_mtx);
1555 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1556 swapout_flags = vm_pageout_req_swapout;
1557 vm_pageout_req_swapout = 0;
1558 mtx_unlock(&vm_daemon_mtx);
1559 if (swapout_flags)
1560 swapout_procs(swapout_flags);
1561
1562 /*
1563 * scan the processes for exceeding their rlimits or if
1564 * process is swapped out -- deactivate pages
1565 */
1566 sx_slock(&allproc_lock);
1567 FOREACH_PROC_IN_SYSTEM(p) {
1568 vm_pindex_t limit, size;
1569
1570 /*
1571 * if this is a system process or if we have already
1572 * looked at this process, skip it.
1573 */
1574 PROC_LOCK(p);
1575 if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1576 PROC_UNLOCK(p);
1577 continue;
1578 }
1579 /*
1580 * if the process is in a non-running type state,
1581 * don't touch it.
1582 */
1583 PROC_SLOCK(p);
1584 breakout = 0;
1585 FOREACH_THREAD_IN_PROC(p, td) {
1586 thread_lock(td);
1587 if (!TD_ON_RUNQ(td) &&
1588 !TD_IS_RUNNING(td) &&
1589 !TD_IS_SLEEPING(td)) {
1590 thread_unlock(td);
1591 breakout = 1;
1592 break;
1593 }
1594 thread_unlock(td);
1595 }
1596 PROC_SUNLOCK(p);
1597 if (breakout) {
1598 PROC_UNLOCK(p);
1599 continue;
1600 }
1601 /*
1602 * get a limit
1603 */
1604 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1605 limit = OFF_TO_IDX(
1606 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1607
1608 /*
1609 * let processes that are swapped out really be
1610 * swapped out set the limit to nothing (will force a
1611 * swap-out.)
1612 */
1613 if ((p->p_flag & P_INMEM) == 0)
1614 limit = 0; /* XXX */
1615 vm = vmspace_acquire_ref(p);
1616 PROC_UNLOCK(p);
1617 if (vm == NULL)
1618 continue;
1619
1620 size = vmspace_resident_count(vm);
1621 if (limit >= 0 && size >= limit) {
1622 vm_pageout_map_deactivate_pages(
1623 &vm->vm_map, limit);
1624 }
1625 vmspace_free(vm);
1626 }
1627 sx_sunlock(&allproc_lock);
1628 }
1629 }
1630 #endif /* !defined(NO_SWAPPING) */
Cache object: a91cdc0c63bd9340a685205e7b1f89bd
|