kernel - fix panic on reboot when swap populated
[dragonfly.git] / sys / vm / vm_pageout.c
CommitLineData
984263bc
MD
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 * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
4ecf7cc9 69 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.36 2008/07/01 02:02:56 dillon Exp $
984263bc
MD
70 */
71
72/*
73 * The proverbial page-out daemon.
74 */
75
76#include "opt_vm.h"
77#include <sys/param.h>
78#include <sys/systm.h>
79#include <sys/kernel.h>
80#include <sys/proc.h>
81#include <sys/kthread.h>
82#include <sys/resourcevar.h>
83#include <sys/signalvar.h>
84#include <sys/vnode.h>
85#include <sys/vmmeter.h>
86#include <sys/sysctl.h>
87
88#include <vm/vm.h>
89#include <vm/vm_param.h>
90#include <sys/lock.h>
91#include <vm/vm_object.h>
92#include <vm/vm_page.h>
93#include <vm/vm_map.h>
94#include <vm/vm_pageout.h>
95#include <vm/vm_pager.h>
96#include <vm/swap_pager.h>
97#include <vm/vm_extern.h>
5fd012e0
MD
98
99#include <sys/thread2.h>
12e4aaff 100#include <vm/vm_page2.h>
984263bc
MD
101
102/*
103 * System initialization
104 */
105
106/* the kernel process "vm_pageout"*/
1388df65
RG
107static void vm_pageout (void);
108static int vm_pageout_clean (vm_page_t);
20479584 109static int vm_pageout_scan (int pass);
1388df65 110static int vm_pageout_free_page_calc (vm_size_t count);
bc6dffab 111struct thread *pagethread;
984263bc
MD
112
113static struct kproc_desc page_kp = {
114 "pagedaemon",
115 vm_pageout,
bc6dffab 116 &pagethread
984263bc
MD
117};
118SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
119
120#if !defined(NO_SWAPPING)
121/* the kernel process "vm_daemon"*/
1388df65 122static void vm_daemon (void);
bc6dffab 123static struct thread *vmthread;
984263bc
MD
124
125static struct kproc_desc vm_kp = {
126 "vmdaemon",
127 vm_daemon,
bc6dffab 128 &vmthread
984263bc
MD
129};
130SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
131#endif
132
133
134int vm_pages_needed=0; /* Event on which pageout daemon sleeps */
135int vm_pageout_deficit=0; /* Estimated number of pages deficit */
136int vm_pageout_pages_needed=0; /* flag saying that the pageout daemon needs pages */
137
138#if !defined(NO_SWAPPING)
139static int vm_pageout_req_swapout; /* XXX */
140static int vm_daemon_needed;
141#endif
984263bc
MD
142static int vm_max_launder = 32;
143static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
144static int vm_pageout_full_stats_interval = 0;
145static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
146static int defer_swap_pageouts=0;
147static int disable_swap_pageouts=0;
148
149#if defined(NO_SWAPPING)
150static int vm_swap_enabled=0;
151static int vm_swap_idle_enabled=0;
152#else
153static int vm_swap_enabled=1;
154static int vm_swap_idle_enabled=0;
155#endif
156
157SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
158 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
159
160SYSCTL_INT(_vm, OID_AUTO, max_launder,
161 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
162
163SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
164 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
165
166SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
167 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
168
169SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
170 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
171
172SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
173 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
174
175#if defined(NO_SWAPPING)
176SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
177 CTLFLAG_RD, &vm_swap_enabled, 0, "");
178SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
179 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
180#else
181SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
182 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
183SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
184 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
185#endif
186
187SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
188 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
189
190SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
191 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
192
193static int pageout_lock_miss;
194SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
195 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
196
46311ac2
MD
197int vm_load;
198SYSCTL_INT(_vm, OID_AUTO, vm_load,
199 CTLFLAG_RD, &vm_load, 0, "load on the VM system");
200int vm_load_enable = 1;
201SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
202 CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
203#ifdef INVARIANTS
204int vm_load_debug;
205SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
206 CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
207#endif
208
984263bc
MD
209#define VM_PAGEOUT_PAGE_COUNT 16
210int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
211
212int vm_page_max_wired; /* XXX max # of wired pages system-wide */
213
214#if !defined(NO_SWAPPING)
1388df65
RG
215typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
216static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
984263bc 217static freeer_fcn_t vm_pageout_object_deactivate_pages;
1388df65 218static void vm_req_vmdaemon (void);
984263bc
MD
219#endif
220static void vm_pageout_page_stats(void);
221
46311ac2 222/*
20479584 223 * Update vm_load to slow down faulting processes.
46311ac2
MD
224 */
225void
226vm_fault_ratecheck(void)
227{
228 if (vm_pages_needed) {
229 if (vm_load < 1000)
230 ++vm_load;
231 } else {
232 if (vm_load > 0)
233 --vm_load;
234 }
235}
236
984263bc
MD
237/*
238 * vm_pageout_clean:
239 *
06ecca5a
MD
240 * Clean the page and remove it from the laundry. The page must not be
241 * busy on-call.
984263bc
MD
242 *
243 * We set the busy bit to cause potential page faults on this page to
244 * block. Note the careful timing, however, the busy bit isn't set till
245 * late and we cannot do anything that will mess with the page.
246 */
247
248static int
57e43348 249vm_pageout_clean(vm_page_t m)
984263bc 250{
5f910b2f 251 vm_object_t object;
984263bc
MD
252 vm_page_t mc[2*vm_pageout_page_count];
253 int pageout_count;
254 int ib, is, page_base;
255 vm_pindex_t pindex = m->pindex;
256
257 object = m->object;
258
259 /*
260 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
261 * with the new swapper, but we could have serious problems paging
262 * out other object types if there is insufficient memory.
263 *
264 * Unfortunately, checking free memory here is far too late, so the
265 * check has been moved up a procedural level.
266 */
267
268 /*
269 * Don't mess with the page if it's busy, held, or special
270 */
271 if ((m->hold_count != 0) ||
272 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
273 return 0;
274 }
275
276 mc[vm_pageout_page_count] = m;
277 pageout_count = 1;
278 page_base = vm_pageout_page_count;
279 ib = 1;
280 is = 1;
281
282 /*
283 * Scan object for clusterable pages.
284 *
285 * We can cluster ONLY if: ->> the page is NOT
286 * clean, wired, busy, held, or mapped into a
287 * buffer, and one of the following:
288 * 1) The page is inactive, or a seldom used
289 * active page.
290 * -or-
291 * 2) we force the issue.
292 *
293 * During heavy mmap/modification loads the pageout
294 * daemon can really fragment the underlying file
295 * due to flushing pages out of order and not trying
296 * align the clusters (which leave sporatic out-of-order
297 * holes). To solve this problem we do the reverse scan
298 * first and attempt to align our cluster, then do a
299 * forward scan if room remains.
300 */
301
302more:
303 while (ib && pageout_count < vm_pageout_page_count) {
304 vm_page_t p;
305
306 if (ib > pindex) {
307 ib = 0;
308 break;
309 }
310
311 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
312 ib = 0;
313 break;
314 }
315 if (((p->queue - p->pc) == PQ_CACHE) ||
316 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
317 ib = 0;
318 break;
319 }
320 vm_page_test_dirty(p);
321 if ((p->dirty & p->valid) == 0 ||
322 p->queue != PQ_INACTIVE ||
323 p->wire_count != 0 || /* may be held by buf cache */
324 p->hold_count != 0) { /* may be undergoing I/O */
325 ib = 0;
326 break;
327 }
328 mc[--page_base] = p;
329 ++pageout_count;
330 ++ib;
331 /*
332 * alignment boundry, stop here and switch directions. Do
333 * not clear ib.
334 */
335 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
336 break;
337 }
338
339 while (pageout_count < vm_pageout_page_count &&
340 pindex + is < object->size) {
341 vm_page_t p;
342
343 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
344 break;
345 if (((p->queue - p->pc) == PQ_CACHE) ||
346 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
347 break;
348 }
349 vm_page_test_dirty(p);
350 if ((p->dirty & p->valid) == 0 ||
351 p->queue != PQ_INACTIVE ||
352 p->wire_count != 0 || /* may be held by buf cache */
353 p->hold_count != 0) { /* may be undergoing I/O */
354 break;
355 }
356 mc[page_base + pageout_count] = p;
357 ++pageout_count;
358 ++is;
359 }
360
361 /*
362 * If we exhausted our forward scan, continue with the reverse scan
363 * when possible, even past a page boundry. This catches boundry
364 * conditions.
365 */
366 if (ib && pageout_count < vm_pageout_page_count)
367 goto more;
368
369 /*
370 * we allow reads during pageouts...
371 */
372 return vm_pageout_flush(&mc[page_base], pageout_count, 0);
373}
374
375/*
376 * vm_pageout_flush() - launder the given pages
377 *
378 * The given pages are laundered. Note that we setup for the start of
379 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
380 * reference count all in here rather then in the parent. If we want
381 * the parent to do more sophisticated things we may have to change
382 * the ordering.
383 */
984263bc 384int
57e43348 385vm_pageout_flush(vm_page_t *mc, int count, int flags)
984263bc 386{
5f910b2f 387 vm_object_t object;
984263bc
MD
388 int pageout_status[count];
389 int numpagedout = 0;
390 int i;
391
392 /*
17cde63e
MD
393 * Initiate I/O. Bump the vm_page_t->busy counter.
394 */
395 for (i = 0; i < count; i++) {
396 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
397 vm_page_io_start(mc[i]);
398 }
399
400 /*
4530a3aa
MD
401 * We must make the pages read-only. This will also force the
402 * modified bit in the related pmaps to be cleared. The pager
403 * cannot clear the bit for us since the I/O completion code
404 * typically runs from an interrupt. The act of making the page
405 * read-only handles the case for us.
984263bc 406 */
984263bc 407 for (i = 0; i < count; i++) {
984263bc
MD
408 vm_page_protect(mc[i], VM_PROT_READ);
409 }
410
411 object = mc[0]->object;
412 vm_object_pip_add(object, count);
413
414 vm_pager_put_pages(object, mc, count,
c439ad8f 415 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
984263bc
MD
416 pageout_status);
417
418 for (i = 0; i < count; i++) {
419 vm_page_t mt = mc[i];
420
421 switch (pageout_status[i]) {
422 case VM_PAGER_OK:
423 numpagedout++;
424 break;
425 case VM_PAGER_PEND:
426 numpagedout++;
427 break;
428 case VM_PAGER_BAD:
429 /*
430 * Page outside of range of object. Right now we
431 * essentially lose the changes by pretending it
432 * worked.
433 */
434 pmap_clear_modify(mt);
435 vm_page_undirty(mt);
436 break;
437 case VM_PAGER_ERROR:
438 case VM_PAGER_FAIL:
439 /*
c84c24da
MD
440 * A page typically cannot be paged out when we
441 * have run out of swap. We leave the page
442 * marked inactive and will try to page it out
443 * again later.
444 *
445 * Starvation of the active page list is used to
446 * determine when the system is massively memory
447 * starved.
984263bc 448 */
984263bc
MD
449 break;
450 case VM_PAGER_AGAIN:
451 break;
452 }
453
454 /*
455 * If the operation is still going, leave the page busy to
456 * block all other accesses. Also, leave the paging in
457 * progress indicator set so that we don't attempt an object
458 * collapse.
93afe6be
MD
459 *
460 * For any pages which have completed synchronously,
461 * deactivate the page if we are under a severe deficit.
462 * Do not try to enter them into the cache, though, they
463 * might still be read-heavy.
984263bc
MD
464 */
465 if (pageout_status[i] != VM_PAGER_PEND) {
466 vm_object_pip_wakeup(object);
467 vm_page_io_finish(mt);
93afe6be
MD
468 if (vm_page_count_severe())
469 vm_page_deactivate(mt);
470#if 0
984263bc
MD
471 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
472 vm_page_protect(mt, VM_PROT_READ);
93afe6be 473#endif
984263bc
MD
474 }
475 }
476 return numpagedout;
477}
478
479#if !defined(NO_SWAPPING)
480/*
481 * vm_pageout_object_deactivate_pages
482 *
483 * deactivate enough pages to satisfy the inactive target
484 * requirements or if vm_page_proc_limit is set, then
485 * deactivate all of the pages in the object and its
486 * backing_objects.
487 *
488 * The object and map must be locked.
489 */
1f804340
MD
490static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
491
984263bc 492static void
57e43348 493vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
06ecca5a 494 vm_pindex_t desired, int map_remove_only)
984263bc 495{
1f804340 496 struct rb_vm_page_scan_info info;
984263bc 497 int remove_mode;
984263bc
MD
498
499 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
500 return;
501
502 while (object) {
503 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
504 return;
505 if (object->paging_in_progress)
506 return;
507
508 remove_mode = map_remove_only;
509 if (object->shadow_count > 1)
510 remove_mode = 1;
06ecca5a
MD
511
512 /*
513 * scan the objects entire memory queue. spl protection is
514 * required to avoid an interrupt unbusy/free race against
515 * our busy check.
516 */
5fd012e0 517 crit_enter();
1f804340
MD
518 info.limit = remove_mode;
519 info.map = map;
520 info.desired = desired;
521 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
522 vm_pageout_object_deactivate_pages_callback,
523 &info
524 );
525 crit_exit();
526 object = object->backing_object;
527 }
528}
529
530static int
531vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
532{
533 struct rb_vm_page_scan_info *info = data;
534 int actcount;
984263bc 535
1f804340
MD
536 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
537 return(-1);
538 }
539 mycpu->gd_cnt.v_pdpages++;
540 if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 ||
541 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
542 !pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
543 return(0);
544 }
984263bc 545
1f804340
MD
546 actcount = pmap_ts_referenced(p);
547 if (actcount) {
548 vm_page_flag_set(p, PG_REFERENCED);
549 } else if (p->flags & PG_REFERENCED) {
550 actcount = 1;
551 }
552
553 if ((p->queue != PQ_ACTIVE) &&
554 (p->flags & PG_REFERENCED)) {
555 vm_page_activate(p);
556 p->act_count += actcount;
557 vm_page_flag_clear(p, PG_REFERENCED);
558 } else if (p->queue == PQ_ACTIVE) {
559 if ((p->flags & PG_REFERENCED) == 0) {
560 p->act_count -= min(p->act_count, ACT_DECLINE);
561 if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) {
17cde63e 562 vm_page_busy(p);
984263bc 563 vm_page_protect(p, VM_PROT_NONE);
17cde63e 564 vm_page_wakeup(p);
1f804340
MD
565 vm_page_deactivate(p);
566 } else {
567 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
568 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
984263bc 569 }
1f804340
MD
570 } else {
571 vm_page_activate(p);
572 vm_page_flag_clear(p, PG_REFERENCED);
573 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
574 p->act_count += ACT_ADVANCE;
575 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
576 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
984263bc 577 }
1f804340 578 } else if (p->queue == PQ_INACTIVE) {
17cde63e 579 vm_page_busy(p);
1f804340 580 vm_page_protect(p, VM_PROT_NONE);
17cde63e 581 vm_page_wakeup(p);
984263bc 582 }
1f804340 583 return(0);
984263bc
MD
584}
585
586/*
587 * deactivate some number of pages in a map, try to do it fairly, but
588 * that is really hard to do.
589 */
590static void
57e43348 591vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
984263bc
MD
592{
593 vm_map_entry_t tmpe;
594 vm_object_t obj, bigobj;
595 int nothingwired;
596
df4f70a6 597 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
984263bc
MD
598 return;
599 }
600
601 bigobj = NULL;
602 nothingwired = TRUE;
603
604 /*
605 * first, search out the biggest object, and try to free pages from
606 * that.
607 */
608 tmpe = map->header.next;
609 while (tmpe != &map->header) {
1b874851
MD
610 switch(tmpe->maptype) {
611 case VM_MAPTYPE_NORMAL:
612 case VM_MAPTYPE_VPAGETABLE:
984263bc
MD
613 obj = tmpe->object.vm_object;
614 if ((obj != NULL) && (obj->shadow_count <= 1) &&
615 ((bigobj == NULL) ||
616 (bigobj->resident_page_count < obj->resident_page_count))) {
617 bigobj = obj;
618 }
1b874851
MD
619 break;
620 default:
621 break;
984263bc
MD
622 }
623 if (tmpe->wired_count > 0)
624 nothingwired = FALSE;
625 tmpe = tmpe->next;
626 }
627
628 if (bigobj)
629 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
630
631 /*
632 * Next, hunt around for other pages to deactivate. We actually
633 * do this search sort of wrong -- .text first is not the best idea.
634 */
635 tmpe = map->header.next;
636 while (tmpe != &map->header) {
637 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
638 break;
1b874851
MD
639 switch(tmpe->maptype) {
640 case VM_MAPTYPE_NORMAL:
641 case VM_MAPTYPE_VPAGETABLE:
984263bc
MD
642 obj = tmpe->object.vm_object;
643 if (obj)
644 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
1b874851
MD
645 break;
646 default:
647 break;
984263bc
MD
648 }
649 tmpe = tmpe->next;
650 };
651
652 /*
653 * Remove all mappings if a process is swapped out, this will free page
654 * table pages.
655 */
656 if (desired == 0 && nothingwired)
657 pmap_remove(vm_map_pmap(map),
88181b08 658 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
984263bc 659 vm_map_unlock(map);
984263bc
MD
660}
661#endif
662
663/*
a11aaa81
MD
664 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We
665 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
666 * be trivially freed.
984263bc 667 */
984263bc 668void
95813af0
MD
669vm_pageout_page_free(vm_page_t m)
670{
984263bc
MD
671 vm_object_t object = m->object;
672 int type = object->type;
673
674 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
675 vm_object_reference(object);
676 vm_page_busy(m);
677 vm_page_protect(m, VM_PROT_NONE);
678 vm_page_free(m);
679 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
680 vm_object_deallocate(object);
681}
682
683/*
20479584 684 * vm_pageout_scan does the dirty work for the pageout daemon.
984263bc 685 */
8fa76237
MD
686struct vm_pageout_scan_info {
687 struct proc *bigproc;
688 vm_offset_t bigsize;
689};
690
691static int vm_pageout_scan_callback(struct proc *p, void *data);
692
20479584 693static int
984263bc
MD
694vm_pageout_scan(int pass)
695{
8fa76237 696 struct vm_pageout_scan_info info;
984263bc
MD
697 vm_page_t m, next;
698 struct vm_page marker;
20479584
MD
699 int maxscan, pcount;
700 int recycle_count;
701 int inactive_shortage, active_shortage;
51db7ca2 702 int inactive_original_shortage;
984263bc
MD
703 vm_object_t object;
704 int actcount;
705 int vnodes_skipped = 0;
706 int maxlaunder;
984263bc
MD
707
708 /*
709 * Do whatever cleanup that the pmap code can.
710 */
711 pmap_collect();
712
984263bc 713 /*
20479584
MD
714 * Calculate our target for the number of free+cache pages we
715 * want to get to. This is higher then the number that causes
716 * allocations to stall (severe) in order to provide hysteresis,
717 * and if we don't make it all the way but get to the minimum
718 * we're happy.
984263bc 719 */
20479584 720 inactive_shortage = vm_paging_target() + vm_pageout_deficit;
51db7ca2 721 inactive_original_shortage = inactive_shortage;
20479584 722 vm_pageout_deficit = 0;
984263bc
MD
723
724 /*
725 * Initialize our marker
726 */
727 bzero(&marker, sizeof(marker));
728 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
729 marker.queue = PQ_INACTIVE;
730 marker.wire_count = 1;
731
732 /*
733 * Start scanning the inactive queue for pages we can move to the
734 * cache or free. The scan will stop when the target is reached or
735 * we have scanned the entire inactive queue. Note that m->act_count
736 * is not used to form decisions for the inactive queue, only for the
737 * active queue.
738 *
739 * maxlaunder limits the number of dirty pages we flush per scan.
740 * For most systems a smaller value (16 or 32) is more robust under
741 * extreme memory and disk pressure because any unnecessary writes
742 * to disk can result in extreme performance degredation. However,
743 * systems with excessive dirty pages (especially when MAP_NOSYNC is
744 * used) will die horribly with limited laundering. If the pageout
745 * daemon cannot clean enough pages in the first pass, we let it go
746 * all out in succeeding passes.
747 */
748 if ((maxlaunder = vm_max_launder) <= 1)
749 maxlaunder = 1;
750 if (pass)
751 maxlaunder = 10000;
752
06ecca5a 753 /*
5fd012e0
MD
754 * We will generally be in a critical section throughout the
755 * scan, but we can release it temporarily when we are sitting on a
756 * non-busy page without fear. this is required to prevent an
757 * interrupt from unbusying or freeing a page prior to our busy
758 * check, leaving us on the wrong queue or checking the wrong
759 * page.
06ecca5a 760 */
5fd012e0 761 crit_enter();
984263bc 762rescan0:
12e4aaff 763 maxscan = vmstats.v_inactive_count;
984263bc 764 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
20479584 765 m != NULL && maxscan-- > 0 && inactive_shortage > 0;
06ecca5a
MD
766 m = next
767 ) {
12e4aaff 768 mycpu->gd_cnt.v_pdpages++;
984263bc 769
06ecca5a
MD
770 /*
771 * Give interrupts a chance
772 */
5fd012e0
MD
773 crit_exit();
774 crit_enter();
984263bc 775
06ecca5a
MD
776 /*
777 * It's easier for some of the conditions below to just loop
778 * and catch queue changes here rather then check everywhere
779 * else.
780 */
781 if (m->queue != PQ_INACTIVE)
782 goto rescan0;
984263bc
MD
783 next = TAILQ_NEXT(m, pageq);
784
785 /*
786 * skip marker pages
787 */
788 if (m->flags & PG_MARKER)
789 continue;
790
791 /*
792 * A held page may be undergoing I/O, so skip it.
793 */
794 if (m->hold_count) {
984263bc
MD
795 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
796 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
797 continue;
798 }
06ecca5a 799
984263bc
MD
800 /*
801 * Dont mess with busy pages, keep in the front of the
802 * queue, most likely are being paged out.
803 */
804 if (m->busy || (m->flags & PG_BUSY)) {
984263bc
MD
805 continue;
806 }
807
984263bc 808 if (m->object->ref_count == 0) {
06ecca5a
MD
809 /*
810 * If the object is not being used, we ignore previous
811 * references.
812 */
984263bc
MD
813 vm_page_flag_clear(m, PG_REFERENCED);
814 pmap_clear_reference(m);
815
984263bc 816 } else if (((m->flags & PG_REFERENCED) == 0) &&
06ecca5a
MD
817 (actcount = pmap_ts_referenced(m))) {
818 /*
819 * Otherwise, if the page has been referenced while
820 * in the inactive queue, we bump the "activation
821 * count" upwards, making it less likely that the
822 * page will be added back to the inactive queue
823 * prematurely again. Here we check the page tables
824 * (or emulated bits, if any), given the upper level
825 * VM system not knowing anything about existing
826 * references.
827 */
984263bc
MD
828 vm_page_activate(m);
829 m->act_count += (actcount + ACT_ADVANCE);
830 continue;
831 }
832
833 /*
834 * If the upper level VM system knows about any page
835 * references, we activate the page. We also set the
836 * "activation count" higher than normal so that we will less
837 * likely place pages back onto the inactive queue again.
838 */
839 if ((m->flags & PG_REFERENCED) != 0) {
840 vm_page_flag_clear(m, PG_REFERENCED);
841 actcount = pmap_ts_referenced(m);
842 vm_page_activate(m);
843 m->act_count += (actcount + ACT_ADVANCE + 1);
844 continue;
845 }
846
847 /*
848 * If the upper level VM system doesn't know anything about
849 * the page being dirty, we have to check for it again. As
850 * far as the VM code knows, any partially dirty pages are
851 * fully dirty.
41a01a4d
MD
852 *
853 * Pages marked PG_WRITEABLE may be mapped into the user
854 * address space of a process running on another cpu. A
855 * user process (without holding the MP lock) running on
856 * another cpu may be able to touch the page while we are
17cde63e
MD
857 * trying to remove it. vm_page_cache() will handle this
858 * case for us.
984263bc
MD
859 */
860 if (m->dirty == 0) {
861 vm_page_test_dirty(m);
862 } else {
863 vm_page_dirty(m);
864 }
865
984263bc 866 if (m->valid == 0) {
41a01a4d
MD
867 /*
868 * Invalid pages can be easily freed
869 */
984263bc 870 vm_pageout_page_free(m);
12e4aaff 871 mycpu->gd_cnt.v_dfree++;
20479584 872 --inactive_shortage;
984263bc
MD
873 } else if (m->dirty == 0) {
874 /*
41a01a4d
MD
875 * Clean pages can be placed onto the cache queue.
876 * This effectively frees them.
984263bc
MD
877 */
878 vm_page_cache(m);
20479584 879 --inactive_shortage;
984263bc
MD
880 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
881 /*
882 * Dirty pages need to be paged out, but flushing
883 * a page is extremely expensive verses freeing
884 * a clean page. Rather then artificially limiting
885 * the number of pages we can flush, we instead give
886 * dirty pages extra priority on the inactive queue
887 * by forcing them to be cycled through the queue
888 * twice before being flushed, after which the
889 * (now clean) page will cycle through once more
890 * before being freed. This significantly extends
891 * the thrash point for a heavily loaded machine.
892 */
984263bc
MD
893 vm_page_flag_set(m, PG_WINATCFLS);
894 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
895 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
896 } else if (maxlaunder > 0) {
897 /*
898 * We always want to try to flush some dirty pages if
899 * we encounter them, to keep the system stable.
900 * Normally this number is small, but under extreme
901 * pressure where there are insufficient clean pages
902 * on the inactive queue, we may have to go all out.
903 */
904 int swap_pageouts_ok;
905 struct vnode *vp = NULL;
906
907 object = m->object;
908
909 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
910 swap_pageouts_ok = 1;
911 } else {
912 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
913 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
20479584 914 vm_page_count_min(0));
984263bc
MD
915
916 }
917
918 /*
919 * We don't bother paging objects that are "dead".
920 * Those objects are in a "rundown" state.
921 */
922 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
984263bc
MD
923 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
924 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
925 continue;
926 }
927
928 /*
929 * The object is already known NOT to be dead. It
930 * is possible for the vget() to block the whole
931 * pageout daemon, but the new low-memory handling
932 * code should prevent it.
933 *
934 * The previous code skipped locked vnodes and, worse,
935 * reordered pages in the queue. This results in
936 * completely non-deterministic operation because,
937 * quite often, a vm_fault has initiated an I/O and
938 * is holding a locked vnode at just the point where
939 * the pageout daemon is woken up.
940 *
941 * We can't wait forever for the vnode lock, we might
942 * deadlock due to a vn_read() getting stuck in
943 * vm_wait while holding this vnode. We skip the
944 * vnode if we can't get it in a reasonable amount
945 * of time.
946 */
947
948 if (object->type == OBJT_VNODE) {
949 vp = object->handle;
950
87de5057 951 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) {
984263bc
MD
952 ++pageout_lock_miss;
953 if (object->flags & OBJ_MIGHTBEDIRTY)
954 vnodes_skipped++;
955 continue;
956 }
957
958 /*
959 * The page might have been moved to another
960 * queue during potential blocking in vget()
961 * above. The page might have been freed and
962 * reused for another vnode. The object might
963 * have been reused for another vnode.
964 */
965 if (m->queue != PQ_INACTIVE ||
966 m->object != object ||
967 object->handle != vp) {
968 if (object->flags & OBJ_MIGHTBEDIRTY)
969 vnodes_skipped++;
970 vput(vp);
971 continue;
972 }
973
974 /*
975 * The page may have been busied during the
976 * blocking in vput(); We don't move the
977 * page back onto the end of the queue so that
978 * statistics are more correct if we don't.
979 */
980 if (m->busy || (m->flags & PG_BUSY)) {
981 vput(vp);
982 continue;
983 }
984
985 /*
986 * If the page has become held it might
987 * be undergoing I/O, so skip it
988 */
989 if (m->hold_count) {
984263bc
MD
990 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
991 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
992 if (object->flags & OBJ_MIGHTBEDIRTY)
993 vnodes_skipped++;
994 vput(vp);
995 continue;
996 }
997 }
998
999 /*
1000 * If a page is dirty, then it is either being washed
1001 * (but not yet cleaned) or it is still in the
1002 * laundry. If it is still in the laundry, then we
1003 * start the cleaning operation.
1004 *
1005 * This operation may cluster, invalidating the 'next'
1006 * pointer. To prevent an inordinate number of
1007 * restarts we use our marker to remember our place.
1008 *
20479584
MD
1009 * decrement inactive_shortage on success to account
1010 * for the (future) cleaned page. Otherwise we
1011 * could wind up laundering or cleaning too many
1012 * pages.
984263bc 1013 */
984263bc 1014 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
984263bc 1015 if (vm_pageout_clean(m) != 0) {
20479584 1016 --inactive_shortage;
984263bc 1017 --maxlaunder;
c84c24da 1018 }
984263bc
MD
1019 next = TAILQ_NEXT(&marker, pageq);
1020 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
984263bc
MD
1021 if (vp != NULL)
1022 vput(vp);
1023 }
1024 }
1025
1026 /*
20479584
MD
1027 * We want to move pages from the active queue to the inactive
1028 * queue to get the inactive queue to the inactive target. If
1029 * we still have a page shortage from above we try to directly free
1030 * clean pages instead of moving them.
06ecca5a 1031 *
20479584
MD
1032 * If we do still have a shortage we keep track of the number of
1033 * pages we free or cache (recycle_count) as a measure of thrashing
1034 * between the active and inactive queues.
1035 *
51db7ca2
MD
1036 * If we were able to completely satisfy the free+cache targets
1037 * from the inactive pool we limit the number of pages we move
1038 * from the active pool to the inactive pool to 2x the pages we
1039 * had removed from the inactive pool. If we were not able to
1040 * completel satisfy the free+cache targets we go for the whole
1041 * target aggressively.
20479584
MD
1042 *
1043 * NOTE: Both variables can end up negative.
1044 * NOTE: We are still in a critical section.
984263bc 1045 */
20479584 1046 active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count;
51db7ca2
MD
1047 if (inactive_shortage <= 0 &&
1048 active_shortage > inactive_original_shortage * 2) {
1049 active_shortage = inactive_original_shortage * 2;
1050 }
20479584 1051
12e4aaff 1052 pcount = vmstats.v_active_count;
20479584 1053 recycle_count = 0;
984263bc
MD
1054 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1055
20479584
MD
1056 while ((m != NULL) && (pcount-- > 0) &&
1057 (inactive_shortage > 0 || active_shortage > 0)
1058 ) {
06ecca5a
MD
1059 /*
1060 * Give interrupts a chance.
1061 */
5fd012e0
MD
1062 crit_exit();
1063 crit_enter();
984263bc
MD
1064
1065 /*
06ecca5a 1066 * If the page was ripped out from under us, just stop.
984263bc 1067 */
06ecca5a 1068 if (m->queue != PQ_ACTIVE)
984263bc 1069 break;
984263bc 1070 next = TAILQ_NEXT(m, pageq);
06ecca5a 1071
984263bc
MD
1072 /*
1073 * Don't deactivate pages that are busy.
1074 */
1075 if ((m->busy != 0) ||
1076 (m->flags & PG_BUSY) ||
1077 (m->hold_count != 0)) {
984263bc
MD
1078 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1079 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1080 m = next;
1081 continue;
1082 }
1083
1084 /*
1085 * The count for pagedaemon pages is done after checking the
1086 * page for eligibility...
1087 */
12e4aaff 1088 mycpu->gd_cnt.v_pdpages++;
984263bc
MD
1089
1090 /*
20479584
MD
1091 * Check to see "how much" the page has been used and clear
1092 * the tracking access bits. If the object has no references
1093 * don't bother paying the expense.
984263bc
MD
1094 */
1095 actcount = 0;
1096 if (m->object->ref_count != 0) {
20479584
MD
1097 if (m->flags & PG_REFERENCED)
1098 ++actcount;
984263bc
MD
1099 actcount += pmap_ts_referenced(m);
1100 if (actcount) {
1101 m->act_count += ACT_ADVANCE + actcount;
1102 if (m->act_count > ACT_MAX)
1103 m->act_count = ACT_MAX;
1104 }
1105 }
984263bc
MD
1106 vm_page_flag_clear(m, PG_REFERENCED);
1107
1108 /*
20479584 1109 * actcount is only valid if the object ref_count is non-zero.
984263bc 1110 */
20479584 1111 if (actcount && m->object->ref_count != 0) {
984263bc
MD
1112 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1113 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1114 } else {
1115 m->act_count -= min(m->act_count, ACT_DECLINE);
1116 if (vm_pageout_algorithm ||
1117 m->object->ref_count == 0 ||
20479584
MD
1118 m->act_count < pass + 1
1119 ) {
1120 /*
1121 * Deactivate the page. If we had a
1122 * shortage from our inactive scan try to
1123 * free (cache) the page instead.
1124 */
1125 --active_shortage;
1126 if (inactive_shortage > 0 ||
1127 m->object->ref_count == 0) {
1128 if (inactive_shortage > 0)
1129 ++recycle_count;
17cde63e 1130 vm_page_busy(m);
984263bc 1131 vm_page_protect(m, VM_PROT_NONE);
17cde63e 1132 vm_page_wakeup(m);
c84c24da 1133 if (m->dirty == 0) {
20479584 1134 --inactive_shortage;
984263bc 1135 vm_page_cache(m);
c84c24da 1136 } else {
984263bc 1137 vm_page_deactivate(m);
c84c24da 1138 }
984263bc
MD
1139 } else {
1140 vm_page_deactivate(m);
1141 }
1142 } else {
984263bc
MD
1143 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1144 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1145 }
1146 }
1147 m = next;
1148 }
1149
984263bc
MD
1150 /*
1151 * We try to maintain some *really* free pages, this allows interrupt
1152 * code to be guaranteed space. Since both cache and free queues
1153 * are considered basically 'free', moving pages from cache to free
1154 * does not effect other calculations.
06ecca5a 1155 *
5fd012e0 1156 * NOTE: we are still in a critical section.
c84c24da
MD
1157 *
1158 * Pages moved from PQ_CACHE to totally free are not counted in the
1159 * pages_freed counter.
984263bc 1160 */
12e4aaff 1161 while (vmstats.v_free_count < vmstats.v_free_reserved) {
984263bc
MD
1162 static int cache_rover = 0;
1163 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
20479584 1164 if (m == NULL)
984263bc
MD
1165 break;
1166 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1167 m->busy ||
1168 m->hold_count ||
1169 m->wire_count) {
1170#ifdef INVARIANTS
086c1d7e 1171 kprintf("Warning: busy page %p found in cache\n", m);
984263bc
MD
1172#endif
1173 vm_page_deactivate(m);
1174 continue;
1175 }
17cde63e
MD
1176 KKASSERT((m->flags & PG_MAPPED) == 0);
1177 KKASSERT(m->dirty == 0);
984263bc
MD
1178 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1179 vm_pageout_page_free(m);
12e4aaff 1180 mycpu->gd_cnt.v_dfree++;
984263bc 1181 }
06ecca5a 1182
5fd012e0 1183 crit_exit();
984263bc
MD
1184
1185#if !defined(NO_SWAPPING)
1186 /*
1187 * Idle process swapout -- run once per second.
1188 */
1189 if (vm_swap_idle_enabled) {
1190 static long lsec;
1191 if (time_second != lsec) {
1192 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1193 vm_req_vmdaemon();
1194 lsec = time_second;
1195 }
1196 }
1197#endif
1198
1199 /*
1200 * If we didn't get enough free pages, and we have skipped a vnode
1201 * in a writeable object, wakeup the sync daemon. And kick swapout
1202 * if we did not get enough free pages.
1203 */
1204 if (vm_paging_target() > 0) {
20479584 1205 if (vnodes_skipped && vm_page_count_min(0))
418ff780 1206 speedup_syncer();
984263bc
MD
1207#if !defined(NO_SWAPPING)
1208 if (vm_swap_enabled && vm_page_count_target()) {
1209 vm_req_vmdaemon();
1210 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1211 }
1212#endif
1213 }
1214
1215 /*
20479584
MD
1216 * Handle catastrophic conditions. Under good conditions we should
1217 * be at the target, well beyond our minimum. If we could not even
1218 * reach our minimum the system is under heavy stress.
1219 *
1220 * Determine whether we have run out of memory. This occurs when
1221 * swap_pager_full is TRUE and the only pages left in the page
1222 * queues are dirty. We will still likely have page shortages.
c84c24da
MD
1223 *
1224 * - swap_pager_full is set if insufficient swap was
1225 * available to satisfy a requested pageout.
1226 *
20479584
MD
1227 * - the inactive queue is bloated (4 x size of active queue),
1228 * meaning it is unable to get rid of dirty pages and.
c84c24da 1229 *
20479584
MD
1230 * - vm_page_count_min() without counting pages recycled from the
1231 * active queue (recycle_count) means we could not recover
1232 * enough pages to meet bare minimum needs. This test only
1233 * works if the inactive queue is bloated.
c84c24da 1234 *
20479584
MD
1235 * - due to a positive inactive_shortage we shifted the remaining
1236 * dirty pages from the active queue to the inactive queue
1237 * trying to find clean ones to free.
984263bc 1238 */
20479584 1239 if (swap_pager_full && vm_page_count_min(recycle_count))
c84c24da 1240 kprintf("Warning: system low on memory+swap!\n");
20479584
MD
1241 if (swap_pager_full && vm_page_count_min(recycle_count) &&
1242 vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1243 inactive_shortage > 0) {
1244 /*
1245 * Kill something.
1246 */
8fa76237
MD
1247 info.bigproc = NULL;
1248 info.bigsize = 0;
1249 allproc_scan(vm_pageout_scan_callback, &info);
1250 if (info.bigproc != NULL) {
1251 killproc(info.bigproc, "out of swap space");
1252 info.bigproc->p_nice = PRIO_MIN;
08f2f1bb
SS
1253 info.bigproc->p_usched->resetpriority(
1254 FIRST_LWP_IN_PROC(info.bigproc));
12e4aaff 1255 wakeup(&vmstats.v_free_count);
8fa76237 1256 PRELE(info.bigproc);
984263bc
MD
1257 }
1258 }
20479584 1259 return(inactive_shortage);
984263bc
MD
1260}
1261
8fa76237
MD
1262static int
1263vm_pageout_scan_callback(struct proc *p, void *data)
1264{
1265 struct vm_pageout_scan_info *info = data;
1266 vm_offset_t size;
1267
1268 /*
20479584
MD
1269 * Never kill system processes or init. If we have configured swap
1270 * then try to avoid killing low-numbered pids.
8fa76237
MD
1271 */
1272 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1273 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1274 return (0);
1275 }
1276
1277 /*
1278 * if the process is in a non-running type state,
1279 * don't touch it.
1280 */
20479584 1281 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
8fa76237 1282 return (0);
8fa76237
MD
1283
1284 /*
20479584
MD
1285 * Get the approximate process size. Note that anonymous pages
1286 * with backing swap will be counted twice, but there should not
1287 * be too many such pages due to the stress the VM system is
1288 * under at this point.
8fa76237 1289 */
20479584 1290 size = vmspace_anonymous_count(p->p_vmspace) +
8fa76237
MD
1291 vmspace_swap_count(p->p_vmspace);
1292
1293 /*
1294 * If the this process is bigger than the biggest one
1295 * remember it.
1296 */
20479584 1297 if (info->bigsize < size) {
8fa76237
MD
1298 if (info->bigproc)
1299 PRELE(info->bigproc);
1300 PHOLD(p);
1301 info->bigproc = p;
1302 info->bigsize = size;
1303 }
1304 return(0);
1305}
1306
984263bc
MD
1307/*
1308 * This routine tries to maintain the pseudo LRU active queue,
1309 * so that during long periods of time where there is no paging,
1310 * that some statistic accumulation still occurs. This code
1311 * helps the situation where paging just starts to occur.
1312 */
1313static void
57e43348 1314vm_pageout_page_stats(void)
984263bc 1315{
984263bc
MD
1316 vm_page_t m,next;
1317 int pcount,tpcount; /* Number of pages to check */
1318 static int fullintervalcount = 0;
1319 int page_shortage;
984263bc
MD
1320
1321 page_shortage =
12e4aaff
MD
1322 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1323 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
984263bc
MD
1324
1325 if (page_shortage <= 0)
1326 return;
1327
5fd012e0 1328 crit_enter();
984263bc 1329
12e4aaff 1330 pcount = vmstats.v_active_count;
984263bc
MD
1331 fullintervalcount += vm_pageout_stats_interval;
1332 if (fullintervalcount < vm_pageout_full_stats_interval) {
12e4aaff 1333 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
984263bc
MD
1334 if (pcount > tpcount)
1335 pcount = tpcount;
1336 } else {
1337 fullintervalcount = 0;
1338 }
1339
1340 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1341 while ((m != NULL) && (pcount-- > 0)) {
1342 int actcount;
1343
1344 if (m->queue != PQ_ACTIVE) {
1345 break;
1346 }
1347
1348 next = TAILQ_NEXT(m, pageq);
1349 /*
1350 * Don't deactivate pages that are busy.
1351 */
1352 if ((m->busy != 0) ||
1353 (m->flags & PG_BUSY) ||
1354 (m->hold_count != 0)) {
984263bc
MD
1355 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1356 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1357 m = next;
1358 continue;
1359 }
1360
1361 actcount = 0;
1362 if (m->flags & PG_REFERENCED) {
1363 vm_page_flag_clear(m, PG_REFERENCED);
1364 actcount += 1;
1365 }
1366
1367 actcount += pmap_ts_referenced(m);
1368 if (actcount) {
1369 m->act_count += ACT_ADVANCE + actcount;
1370 if (m->act_count > ACT_MAX)
1371 m->act_count = ACT_MAX;
984263bc
MD
1372 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1373 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1374 } else {
1375 if (m->act_count == 0) {
1376 /*
1377 * We turn off page access, so that we have
1378 * more accurate RSS stats. We don't do this
1379 * in the normal page deactivation when the
1380 * system is loaded VM wise, because the
1381 * cost of the large number of page protect
1382 * operations would be higher than the value
1383 * of doing the operation.
1384 */
17cde63e 1385 vm_page_busy(m);
984263bc 1386 vm_page_protect(m, VM_PROT_NONE);
17cde63e 1387 vm_page_wakeup(m);
984263bc
MD
1388 vm_page_deactivate(m);
1389 } else {
1390 m->act_count -= min(m->act_count, ACT_DECLINE);
984263bc
MD
1391 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1392 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1393 }
1394 }
1395
1396 m = next;
1397 }
5fd012e0 1398 crit_exit();
984263bc
MD
1399}
1400
1401static int
57e43348 1402vm_pageout_free_page_calc(vm_size_t count)
984263bc 1403{
12e4aaff 1404 if (count < vmstats.v_page_count)
984263bc
MD
1405 return 0;
1406 /*
1407 * free_reserved needs to include enough for the largest swap pager
1408 * structures plus enough for any pv_entry structs when paging.
1409 */
12e4aaff
MD
1410 if (vmstats.v_page_count > 1024)
1411 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
984263bc 1412 else
12e4aaff
MD
1413 vmstats.v_free_min = 4;
1414 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1415 vmstats.v_interrupt_free_min;
1416 vmstats.v_free_reserved = vm_pageout_page_count +
1417 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1418 vmstats.v_free_severe = vmstats.v_free_min / 2;
1419 vmstats.v_free_min += vmstats.v_free_reserved;
1420 vmstats.v_free_severe += vmstats.v_free_reserved;
984263bc
MD
1421 return 1;
1422}
1423
1424
1425/*
20479584 1426 * vm_pageout is the high level pageout daemon.
984263bc
MD
1427 */
1428static void
57e43348 1429vm_pageout(void)
984263bc
MD
1430{
1431 int pass;
20479584 1432 int inactive_shortage;
984263bc
MD
1433
1434 /*
1435 * Initialize some paging parameters.
1436 */
4ecf7cc9 1437 curthread->td_flags |= TDF_SYSTHREAD;
984263bc 1438
12e4aaff
MD
1439 vmstats.v_interrupt_free_min = 2;
1440 if (vmstats.v_page_count < 2000)
984263bc
MD
1441 vm_pageout_page_count = 8;
1442
12e4aaff 1443 vm_pageout_free_page_calc(vmstats.v_page_count);
20479584 1444
984263bc
MD
1445 /*
1446 * v_free_target and v_cache_min control pageout hysteresis. Note
1447 * that these are more a measure of the VM cache queue hysteresis
1448 * then the VM free queue. Specifically, v_free_target is the
1449 * high water mark (free+cache pages).
1450 *
1451 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1452 * low water mark, while v_free_min is the stop. v_cache_min must
1453 * be big enough to handle memory needs while the pageout daemon
1454 * is signalled and run to free more pages.
1455 */
12e4aaff
MD
1456 if (vmstats.v_free_count > 6144)
1457 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
984263bc 1458 else
12e4aaff 1459 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
984263bc 1460
0e8bd897
MD
1461 /*
1462 * NOTE: With the new buffer cache b_act_count we want the default
1463 * inactive target to be a percentage of available memory.
1464 *
1465 * The inactive target essentially determines the minimum
1466 * number of 'temporary' pages capable of caching one-time-use
1467 * files when the VM system is otherwise full of pages
1468 * belonging to multi-time-use files or active program data.
51db7ca2
MD
1469 *
1470 * NOTE: The inactive target is aggressively persued only if the
1471 * inactive queue becomes too small. If the inactive queue
1472 * is large enough to satisfy page movement to free+cache
1473 * then it is repopulated more slowly from the active queue.
1474 * This allows a generate inactive_target default to be set.
1475 *
1476 * There is an issue here for processes which sit mostly idle
1477 * 'overnight', such as sshd, tcsh, and X. Any movement from
1478 * the active queue will eventually cause such pages to
1479 * recycle eventually causing a lot of paging in the morning.
1480 * To reduce the incidence of this pages cycled out of the
1481 * buffer cache are moved directly to the inactive queue if
1482 * they were only used once or twice. The vfs.vm_cycle_point
1483 * sysctl can be used to adjust this.
0e8bd897 1484 */
12e4aaff
MD
1485 if (vmstats.v_free_count > 2048) {
1486 vmstats.v_cache_min = vmstats.v_free_target;
1487 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
984263bc 1488 } else {
12e4aaff
MD
1489 vmstats.v_cache_min = 0;
1490 vmstats.v_cache_max = 0;
984263bc 1491 }
51db7ca2 1492 vmstats.v_inactive_target = vmstats.v_free_count / 2;
984263bc
MD
1493
1494 /* XXX does not really belong here */
1495 if (vm_page_max_wired == 0)
12e4aaff 1496 vm_page_max_wired = vmstats.v_free_count / 3;
984263bc
MD
1497
1498 if (vm_pageout_stats_max == 0)
12e4aaff 1499 vm_pageout_stats_max = vmstats.v_free_target;
984263bc
MD
1500
1501 /*
1502 * Set interval in seconds for stats scan.
1503 */
1504 if (vm_pageout_stats_interval == 0)
1505 vm_pageout_stats_interval = 5;
1506 if (vm_pageout_full_stats_interval == 0)
1507 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1508
1509
1510 /*
1511 * Set maximum free per pass
1512 */
1513 if (vm_pageout_stats_free_max == 0)
1514 vm_pageout_stats_free_max = 5;
1515
1516 swap_pager_swap_init();
1517 pass = 0;
20479584 1518
984263bc
MD
1519 /*
1520 * The pageout daemon is never done, so loop forever.
1521 */
1522 while (TRUE) {
1523 int error;
984263bc 1524
12d8aca7
MD
1525 /*
1526 * Wait for an action request
1527 */
1528 crit_enter();
20479584 1529 if (vm_pages_needed == 0) {
984263bc 1530 error = tsleep(&vm_pages_needed,
20479584
MD
1531 0, "psleep",
1532 vm_pageout_stats_interval * hz);
1533 if (error && vm_pages_needed == 0) {
984263bc
MD
1534 vm_pageout_page_stats();
1535 continue;
1536 }
20479584 1537 vm_pages_needed = 1;
984263bc 1538 }
12d8aca7 1539 crit_exit();
984263bc 1540
20479584
MD
1541 /*
1542 * If we have enough free memory, wakeup waiters.
12d8aca7 1543 * (This is optional here)
20479584
MD
1544 */
1545 crit_enter();
1546 if (!vm_page_count_min(0))
1547 wakeup(&vmstats.v_free_count);
1548 mycpu->gd_cnt.v_pdwakeups++;
5fd012e0 1549 crit_exit();
20479584
MD
1550
1551 /*
12d8aca7
MD
1552 * Scan for pageout. Try to avoid thrashing the system
1553 * with activity.
20479584 1554 */
12d8aca7 1555 inactive_shortage = vm_pageout_scan(pass);
20479584
MD
1556 if (inactive_shortage > 0) {
1557 ++pass;
1558 if (swap_pager_full) {
1559 /*
1560 * Running out of memory, catastrophic back-off
1561 * to one-second intervals.
1562 */
1563 tsleep(&vm_pages_needed, 0, "pdelay", hz);
1564 } else if (pass < 10 && vm_pages_needed > 1) {
1565 /*
1566 * Normal operation, additional processes
1567 * have already kicked us. Retry immediately.
1568 */
1569 } else if (pass < 10) {
1570 /*
1571 * Normal operation, fewer processes. Delay
1572 * a bit but allow wakeups.
1573 */
1574 vm_pages_needed = 0;
1575 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1576 vm_pages_needed = 1;
1577 } else {
1578 /*
1579 * We've taken too many passes, forced delay.
1580 */
1581 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1582 }
1583 } else {
12d8aca7
MD
1584 /*
1585 * Interlocked wakeup of waiters (non-optional)
1586 */
20479584 1587 pass = 0;
12d8aca7
MD
1588 if (vm_pages_needed && !vm_page_count_min(0)) {
1589 wakeup(&vmstats.v_free_count);
1590 vm_pages_needed = 0;
1591 }
20479584 1592 }
984263bc
MD
1593 }
1594}
1595
20479584
MD
1596/*
1597 * Called after allocating a page out of the cache or free queue
1598 * to possibly wake the pagedaemon up to replentish our supply.
1599 *
1600 * We try to generate some hysteresis by waking the pagedaemon up
1601 * when our free+cache pages go below the severe level. The pagedaemon
1602 * tries to get the count back up to at least the minimum, and through
1603 * to the target level if possible.
1604 *
1605 * If the pagedaemon is already active bump vm_pages_needed as a hint
1606 * that there are even more requests pending.
1607 */
984263bc 1608void
57e43348 1609pagedaemon_wakeup(void)
984263bc 1610{
20479584
MD
1611 if (vm_page_count_severe() && curthread != pagethread) {
1612 if (vm_pages_needed == 0) {
1613 vm_pages_needed = 1;
1614 wakeup(&vm_pages_needed);
1615 } else if (vm_page_count_min(0)) {
1616 ++vm_pages_needed;
1617 }
984263bc
MD
1618 }
1619}
1620
1621#if !defined(NO_SWAPPING)
1622static void
57e43348 1623vm_req_vmdaemon(void)
984263bc
MD
1624{
1625 static int lastrun = 0;
1626
1627 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1628 wakeup(&vm_daemon_needed);
1629 lastrun = ticks;
1630 }
1631}
1632
8fa76237
MD
1633static int vm_daemon_callback(struct proc *p, void *data __unused);
1634
984263bc 1635static void
57e43348 1636vm_daemon(void)
984263bc 1637{
984263bc 1638 while (TRUE) {
377d4740 1639 tsleep(&vm_daemon_needed, 0, "psleep", 0);
984263bc
MD
1640 if (vm_pageout_req_swapout) {
1641 swapout_procs(vm_pageout_req_swapout);
1642 vm_pageout_req_swapout = 0;
1643 }
1644 /*
1645 * scan the processes for exceeding their rlimits or if
1646 * process is swapped out -- deactivate pages
1647 */
8fa76237
MD
1648 allproc_scan(vm_daemon_callback, NULL);
1649 }
1650}
984263bc 1651
8fa76237
MD
1652static int
1653vm_daemon_callback(struct proc *p, void *data __unused)
1654{
1655 vm_pindex_t limit, size;
984263bc 1656
8fa76237
MD
1657 /*
1658 * if this is a system process or if we have already
1659 * looked at this process, skip it.
1660 */
1661 if (p->p_flag & (P_SYSTEM | P_WEXIT))
1662 return (0);
984263bc 1663
8fa76237
MD
1664 /*
1665 * if the process is in a non-running type state,
1666 * don't touch it.
1667 */
164b8401 1668 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
8fa76237 1669 return (0);
984263bc 1670
8fa76237
MD
1671 /*
1672 * get a limit
1673 */
1674 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1675 p->p_rlimit[RLIMIT_RSS].rlim_max));
1676
1677 /*
1678 * let processes that are swapped out really be
1679 * swapped out. Set the limit to nothing to get as
1680 * many pages out to swap as possible.
1681 */
1682 if (p->p_flag & P_SWAPPEDOUT)
1683 limit = 0;
1684
1685 size = vmspace_resident_count(p->p_vmspace);
1686 if (limit >= 0 && size >= limit) {
1687 vm_pageout_map_deactivate_pages(
1688 &p->p_vmspace->vm_map, limit);
984263bc 1689 }
8fa76237 1690 return (0);
984263bc 1691}
8fa76237 1692
984263bc 1693#endif