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 *
9 * This code is derived from software contributed to Berkeley by
10 * The Mach Operating System project at Carnegie-Mellon University.
11 *
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
14 * are met:
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * SUCH DAMAGE.
39 *
40 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91
41 *
42 *
43 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44 * All rights reserved.
45 *
46 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47 *
48 * Permission to use, copy, modify and distribute this software and
49 * its documentation is hereby granted, provided that both the copyright
50 * notice and this permission notice appear in all copies of the
51 * software, derivative works or modified versions, and any portions
52 * thereof, and that both notices appear in supporting documentation.
53 *
54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57 *
58 * Carnegie Mellon requests users of this software to return to
59 *
60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
61 * School of Computer Science
62 * Carnegie Mellon University
63 * Pittsburgh PA 15213-3890
64 *
65 * any improvements or extensions that they make and grant Carnegie the
66 * rights to redistribute these changes.
67 */
68
69 /*
70 * The proverbial page-out daemon.
71 */
72
73 #include <sys/cdefs.h>
74 __FBSDID("$FreeBSD: releng/5.4/sys/vm/vm_pageout.c 142737 2005-02-28 01:29:46Z alc $");
75
76 #include "opt_vm.h"
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/eventhandler.h>
81 #include <sys/lock.h>
82 #include <sys/mutex.h>
83 #include <sys/proc.h>
84 #include <sys/kthread.h>
85 #include <sys/ktr.h>
86 #include <sys/resourcevar.h>
87 #include <sys/sched.h>
88 #include <sys/signalvar.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 #include <sys/sx.h>
92 #include <sys/sysctl.h>
93
94 #include <vm/vm.h>
95 #include <vm/vm_param.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_pageout.h>
100 #include <vm/vm_pager.h>
101 #include <vm/swap_pager.h>
102 #include <vm/vm_extern.h>
103 #include <vm/uma.h>
104
105 #include <machine/mutex.h>
106
107 /*
108 * System initialization
109 */
110
111 /* the kernel process "vm_pageout"*/
112 static void vm_pageout(void);
113 static int vm_pageout_clean(vm_page_t);
114 static void vm_pageout_pmap_collect(void);
115 static void vm_pageout_scan(int pass);
116
117 struct proc *pageproc;
118
119 static struct kproc_desc page_kp = {
120 "pagedaemon",
121 vm_pageout,
122 &pageproc
123 };
124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
125
126 #if !defined(NO_SWAPPING)
127 /* the kernel process "vm_daemon"*/
128 static void vm_daemon(void);
129 static struct proc *vmproc;
130
131 static struct kproc_desc vm_kp = {
132 "vmdaemon",
133 vm_daemon,
134 &vmproc
135 };
136 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
137 #endif
138
139
140 int vm_pages_needed; /* Event on which pageout daemon sleeps */
141 int vm_pageout_deficit; /* Estimated number of pages deficit */
142 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */
143
144 #if !defined(NO_SWAPPING)
145 static int vm_pageout_req_swapout; /* XXX */
146 static int vm_daemon_needed;
147 #endif
148 static int vm_max_launder = 32;
149 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
150 static int vm_pageout_full_stats_interval = 0;
151 static int vm_pageout_algorithm=0;
152 static int defer_swap_pageouts=0;
153 static int disable_swap_pageouts=0;
154
155 #if defined(NO_SWAPPING)
156 static int vm_swap_enabled=0;
157 static int vm_swap_idle_enabled=0;
158 #else
159 static int vm_swap_enabled=1;
160 static int vm_swap_idle_enabled=0;
161 #endif
162
163 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
164 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
165
166 SYSCTL_INT(_vm, OID_AUTO, max_launder,
167 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
168
169 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
170 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
171
172 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
173 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
174
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
176 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
177
178 #if defined(NO_SWAPPING)
179 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
180 CTLFLAG_RD, &vm_swap_enabled, 0, "");
181 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
182 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
183 #else
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 #endif
189
190 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
191 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
192
193 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
194 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
195
196 static int pageout_lock_miss;
197 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
198 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
199
200 #define VM_PAGEOUT_PAGE_COUNT 16
201 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
202
203 int vm_page_max_wired; /* XXX max # of wired pages system-wide */
204
205 #if !defined(NO_SWAPPING)
206 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
207 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
208 static void vm_req_vmdaemon(void);
209 #endif
210 static void vm_pageout_page_stats(void);
211
212 /*
213 * vm_pageout_clean:
214 *
215 * Clean the page and remove it from the laundry.
216 *
217 * We set the busy bit to cause potential page faults on this page to
218 * block. Note the careful timing, however, the busy bit isn't set till
219 * late and we cannot do anything that will mess with the page.
220 */
221 static int
222 vm_pageout_clean(m)
223 vm_page_t m;
224 {
225 vm_object_t object;
226 vm_page_t mc[2*vm_pageout_page_count];
227 int pageout_count;
228 int ib, is, page_base;
229 vm_pindex_t pindex = m->pindex;
230
231 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
232 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
233
234 /*
235 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
236 * with the new swapper, but we could have serious problems paging
237 * out other object types if there is insufficient memory.
238 *
239 * Unfortunately, checking free memory here is far too late, so the
240 * check has been moved up a procedural level.
241 */
242
243 /*
244 * Don't mess with the page if it's busy, held, or special
245 */
246 if ((m->hold_count != 0) ||
247 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
248 return 0;
249 }
250
251 mc[vm_pageout_page_count] = m;
252 pageout_count = 1;
253 page_base = vm_pageout_page_count;
254 ib = 1;
255 is = 1;
256
257 /*
258 * Scan object for clusterable pages.
259 *
260 * We can cluster ONLY if: ->> the page is NOT
261 * clean, wired, busy, held, or mapped into a
262 * buffer, and one of the following:
263 * 1) The page is inactive, or a seldom used
264 * active page.
265 * -or-
266 * 2) we force the issue.
267 *
268 * During heavy mmap/modification loads the pageout
269 * daemon can really fragment the underlying file
270 * due to flushing pages out of order and not trying
271 * align the clusters (which leave sporatic out-of-order
272 * holes). To solve this problem we do the reverse scan
273 * first and attempt to align our cluster, then do a
274 * forward scan if room remains.
275 */
276 object = m->object;
277 more:
278 while (ib && pageout_count < vm_pageout_page_count) {
279 vm_page_t p;
280
281 if (ib > pindex) {
282 ib = 0;
283 break;
284 }
285
286 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
287 ib = 0;
288 break;
289 }
290 if (((p->queue - p->pc) == PQ_CACHE) ||
291 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
292 ib = 0;
293 break;
294 }
295 vm_page_test_dirty(p);
296 if ((p->dirty & p->valid) == 0 ||
297 p->queue != PQ_INACTIVE ||
298 p->wire_count != 0 || /* may be held by buf cache */
299 p->hold_count != 0) { /* may be undergoing I/O */
300 ib = 0;
301 break;
302 }
303 mc[--page_base] = p;
304 ++pageout_count;
305 ++ib;
306 /*
307 * alignment boundry, stop here and switch directions. Do
308 * not clear ib.
309 */
310 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
311 break;
312 }
313
314 while (pageout_count < vm_pageout_page_count &&
315 pindex + is < object->size) {
316 vm_page_t p;
317
318 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
319 break;
320 if (((p->queue - p->pc) == PQ_CACHE) ||
321 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
322 break;
323 }
324 vm_page_test_dirty(p);
325 if ((p->dirty & p->valid) == 0 ||
326 p->queue != PQ_INACTIVE ||
327 p->wire_count != 0 || /* may be held by buf cache */
328 p->hold_count != 0) { /* may be undergoing I/O */
329 break;
330 }
331 mc[page_base + pageout_count] = p;
332 ++pageout_count;
333 ++is;
334 }
335
336 /*
337 * If we exhausted our forward scan, continue with the reverse scan
338 * when possible, even past a page boundry. This catches boundry
339 * conditions.
340 */
341 if (ib && pageout_count < vm_pageout_page_count)
342 goto more;
343
344 /*
345 * we allow reads during pageouts...
346 */
347 return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
348 }
349
350 /*
351 * vm_pageout_flush() - launder the given pages
352 *
353 * The given pages are laundered. Note that we setup for the start of
354 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
355 * reference count all in here rather then in the parent. If we want
356 * the parent to do more sophisticated things we may have to change
357 * the ordering.
358 */
359 int
360 vm_pageout_flush(vm_page_t *mc, int count, int flags)
361 {
362 vm_object_t object = mc[0]->object;
363 int pageout_status[count];
364 int numpagedout = 0;
365 int i;
366
367 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
368 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
369 /*
370 * Initiate I/O. Bump the vm_page_t->busy counter and
371 * mark the pages read-only.
372 *
373 * We do not have to fixup the clean/dirty bits here... we can
374 * allow the pager to do it after the I/O completes.
375 *
376 * NOTE! mc[i]->dirty may be partial or fragmented due to an
377 * edge case with file fragments.
378 */
379 for (i = 0; i < count; i++) {
380 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
381 ("vm_pageout_flush: partially invalid page %p index %d/%d",
382 mc[i], i, count));
383 vm_page_io_start(mc[i]);
384 pmap_page_protect(mc[i], VM_PROT_READ);
385 }
386 vm_page_unlock_queues();
387 vm_object_pip_add(object, count);
388
389 vm_pager_put_pages(object, mc, count,
390 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
391 pageout_status);
392
393 vm_page_lock_queues();
394 for (i = 0; i < count; i++) {
395 vm_page_t mt = mc[i];
396
397 KASSERT((mt->flags & PG_WRITEABLE) == 0,
398 ("vm_pageout_flush: page %p is not write protected", mt));
399 switch (pageout_status[i]) {
400 case VM_PAGER_OK:
401 case VM_PAGER_PEND:
402 numpagedout++;
403 break;
404 case VM_PAGER_BAD:
405 /*
406 * Page outside of range of object. Right now we
407 * essentially lose the changes by pretending it
408 * worked.
409 */
410 pmap_clear_modify(mt);
411 vm_page_undirty(mt);
412 break;
413 case VM_PAGER_ERROR:
414 case VM_PAGER_FAIL:
415 /*
416 * If page couldn't be paged out, then reactivate the
417 * page so it doesn't clog the inactive list. (We
418 * will try paging out it again later).
419 */
420 vm_page_activate(mt);
421 break;
422 case VM_PAGER_AGAIN:
423 break;
424 }
425
426 /*
427 * If the operation is still going, leave the page busy to
428 * block all other accesses. Also, leave the paging in
429 * progress indicator set so that we don't attempt an object
430 * collapse.
431 */
432 if (pageout_status[i] != VM_PAGER_PEND) {
433 vm_object_pip_wakeup(object);
434 vm_page_io_finish(mt);
435 if (vm_page_count_severe())
436 vm_page_try_to_cache(mt);
437 }
438 }
439 return numpagedout;
440 }
441
442 #if !defined(NO_SWAPPING)
443 /*
444 * vm_pageout_object_deactivate_pages
445 *
446 * deactivate enough pages to satisfy the inactive target
447 * requirements or if vm_page_proc_limit is set, then
448 * deactivate all of the pages in the object and its
449 * backing_objects.
450 *
451 * The object and map must be locked.
452 */
453 static void
454 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
455 pmap_t pmap;
456 vm_object_t first_object;
457 long desired;
458 {
459 vm_object_t backing_object, object;
460 vm_page_t p, next;
461 int actcount, rcount, remove_mode;
462
463 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
464 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
465 return;
466 for (object = first_object;; object = backing_object) {
467 if (pmap_resident_count(pmap) <= desired)
468 goto unlock_return;
469 if (object->paging_in_progress)
470 goto unlock_return;
471
472 remove_mode = 0;
473 if (object->shadow_count > 1)
474 remove_mode = 1;
475 /*
476 * scan the objects entire memory queue
477 */
478 rcount = object->resident_page_count;
479 p = TAILQ_FIRST(&object->memq);
480 vm_page_lock_queues();
481 while (p && (rcount-- > 0)) {
482 if (pmap_resident_count(pmap) <= desired) {
483 vm_page_unlock_queues();
484 goto unlock_return;
485 }
486 next = TAILQ_NEXT(p, listq);
487 cnt.v_pdpages++;
488 if (p->wire_count != 0 ||
489 p->hold_count != 0 ||
490 p->busy != 0 ||
491 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
492 !pmap_page_exists_quick(pmap, p)) {
493 p = next;
494 continue;
495 }
496 actcount = pmap_ts_referenced(p);
497 if (actcount) {
498 vm_page_flag_set(p, PG_REFERENCED);
499 } else if (p->flags & PG_REFERENCED) {
500 actcount = 1;
501 }
502 if ((p->queue != PQ_ACTIVE) &&
503 (p->flags & PG_REFERENCED)) {
504 vm_page_activate(p);
505 p->act_count += actcount;
506 vm_page_flag_clear(p, PG_REFERENCED);
507 } else if (p->queue == PQ_ACTIVE) {
508 if ((p->flags & PG_REFERENCED) == 0) {
509 p->act_count -= min(p->act_count, ACT_DECLINE);
510 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
511 pmap_remove_all(p);
512 vm_page_deactivate(p);
513 } else {
514 vm_pageq_requeue(p);
515 }
516 } else {
517 vm_page_activate(p);
518 vm_page_flag_clear(p, PG_REFERENCED);
519 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
520 p->act_count += ACT_ADVANCE;
521 vm_pageq_requeue(p);
522 }
523 } else if (p->queue == PQ_INACTIVE) {
524 pmap_remove_all(p);
525 }
526 p = next;
527 }
528 vm_page_unlock_queues();
529 if ((backing_object = object->backing_object) == NULL)
530 goto unlock_return;
531 VM_OBJECT_LOCK(backing_object);
532 if (object != first_object)
533 VM_OBJECT_UNLOCK(object);
534 }
535 unlock_return:
536 if (object != first_object)
537 VM_OBJECT_UNLOCK(object);
538 }
539
540 /*
541 * deactivate some number of pages in a map, try to do it fairly, but
542 * that is really hard to do.
543 */
544 static void
545 vm_pageout_map_deactivate_pages(map, desired)
546 vm_map_t map;
547 long desired;
548 {
549 vm_map_entry_t tmpe;
550 vm_object_t obj, bigobj;
551 int nothingwired;
552
553 if (!vm_map_trylock(map))
554 return;
555
556 bigobj = NULL;
557 nothingwired = TRUE;
558
559 /*
560 * first, search out the biggest object, and try to free pages from
561 * that.
562 */
563 tmpe = map->header.next;
564 while (tmpe != &map->header) {
565 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
566 obj = tmpe->object.vm_object;
567 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
568 if (obj->shadow_count <= 1 &&
569 (bigobj == NULL ||
570 bigobj->resident_page_count < obj->resident_page_count)) {
571 if (bigobj != NULL)
572 VM_OBJECT_UNLOCK(bigobj);
573 bigobj = obj;
574 } else
575 VM_OBJECT_UNLOCK(obj);
576 }
577 }
578 if (tmpe->wired_count > 0)
579 nothingwired = FALSE;
580 tmpe = tmpe->next;
581 }
582
583 if (bigobj != NULL) {
584 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
585 VM_OBJECT_UNLOCK(bigobj);
586 }
587 /*
588 * Next, hunt around for other pages to deactivate. We actually
589 * do this search sort of wrong -- .text first is not the best idea.
590 */
591 tmpe = map->header.next;
592 while (tmpe != &map->header) {
593 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
594 break;
595 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
596 obj = tmpe->object.vm_object;
597 if (obj != NULL) {
598 VM_OBJECT_LOCK(obj);
599 vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
600 VM_OBJECT_UNLOCK(obj);
601 }
602 }
603 tmpe = tmpe->next;
604 }
605
606 /*
607 * Remove all mappings if a process is swapped out, this will free page
608 * table pages.
609 */
610 if (desired == 0 && nothingwired) {
611 pmap_remove(vm_map_pmap(map), vm_map_min(map),
612 vm_map_max(map));
613 }
614 vm_map_unlock(map);
615 }
616 #endif /* !defined(NO_SWAPPING) */
617
618 /*
619 * This routine is very drastic, but can save the system
620 * in a pinch.
621 */
622 static void
623 vm_pageout_pmap_collect(void)
624 {
625 int i;
626 vm_page_t m;
627 static int warningdone;
628
629 if (pmap_pagedaemon_waken == 0)
630 return;
631 if (warningdone < 5) {
632 printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
633 warningdone++;
634 }
635 vm_page_lock_queues();
636 for (i = 0; i < vm_page_array_size; i++) {
637 m = &vm_page_array[i];
638 if (m->wire_count || m->hold_count || m->busy ||
639 (m->flags & (PG_BUSY | PG_UNMANAGED)))
640 continue;
641 pmap_remove_all(m);
642 }
643 vm_page_unlock_queues();
644 pmap_pagedaemon_waken = 0;
645 }
646
647 /*
648 * vm_pageout_scan does the dirty work for the pageout daemon.
649 */
650 static void
651 vm_pageout_scan(int pass)
652 {
653 vm_page_t m, next;
654 struct vm_page marker;
655 int page_shortage, maxscan, pcount;
656 int addl_page_shortage, addl_page_shortage_init;
657 struct proc *p, *bigproc;
658 struct thread *td;
659 vm_offset_t size, bigsize;
660 vm_object_t object;
661 int actcount, cache_cur, cache_first_failure;
662 static int cache_last_free;
663 int vnodes_skipped = 0;
664 int maxlaunder;
665
666 mtx_lock(&Giant);
667 /*
668 * Decrease registered cache sizes.
669 */
670 EVENTHANDLER_INVOKE(vm_lowmem, 0);
671 /*
672 * We do this explicitly after the caches have been drained above.
673 */
674 uma_reclaim();
675 /*
676 * Do whatever cleanup that the pmap code can.
677 */
678 vm_pageout_pmap_collect();
679
680 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
681
682 /*
683 * Calculate the number of pages we want to either free or move
684 * to the cache.
685 */
686 page_shortage = vm_paging_target() + addl_page_shortage_init;
687
688 /*
689 * Initialize our marker
690 */
691 bzero(&marker, sizeof(marker));
692 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
693 marker.queue = PQ_INACTIVE;
694 marker.wire_count = 1;
695
696 /*
697 * Start scanning the inactive queue for pages we can move to the
698 * cache or free. The scan will stop when the target is reached or
699 * we have scanned the entire inactive queue. Note that m->act_count
700 * is not used to form decisions for the inactive queue, only for the
701 * active queue.
702 *
703 * maxlaunder limits the number of dirty pages we flush per scan.
704 * For most systems a smaller value (16 or 32) is more robust under
705 * extreme memory and disk pressure because any unnecessary writes
706 * to disk can result in extreme performance degredation. However,
707 * systems with excessive dirty pages (especially when MAP_NOSYNC is
708 * used) will die horribly with limited laundering. If the pageout
709 * daemon cannot clean enough pages in the first pass, we let it go
710 * all out in succeeding passes.
711 */
712 if ((maxlaunder = vm_max_launder) <= 1)
713 maxlaunder = 1;
714 if (pass)
715 maxlaunder = 10000;
716 vm_page_lock_queues();
717 rescan0:
718 addl_page_shortage = addl_page_shortage_init;
719 maxscan = cnt.v_inactive_count;
720
721 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
722 m != NULL && maxscan-- > 0 && page_shortage > 0;
723 m = next) {
724
725 cnt.v_pdpages++;
726
727 if (m->queue != PQ_INACTIVE) {
728 goto rescan0;
729 }
730
731 next = TAILQ_NEXT(m, pageq);
732 object = m->object;
733
734 /*
735 * skip marker pages
736 */
737 if (m->flags & PG_MARKER)
738 continue;
739
740 /*
741 * A held page may be undergoing I/O, so skip it.
742 */
743 if (m->hold_count) {
744 vm_pageq_requeue(m);
745 addl_page_shortage++;
746 continue;
747 }
748 /*
749 * Don't mess with busy pages, keep in the front of the
750 * queue, most likely are being paged out.
751 */
752 if (!VM_OBJECT_TRYLOCK(object)) {
753 addl_page_shortage++;
754 continue;
755 }
756 if (m->busy || (m->flags & PG_BUSY)) {
757 VM_OBJECT_UNLOCK(object);
758 addl_page_shortage++;
759 continue;
760 }
761
762 /*
763 * If the object is not being used, we ignore previous
764 * references.
765 */
766 if (object->ref_count == 0) {
767 vm_page_flag_clear(m, PG_REFERENCED);
768 pmap_clear_reference(m);
769
770 /*
771 * Otherwise, if the page has been referenced while in the
772 * inactive queue, we bump the "activation count" upwards,
773 * making it less likely that the page will be added back to
774 * the inactive queue prematurely again. Here we check the
775 * page tables (or emulated bits, if any), given the upper
776 * level VM system not knowing anything about existing
777 * references.
778 */
779 } else if (((m->flags & PG_REFERENCED) == 0) &&
780 (actcount = pmap_ts_referenced(m))) {
781 vm_page_activate(m);
782 VM_OBJECT_UNLOCK(object);
783 m->act_count += (actcount + ACT_ADVANCE);
784 continue;
785 }
786
787 /*
788 * If the upper level VM system knows about any page
789 * references, we activate the page. We also set the
790 * "activation count" higher than normal so that we will less
791 * likely place pages back onto the inactive queue again.
792 */
793 if ((m->flags & PG_REFERENCED) != 0) {
794 vm_page_flag_clear(m, PG_REFERENCED);
795 actcount = pmap_ts_referenced(m);
796 vm_page_activate(m);
797 VM_OBJECT_UNLOCK(object);
798 m->act_count += (actcount + ACT_ADVANCE + 1);
799 continue;
800 }
801
802 /*
803 * If the upper level VM system doesn't know anything about
804 * the page being dirty, we have to check for it again. As
805 * far as the VM code knows, any partially dirty pages are
806 * fully dirty.
807 */
808 if (m->dirty == 0 && !pmap_is_modified(m)) {
809 /*
810 * Avoid a race condition: Unless write access is
811 * removed from the page, another processor could
812 * modify it before all access is removed by the call
813 * to vm_page_cache() below. If vm_page_cache() finds
814 * that the page has been modified when it removes all
815 * access, it panics because it cannot cache dirty
816 * pages. In principle, we could eliminate just write
817 * access here rather than all access. In the expected
818 * case, when there are no last instant modifications
819 * to the page, removing all access will be cheaper
820 * overall.
821 */
822 if ((m->flags & PG_WRITEABLE) != 0)
823 pmap_remove_all(m);
824 } else {
825 vm_page_dirty(m);
826 }
827
828 if (m->valid == 0) {
829 /*
830 * Invalid pages can be easily freed
831 */
832 pmap_remove_all(m);
833 vm_page_free(m);
834 cnt.v_dfree++;
835 --page_shortage;
836 } else if (m->dirty == 0) {
837 /*
838 * Clean pages can be placed onto the cache queue.
839 * This effectively frees them.
840 */
841 vm_page_cache(m);
842 --page_shortage;
843 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
844 /*
845 * Dirty pages need to be paged out, but flushing
846 * a page is extremely expensive verses freeing
847 * a clean page. Rather then artificially limiting
848 * the number of pages we can flush, we instead give
849 * dirty pages extra priority on the inactive queue
850 * by forcing them to be cycled through the queue
851 * twice before being flushed, after which the
852 * (now clean) page will cycle through once more
853 * before being freed. This significantly extends
854 * the thrash point for a heavily loaded machine.
855 */
856 vm_page_flag_set(m, PG_WINATCFLS);
857 vm_pageq_requeue(m);
858 } else if (maxlaunder > 0) {
859 /*
860 * We always want to try to flush some dirty pages if
861 * we encounter them, to keep the system stable.
862 * Normally this number is small, but under extreme
863 * pressure where there are insufficient clean pages
864 * on the inactive queue, we may have to go all out.
865 */
866 int swap_pageouts_ok;
867 struct vnode *vp = NULL;
868 struct mount *mp;
869
870 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
871 swap_pageouts_ok = 1;
872 } else {
873 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
874 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
875 vm_page_count_min());
876
877 }
878
879 /*
880 * We don't bother paging objects that are "dead".
881 * Those objects are in a "rundown" state.
882 */
883 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
884 VM_OBJECT_UNLOCK(object);
885 vm_pageq_requeue(m);
886 continue;
887 }
888
889 /*
890 * The object is already known NOT to be dead. It
891 * is possible for the vget() to block the whole
892 * pageout daemon, but the new low-memory handling
893 * code should prevent it.
894 *
895 * The previous code skipped locked vnodes and, worse,
896 * reordered pages in the queue. This results in
897 * completely non-deterministic operation and, on a
898 * busy system, can lead to extremely non-optimal
899 * pageouts. For example, it can cause clean pages
900 * to be freed and dirty pages to be moved to the end
901 * of the queue. Since dirty pages are also moved to
902 * the end of the queue once-cleaned, this gives
903 * way too large a weighting to defering the freeing
904 * of dirty pages.
905 *
906 * We can't wait forever for the vnode lock, we might
907 * deadlock due to a vn_read() getting stuck in
908 * vm_wait while holding this vnode. We skip the
909 * vnode if we can't get it in a reasonable amount
910 * of time.
911 */
912 if (object->type == OBJT_VNODE) {
913 vp = object->handle;
914 mp = NULL;
915 if (vp->v_type == VREG)
916 vn_start_write(vp, &mp, V_NOWAIT);
917 vm_page_unlock_queues();
918 VI_LOCK(vp);
919 VM_OBJECT_UNLOCK(object);
920 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK |
921 LK_TIMELOCK, curthread)) {
922 VM_OBJECT_LOCK(object);
923 vm_page_lock_queues();
924 ++pageout_lock_miss;
925 vn_finished_write(mp);
926 if (object->flags & OBJ_MIGHTBEDIRTY)
927 vnodes_skipped++;
928 VM_OBJECT_UNLOCK(object);
929 continue;
930 }
931 VM_OBJECT_LOCK(object);
932 vm_page_lock_queues();
933 /*
934 * The page might have been moved to another
935 * queue during potential blocking in vget()
936 * above. The page might have been freed and
937 * reused for another vnode. The object might
938 * have been reused for another vnode.
939 */
940 if (m->queue != PQ_INACTIVE ||
941 m->object != object ||
942 object->handle != vp) {
943 if (object->flags & OBJ_MIGHTBEDIRTY)
944 vnodes_skipped++;
945 goto unlock_and_continue;
946 }
947
948 /*
949 * The page may have been busied during the
950 * blocking in vput(); We don't move the
951 * page back onto the end of the queue so that
952 * statistics are more correct if we don't.
953 */
954 if (m->busy || (m->flags & PG_BUSY)) {
955 goto unlock_and_continue;
956 }
957
958 /*
959 * If the page has become held it might
960 * be undergoing I/O, so skip it
961 */
962 if (m->hold_count) {
963 vm_pageq_requeue(m);
964 if (object->flags & OBJ_MIGHTBEDIRTY)
965 vnodes_skipped++;
966 goto unlock_and_continue;
967 }
968 }
969
970 /*
971 * If a page is dirty, then it is either being washed
972 * (but not yet cleaned) or it is still in the
973 * laundry. If it is still in the laundry, then we
974 * start the cleaning operation.
975 *
976 * This operation may cluster, invalidating the 'next'
977 * pointer. To prevent an inordinate number of
978 * restarts we use our marker to remember our place.
979 *
980 * decrement page_shortage on success to account for
981 * the (future) cleaned page. Otherwise we could wind
982 * up laundering or cleaning too many pages.
983 */
984 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
985 if (vm_pageout_clean(m) != 0) {
986 --page_shortage;
987 --maxlaunder;
988 }
989 next = TAILQ_NEXT(&marker, pageq);
990 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
991 unlock_and_continue:
992 VM_OBJECT_UNLOCK(object);
993 if (vp) {
994 vm_page_unlock_queues();
995 vput(vp);
996 vn_finished_write(mp);
997 vm_page_lock_queues();
998 }
999 continue;
1000 }
1001 VM_OBJECT_UNLOCK(object);
1002 }
1003
1004 /*
1005 * Compute the number of pages we want to try to move from the
1006 * active queue to the inactive queue.
1007 */
1008 page_shortage = vm_paging_target() +
1009 cnt.v_inactive_target - cnt.v_inactive_count;
1010 page_shortage += addl_page_shortage;
1011
1012 /*
1013 * Scan the active queue for things we can deactivate. We nominally
1014 * track the per-page activity counter and use it to locate
1015 * deactivation candidates.
1016 */
1017 pcount = cnt.v_active_count;
1018 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1019
1020 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1021
1022 KASSERT(m->queue == PQ_ACTIVE,
1023 ("vm_pageout_scan: page %p isn't active", m));
1024
1025 next = TAILQ_NEXT(m, pageq);
1026 object = m->object;
1027 if (!VM_OBJECT_TRYLOCK(object)) {
1028 vm_pageq_requeue(m);
1029 m = next;
1030 continue;
1031 }
1032
1033 /*
1034 * Don't deactivate pages that are busy.
1035 */
1036 if ((m->busy != 0) ||
1037 (m->flags & PG_BUSY) ||
1038 (m->hold_count != 0)) {
1039 VM_OBJECT_UNLOCK(object);
1040 vm_pageq_requeue(m);
1041 m = next;
1042 continue;
1043 }
1044
1045 /*
1046 * The count for pagedaemon pages is done after checking the
1047 * page for eligibility...
1048 */
1049 cnt.v_pdpages++;
1050
1051 /*
1052 * Check to see "how much" the page has been used.
1053 */
1054 actcount = 0;
1055 if (object->ref_count != 0) {
1056 if (m->flags & PG_REFERENCED) {
1057 actcount += 1;
1058 }
1059 actcount += pmap_ts_referenced(m);
1060 if (actcount) {
1061 m->act_count += ACT_ADVANCE + actcount;
1062 if (m->act_count > ACT_MAX)
1063 m->act_count = ACT_MAX;
1064 }
1065 }
1066
1067 /*
1068 * Since we have "tested" this bit, we need to clear it now.
1069 */
1070 vm_page_flag_clear(m, PG_REFERENCED);
1071
1072 /*
1073 * Only if an object is currently being used, do we use the
1074 * page activation count stats.
1075 */
1076 if (actcount && (object->ref_count != 0)) {
1077 vm_pageq_requeue(m);
1078 } else {
1079 m->act_count -= min(m->act_count, ACT_DECLINE);
1080 if (vm_pageout_algorithm ||
1081 object->ref_count == 0 ||
1082 m->act_count == 0) {
1083 page_shortage--;
1084 if (object->ref_count == 0) {
1085 pmap_remove_all(m);
1086 if (m->dirty == 0)
1087 vm_page_cache(m);
1088 else
1089 vm_page_deactivate(m);
1090 } else {
1091 vm_page_deactivate(m);
1092 }
1093 } else {
1094 vm_pageq_requeue(m);
1095 }
1096 }
1097 VM_OBJECT_UNLOCK(object);
1098 m = next;
1099 }
1100
1101 /*
1102 * We try to maintain some *really* free pages, this allows interrupt
1103 * code to be guaranteed space. Since both cache and free queues
1104 * are considered basically 'free', moving pages from cache to free
1105 * does not effect other calculations.
1106 */
1107 cache_cur = cache_last_free;
1108 cache_first_failure = -1;
1109 while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur =
1110 (cache_cur + PQ_PRIME2) & PQ_L2_MASK) != cache_first_failure) {
1111 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl,
1112 pageq) {
1113 KASSERT(m->dirty == 0,
1114 ("Found dirty cache page %p", m));
1115 KASSERT(!pmap_page_is_mapped(m),
1116 ("Found mapped cache page %p", m));
1117 KASSERT((m->flags & PG_UNMANAGED) == 0,
1118 ("Found unmanaged cache page %p", m));
1119 KASSERT(m->wire_count == 0,
1120 ("Found wired cache page %p", m));
1121 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1122 m->object)) {
1123 KASSERT((m->flags & PG_BUSY) == 0 &&
1124 m->busy == 0, ("Found busy cache page %p",
1125 m));
1126 vm_page_free(m);
1127 VM_OBJECT_UNLOCK(object);
1128 cnt.v_dfree++;
1129 cache_last_free = cache_cur;
1130 cache_first_failure = -1;
1131 break;
1132 }
1133 }
1134 if (m == NULL && cache_first_failure == -1)
1135 cache_first_failure = cache_cur;
1136 }
1137 vm_page_unlock_queues();
1138 #if !defined(NO_SWAPPING)
1139 /*
1140 * Idle process swapout -- run once per second.
1141 */
1142 if (vm_swap_idle_enabled) {
1143 static long lsec;
1144 if (time_second != lsec) {
1145 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1146 vm_req_vmdaemon();
1147 lsec = time_second;
1148 }
1149 }
1150 #endif
1151
1152 /*
1153 * If we didn't get enough free pages, and we have skipped a vnode
1154 * in a writeable object, wakeup the sync daemon. And kick swapout
1155 * if we did not get enough free pages.
1156 */
1157 if (vm_paging_target() > 0) {
1158 if (vnodes_skipped && vm_page_count_min())
1159 (void) speedup_syncer();
1160 #if !defined(NO_SWAPPING)
1161 if (vm_swap_enabled && vm_page_count_target()) {
1162 vm_req_vmdaemon();
1163 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1164 }
1165 #endif
1166 }
1167
1168 /*
1169 * If we are critically low on one of RAM or swap and low on
1170 * the other, kill the largest process. However, we avoid
1171 * doing this on the first pass in order to give ourselves a
1172 * chance to flush out dirty vnode-backed pages and to allow
1173 * active pages to be moved to the inactive queue and reclaimed.
1174 *
1175 * We keep the process bigproc locked once we find it to keep anyone
1176 * from messing with it; however, there is a possibility of
1177 * deadlock if process B is bigproc and one of it's child processes
1178 * attempts to propagate a signal to B while we are waiting for A's
1179 * lock while walking this list. To avoid this, we don't block on
1180 * the process lock but just skip a process if it is already locked.
1181 */
1182 if (pass != 0 &&
1183 ((swap_pager_avail < 64 && vm_page_count_min()) ||
1184 (swap_pager_full && vm_paging_target() > 0))) {
1185 bigproc = NULL;
1186 bigsize = 0;
1187 sx_slock(&allproc_lock);
1188 FOREACH_PROC_IN_SYSTEM(p) {
1189 int breakout;
1190
1191 if (PROC_TRYLOCK(p) == 0)
1192 continue;
1193 /*
1194 * If this is a system or protected process, skip it.
1195 */
1196 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1197 (p->p_flag & P_PROTECTED) ||
1198 ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1199 PROC_UNLOCK(p);
1200 continue;
1201 }
1202 /*
1203 * If the process is in a non-running type state,
1204 * don't touch it. Check all the threads individually.
1205 */
1206 mtx_lock_spin(&sched_lock);
1207 breakout = 0;
1208 FOREACH_THREAD_IN_PROC(p, td) {
1209 if (!TD_ON_RUNQ(td) &&
1210 !TD_IS_RUNNING(td) &&
1211 !TD_IS_SLEEPING(td)) {
1212 breakout = 1;
1213 break;
1214 }
1215 }
1216 if (breakout) {
1217 mtx_unlock_spin(&sched_lock);
1218 PROC_UNLOCK(p);
1219 continue;
1220 }
1221 mtx_unlock_spin(&sched_lock);
1222 /*
1223 * get the process size
1224 */
1225 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1226 PROC_UNLOCK(p);
1227 continue;
1228 }
1229 size = vmspace_swap_count(p->p_vmspace);
1230 vm_map_unlock_read(&p->p_vmspace->vm_map);
1231 size += vmspace_resident_count(p->p_vmspace);
1232 /*
1233 * if the this process is bigger than the biggest one
1234 * remember it.
1235 */
1236 if (size > bigsize) {
1237 if (bigproc != NULL)
1238 PROC_UNLOCK(bigproc);
1239 bigproc = p;
1240 bigsize = size;
1241 } else
1242 PROC_UNLOCK(p);
1243 }
1244 sx_sunlock(&allproc_lock);
1245 if (bigproc != NULL) {
1246 killproc(bigproc, "out of swap space");
1247 mtx_lock_spin(&sched_lock);
1248 sched_nice(bigproc, PRIO_MIN);
1249 mtx_unlock_spin(&sched_lock);
1250 PROC_UNLOCK(bigproc);
1251 wakeup(&cnt.v_free_count);
1252 }
1253 }
1254 mtx_unlock(&Giant);
1255 }
1256
1257 /*
1258 * This routine tries to maintain the pseudo LRU active queue,
1259 * so that during long periods of time where there is no paging,
1260 * that some statistic accumulation still occurs. This code
1261 * helps the situation where paging just starts to occur.
1262 */
1263 static void
1264 vm_pageout_page_stats()
1265 {
1266 vm_object_t object;
1267 vm_page_t m,next;
1268 int pcount,tpcount; /* Number of pages to check */
1269 static int fullintervalcount = 0;
1270 int page_shortage;
1271
1272 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1273 page_shortage =
1274 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1275 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1276
1277 if (page_shortage <= 0)
1278 return;
1279
1280 pcount = cnt.v_active_count;
1281 fullintervalcount += vm_pageout_stats_interval;
1282 if (fullintervalcount < vm_pageout_full_stats_interval) {
1283 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1284 if (pcount > tpcount)
1285 pcount = tpcount;
1286 } else {
1287 fullintervalcount = 0;
1288 }
1289
1290 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1291 while ((m != NULL) && (pcount-- > 0)) {
1292 int actcount;
1293
1294 KASSERT(m->queue == PQ_ACTIVE,
1295 ("vm_pageout_page_stats: page %p isn't active", m));
1296
1297 next = TAILQ_NEXT(m, pageq);
1298 object = m->object;
1299 if (!VM_OBJECT_TRYLOCK(object)) {
1300 vm_pageq_requeue(m);
1301 m = next;
1302 continue;
1303 }
1304
1305 /*
1306 * Don't deactivate pages that are busy.
1307 */
1308 if ((m->busy != 0) ||
1309 (m->flags & PG_BUSY) ||
1310 (m->hold_count != 0)) {
1311 VM_OBJECT_UNLOCK(object);
1312 vm_pageq_requeue(m);
1313 m = next;
1314 continue;
1315 }
1316
1317 actcount = 0;
1318 if (m->flags & PG_REFERENCED) {
1319 vm_page_flag_clear(m, PG_REFERENCED);
1320 actcount += 1;
1321 }
1322
1323 actcount += pmap_ts_referenced(m);
1324 if (actcount) {
1325 m->act_count += ACT_ADVANCE + actcount;
1326 if (m->act_count > ACT_MAX)
1327 m->act_count = ACT_MAX;
1328 vm_pageq_requeue(m);
1329 } else {
1330 if (m->act_count == 0) {
1331 /*
1332 * We turn off page access, so that we have
1333 * more accurate RSS stats. We don't do this
1334 * in the normal page deactivation when the
1335 * system is loaded VM wise, because the
1336 * cost of the large number of page protect
1337 * operations would be higher than the value
1338 * of doing the operation.
1339 */
1340 pmap_remove_all(m);
1341 vm_page_deactivate(m);
1342 } else {
1343 m->act_count -= min(m->act_count, ACT_DECLINE);
1344 vm_pageq_requeue(m);
1345 }
1346 }
1347 VM_OBJECT_UNLOCK(object);
1348 m = next;
1349 }
1350 }
1351
1352 /*
1353 * vm_pageout is the high level pageout daemon.
1354 */
1355 static void
1356 vm_pageout()
1357 {
1358 int error, pass;
1359
1360 /*
1361 * Initialize some paging parameters.
1362 */
1363 cnt.v_interrupt_free_min = 2;
1364 if (cnt.v_page_count < 2000)
1365 vm_pageout_page_count = 8;
1366
1367 /*
1368 * v_free_reserved needs to include enough for the largest
1369 * swap pager structures plus enough for any pv_entry structs
1370 * when paging.
1371 */
1372 if (cnt.v_page_count > 1024)
1373 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1374 else
1375 cnt.v_free_min = 4;
1376 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1377 cnt.v_interrupt_free_min;
1378 cnt.v_free_reserved = vm_pageout_page_count +
1379 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_L2_SIZE;
1380 cnt.v_free_severe = cnt.v_free_min / 2;
1381 cnt.v_free_min += cnt.v_free_reserved;
1382 cnt.v_free_severe += cnt.v_free_reserved;
1383
1384 /*
1385 * v_free_target and v_cache_min control pageout hysteresis. Note
1386 * that these are more a measure of the VM cache queue hysteresis
1387 * then the VM free queue. Specifically, v_free_target is the
1388 * high water mark (free+cache pages).
1389 *
1390 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1391 * low water mark, while v_free_min is the stop. v_cache_min must
1392 * be big enough to handle memory needs while the pageout daemon
1393 * is signalled and run to free more pages.
1394 */
1395 if (cnt.v_free_count > 6144)
1396 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1397 else
1398 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1399
1400 if (cnt.v_free_count > 2048) {
1401 cnt.v_cache_min = cnt.v_free_target;
1402 cnt.v_cache_max = 2 * cnt.v_cache_min;
1403 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1404 } else {
1405 cnt.v_cache_min = 0;
1406 cnt.v_cache_max = 0;
1407 cnt.v_inactive_target = cnt.v_free_count / 4;
1408 }
1409 if (cnt.v_inactive_target > cnt.v_free_count / 3)
1410 cnt.v_inactive_target = cnt.v_free_count / 3;
1411
1412 /* XXX does not really belong here */
1413 if (vm_page_max_wired == 0)
1414 vm_page_max_wired = cnt.v_free_count / 3;
1415
1416 if (vm_pageout_stats_max == 0)
1417 vm_pageout_stats_max = cnt.v_free_target;
1418
1419 /*
1420 * Set interval in seconds for stats scan.
1421 */
1422 if (vm_pageout_stats_interval == 0)
1423 vm_pageout_stats_interval = 5;
1424 if (vm_pageout_full_stats_interval == 0)
1425 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1426
1427 swap_pager_swap_init();
1428 pass = 0;
1429 /*
1430 * The pageout daemon is never done, so loop forever.
1431 */
1432 while (TRUE) {
1433 vm_page_lock_queues();
1434 /*
1435 * If we have enough free memory, wakeup waiters. Do
1436 * not clear vm_pages_needed until we reach our target,
1437 * otherwise we may be woken up over and over again and
1438 * waste a lot of cpu.
1439 */
1440 if (vm_pages_needed && !vm_page_count_min()) {
1441 if (!vm_paging_needed())
1442 vm_pages_needed = 0;
1443 wakeup(&cnt.v_free_count);
1444 }
1445 if (vm_pages_needed) {
1446 /*
1447 * Still not done, take a second pass without waiting
1448 * (unlimited dirty cleaning), otherwise sleep a bit
1449 * and try again.
1450 */
1451 ++pass;
1452 if (pass > 1)
1453 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1454 "psleep", hz/2);
1455 } else {
1456 /*
1457 * Good enough, sleep & handle stats. Prime the pass
1458 * for the next run.
1459 */
1460 if (pass > 1)
1461 pass = 1;
1462 else
1463 pass = 0;
1464 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1465 "psleep", vm_pageout_stats_interval * hz);
1466 if (error && !vm_pages_needed) {
1467 pass = 0;
1468 vm_pageout_page_stats();
1469 vm_page_unlock_queues();
1470 continue;
1471 }
1472 }
1473 if (vm_pages_needed)
1474 cnt.v_pdwakeups++;
1475 vm_page_unlock_queues();
1476 vm_pageout_scan(pass);
1477 }
1478 }
1479
1480 /*
1481 * Unless the page queue lock is held by the caller, this function
1482 * should be regarded as advisory. Specifically, the caller should
1483 * not msleep() on &cnt.v_free_count following this function unless
1484 * the page queue lock is held until the msleep() is performed.
1485 */
1486 void
1487 pagedaemon_wakeup()
1488 {
1489
1490 if (!vm_pages_needed && curthread->td_proc != pageproc) {
1491 vm_pages_needed = 1;
1492 wakeup(&vm_pages_needed);
1493 }
1494 }
1495
1496 #if !defined(NO_SWAPPING)
1497 static void
1498 vm_req_vmdaemon()
1499 {
1500 static int lastrun = 0;
1501
1502 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1503 wakeup(&vm_daemon_needed);
1504 lastrun = ticks;
1505 }
1506 }
1507
1508 static void
1509 vm_daemon()
1510 {
1511 struct rlimit rsslim;
1512 struct proc *p;
1513 struct thread *td;
1514 int breakout;
1515
1516 mtx_lock(&Giant);
1517 while (TRUE) {
1518 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1519 if (vm_pageout_req_swapout) {
1520 swapout_procs(vm_pageout_req_swapout);
1521 vm_pageout_req_swapout = 0;
1522 }
1523 /*
1524 * scan the processes for exceeding their rlimits or if
1525 * process is swapped out -- deactivate pages
1526 */
1527 sx_slock(&allproc_lock);
1528 LIST_FOREACH(p, &allproc, p_list) {
1529 vm_pindex_t limit, size;
1530
1531 /*
1532 * if this is a system process or if we have already
1533 * looked at this process, skip it.
1534 */
1535 PROC_LOCK(p);
1536 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1537 PROC_UNLOCK(p);
1538 continue;
1539 }
1540 /*
1541 * if the process is in a non-running type state,
1542 * don't touch it.
1543 */
1544 mtx_lock_spin(&sched_lock);
1545 breakout = 0;
1546 FOREACH_THREAD_IN_PROC(p, td) {
1547 if (!TD_ON_RUNQ(td) &&
1548 !TD_IS_RUNNING(td) &&
1549 !TD_IS_SLEEPING(td)) {
1550 breakout = 1;
1551 break;
1552 }
1553 }
1554 mtx_unlock_spin(&sched_lock);
1555 if (breakout) {
1556 PROC_UNLOCK(p);
1557 continue;
1558 }
1559 /*
1560 * get a limit
1561 */
1562 lim_rlimit(p, RLIMIT_RSS, &rsslim);
1563 limit = OFF_TO_IDX(
1564 qmin(rsslim.rlim_cur, rsslim.rlim_max));
1565
1566 /*
1567 * let processes that are swapped out really be
1568 * swapped out set the limit to nothing (will force a
1569 * swap-out.)
1570 */
1571 if ((p->p_sflag & PS_INMEM) == 0)
1572 limit = 0; /* XXX */
1573 PROC_UNLOCK(p);
1574
1575 size = vmspace_resident_count(p->p_vmspace);
1576 if (limit >= 0 && size >= limit) {
1577 vm_pageout_map_deactivate_pages(
1578 &p->p_vmspace->vm_map, limit);
1579 }
1580 }
1581 sx_sunlock(&allproc_lock);
1582 }
1583 }
1584 #endif /* !defined(NO_SWAPPING) */
Cache object: 1f26782acf96758b1ae988c49d09c5c7
|