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/10.1/sys/vm/vm_pageout.c 272221 2014-09-27 18:20:45Z smh $");
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/racct.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sched.h>
92 #include <sys/signalvar.h>
93 #include <sys/smp.h>
94 #include <sys/vnode.h>
95 #include <sys/vmmeter.h>
96 #include <sys/rwlock.h>
97 #include <sys/sx.h>
98 #include <sys/sysctl.h>
99
100 #include <vm/vm.h>
101 #include <vm/vm_param.h>
102 #include <vm/vm_object.h>
103 #include <vm/vm_page.h>
104 #include <vm/vm_map.h>
105 #include <vm/vm_pageout.h>
106 #include <vm/vm_pager.h>
107 #include <vm/vm_phys.h>
108 #include <vm/swap_pager.h>
109 #include <vm/vm_extern.h>
110 #include <vm/uma.h>
111
112 /*
113 * System initialization
114 */
115
116 /* the kernel process "vm_pageout"*/
117 static void vm_pageout(void);
118 static int vm_pageout_clean(vm_page_t);
119 static void vm_pageout_scan(struct vm_domain *vmd, int pass);
120 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass);
121
122 struct proc *pageproc;
123
124 static struct kproc_desc page_kp = {
125 "pagedaemon",
126 vm_pageout,
127 &pageproc
128 };
129 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
130 &page_kp);
131
132 #if !defined(NO_SWAPPING)
133 /* the kernel process "vm_daemon"*/
134 static void vm_daemon(void);
135 static struct proc *vmproc;
136
137 static struct kproc_desc vm_kp = {
138 "vmdaemon",
139 vm_daemon,
140 &vmproc
141 };
142 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
143 #endif
144
145
146 int vm_pages_needed; /* Event on which pageout daemon sleeps */
147 int vm_pageout_deficit; /* Estimated number of pages deficit */
148 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
149 int vm_pageout_wakeup_thresh;
150
151 #if !defined(NO_SWAPPING)
152 static int vm_pageout_req_swapout; /* XXX */
153 static int vm_daemon_needed;
154 static struct mtx vm_daemon_mtx;
155 /* Allow for use by vm_pageout before vm_daemon is initialized. */
156 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
157 #endif
158 static int vm_max_launder = 32;
159 static int vm_pageout_update_period;
160 static int defer_swap_pageouts;
161 static int disable_swap_pageouts;
162 static int lowmem_period = 10;
163 static int lowmem_ticks;
164
165 #if defined(NO_SWAPPING)
166 static int vm_swap_enabled = 0;
167 static int vm_swap_idle_enabled = 0;
168 #else
169 static int vm_swap_enabled = 1;
170 static int vm_swap_idle_enabled = 0;
171 #endif
172
173 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
174 CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
175 "free page threshold for waking up the pageout daemon");
176
177 SYSCTL_INT(_vm, OID_AUTO, max_launder,
178 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
179
180 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
181 CTLFLAG_RW, &vm_pageout_update_period, 0,
182 "Maximum active LRU update period");
183
184 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
185 "Low memory callback period");
186
187 #if defined(NO_SWAPPING)
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
192 #else
193 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
194 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
195 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
196 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
197 #endif
198
199 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
200 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
201
202 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
203 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
204
205 static int pageout_lock_miss;
206 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
207 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
208
209 #define VM_PAGEOUT_PAGE_COUNT 16
210 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
211
212 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
213 SYSCTL_INT(_vm, OID_AUTO, max_wired,
214 CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
215
216 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
217 static boolean_t vm_pageout_launder(struct vm_pagequeue *pq, int, vm_paddr_t,
218 vm_paddr_t);
219 #if !defined(NO_SWAPPING)
220 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
221 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
222 static void vm_req_vmdaemon(int req);
223 #endif
224 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
225
226 /*
227 * Initialize a dummy page for marking the caller's place in the specified
228 * paging queue. In principle, this function only needs to set the flag
229 * PG_MARKER. Nonetheless, it wirte busies and initializes the hold count
230 * to one as safety precautions.
231 */
232 static void
233 vm_pageout_init_marker(vm_page_t marker, u_short queue)
234 {
235
236 bzero(marker, sizeof(*marker));
237 marker->flags = PG_MARKER;
238 marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
239 marker->queue = queue;
240 marker->hold_count = 1;
241 }
242
243 /*
244 * vm_pageout_fallback_object_lock:
245 *
246 * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
247 * known to have failed and page queue must be either PQ_ACTIVE or
248 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues
249 * while locking the vm object. Use marker page to detect page queue
250 * changes and maintain notion of next page on page queue. Return
251 * TRUE if no changes were detected, FALSE otherwise. vm object is
252 * locked on return.
253 *
254 * This function depends on both the lock portion of struct vm_object
255 * and normal struct vm_page being type stable.
256 */
257 static boolean_t
258 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
259 {
260 struct vm_page marker;
261 struct vm_pagequeue *pq;
262 boolean_t unchanged;
263 u_short queue;
264 vm_object_t object;
265
266 queue = m->queue;
267 vm_pageout_init_marker(&marker, queue);
268 pq = vm_page_pagequeue(m);
269 object = m->object;
270
271 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
272 vm_pagequeue_unlock(pq);
273 vm_page_unlock(m);
274 VM_OBJECT_WLOCK(object);
275 vm_page_lock(m);
276 vm_pagequeue_lock(pq);
277
278 /* Page queue might have changed. */
279 *next = TAILQ_NEXT(&marker, plinks.q);
280 unchanged = (m->queue == queue &&
281 m->object == object &&
282 &marker == TAILQ_NEXT(m, plinks.q));
283 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
284 return (unchanged);
285 }
286
287 /*
288 * Lock the page while holding the page queue lock. Use marker page
289 * to detect page queue changes and maintain notion of next page on
290 * page queue. Return TRUE if no changes were detected, FALSE
291 * otherwise. The page is locked on return. The page queue lock might
292 * be dropped and reacquired.
293 *
294 * This function depends on normal struct vm_page being type stable.
295 */
296 static boolean_t
297 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
298 {
299 struct vm_page marker;
300 struct vm_pagequeue *pq;
301 boolean_t unchanged;
302 u_short queue;
303
304 vm_page_lock_assert(m, MA_NOTOWNED);
305 if (vm_page_trylock(m))
306 return (TRUE);
307
308 queue = m->queue;
309 vm_pageout_init_marker(&marker, queue);
310 pq = vm_page_pagequeue(m);
311
312 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
313 vm_pagequeue_unlock(pq);
314 vm_page_lock(m);
315 vm_pagequeue_lock(pq);
316
317 /* Page queue might have changed. */
318 *next = TAILQ_NEXT(&marker, plinks.q);
319 unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, plinks.q));
320 TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
321 return (unchanged);
322 }
323
324 /*
325 * vm_pageout_clean:
326 *
327 * Clean the page and remove it from the laundry.
328 *
329 * We set the busy bit to cause potential page faults on this page to
330 * block. Note the careful timing, however, the busy bit isn't set till
331 * late and we cannot do anything that will mess with the page.
332 */
333 static int
334 vm_pageout_clean(vm_page_t m)
335 {
336 vm_object_t object;
337 vm_page_t mc[2*vm_pageout_page_count], pb, ps;
338 int pageout_count;
339 int ib, is, page_base;
340 vm_pindex_t pindex = m->pindex;
341
342 vm_page_lock_assert(m, MA_OWNED);
343 object = m->object;
344 VM_OBJECT_ASSERT_WLOCKED(object);
345
346 /*
347 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
348 * with the new swapper, but we could have serious problems paging
349 * out other object types if there is insufficient memory.
350 *
351 * Unfortunately, checking free memory here is far too late, so the
352 * check has been moved up a procedural level.
353 */
354
355 /*
356 * Can't clean the page if it's busy or held.
357 */
358 vm_page_assert_unbusied(m);
359 KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
360 vm_page_unlock(m);
361
362 mc[vm_pageout_page_count] = pb = ps = m;
363 pageout_count = 1;
364 page_base = vm_pageout_page_count;
365 ib = 1;
366 is = 1;
367
368 /*
369 * Scan object for clusterable pages.
370 *
371 * We can cluster ONLY if: ->> the page is NOT
372 * clean, wired, busy, held, or mapped into a
373 * buffer, and one of the following:
374 * 1) The page is inactive, or a seldom used
375 * active page.
376 * -or-
377 * 2) we force the issue.
378 *
379 * During heavy mmap/modification loads the pageout
380 * daemon can really fragment the underlying file
381 * due to flushing pages out of order and not trying
382 * align the clusters (which leave sporatic out-of-order
383 * holes). To solve this problem we do the reverse scan
384 * first and attempt to align our cluster, then do a
385 * forward scan if room remains.
386 */
387 more:
388 while (ib && pageout_count < vm_pageout_page_count) {
389 vm_page_t p;
390
391 if (ib > pindex) {
392 ib = 0;
393 break;
394 }
395
396 if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
397 ib = 0;
398 break;
399 }
400 vm_page_lock(p);
401 vm_page_test_dirty(p);
402 if (p->dirty == 0 ||
403 p->queue != PQ_INACTIVE ||
404 p->hold_count != 0) { /* may be undergoing I/O */
405 vm_page_unlock(p);
406 ib = 0;
407 break;
408 }
409 vm_page_unlock(p);
410 mc[--page_base] = pb = p;
411 ++pageout_count;
412 ++ib;
413 /*
414 * alignment boundry, stop here and switch directions. Do
415 * not clear ib.
416 */
417 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
418 break;
419 }
420
421 while (pageout_count < vm_pageout_page_count &&
422 pindex + is < object->size) {
423 vm_page_t p;
424
425 if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
426 break;
427 vm_page_lock(p);
428 vm_page_test_dirty(p);
429 if (p->dirty == 0 ||
430 p->queue != PQ_INACTIVE ||
431 p->hold_count != 0) { /* may be undergoing I/O */
432 vm_page_unlock(p);
433 break;
434 }
435 vm_page_unlock(p);
436 mc[page_base + pageout_count] = ps = p;
437 ++pageout_count;
438 ++is;
439 }
440
441 /*
442 * If we exhausted our forward scan, continue with the reverse scan
443 * when possible, even past a page boundry. This catches boundry
444 * conditions.
445 */
446 if (ib && pageout_count < vm_pageout_page_count)
447 goto more;
448
449 /*
450 * we allow reads during pageouts...
451 */
452 return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
453 NULL));
454 }
455
456 /*
457 * vm_pageout_flush() - launder the given pages
458 *
459 * The given pages are laundered. Note that we setup for the start of
460 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
461 * reference count all in here rather then in the parent. If we want
462 * the parent to do more sophisticated things we may have to change
463 * the ordering.
464 *
465 * Returned runlen is the count of pages between mreq and first
466 * page after mreq with status VM_PAGER_AGAIN.
467 * *eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
468 * for any page in runlen set.
469 */
470 int
471 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
472 boolean_t *eio)
473 {
474 vm_object_t object = mc[0]->object;
475 int pageout_status[count];
476 int numpagedout = 0;
477 int i, runlen;
478
479 VM_OBJECT_ASSERT_WLOCKED(object);
480
481 /*
482 * Initiate I/O. Bump the vm_page_t->busy counter and
483 * mark the pages read-only.
484 *
485 * We do not have to fixup the clean/dirty bits here... we can
486 * allow the pager to do it after the I/O completes.
487 *
488 * NOTE! mc[i]->dirty may be partial or fragmented due to an
489 * edge case with file fragments.
490 */
491 for (i = 0; i < count; i++) {
492 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
493 ("vm_pageout_flush: partially invalid page %p index %d/%d",
494 mc[i], i, count));
495 vm_page_sbusy(mc[i]);
496 pmap_remove_write(mc[i]);
497 }
498 vm_object_pip_add(object, count);
499
500 vm_pager_put_pages(object, mc, count, flags, pageout_status);
501
502 runlen = count - mreq;
503 if (eio != NULL)
504 *eio = FALSE;
505 for (i = 0; i < count; i++) {
506 vm_page_t mt = mc[i];
507
508 KASSERT(pageout_status[i] == VM_PAGER_PEND ||
509 !pmap_page_is_write_mapped(mt),
510 ("vm_pageout_flush: page %p is not write protected", mt));
511 switch (pageout_status[i]) {
512 case VM_PAGER_OK:
513 case VM_PAGER_PEND:
514 numpagedout++;
515 break;
516 case VM_PAGER_BAD:
517 /*
518 * Page outside of range of object. Right now we
519 * essentially lose the changes by pretending it
520 * worked.
521 */
522 vm_page_undirty(mt);
523 break;
524 case VM_PAGER_ERROR:
525 case VM_PAGER_FAIL:
526 /*
527 * If page couldn't be paged out, then reactivate the
528 * page so it doesn't clog the inactive list. (We
529 * will try paging out it again later).
530 */
531 vm_page_lock(mt);
532 vm_page_activate(mt);
533 vm_page_unlock(mt);
534 if (eio != NULL && i >= mreq && i - mreq < runlen)
535 *eio = TRUE;
536 break;
537 case VM_PAGER_AGAIN:
538 if (i >= mreq && i - mreq < runlen)
539 runlen = i - mreq;
540 break;
541 }
542
543 /*
544 * If the operation is still going, leave the page busy to
545 * block all other accesses. Also, leave the paging in
546 * progress indicator set so that we don't attempt an object
547 * collapse.
548 */
549 if (pageout_status[i] != VM_PAGER_PEND) {
550 vm_object_pip_wakeup(object);
551 vm_page_sunbusy(mt);
552 if (vm_page_count_severe()) {
553 vm_page_lock(mt);
554 vm_page_try_to_cache(mt);
555 vm_page_unlock(mt);
556 }
557 }
558 }
559 if (prunlen != NULL)
560 *prunlen = runlen;
561 return (numpagedout);
562 }
563
564 static boolean_t
565 vm_pageout_launder(struct vm_pagequeue *pq, int tries, vm_paddr_t low,
566 vm_paddr_t high)
567 {
568 struct mount *mp;
569 struct vnode *vp;
570 vm_object_t object;
571 vm_paddr_t pa;
572 vm_page_t m, m_tmp, next;
573 int lockmode;
574
575 vm_pagequeue_lock(pq);
576 TAILQ_FOREACH_SAFE(m, &pq->pq_pl, plinks.q, next) {
577 if ((m->flags & PG_MARKER) != 0)
578 continue;
579 pa = VM_PAGE_TO_PHYS(m);
580 if (pa < low || pa + PAGE_SIZE > high)
581 continue;
582 if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
583 vm_page_unlock(m);
584 continue;
585 }
586 object = m->object;
587 if ((!VM_OBJECT_TRYWLOCK(object) &&
588 (!vm_pageout_fallback_object_lock(m, &next) ||
589 m->hold_count != 0)) || vm_page_busied(m)) {
590 vm_page_unlock(m);
591 VM_OBJECT_WUNLOCK(object);
592 continue;
593 }
594 vm_page_test_dirty(m);
595 if (m->dirty == 0 && object->ref_count != 0)
596 pmap_remove_all(m);
597 if (m->dirty != 0) {
598 vm_page_unlock(m);
599 if (tries == 0 || (object->flags & OBJ_DEAD) != 0) {
600 VM_OBJECT_WUNLOCK(object);
601 continue;
602 }
603 if (object->type == OBJT_VNODE) {
604 vm_pagequeue_unlock(pq);
605 vp = object->handle;
606 vm_object_reference_locked(object);
607 VM_OBJECT_WUNLOCK(object);
608 (void)vn_start_write(vp, &mp, V_WAIT);
609 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
610 LK_SHARED : LK_EXCLUSIVE;
611 vn_lock(vp, lockmode | LK_RETRY);
612 VM_OBJECT_WLOCK(object);
613 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
614 VM_OBJECT_WUNLOCK(object);
615 VOP_UNLOCK(vp, 0);
616 vm_object_deallocate(object);
617 vn_finished_write(mp);
618 return (TRUE);
619 } else if (object->type == OBJT_SWAP ||
620 object->type == OBJT_DEFAULT) {
621 vm_pagequeue_unlock(pq);
622 m_tmp = m;
623 vm_pageout_flush(&m_tmp, 1, VM_PAGER_PUT_SYNC,
624 0, NULL, NULL);
625 VM_OBJECT_WUNLOCK(object);
626 return (TRUE);
627 }
628 } else {
629 /*
630 * Dequeue here to prevent lock recursion in
631 * vm_page_cache().
632 */
633 vm_page_dequeue_locked(m);
634 vm_page_cache(m);
635 vm_page_unlock(m);
636 }
637 VM_OBJECT_WUNLOCK(object);
638 }
639 vm_pagequeue_unlock(pq);
640 return (FALSE);
641 }
642
643 /*
644 * Increase the number of cached pages. The specified value, "tries",
645 * determines which categories of pages are cached:
646 *
647 * 0: All clean, inactive pages within the specified physical address range
648 * are cached. Will not sleep.
649 * 1: The vm_lowmem handlers are called. All inactive pages within
650 * the specified physical address range are cached. May sleep.
651 * 2: The vm_lowmem handlers are called. All inactive and active pages
652 * within the specified physical address range are cached. May sleep.
653 */
654 void
655 vm_pageout_grow_cache(int tries, vm_paddr_t low, vm_paddr_t high)
656 {
657 int actl, actmax, inactl, inactmax, dom, initial_dom;
658 static int start_dom = 0;
659
660 if (tries > 0) {
661 /*
662 * Decrease registered cache sizes. The vm_lowmem handlers
663 * may acquire locks and/or sleep, so they can only be invoked
664 * when "tries" is greater than zero.
665 */
666 EVENTHANDLER_INVOKE(vm_lowmem, 0);
667
668 /*
669 * We do this explicitly after the caches have been drained
670 * above.
671 */
672 uma_reclaim();
673 }
674
675 /*
676 * Make the next scan start on the next domain.
677 */
678 initial_dom = atomic_fetchadd_int(&start_dom, 1) % vm_ndomains;
679
680 inactl = 0;
681 inactmax = cnt.v_inactive_count;
682 actl = 0;
683 actmax = tries < 2 ? 0 : cnt.v_active_count;
684 dom = initial_dom;
685
686 /*
687 * Scan domains in round-robin order, first inactive queues,
688 * then active. Since domain usually owns large physically
689 * contiguous chunk of memory, it makes sense to completely
690 * exhaust one domain before switching to next, while growing
691 * the pool of contiguous physical pages.
692 *
693 * Do not even start launder a domain which cannot contain
694 * the specified address range, as indicated by segments
695 * constituting the domain.
696 */
697 again:
698 if (inactl < inactmax) {
699 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
700 low, high) &&
701 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_INACTIVE],
702 tries, low, high)) {
703 inactl++;
704 goto again;
705 }
706 if (++dom == vm_ndomains)
707 dom = 0;
708 if (dom != initial_dom)
709 goto again;
710 }
711 if (actl < actmax) {
712 if (vm_phys_domain_intersects(vm_dom[dom].vmd_segs,
713 low, high) &&
714 vm_pageout_launder(&vm_dom[dom].vmd_pagequeues[PQ_ACTIVE],
715 tries, low, high)) {
716 actl++;
717 goto again;
718 }
719 if (++dom == vm_ndomains)
720 dom = 0;
721 if (dom != initial_dom)
722 goto again;
723 }
724 }
725
726 #if !defined(NO_SWAPPING)
727 /*
728 * vm_pageout_object_deactivate_pages
729 *
730 * Deactivate enough pages to satisfy the inactive target
731 * requirements.
732 *
733 * The object and map must be locked.
734 */
735 static void
736 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
737 long desired)
738 {
739 vm_object_t backing_object, object;
740 vm_page_t p;
741 int act_delta, remove_mode;
742
743 VM_OBJECT_ASSERT_LOCKED(first_object);
744 if ((first_object->flags & OBJ_FICTITIOUS) != 0)
745 return;
746 for (object = first_object;; object = backing_object) {
747 if (pmap_resident_count(pmap) <= desired)
748 goto unlock_return;
749 VM_OBJECT_ASSERT_LOCKED(object);
750 if ((object->flags & OBJ_UNMANAGED) != 0 ||
751 object->paging_in_progress != 0)
752 goto unlock_return;
753
754 remove_mode = 0;
755 if (object->shadow_count > 1)
756 remove_mode = 1;
757 /*
758 * Scan the object's entire memory queue.
759 */
760 TAILQ_FOREACH(p, &object->memq, listq) {
761 if (pmap_resident_count(pmap) <= desired)
762 goto unlock_return;
763 if (vm_page_busied(p))
764 continue;
765 PCPU_INC(cnt.v_pdpages);
766 vm_page_lock(p);
767 if (p->wire_count != 0 || p->hold_count != 0 ||
768 !pmap_page_exists_quick(pmap, p)) {
769 vm_page_unlock(p);
770 continue;
771 }
772 act_delta = pmap_ts_referenced(p);
773 if ((p->aflags & PGA_REFERENCED) != 0) {
774 if (act_delta == 0)
775 act_delta = 1;
776 vm_page_aflag_clear(p, PGA_REFERENCED);
777 }
778 if (p->queue != PQ_ACTIVE && act_delta != 0) {
779 vm_page_activate(p);
780 p->act_count += act_delta;
781 } else if (p->queue == PQ_ACTIVE) {
782 if (act_delta == 0) {
783 p->act_count -= min(p->act_count,
784 ACT_DECLINE);
785 if (!remove_mode && p->act_count == 0) {
786 pmap_remove_all(p);
787 vm_page_deactivate(p);
788 } else
789 vm_page_requeue(p);
790 } else {
791 vm_page_activate(p);
792 if (p->act_count < ACT_MAX -
793 ACT_ADVANCE)
794 p->act_count += ACT_ADVANCE;
795 vm_page_requeue(p);
796 }
797 } else if (p->queue == PQ_INACTIVE)
798 pmap_remove_all(p);
799 vm_page_unlock(p);
800 }
801 if ((backing_object = object->backing_object) == NULL)
802 goto unlock_return;
803 VM_OBJECT_RLOCK(backing_object);
804 if (object != first_object)
805 VM_OBJECT_RUNLOCK(object);
806 }
807 unlock_return:
808 if (object != first_object)
809 VM_OBJECT_RUNLOCK(object);
810 }
811
812 /*
813 * deactivate some number of pages in a map, try to do it fairly, but
814 * that is really hard to do.
815 */
816 static void
817 vm_pageout_map_deactivate_pages(map, desired)
818 vm_map_t map;
819 long desired;
820 {
821 vm_map_entry_t tmpe;
822 vm_object_t obj, bigobj;
823 int nothingwired;
824
825 if (!vm_map_trylock(map))
826 return;
827
828 bigobj = NULL;
829 nothingwired = TRUE;
830
831 /*
832 * first, search out the biggest object, and try to free pages from
833 * that.
834 */
835 tmpe = map->header.next;
836 while (tmpe != &map->header) {
837 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
838 obj = tmpe->object.vm_object;
839 if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
840 if (obj->shadow_count <= 1 &&
841 (bigobj == NULL ||
842 bigobj->resident_page_count < obj->resident_page_count)) {
843 if (bigobj != NULL)
844 VM_OBJECT_RUNLOCK(bigobj);
845 bigobj = obj;
846 } else
847 VM_OBJECT_RUNLOCK(obj);
848 }
849 }
850 if (tmpe->wired_count > 0)
851 nothingwired = FALSE;
852 tmpe = tmpe->next;
853 }
854
855 if (bigobj != NULL) {
856 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
857 VM_OBJECT_RUNLOCK(bigobj);
858 }
859 /*
860 * Next, hunt around for other pages to deactivate. We actually
861 * do this search sort of wrong -- .text first is not the best idea.
862 */
863 tmpe = map->header.next;
864 while (tmpe != &map->header) {
865 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
866 break;
867 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
868 obj = tmpe->object.vm_object;
869 if (obj != NULL) {
870 VM_OBJECT_RLOCK(obj);
871 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
872 VM_OBJECT_RUNLOCK(obj);
873 }
874 }
875 tmpe = tmpe->next;
876 }
877
878 #ifdef __ia64__
879 /*
880 * Remove all non-wired, managed mappings if a process is swapped out.
881 * This will free page table pages.
882 */
883 if (desired == 0)
884 pmap_remove_pages(map->pmap);
885 #else
886 /*
887 * Remove all mappings if a process is swapped out, this will free page
888 * table pages.
889 */
890 if (desired == 0 && nothingwired) {
891 pmap_remove(vm_map_pmap(map), vm_map_min(map),
892 vm_map_max(map));
893 }
894 #endif
895
896 vm_map_unlock(map);
897 }
898 #endif /* !defined(NO_SWAPPING) */
899
900 /*
901 * vm_pageout_scan does the dirty work for the pageout daemon.
902 *
903 * pass 0 - Update active LRU/deactivate pages
904 * pass 1 - Move inactive to cache or free
905 * pass 2 - Launder dirty pages
906 */
907 static void
908 vm_pageout_scan(struct vm_domain *vmd, int pass)
909 {
910 vm_page_t m, next;
911 struct vm_pagequeue *pq;
912 vm_object_t object;
913 int act_delta, addl_page_shortage, deficit, maxscan, page_shortage;
914 int vnodes_skipped = 0;
915 int maxlaunder;
916 int lockmode;
917 boolean_t queues_locked;
918
919 /*
920 * If we need to reclaim memory ask kernel caches to return
921 * some. We rate limit to avoid thrashing.
922 */
923 if (vmd == &vm_dom[0] && pass > 0 &&
924 (ticks - lowmem_ticks) / hz >= lowmem_period) {
925 /*
926 * Decrease registered cache sizes.
927 */
928 EVENTHANDLER_INVOKE(vm_lowmem, 0);
929 /*
930 * We do this explicitly after the caches have been
931 * drained above.
932 */
933 uma_reclaim();
934 lowmem_ticks = ticks;
935 }
936
937 /*
938 * The addl_page_shortage is the number of temporarily
939 * stuck pages in the inactive queue. In other words, the
940 * number of pages from the inactive count that should be
941 * discounted in setting the target for the active queue scan.
942 */
943 addl_page_shortage = 0;
944
945 /*
946 * Calculate the number of pages we want to either free or move
947 * to the cache.
948 */
949 if (pass > 0) {
950 deficit = atomic_readandclear_int(&vm_pageout_deficit);
951 page_shortage = vm_paging_target() + deficit;
952 } else
953 page_shortage = deficit = 0;
954
955 /*
956 * maxlaunder limits the number of dirty pages we flush per scan.
957 * For most systems a smaller value (16 or 32) is more robust under
958 * extreme memory and disk pressure because any unnecessary writes
959 * to disk can result in extreme performance degredation. However,
960 * systems with excessive dirty pages (especially when MAP_NOSYNC is
961 * used) will die horribly with limited laundering. If the pageout
962 * daemon cannot clean enough pages in the first pass, we let it go
963 * all out in succeeding passes.
964 */
965 if ((maxlaunder = vm_max_launder) <= 1)
966 maxlaunder = 1;
967 if (pass > 1)
968 maxlaunder = 10000;
969
970 /*
971 * Start scanning the inactive queue for pages we can move to the
972 * cache or free. The scan will stop when the target is reached or
973 * we have scanned the entire inactive queue. Note that m->act_count
974 * is not used to form decisions for the inactive queue, only for the
975 * active queue.
976 */
977 pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
978 maxscan = pq->pq_cnt;
979 vm_pagequeue_lock(pq);
980 queues_locked = TRUE;
981 for (m = TAILQ_FIRST(&pq->pq_pl);
982 m != NULL && maxscan-- > 0 && page_shortage > 0;
983 m = next) {
984 vm_pagequeue_assert_locked(pq);
985 KASSERT(queues_locked, ("unlocked queues"));
986 KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
987
988 PCPU_INC(cnt.v_pdpages);
989 next = TAILQ_NEXT(m, plinks.q);
990
991 /*
992 * skip marker pages
993 */
994 if (m->flags & PG_MARKER)
995 continue;
996
997 KASSERT((m->flags & PG_FICTITIOUS) == 0,
998 ("Fictitious page %p cannot be in inactive queue", m));
999 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1000 ("Unmanaged page %p cannot be in inactive queue", m));
1001
1002 /*
1003 * The page or object lock acquisitions fail if the
1004 * page was removed from the queue or moved to a
1005 * different position within the queue. In either
1006 * case, addl_page_shortage should not be incremented.
1007 */
1008 if (!vm_pageout_page_lock(m, &next)) {
1009 vm_page_unlock(m);
1010 continue;
1011 }
1012 object = m->object;
1013 if (!VM_OBJECT_TRYWLOCK(object) &&
1014 !vm_pageout_fallback_object_lock(m, &next)) {
1015 vm_page_unlock(m);
1016 VM_OBJECT_WUNLOCK(object);
1017 continue;
1018 }
1019
1020 /*
1021 * Don't mess with busy pages, keep them at at the
1022 * front of the queue, most likely they are being
1023 * paged out. Increment addl_page_shortage for busy
1024 * pages, because they may leave the inactive queue
1025 * shortly after page scan is finished.
1026 */
1027 if (vm_page_busied(m)) {
1028 vm_page_unlock(m);
1029 VM_OBJECT_WUNLOCK(object);
1030 addl_page_shortage++;
1031 continue;
1032 }
1033
1034 /*
1035 * We unlock the inactive page queue, invalidating the
1036 * 'next' pointer. Use our marker to remember our
1037 * place.
1038 */
1039 TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1040 vm_pagequeue_unlock(pq);
1041 queues_locked = FALSE;
1042
1043 /*
1044 * We bump the activation count if the page has been
1045 * referenced while in the inactive queue. This makes
1046 * it less likely that the page will be added back to the
1047 * inactive queue prematurely again. Here we check the
1048 * page tables (or emulated bits, if any), given the upper
1049 * level VM system not knowing anything about existing
1050 * references.
1051 */
1052 act_delta = 0;
1053 if ((m->aflags & PGA_REFERENCED) != 0) {
1054 vm_page_aflag_clear(m, PGA_REFERENCED);
1055 act_delta = 1;
1056 }
1057 if (object->ref_count != 0) {
1058 act_delta += pmap_ts_referenced(m);
1059 } else {
1060 KASSERT(!pmap_page_is_mapped(m),
1061 ("vm_pageout_scan: page %p is mapped", m));
1062 }
1063
1064 /*
1065 * If the upper level VM system knows about any page
1066 * references, we reactivate the page or requeue it.
1067 */
1068 if (act_delta != 0) {
1069 if (object->ref_count) {
1070 vm_page_activate(m);
1071 m->act_count += act_delta + ACT_ADVANCE;
1072 } else {
1073 vm_pagequeue_lock(pq);
1074 queues_locked = TRUE;
1075 vm_page_requeue_locked(m);
1076 }
1077 VM_OBJECT_WUNLOCK(object);
1078 vm_page_unlock(m);
1079 goto relock_queues;
1080 }
1081
1082 if (m->hold_count != 0) {
1083 vm_page_unlock(m);
1084 VM_OBJECT_WUNLOCK(object);
1085
1086 /*
1087 * Held pages are essentially stuck in the
1088 * queue. So, they ought to be discounted
1089 * from the inactive count. See the
1090 * calculation of the page_shortage for the
1091 * loop over the active queue below.
1092 */
1093 addl_page_shortage++;
1094 goto relock_queues;
1095 }
1096
1097 /*
1098 * If the page appears to be clean at the machine-independent
1099 * layer, then remove all of its mappings from the pmap in
1100 * anticipation of placing it onto the cache queue. If,
1101 * however, any of the page's mappings allow write access,
1102 * then the page may still be modified until the last of those
1103 * mappings are removed.
1104 */
1105 vm_page_test_dirty(m);
1106 if (m->dirty == 0 && object->ref_count != 0)
1107 pmap_remove_all(m);
1108
1109 if (m->valid == 0) {
1110 /*
1111 * Invalid pages can be easily freed
1112 */
1113 vm_page_free(m);
1114 PCPU_INC(cnt.v_dfree);
1115 --page_shortage;
1116 } else if (m->dirty == 0) {
1117 /*
1118 * Clean pages can be placed onto the cache queue.
1119 * This effectively frees them.
1120 */
1121 vm_page_cache(m);
1122 --page_shortage;
1123 } else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1124 /*
1125 * Dirty pages need to be paged out, but flushing
1126 * a page is extremely expensive verses freeing
1127 * a clean page. Rather then artificially limiting
1128 * the number of pages we can flush, we instead give
1129 * dirty pages extra priority on the inactive queue
1130 * by forcing them to be cycled through the queue
1131 * twice before being flushed, after which the
1132 * (now clean) page will cycle through once more
1133 * before being freed. This significantly extends
1134 * the thrash point for a heavily loaded machine.
1135 */
1136 m->flags |= PG_WINATCFLS;
1137 vm_pagequeue_lock(pq);
1138 queues_locked = TRUE;
1139 vm_page_requeue_locked(m);
1140 } else if (maxlaunder > 0) {
1141 /*
1142 * We always want to try to flush some dirty pages if
1143 * we encounter them, to keep the system stable.
1144 * Normally this number is small, but under extreme
1145 * pressure where there are insufficient clean pages
1146 * on the inactive queue, we may have to go all out.
1147 */
1148 int swap_pageouts_ok;
1149 struct vnode *vp = NULL;
1150 struct mount *mp = NULL;
1151
1152 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
1153 swap_pageouts_ok = 1;
1154 } else {
1155 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
1156 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
1157 vm_page_count_min());
1158
1159 }
1160
1161 /*
1162 * We don't bother paging objects that are "dead".
1163 * Those objects are in a "rundown" state.
1164 */
1165 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
1166 vm_pagequeue_lock(pq);
1167 vm_page_unlock(m);
1168 VM_OBJECT_WUNLOCK(object);
1169 queues_locked = TRUE;
1170 vm_page_requeue_locked(m);
1171 goto relock_queues;
1172 }
1173
1174 /*
1175 * The object is already known NOT to be dead. It
1176 * is possible for the vget() to block the whole
1177 * pageout daemon, but the new low-memory handling
1178 * code should prevent it.
1179 *
1180 * The previous code skipped locked vnodes and, worse,
1181 * reordered pages in the queue. This results in
1182 * completely non-deterministic operation and, on a
1183 * busy system, can lead to extremely non-optimal
1184 * pageouts. For example, it can cause clean pages
1185 * to be freed and dirty pages to be moved to the end
1186 * of the queue. Since dirty pages are also moved to
1187 * the end of the queue once-cleaned, this gives
1188 * way too large a weighting to defering the freeing
1189 * of dirty pages.
1190 *
1191 * We can't wait forever for the vnode lock, we might
1192 * deadlock due to a vn_read() getting stuck in
1193 * vm_wait while holding this vnode. We skip the
1194 * vnode if we can't get it in a reasonable amount
1195 * of time.
1196 */
1197 if (object->type == OBJT_VNODE) {
1198 vm_page_unlock(m);
1199 vp = object->handle;
1200 if (vp->v_type == VREG &&
1201 vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1202 mp = NULL;
1203 ++pageout_lock_miss;
1204 if (object->flags & OBJ_MIGHTBEDIRTY)
1205 vnodes_skipped++;
1206 goto unlock_and_continue;
1207 }
1208 KASSERT(mp != NULL,
1209 ("vp %p with NULL v_mount", vp));
1210 vm_object_reference_locked(object);
1211 VM_OBJECT_WUNLOCK(object);
1212 lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
1213 LK_SHARED : LK_EXCLUSIVE;
1214 if (vget(vp, lockmode | LK_TIMELOCK,
1215 curthread)) {
1216 VM_OBJECT_WLOCK(object);
1217 ++pageout_lock_miss;
1218 if (object->flags & OBJ_MIGHTBEDIRTY)
1219 vnodes_skipped++;
1220 vp = NULL;
1221 goto unlock_and_continue;
1222 }
1223 VM_OBJECT_WLOCK(object);
1224 vm_page_lock(m);
1225 vm_pagequeue_lock(pq);
1226 queues_locked = TRUE;
1227 /*
1228 * The page might have been moved to another
1229 * queue during potential blocking in vget()
1230 * above. The page might have been freed and
1231 * reused for another vnode.
1232 */
1233 if (m->queue != PQ_INACTIVE ||
1234 m->object != object ||
1235 TAILQ_NEXT(m, plinks.q) != &vmd->vmd_marker) {
1236 vm_page_unlock(m);
1237 if (object->flags & OBJ_MIGHTBEDIRTY)
1238 vnodes_skipped++;
1239 goto unlock_and_continue;
1240 }
1241
1242 /*
1243 * The page may have been busied during the
1244 * blocking in vget(). We don't move the
1245 * page back onto the end of the queue so that
1246 * statistics are more correct if we don't.
1247 */
1248 if (vm_page_busied(m)) {
1249 vm_page_unlock(m);
1250 addl_page_shortage++;
1251 goto unlock_and_continue;
1252 }
1253
1254 /*
1255 * If the page has become held it might
1256 * be undergoing I/O, so skip it
1257 */
1258 if (m->hold_count != 0) {
1259 vm_page_unlock(m);
1260 addl_page_shortage++;
1261 if (object->flags & OBJ_MIGHTBEDIRTY)
1262 vnodes_skipped++;
1263 goto unlock_and_continue;
1264 }
1265 vm_pagequeue_unlock(pq);
1266 queues_locked = FALSE;
1267 }
1268
1269 /*
1270 * If a page is dirty, then it is either being washed
1271 * (but not yet cleaned) or it is still in the
1272 * laundry. If it is still in the laundry, then we
1273 * start the cleaning operation.
1274 *
1275 * decrement page_shortage on success to account for
1276 * the (future) cleaned page. Otherwise we could wind
1277 * up laundering or cleaning too many pages.
1278 */
1279 if (vm_pageout_clean(m) != 0) {
1280 --page_shortage;
1281 --maxlaunder;
1282 }
1283 unlock_and_continue:
1284 vm_page_lock_assert(m, MA_NOTOWNED);
1285 VM_OBJECT_WUNLOCK(object);
1286 if (mp != NULL) {
1287 if (queues_locked) {
1288 vm_pagequeue_unlock(pq);
1289 queues_locked = FALSE;
1290 }
1291 if (vp != NULL)
1292 vput(vp);
1293 vm_object_deallocate(object);
1294 vn_finished_write(mp);
1295 }
1296 vm_page_lock_assert(m, MA_NOTOWNED);
1297 goto relock_queues;
1298 }
1299 vm_page_unlock(m);
1300 VM_OBJECT_WUNLOCK(object);
1301 relock_queues:
1302 if (!queues_locked) {
1303 vm_pagequeue_lock(pq);
1304 queues_locked = TRUE;
1305 }
1306 next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1307 TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1308 }
1309 vm_pagequeue_unlock(pq);
1310
1311 #if !defined(NO_SWAPPING)
1312 /*
1313 * Wakeup the swapout daemon if we didn't cache or free the targeted
1314 * number of pages.
1315 */
1316 if (vm_swap_enabled && page_shortage > 0)
1317 vm_req_vmdaemon(VM_SWAP_NORMAL);
1318 #endif
1319
1320 /*
1321 * Wakeup the sync daemon if we skipped a vnode in a writeable object
1322 * and we didn't cache or free enough pages.
1323 */
1324 if (vnodes_skipped > 0 && page_shortage > cnt.v_free_target -
1325 cnt.v_free_min)
1326 (void)speedup_syncer();
1327
1328 /*
1329 * Compute the number of pages we want to try to move from the
1330 * active queue to the inactive queue.
1331 */
1332 page_shortage = cnt.v_inactive_target - cnt.v_inactive_count +
1333 vm_paging_target() + deficit + addl_page_shortage;
1334
1335 pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1336 vm_pagequeue_lock(pq);
1337 maxscan = pq->pq_cnt;
1338
1339 /*
1340 * If we're just idle polling attempt to visit every
1341 * active page within 'update_period' seconds.
1342 */
1343 if (pass == 0 && vm_pageout_update_period != 0) {
1344 maxscan /= vm_pageout_update_period;
1345 page_shortage = maxscan;
1346 }
1347
1348 /*
1349 * Scan the active queue for things we can deactivate. We nominally
1350 * track the per-page activity counter and use it to locate
1351 * deactivation candidates.
1352 */
1353 m = TAILQ_FIRST(&pq->pq_pl);
1354 while (m != NULL && maxscan-- > 0 && page_shortage > 0) {
1355
1356 KASSERT(m->queue == PQ_ACTIVE,
1357 ("vm_pageout_scan: page %p isn't active", m));
1358
1359 next = TAILQ_NEXT(m, plinks.q);
1360 if ((m->flags & PG_MARKER) != 0) {
1361 m = next;
1362 continue;
1363 }
1364 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1365 ("Fictitious page %p cannot be in active queue", m));
1366 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1367 ("Unmanaged page %p cannot be in active queue", m));
1368 if (!vm_pageout_page_lock(m, &next)) {
1369 vm_page_unlock(m);
1370 m = next;
1371 continue;
1372 }
1373
1374 /*
1375 * The count for pagedaemon pages is done after checking the
1376 * page for eligibility...
1377 */
1378 PCPU_INC(cnt.v_pdpages);
1379
1380 /*
1381 * Check to see "how much" the page has been used.
1382 */
1383 act_delta = 0;
1384 if (m->aflags & PGA_REFERENCED) {
1385 vm_page_aflag_clear(m, PGA_REFERENCED);
1386 act_delta += 1;
1387 }
1388 /*
1389 * Unlocked object ref count check. Two races are possible.
1390 * 1) The ref was transitioning to zero and we saw non-zero,
1391 * the pmap bits will be checked unnecessarily.
1392 * 2) The ref was transitioning to one and we saw zero.
1393 * The page lock prevents a new reference to this page so
1394 * we need not check the reference bits.
1395 */
1396 if (m->object->ref_count != 0)
1397 act_delta += pmap_ts_referenced(m);
1398
1399 /*
1400 * Advance or decay the act_count based on recent usage.
1401 */
1402 if (act_delta) {
1403 m->act_count += ACT_ADVANCE + act_delta;
1404 if (m->act_count > ACT_MAX)
1405 m->act_count = ACT_MAX;
1406 } else {
1407 m->act_count -= min(m->act_count, ACT_DECLINE);
1408 act_delta = m->act_count;
1409 }
1410
1411 /*
1412 * Move this page to the tail of the active or inactive
1413 * queue depending on usage.
1414 */
1415 if (act_delta == 0) {
1416 /* Dequeue to avoid later lock recursion. */
1417 vm_page_dequeue_locked(m);
1418 vm_page_deactivate(m);
1419 page_shortage--;
1420 } else
1421 vm_page_requeue_locked(m);
1422 vm_page_unlock(m);
1423 m = next;
1424 }
1425 vm_pagequeue_unlock(pq);
1426 #if !defined(NO_SWAPPING)
1427 /*
1428 * Idle process swapout -- run once per second.
1429 */
1430 if (vm_swap_idle_enabled) {
1431 static long lsec;
1432 if (time_second != lsec) {
1433 vm_req_vmdaemon(VM_SWAP_IDLE);
1434 lsec = time_second;
1435 }
1436 }
1437 #endif
1438
1439 /*
1440 * If we are critically low on one of RAM or swap and low on
1441 * the other, kill the largest process. However, we avoid
1442 * doing this on the first pass in order to give ourselves a
1443 * chance to flush out dirty vnode-backed pages and to allow
1444 * active pages to be moved to the inactive queue and reclaimed.
1445 */
1446 vm_pageout_mightbe_oom(vmd, pass);
1447 }
1448
1449 static int vm_pageout_oom_vote;
1450
1451 /*
1452 * The pagedaemon threads randlomly select one to perform the
1453 * OOM. Trying to kill processes before all pagedaemons
1454 * failed to reach free target is premature.
1455 */
1456 static void
1457 vm_pageout_mightbe_oom(struct vm_domain *vmd, int pass)
1458 {
1459 int old_vote;
1460
1461 if (pass <= 1 || !((swap_pager_avail < 64 && vm_page_count_min()) ||
1462 (swap_pager_full && vm_paging_target() > 0))) {
1463 if (vmd->vmd_oom) {
1464 vmd->vmd_oom = FALSE;
1465 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1466 }
1467 return;
1468 }
1469
1470 if (vmd->vmd_oom)
1471 return;
1472
1473 vmd->vmd_oom = TRUE;
1474 old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1475 if (old_vote != vm_ndomains - 1)
1476 return;
1477
1478 /*
1479 * The current pagedaemon thread is the last in the quorum to
1480 * start OOM. Initiate the selection and signaling of the
1481 * victim.
1482 */
1483 vm_pageout_oom(VM_OOM_MEM);
1484
1485 /*
1486 * After one round of OOM terror, recall our vote. On the
1487 * next pass, current pagedaemon would vote again if the low
1488 * memory condition is still there, due to vmd_oom being
1489 * false.
1490 */
1491 vmd->vmd_oom = FALSE;
1492 atomic_subtract_int(&vm_pageout_oom_vote, 1);
1493 }
1494
1495 void
1496 vm_pageout_oom(int shortage)
1497 {
1498 struct proc *p, *bigproc;
1499 vm_offset_t size, bigsize;
1500 struct thread *td;
1501 struct vmspace *vm;
1502
1503 /*
1504 * We keep the process bigproc locked once we find it to keep anyone
1505 * from messing with it; however, there is a possibility of
1506 * deadlock if process B is bigproc and one of it's child processes
1507 * attempts to propagate a signal to B while we are waiting for A's
1508 * lock while walking this list. To avoid this, we don't block on
1509 * the process lock but just skip a process if it is already locked.
1510 */
1511 bigproc = NULL;
1512 bigsize = 0;
1513 sx_slock(&allproc_lock);
1514 FOREACH_PROC_IN_SYSTEM(p) {
1515 int breakout;
1516
1517 if (PROC_TRYLOCK(p) == 0)
1518 continue;
1519 /*
1520 * If this is a system, protected or killed process, skip it.
1521 */
1522 if (p->p_state != PRS_NORMAL ||
1523 (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1524 (p->p_pid == 1) || P_KILLED(p) ||
1525 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1526 PROC_UNLOCK(p);
1527 continue;
1528 }
1529 /*
1530 * If the process is in a non-running type state,
1531 * don't touch it. Check all the threads individually.
1532 */
1533 breakout = 0;
1534 FOREACH_THREAD_IN_PROC(p, td) {
1535 thread_lock(td);
1536 if (!TD_ON_RUNQ(td) &&
1537 !TD_IS_RUNNING(td) &&
1538 !TD_IS_SLEEPING(td) &&
1539 !TD_IS_SUSPENDED(td)) {
1540 thread_unlock(td);
1541 breakout = 1;
1542 break;
1543 }
1544 thread_unlock(td);
1545 }
1546 if (breakout) {
1547 PROC_UNLOCK(p);
1548 continue;
1549 }
1550 /*
1551 * get the process size
1552 */
1553 vm = vmspace_acquire_ref(p);
1554 if (vm == NULL) {
1555 PROC_UNLOCK(p);
1556 continue;
1557 }
1558 if (!vm_map_trylock_read(&vm->vm_map)) {
1559 vmspace_free(vm);
1560 PROC_UNLOCK(p);
1561 continue;
1562 }
1563 size = vmspace_swap_count(vm);
1564 vm_map_unlock_read(&vm->vm_map);
1565 if (shortage == VM_OOM_MEM)
1566 size += vmspace_resident_count(vm);
1567 vmspace_free(vm);
1568 /*
1569 * if the this process is bigger than the biggest one
1570 * remember it.
1571 */
1572 if (size > bigsize) {
1573 if (bigproc != NULL)
1574 PROC_UNLOCK(bigproc);
1575 bigproc = p;
1576 bigsize = size;
1577 } else
1578 PROC_UNLOCK(p);
1579 }
1580 sx_sunlock(&allproc_lock);
1581 if (bigproc != NULL) {
1582 killproc(bigproc, "out of swap space");
1583 sched_nice(bigproc, PRIO_MIN);
1584 PROC_UNLOCK(bigproc);
1585 wakeup(&cnt.v_free_count);
1586 }
1587 }
1588
1589 static void
1590 vm_pageout_worker(void *arg)
1591 {
1592 struct vm_domain *domain;
1593 int domidx;
1594
1595 domidx = (uintptr_t)arg;
1596 domain = &vm_dom[domidx];
1597
1598 /*
1599 * XXXKIB It could be useful to bind pageout daemon threads to
1600 * the cores belonging to the domain, from which vm_page_array
1601 * is allocated.
1602 */
1603
1604 KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1605 vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1606
1607 /*
1608 * The pageout daemon worker is never done, so loop forever.
1609 */
1610 while (TRUE) {
1611 /*
1612 * If we have enough free memory, wakeup waiters. Do
1613 * not clear vm_pages_needed until we reach our target,
1614 * otherwise we may be woken up over and over again and
1615 * waste a lot of cpu.
1616 */
1617 mtx_lock(&vm_page_queue_free_mtx);
1618 if (vm_pages_needed && !vm_page_count_min()) {
1619 if (!vm_paging_needed())
1620 vm_pages_needed = 0;
1621 wakeup(&cnt.v_free_count);
1622 }
1623 if (vm_pages_needed) {
1624 /*
1625 * Still not done, take a second pass without waiting
1626 * (unlimited dirty cleaning), otherwise sleep a bit
1627 * and try again.
1628 */
1629 if (domain->vmd_pass > 1)
1630 msleep(&vm_pages_needed,
1631 &vm_page_queue_free_mtx, PVM, "psleep",
1632 hz / 2);
1633 } else {
1634 /*
1635 * Good enough, sleep until required to refresh
1636 * stats.
1637 */
1638 domain->vmd_pass = 0;
1639 msleep(&vm_pages_needed, &vm_page_queue_free_mtx,
1640 PVM, "psleep", hz);
1641
1642 }
1643 if (vm_pages_needed) {
1644 cnt.v_pdwakeups++;
1645 domain->vmd_pass++;
1646 }
1647 mtx_unlock(&vm_page_queue_free_mtx);
1648 vm_pageout_scan(domain, domain->vmd_pass);
1649 }
1650 }
1651
1652 /*
1653 * vm_pageout is the high level pageout daemon.
1654 */
1655 static void
1656 vm_pageout(void)
1657 {
1658 #if MAXMEMDOM > 1
1659 int error, i;
1660 #endif
1661
1662 /*
1663 * Initialize some paging parameters.
1664 */
1665 cnt.v_interrupt_free_min = 2;
1666 if (cnt.v_page_count < 2000)
1667 vm_pageout_page_count = 8;
1668
1669 /*
1670 * v_free_reserved needs to include enough for the largest
1671 * swap pager structures plus enough for any pv_entry structs
1672 * when paging.
1673 */
1674 if (cnt.v_page_count > 1024)
1675 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1676 else
1677 cnt.v_free_min = 4;
1678 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1679 cnt.v_interrupt_free_min;
1680 cnt.v_free_reserved = vm_pageout_page_count +
1681 cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1682 cnt.v_free_severe = cnt.v_free_min / 2;
1683 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1684 cnt.v_free_min += cnt.v_free_reserved;
1685 cnt.v_free_severe += cnt.v_free_reserved;
1686 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1687 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1688 cnt.v_inactive_target = cnt.v_free_count / 3;
1689
1690 /*
1691 * Set the default wakeup threshold to be 10% above the minimum
1692 * page limit. This keeps the steady state out of shortfall.
1693 */
1694 vm_pageout_wakeup_thresh = (cnt.v_free_min / 10) * 11;
1695
1696 /*
1697 * Set interval in seconds for active scan. We want to visit each
1698 * page at least once every ten minutes. This is to prevent worst
1699 * case paging behaviors with stale active LRU.
1700 */
1701 if (vm_pageout_update_period == 0)
1702 vm_pageout_update_period = 600;
1703
1704 /* XXX does not really belong here */
1705 if (vm_page_max_wired == 0)
1706 vm_page_max_wired = cnt.v_free_count / 3;
1707
1708 swap_pager_swap_init();
1709 #if MAXMEMDOM > 1
1710 for (i = 1; i < vm_ndomains; i++) {
1711 error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1712 curproc, NULL, 0, 0, "dom%d", i);
1713 if (error != 0) {
1714 panic("starting pageout for domain %d, error %d\n",
1715 i, error);
1716 }
1717 }
1718 #endif
1719 vm_pageout_worker((void *)(uintptr_t)0);
1720 }
1721
1722 /*
1723 * Unless the free page queue lock is held by the caller, this function
1724 * should be regarded as advisory. Specifically, the caller should
1725 * not msleep() on &cnt.v_free_count following this function unless
1726 * the free page queue lock is held until the msleep() is performed.
1727 */
1728 void
1729 pagedaemon_wakeup(void)
1730 {
1731
1732 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1733 vm_pages_needed = 1;
1734 wakeup(&vm_pages_needed);
1735 }
1736 }
1737
1738 #if !defined(NO_SWAPPING)
1739 static void
1740 vm_req_vmdaemon(int req)
1741 {
1742 static int lastrun = 0;
1743
1744 mtx_lock(&vm_daemon_mtx);
1745 vm_pageout_req_swapout |= req;
1746 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1747 wakeup(&vm_daemon_needed);
1748 lastrun = ticks;
1749 }
1750 mtx_unlock(&vm_daemon_mtx);
1751 }
1752
1753 static void
1754 vm_daemon(void)
1755 {
1756 struct rlimit rsslim;
1757 struct proc *p;
1758 struct thread *td;
1759 struct vmspace *vm;
1760 int breakout, swapout_flags, tryagain, attempts;
1761 #ifdef RACCT
1762 uint64_t rsize, ravailable;
1763 #endif
1764
1765 while (TRUE) {
1766 mtx_lock(&vm_daemon_mtx);
1767 #ifdef RACCT
1768 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz);
1769 #else
1770 msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1771 #endif
1772 swapout_flags = vm_pageout_req_swapout;
1773 vm_pageout_req_swapout = 0;
1774 mtx_unlock(&vm_daemon_mtx);
1775 if (swapout_flags)
1776 swapout_procs(swapout_flags);
1777
1778 /*
1779 * scan the processes for exceeding their rlimits or if
1780 * process is swapped out -- deactivate pages
1781 */
1782 tryagain = 0;
1783 attempts = 0;
1784 again:
1785 attempts++;
1786 sx_slock(&allproc_lock);
1787 FOREACH_PROC_IN_SYSTEM(p) {
1788 vm_pindex_t limit, size;
1789
1790 /*
1791 * if this is a system process or if we have already
1792 * looked at this process, skip it.
1793 */
1794 PROC_LOCK(p);
1795 if (p->p_state != PRS_NORMAL ||
1796 p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1797 PROC_UNLOCK(p);
1798 continue;
1799 }
1800 /*
1801 * if the process is in a non-running type state,
1802 * don't touch it.
1803 */
1804 breakout = 0;
1805 FOREACH_THREAD_IN_PROC(p, td) {
1806 thread_lock(td);
1807 if (!TD_ON_RUNQ(td) &&
1808 !TD_IS_RUNNING(td) &&
1809 !TD_IS_SLEEPING(td) &&
1810 !TD_IS_SUSPENDED(td)) {
1811 thread_unlock(td);
1812 breakout = 1;
1813 break;
1814 }
1815 thread_unlock(td);
1816 }
1817 if (breakout) {
1818 PROC_UNLOCK(p);
1819 continue;
1820 }
1821 /*
1822 * get a limit
1823 */
1824 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1825 limit = OFF_TO_IDX(
1826 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1827
1828 /*
1829 * let processes that are swapped out really be
1830 * swapped out set the limit to nothing (will force a
1831 * swap-out.)
1832 */
1833 if ((p->p_flag & P_INMEM) == 0)
1834 limit = 0; /* XXX */
1835 vm = vmspace_acquire_ref(p);
1836 PROC_UNLOCK(p);
1837 if (vm == NULL)
1838 continue;
1839
1840 size = vmspace_resident_count(vm);
1841 if (size >= limit) {
1842 vm_pageout_map_deactivate_pages(
1843 &vm->vm_map, limit);
1844 }
1845 #ifdef RACCT
1846 rsize = IDX_TO_OFF(size);
1847 PROC_LOCK(p);
1848 racct_set(p, RACCT_RSS, rsize);
1849 ravailable = racct_get_available(p, RACCT_RSS);
1850 PROC_UNLOCK(p);
1851 if (rsize > ravailable) {
1852 /*
1853 * Don't be overly aggressive; this might be
1854 * an innocent process, and the limit could've
1855 * been exceeded by some memory hog. Don't
1856 * try to deactivate more than 1/4th of process'
1857 * resident set size.
1858 */
1859 if (attempts <= 8) {
1860 if (ravailable < rsize - (rsize / 4))
1861 ravailable = rsize - (rsize / 4);
1862 }
1863 vm_pageout_map_deactivate_pages(
1864 &vm->vm_map, OFF_TO_IDX(ravailable));
1865 /* Update RSS usage after paging out. */
1866 size = vmspace_resident_count(vm);
1867 rsize = IDX_TO_OFF(size);
1868 PROC_LOCK(p);
1869 racct_set(p, RACCT_RSS, rsize);
1870 PROC_UNLOCK(p);
1871 if (rsize > ravailable)
1872 tryagain = 1;
1873 }
1874 #endif
1875 vmspace_free(vm);
1876 }
1877 sx_sunlock(&allproc_lock);
1878 if (tryagain != 0 && attempts <= 10)
1879 goto again;
1880 }
1881 }
1882 #endif /* !defined(NO_SWAPPING) */
Cache object: 018b019c3b28c27f242739eba524477a
|