This is a major revamping of the pageout and low-memory handling code.
[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
142extern int vm_swap_size;
143static int vm_max_launder = 32;
144static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
145static int vm_pageout_full_stats_interval = 0;
146static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
147static int defer_swap_pageouts=0;
148static int disable_swap_pageouts=0;
149
150#if defined(NO_SWAPPING)
151static int vm_swap_enabled=0;
152static int vm_swap_idle_enabled=0;
153#else
154static int vm_swap_enabled=1;
155static int vm_swap_idle_enabled=0;
156#endif
157
158SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
159 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
160
161SYSCTL_INT(_vm, OID_AUTO, max_launder,
162 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
163
164SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
165 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
166
167SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
168 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
169
170SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
171 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
172
173SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
174 CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
175
176#if defined(NO_SWAPPING)
177SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
178 CTLFLAG_RD, &vm_swap_enabled, 0, "");
179SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
180 CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
181#else
182SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
183 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
184SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
185 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
186#endif
187
188SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
189 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
190
191SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
192 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
193
194static int pageout_lock_miss;
195SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
196 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
197
46311ac2
MD
198int vm_load;
199SYSCTL_INT(_vm, OID_AUTO, vm_load,
200 CTLFLAG_RD, &vm_load, 0, "load on the VM system");
201int vm_load_enable = 1;
202SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
203 CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
204#ifdef INVARIANTS
205int vm_load_debug;
206SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
207 CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
208#endif
209
984263bc
MD
210#define VM_PAGEOUT_PAGE_COUNT 16
211int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
212
213int vm_page_max_wired; /* XXX max # of wired pages system-wide */
214
215#if !defined(NO_SWAPPING)
1388df65
RG
216typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
217static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
984263bc 218static freeer_fcn_t vm_pageout_object_deactivate_pages;
1388df65 219static void vm_req_vmdaemon (void);
984263bc
MD
220#endif
221static void vm_pageout_page_stats(void);
222
223/*
20479584 224 * Update vm_load to slow down faulting processes.
46311ac2
MD
225 */
226void
227vm_fault_ratecheck(void)
228{
229 if (vm_pages_needed) {
230 if (vm_load < 1000)
231 ++vm_load;
232 } else {
233 if (vm_load > 0)
234 --vm_load;
235 }
236}
237
238/*
984263bc
MD
239 * vm_pageout_clean:
240 *
06ecca5a
MD
241 * Clean the page and remove it from the laundry. The page must not be
242 * busy on-call.
984263bc
MD
243 *
244 * We set the busy bit to cause potential page faults on this page to
245 * block. Note the careful timing, however, the busy bit isn't set till
246 * late and we cannot do anything that will mess with the page.
247 */
248
249static int
57e43348 250vm_pageout_clean(vm_page_t m)
984263bc 251{
5f910b2f 252 vm_object_t object;
984263bc
MD
253 vm_page_t mc[2*vm_pageout_page_count];
254 int pageout_count;
255 int ib, is, page_base;
256 vm_pindex_t pindex = m->pindex;
257
258 object = m->object;
259
260 /*
261 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
262 * with the new swapper, but we could have serious problems paging
263 * out other object types if there is insufficient memory.
264 *
265 * Unfortunately, checking free memory here is far too late, so the
266 * check has been moved up a procedural level.
267 */
268
269 /*
270 * Don't mess with the page if it's busy, held, or special
271 */
272 if ((m->hold_count != 0) ||
273 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
274 return 0;
275 }
276
277 mc[vm_pageout_page_count] = m;
278 pageout_count = 1;
279 page_base = vm_pageout_page_count;
280 ib = 1;
281 is = 1;
282
283 /*
284 * Scan object for clusterable pages.
285 *
286 * We can cluster ONLY if: ->> the page is NOT
287 * clean, wired, busy, held, or mapped into a
288 * buffer, and one of the following:
289 * 1) The page is inactive, or a seldom used
290 * active page.
291 * -or-
292 * 2) we force the issue.
293 *
294 * During heavy mmap/modification loads the pageout
295 * daemon can really fragment the underlying file
296 * due to flushing pages out of order and not trying
297 * align the clusters (which leave sporatic out-of-order
298 * holes). To solve this problem we do the reverse scan
299 * first and attempt to align our cluster, then do a
300 * forward scan if room remains.
301 */
302
303more:
304 while (ib && pageout_count < vm_pageout_page_count) {
305 vm_page_t p;
306
307 if (ib > pindex) {
308 ib = 0;
309 break;
310 }
311
312 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
313 ib = 0;
314 break;
315 }
316 if (((p->queue - p->pc) == PQ_CACHE) ||
317 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
318 ib = 0;
319 break;
320 }
321 vm_page_test_dirty(p);
322 if ((p->dirty & p->valid) == 0 ||
323 p->queue != PQ_INACTIVE ||
324 p->wire_count != 0 || /* may be held by buf cache */
325 p->hold_count != 0) { /* may be undergoing I/O */
326 ib = 0;
327 break;
328 }
329 mc[--page_base] = p;
330 ++pageout_count;
331 ++ib;
332 /*
333 * alignment boundry, stop here and switch directions. Do
334 * not clear ib.
335 */
336 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
337 break;
338 }
339
340 while (pageout_count < vm_pageout_page_count &&
341 pindex + is < object->size) {
342 vm_page_t p;
343
344 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
345 break;
346 if (((p->queue - p->pc) == PQ_CACHE) ||
347 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
348 break;
349 }
350 vm_page_test_dirty(p);
351 if ((p->dirty & p->valid) == 0 ||
352 p->queue != PQ_INACTIVE ||
353 p->wire_count != 0 || /* may be held by buf cache */
354 p->hold_count != 0) { /* may be undergoing I/O */
355 break;
356 }
357 mc[page_base + pageout_count] = p;
358 ++pageout_count;
359 ++is;
360 }
361
362 /*
363 * If we exhausted our forward scan, continue with the reverse scan
364 * when possible, even past a page boundry. This catches boundry
365 * conditions.
366 */
367 if (ib && pageout_count < vm_pageout_page_count)
368 goto more;
369
370 /*
371 * we allow reads during pageouts...
372 */
373 return vm_pageout_flush(&mc[page_base], pageout_count, 0);
374}
375
376/*
377 * vm_pageout_flush() - launder the given pages
378 *
379 * The given pages are laundered. Note that we setup for the start of
380 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
381 * reference count all in here rather then in the parent. If we want
382 * the parent to do more sophisticated things we may have to change
383 * the ordering.
384 */
385
386int
57e43348 387vm_pageout_flush(vm_page_t *mc, int count, int flags)
984263bc 388{
5f910b2f 389 vm_object_t object;
984263bc
MD
390 int pageout_status[count];
391 int numpagedout = 0;
392 int i;
393
394 /*
17cde63e
MD
395 * Initiate I/O. Bump the vm_page_t->busy counter.
396 */
397 for (i = 0; i < count; i++) {
398 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
399 vm_page_io_start(mc[i]);
400 }
401
402 /*
4530a3aa
MD
403 * We must make the pages read-only. This will also force the
404 * modified bit in the related pmaps to be cleared. The pager
405 * cannot clear the bit for us since the I/O completion code
406 * typically runs from an interrupt. The act of making the page
407 * read-only handles the case for us.
984263bc 408 */
984263bc 409 for (i = 0; i < count; i++) {
984263bc
MD
410 vm_page_protect(mc[i], VM_PROT_READ);
411 }
412
413 object = mc[0]->object;
414 vm_object_pip_add(object, count);
415
416 vm_pager_put_pages(object, mc, count,
c439ad8f 417 (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
984263bc
MD
418 pageout_status);
419
420 for (i = 0; i < count; i++) {
421 vm_page_t mt = mc[i];
422
423 switch (pageout_status[i]) {
424 case VM_PAGER_OK:
425 numpagedout++;
426 break;
427 case VM_PAGER_PEND:
428 numpagedout++;
429 break;
430 case VM_PAGER_BAD:
431 /*
432 * Page outside of range of object. Right now we
433 * essentially lose the changes by pretending it
434 * worked.
435 */
436 pmap_clear_modify(mt);
437 vm_page_undirty(mt);
438 break;
439 case VM_PAGER_ERROR:
440 case VM_PAGER_FAIL:
441 /*
c84c24da
MD
442 * A page typically cannot be paged out when we
443 * have run out of swap. We leave the page
444 * marked inactive and will try to page it out
445 * again later.
446 *
447 * Starvation of the active page list is used to
448 * determine when the system is massively memory
449 * starved.
984263bc 450 */
984263bc
MD
451 break;
452 case VM_PAGER_AGAIN:
453 break;
454 }
455
456 /*
457 * If the operation is still going, leave the page busy to
458 * block all other accesses. Also, leave the paging in
459 * progress indicator set so that we don't attempt an object
460 * collapse.
93afe6be
MD
461 *
462 * For any pages which have completed synchronously,
463 * deactivate the page if we are under a severe deficit.
464 * Do not try to enter them into the cache, though, they
465 * might still be read-heavy.
984263bc
MD
466 */
467 if (pageout_status[i] != VM_PAGER_PEND) {
468 vm_object_pip_wakeup(object);
469 vm_page_io_finish(mt);
93afe6be
MD
470 if (vm_page_count_severe())
471 vm_page_deactivate(mt);
472#if 0
984263bc
MD
473 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
474 vm_page_protect(mt, VM_PROT_READ);
93afe6be 475#endif
984263bc
MD
476 }
477 }
478 return numpagedout;
479}
480
481#if !defined(NO_SWAPPING)
482/*
483 * vm_pageout_object_deactivate_pages
484 *
485 * deactivate enough pages to satisfy the inactive target
486 * requirements or if vm_page_proc_limit is set, then
487 * deactivate all of the pages in the object and its
488 * backing_objects.
489 *
490 * The object and map must be locked.
491 */
1f804340
MD
492static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
493
984263bc 494static void
57e43348 495vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
06ecca5a 496 vm_pindex_t desired, int map_remove_only)
984263bc 497{
1f804340 498 struct rb_vm_page_scan_info info;
984263bc 499 int remove_mode;
984263bc
MD
500
501 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
502 return;
503
504 while (object) {
505 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
506 return;
507 if (object->paging_in_progress)
508 return;
509
510 remove_mode = map_remove_only;
511 if (object->shadow_count > 1)
512 remove_mode = 1;
06ecca5a
MD
513
514 /*
515 * scan the objects entire memory queue. spl protection is
516 * required to avoid an interrupt unbusy/free race against
517 * our busy check.
518 */
5fd012e0 519 crit_enter();
1f804340
MD
520 info.limit = remove_mode;
521 info.map = map;
522 info.desired = desired;
523 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
524 vm_pageout_object_deactivate_pages_callback,
525 &info
526 );
527 crit_exit();
528 object = object->backing_object;
529 }
530}
531
532static int
533vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
534{
535 struct rb_vm_page_scan_info *info = data;
536 int actcount;
984263bc 537
1f804340
MD
538 if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
539 return(-1);
540 }
541 mycpu->gd_cnt.v_pdpages++;
542 if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 ||
543 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
544 !pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
545 return(0);
546 }
984263bc 547
1f804340
MD
548 actcount = pmap_ts_referenced(p);
549 if (actcount) {
550 vm_page_flag_set(p, PG_REFERENCED);
551 } else if (p->flags & PG_REFERENCED) {
552 actcount = 1;
553 }
554
555 if ((p->queue != PQ_ACTIVE) &&
556 (p->flags & PG_REFERENCED)) {
557 vm_page_activate(p);
558 p->act_count += actcount;
559 vm_page_flag_clear(p, PG_REFERENCED);
560 } else if (p->queue == PQ_ACTIVE) {
561 if ((p->flags & PG_REFERENCED) == 0) {
562 p->act_count -= min(p->act_count, ACT_DECLINE);
563 if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) {
17cde63e 564 vm_page_busy(p);
984263bc 565 vm_page_protect(p, VM_PROT_NONE);
17cde63e 566 vm_page_wakeup(p);
1f804340
MD
567 vm_page_deactivate(p);
568 } else {
569 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
570 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
984263bc 571 }
1f804340
MD
572 } else {
573 vm_page_activate(p);
574 vm_page_flag_clear(p, PG_REFERENCED);
575 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
576 p->act_count += ACT_ADVANCE;
577 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
578 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
984263bc 579 }
1f804340 580 } else if (p->queue == PQ_INACTIVE) {
17cde63e 581 vm_page_busy(p);
1f804340 582 vm_page_protect(p, VM_PROT_NONE);
17cde63e 583 vm_page_wakeup(p);
984263bc 584 }
1f804340 585 return(0);
984263bc
MD
586}
587
588/*
589 * deactivate some number of pages in a map, try to do it fairly, but
590 * that is really hard to do.
591 */
592static void
57e43348 593vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
984263bc
MD
594{
595 vm_map_entry_t tmpe;
596 vm_object_t obj, bigobj;
597 int nothingwired;
598
df4f70a6 599 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
984263bc
MD
600 return;
601 }
602
603 bigobj = NULL;
604 nothingwired = TRUE;
605
606 /*
607 * first, search out the biggest object, and try to free pages from
608 * that.
609 */
610 tmpe = map->header.next;
611 while (tmpe != &map->header) {
1b874851
MD
612 switch(tmpe->maptype) {
613 case VM_MAPTYPE_NORMAL:
614 case VM_MAPTYPE_VPAGETABLE:
984263bc
MD
615 obj = tmpe->object.vm_object;
616 if ((obj != NULL) && (obj->shadow_count <= 1) &&
617 ((bigobj == NULL) ||
618 (bigobj->resident_page_count < obj->resident_page_count))) {
619 bigobj = obj;
620 }
1b874851
MD
621 break;
622 default:
623 break;
984263bc
MD
624 }
625 if (tmpe->wired_count > 0)
626 nothingwired = FALSE;
627 tmpe = tmpe->next;
628 }
629
630 if (bigobj)
631 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
632
633 /*
634 * Next, hunt around for other pages to deactivate. We actually
635 * do this search sort of wrong -- .text first is not the best idea.
636 */
637 tmpe = map->header.next;
638 while (tmpe != &map->header) {
639 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
640 break;
1b874851
MD
641 switch(tmpe->maptype) {
642 case VM_MAPTYPE_NORMAL:
643 case VM_MAPTYPE_VPAGETABLE:
984263bc
MD
644 obj = tmpe->object.vm_object;
645 if (obj)
646 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
1b874851
MD
647 break;
648 default:
649 break;
984263bc
MD
650 }
651 tmpe = tmpe->next;
652 };
653
654 /*
655 * Remove all mappings if a process is swapped out, this will free page
656 * table pages.
657 */
658 if (desired == 0 && nothingwired)
659 pmap_remove(vm_map_pmap(map),
88181b08 660 VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
984263bc 661 vm_map_unlock(map);
984263bc
MD
662}
663#endif
664
665/*
a11aaa81
MD
666 * Don't try to be fancy - being fancy can lead to vnode deadlocks. We
667 * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
668 * be trivially freed.
984263bc 669 */
984263bc 670void
95813af0
MD
671vm_pageout_page_free(vm_page_t m)
672{
984263bc
MD
673 vm_object_t object = m->object;
674 int type = object->type;
675
676 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
677 vm_object_reference(object);
678 vm_page_busy(m);
679 vm_page_protect(m, VM_PROT_NONE);
680 vm_page_free(m);
681 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
682 vm_object_deallocate(object);
683}
684
685/*
20479584 686 * vm_pageout_scan does the dirty work for the pageout daemon.
984263bc 687 */
8fa76237
MD
688struct vm_pageout_scan_info {
689 struct proc *bigproc;
690 vm_offset_t bigsize;
691};
692
693static int vm_pageout_scan_callback(struct proc *p, void *data);
694
20479584 695static int
984263bc
MD
696vm_pageout_scan(int pass)
697{
8fa76237 698 struct vm_pageout_scan_info info;
984263bc
MD
699 vm_page_t m, next;
700 struct vm_page marker;
20479584
MD
701 int maxscan, pcount;
702 int recycle_count;
703 int inactive_shortage, active_shortage;
984263bc
MD
704 vm_object_t object;
705 int actcount;
706 int vnodes_skipped = 0;
707 int maxlaunder;
984263bc
MD
708
709 /*
710 * Do whatever cleanup that the pmap code can.
711 */
712 pmap_collect();
713
984263bc 714 /*
20479584
MD
715 * Calculate our target for the number of free+cache pages we
716 * want to get to. This is higher then the number that causes
717 * allocations to stall (severe) in order to provide hysteresis,
718 * and if we don't make it all the way but get to the minimum
719 * we're happy.
984263bc 720 */
20479584
MD
721 inactive_shortage = vm_paging_target() + vm_pageout_deficit;
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 *
1036 * We do not do this if we were able to satisfy the requirement
1037 * entirely from the inactive queue.
1038 *
1039 * NOTE: Both variables can end up negative.
1040 * NOTE: We are still in a critical section.
984263bc 1041 */
20479584
MD
1042 active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count;
1043 if (inactive_shortage <= 0)
1044 active_shortage = 0;
1045
12e4aaff 1046 pcount = vmstats.v_active_count;
20479584 1047 recycle_count = 0;
984263bc
MD
1048 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1049
20479584
MD
1050 while ((m != NULL) && (pcount-- > 0) &&
1051 (inactive_shortage > 0 || active_shortage > 0)
1052 ) {
06ecca5a
MD
1053 /*
1054 * Give interrupts a chance.
1055 */
5fd012e0
MD
1056 crit_exit();
1057 crit_enter();
984263bc
MD
1058
1059 /*
06ecca5a 1060 * If the page was ripped out from under us, just stop.
984263bc 1061 */
06ecca5a 1062 if (m->queue != PQ_ACTIVE)
984263bc 1063 break;
984263bc 1064 next = TAILQ_NEXT(m, pageq);
06ecca5a 1065
984263bc
MD
1066 /*
1067 * Don't deactivate pages that are busy.
1068 */
1069 if ((m->busy != 0) ||
1070 (m->flags & PG_BUSY) ||
1071 (m->hold_count != 0)) {
984263bc
MD
1072 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1073 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1074 m = next;
1075 continue;
1076 }
1077
1078 /*
1079 * The count for pagedaemon pages is done after checking the
1080 * page for eligibility...
1081 */
12e4aaff 1082 mycpu->gd_cnt.v_pdpages++;
984263bc
MD
1083
1084 /*
20479584
MD
1085 * Check to see "how much" the page has been used and clear
1086 * the tracking access bits. If the object has no references
1087 * don't bother paying the expense.
984263bc
MD
1088 */
1089 actcount = 0;
1090 if (m->object->ref_count != 0) {
20479584
MD
1091 if (m->flags & PG_REFERENCED)
1092 ++actcount;
984263bc
MD
1093 actcount += pmap_ts_referenced(m);
1094 if (actcount) {
1095 m->act_count += ACT_ADVANCE + actcount;
1096 if (m->act_count > ACT_MAX)
1097 m->act_count = ACT_MAX;
1098 }
1099 }
984263bc
MD
1100 vm_page_flag_clear(m, PG_REFERENCED);
1101
1102 /*
20479584 1103 * actcount is only valid if the object ref_count is non-zero.
984263bc 1104 */
20479584 1105 if (actcount && m->object->ref_count != 0) {
984263bc
MD
1106 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1107 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1108 } else {
1109 m->act_count -= min(m->act_count, ACT_DECLINE);
1110 if (vm_pageout_algorithm ||
1111 m->object->ref_count == 0 ||
20479584
MD
1112 m->act_count < pass + 1
1113 ) {
1114 /*
1115 * Deactivate the page. If we had a
1116 * shortage from our inactive scan try to
1117 * free (cache) the page instead.
1118 */
1119 --active_shortage;
1120 if (inactive_shortage > 0 ||
1121 m->object->ref_count == 0) {
1122 if (inactive_shortage > 0)
1123 ++recycle_count;
17cde63e 1124 vm_page_busy(m);
984263bc 1125 vm_page_protect(m, VM_PROT_NONE);
17cde63e 1126 vm_page_wakeup(m);
c84c24da 1127 if (m->dirty == 0) {
20479584 1128 --inactive_shortage;
984263bc 1129 vm_page_cache(m);
c84c24da 1130 } else {
984263bc 1131 vm_page_deactivate(m);
c84c24da 1132 }
984263bc
MD
1133 } else {
1134 vm_page_deactivate(m);
1135 }
1136 } else {
984263bc
MD
1137 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1138 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1139 }
1140 }
1141 m = next;
1142 }
1143
984263bc
MD
1144 /*
1145 * We try to maintain some *really* free pages, this allows interrupt
1146 * code to be guaranteed space. Since both cache and free queues
1147 * are considered basically 'free', moving pages from cache to free
1148 * does not effect other calculations.
06ecca5a 1149 *
5fd012e0 1150 * NOTE: we are still in a critical section.
c84c24da
MD
1151 *
1152 * Pages moved from PQ_CACHE to totally free are not counted in the
1153 * pages_freed counter.
984263bc 1154 */
12e4aaff 1155 while (vmstats.v_free_count < vmstats.v_free_reserved) {
984263bc
MD
1156 static int cache_rover = 0;
1157 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
20479584 1158 if (m == NULL)
984263bc
MD
1159 break;
1160 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1161 m->busy ||
1162 m->hold_count ||
1163 m->wire_count) {
1164#ifdef INVARIANTS
086c1d7e 1165 kprintf("Warning: busy page %p found in cache\n", m);
984263bc
MD
1166#endif
1167 vm_page_deactivate(m);
1168 continue;
1169 }
17cde63e
MD
1170 KKASSERT((m->flags & PG_MAPPED) == 0);
1171 KKASSERT(m->dirty == 0);
984263bc
MD
1172 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1173 vm_pageout_page_free(m);
12e4aaff 1174 mycpu->gd_cnt.v_dfree++;
984263bc 1175 }
06ecca5a 1176
5fd012e0 1177 crit_exit();
984263bc
MD
1178
1179#if !defined(NO_SWAPPING)
1180 /*
1181 * Idle process swapout -- run once per second.
1182 */
1183 if (vm_swap_idle_enabled) {
1184 static long lsec;
1185 if (time_second != lsec) {
1186 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1187 vm_req_vmdaemon();
1188 lsec = time_second;
1189 }
1190 }
1191#endif
1192
1193 /*
1194 * If we didn't get enough free pages, and we have skipped a vnode
1195 * in a writeable object, wakeup the sync daemon. And kick swapout
1196 * if we did not get enough free pages.
1197 */
1198 if (vm_paging_target() > 0) {
20479584 1199 if (vnodes_skipped && vm_page_count_min(0))
418ff780 1200 speedup_syncer();
984263bc
MD
1201#if !defined(NO_SWAPPING)
1202 if (vm_swap_enabled && vm_page_count_target()) {
1203 vm_req_vmdaemon();
1204 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1205 }
1206#endif
1207 }
1208
1209 /*
20479584
MD
1210 * Handle catastrophic conditions. Under good conditions we should
1211 * be at the target, well beyond our minimum. If we could not even
1212 * reach our minimum the system is under heavy stress.
1213 *
1214 * Determine whether we have run out of memory. This occurs when
1215 * swap_pager_full is TRUE and the only pages left in the page
1216 * queues are dirty. We will still likely have page shortages.
c84c24da
MD
1217 *
1218 * - swap_pager_full is set if insufficient swap was
1219 * available to satisfy a requested pageout.
1220 *
20479584
MD
1221 * - the inactive queue is bloated (4 x size of active queue),
1222 * meaning it is unable to get rid of dirty pages and.
c84c24da 1223 *
20479584
MD
1224 * - vm_page_count_min() without counting pages recycled from the
1225 * active queue (recycle_count) means we could not recover
1226 * enough pages to meet bare minimum needs. This test only
1227 * works if the inactive queue is bloated.
c84c24da 1228 *
20479584
MD
1229 * - due to a positive inactive_shortage we shifted the remaining
1230 * dirty pages from the active queue to the inactive queue
1231 * trying to find clean ones to free.
984263bc 1232 */
20479584 1233 if (swap_pager_full && vm_page_count_min(recycle_count))
c84c24da 1234 kprintf("Warning: system low on memory+swap!\n");
20479584
MD
1235 if (swap_pager_full && vm_page_count_min(recycle_count) &&
1236 vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1237 inactive_shortage > 0) {
1238 /*
1239 * Kill something.
1240 */
8fa76237
MD
1241 info.bigproc = NULL;
1242 info.bigsize = 0;
1243 allproc_scan(vm_pageout_scan_callback, &info);
1244 if (info.bigproc != NULL) {
1245 killproc(info.bigproc, "out of swap space");
1246 info.bigproc->p_nice = PRIO_MIN;
08f2f1bb
SS
1247 info.bigproc->p_usched->resetpriority(
1248 FIRST_LWP_IN_PROC(info.bigproc));
12e4aaff 1249 wakeup(&vmstats.v_free_count);
8fa76237 1250 PRELE(info.bigproc);
984263bc
MD
1251 }
1252 }
20479584 1253 return(inactive_shortage);
984263bc
MD
1254}
1255
8fa76237
MD
1256static int
1257vm_pageout_scan_callback(struct proc *p, void *data)
1258{
1259 struct vm_pageout_scan_info *info = data;
1260 vm_offset_t size;
1261
1262 /*
20479584
MD
1263 * Never kill system processes or init. If we have configured swap
1264 * then try to avoid killing low-numbered pids.
8fa76237
MD
1265 */
1266 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1267 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1268 return (0);
1269 }
1270
1271 /*
1272 * if the process is in a non-running type state,
1273 * don't touch it.
1274 */
20479584 1275 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
8fa76237 1276 return (0);
8fa76237
MD
1277
1278 /*
20479584
MD
1279 * Get the approximate process size. Note that anonymous pages
1280 * with backing swap will be counted twice, but there should not
1281 * be too many such pages due to the stress the VM system is
1282 * under at this point.
8fa76237 1283 */
20479584 1284 size = vmspace_anonymous_count(p->p_vmspace) +
8fa76237
MD
1285 vmspace_swap_count(p->p_vmspace);
1286
1287 /*
1288 * If the this process is bigger than the biggest one
1289 * remember it.
1290 */
20479584 1291 if (info->bigsize < size) {
8fa76237
MD
1292 if (info->bigproc)
1293 PRELE(info->bigproc);
1294 PHOLD(p);
1295 info->bigproc = p;
1296 info->bigsize = size;
1297 }
1298 return(0);
1299}
1300
984263bc
MD
1301/*
1302 * This routine tries to maintain the pseudo LRU active queue,
1303 * so that during long periods of time where there is no paging,
1304 * that some statistic accumulation still occurs. This code
1305 * helps the situation where paging just starts to occur.
1306 */
1307static void
57e43348 1308vm_pageout_page_stats(void)
984263bc 1309{
984263bc
MD
1310 vm_page_t m,next;
1311 int pcount,tpcount; /* Number of pages to check */
1312 static int fullintervalcount = 0;
1313 int page_shortage;
984263bc
MD
1314
1315 page_shortage =
12e4aaff
MD
1316 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1317 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
984263bc
MD
1318
1319 if (page_shortage <= 0)
1320 return;
1321
5fd012e0 1322 crit_enter();
984263bc 1323
12e4aaff 1324 pcount = vmstats.v_active_count;
984263bc
MD
1325 fullintervalcount += vm_pageout_stats_interval;
1326 if (fullintervalcount < vm_pageout_full_stats_interval) {
12e4aaff 1327 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
984263bc
MD
1328 if (pcount > tpcount)
1329 pcount = tpcount;
1330 } else {
1331 fullintervalcount = 0;
1332 }
1333
1334 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1335 while ((m != NULL) && (pcount-- > 0)) {
1336 int actcount;
1337
1338 if (m->queue != PQ_ACTIVE) {
1339 break;
1340 }
1341
1342 next = TAILQ_NEXT(m, pageq);
1343 /*
1344 * Don't deactivate pages that are busy.
1345 */
1346 if ((m->busy != 0) ||
1347 (m->flags & PG_BUSY) ||
1348 (m->hold_count != 0)) {
984263bc
MD
1349 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1350 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1351 m = next;
1352 continue;
1353 }
1354
1355 actcount = 0;
1356 if (m->flags & PG_REFERENCED) {
1357 vm_page_flag_clear(m, PG_REFERENCED);
1358 actcount += 1;
1359 }
1360
1361 actcount += pmap_ts_referenced(m);
1362 if (actcount) {
1363 m->act_count += ACT_ADVANCE + actcount;
1364 if (m->act_count > ACT_MAX)
1365 m->act_count = ACT_MAX;
984263bc
MD
1366 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1367 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1368 } else {
1369 if (m->act_count == 0) {
1370 /*
1371 * We turn off page access, so that we have
1372 * more accurate RSS stats. We don't do this
1373 * in the normal page deactivation when the
1374 * system is loaded VM wise, because the
1375 * cost of the large number of page protect
1376 * operations would be higher than the value
1377 * of doing the operation.
1378 */
17cde63e 1379 vm_page_busy(m);
984263bc 1380 vm_page_protect(m, VM_PROT_NONE);
17cde63e 1381 vm_page_wakeup(m);
984263bc
MD
1382 vm_page_deactivate(m);
1383 } else {
1384 m->act_count -= min(m->act_count, ACT_DECLINE);
984263bc
MD
1385 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1386 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1387 }
1388 }
1389
1390 m = next;
1391 }
5fd012e0 1392 crit_exit();
984263bc
MD
1393}
1394
1395static int
57e43348 1396vm_pageout_free_page_calc(vm_size_t count)
984263bc 1397{
12e4aaff 1398 if (count < vmstats.v_page_count)
984263bc
MD
1399 return 0;
1400 /*
1401 * free_reserved needs to include enough for the largest swap pager
1402 * structures plus enough for any pv_entry structs when paging.
1403 */
12e4aaff
MD
1404 if (vmstats.v_page_count > 1024)
1405 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
984263bc 1406 else
12e4aaff
MD
1407 vmstats.v_free_min = 4;
1408 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1409 vmstats.v_interrupt_free_min;
1410 vmstats.v_free_reserved = vm_pageout_page_count +
1411 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1412 vmstats.v_free_severe = vmstats.v_free_min / 2;
1413 vmstats.v_free_min += vmstats.v_free_reserved;
1414 vmstats.v_free_severe += vmstats.v_free_reserved;
984263bc
MD
1415 return 1;
1416}
1417
1418
1419/*
20479584 1420 * vm_pageout is the high level pageout daemon.
984263bc
MD
1421 */
1422static void
57e43348 1423vm_pageout(void)
984263bc
MD
1424{
1425 int pass;
20479584 1426 int inactive_shortage;
984263bc
MD
1427
1428 /*
1429 * Initialize some paging parameters.
1430 */
4ecf7cc9 1431 curthread->td_flags |= TDF_SYSTHREAD;
984263bc 1432
12e4aaff
MD
1433 vmstats.v_interrupt_free_min = 2;
1434 if (vmstats.v_page_count < 2000)
984263bc
MD
1435 vm_pageout_page_count = 8;
1436
12e4aaff 1437 vm_pageout_free_page_calc(vmstats.v_page_count);
20479584 1438
984263bc
MD
1439 /*
1440 * v_free_target and v_cache_min control pageout hysteresis. Note
1441 * that these are more a measure of the VM cache queue hysteresis
1442 * then the VM free queue. Specifically, v_free_target is the
1443 * high water mark (free+cache pages).
1444 *
1445 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1446 * low water mark, while v_free_min is the stop. v_cache_min must
1447 * be big enough to handle memory needs while the pageout daemon
1448 * is signalled and run to free more pages.
1449 */
12e4aaff
MD
1450 if (vmstats.v_free_count > 6144)
1451 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
984263bc 1452 else
12e4aaff 1453 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
984263bc 1454
12e4aaff
MD
1455 if (vmstats.v_free_count > 2048) {
1456 vmstats.v_cache_min = vmstats.v_free_target;
1457 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1458 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2;
984263bc 1459 } else {
12e4aaff
MD
1460 vmstats.v_cache_min = 0;
1461 vmstats.v_cache_max = 0;
1462 vmstats.v_inactive_target = vmstats.v_free_count / 4;
984263bc 1463 }
12e4aaff
MD
1464 if (vmstats.v_inactive_target > vmstats.v_free_count / 3)
1465 vmstats.v_inactive_target = vmstats.v_free_count / 3;
984263bc
MD
1466
1467 /* XXX does not really belong here */
1468 if (vm_page_max_wired == 0)
12e4aaff 1469 vm_page_max_wired = vmstats.v_free_count / 3;
984263bc
MD
1470
1471 if (vm_pageout_stats_max == 0)
12e4aaff 1472 vm_pageout_stats_max = vmstats.v_free_target;
984263bc
MD
1473
1474 /*
1475 * Set interval in seconds for stats scan.
1476 */
1477 if (vm_pageout_stats_interval == 0)
1478 vm_pageout_stats_interval = 5;
1479 if (vm_pageout_full_stats_interval == 0)
1480 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1481
1482
1483 /*
1484 * Set maximum free per pass
1485 */
1486 if (vm_pageout_stats_free_max == 0)
1487 vm_pageout_stats_free_max = 5;
1488
1489 swap_pager_swap_init();
1490 pass = 0;
20479584 1491
984263bc
MD
1492 /*
1493 * The pageout daemon is never done, so loop forever.
1494 */
1495 while (TRUE) {
1496 int error;
984263bc 1497
20479584 1498 if (vm_pages_needed == 0) {
984263bc 1499 /*
20479584 1500 * Wait for an action request
984263bc 1501 */
984263bc 1502 error = tsleep(&vm_pages_needed,
20479584
MD
1503 0, "psleep",
1504 vm_pageout_stats_interval * hz);
1505 if (error && vm_pages_needed == 0) {
984263bc
MD
1506 vm_pageout_page_stats();
1507 continue;
1508 }
20479584 1509 vm_pages_needed = 1;
984263bc
MD
1510 }
1511
20479584
MD
1512 /*
1513 * If we have enough free memory, wakeup waiters.
1514 */
1515 crit_enter();
1516 if (!vm_page_count_min(0))
1517 wakeup(&vmstats.v_free_count);
1518 mycpu->gd_cnt.v_pdwakeups++;
5fd012e0 1519 crit_exit();
20479584
MD
1520 inactive_shortage = vm_pageout_scan(pass);
1521
1522 /*
1523 * Try to avoid thrashing the system with activity.
1524 */
1525 if (inactive_shortage > 0) {
1526 ++pass;
1527 if (swap_pager_full) {
1528 /*
1529 * Running out of memory, catastrophic back-off
1530 * to one-second intervals.
1531 */
1532 tsleep(&vm_pages_needed, 0, "pdelay", hz);
1533 } else if (pass < 10 && vm_pages_needed > 1) {
1534 /*
1535 * Normal operation, additional processes
1536 * have already kicked us. Retry immediately.
1537 */
1538 } else if (pass < 10) {
1539 /*
1540 * Normal operation, fewer processes. Delay
1541 * a bit but allow wakeups.
1542 */
1543 vm_pages_needed = 0;
1544 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1545 vm_pages_needed = 1;
1546 } else {
1547 /*
1548 * We've taken too many passes, forced delay.
1549 */
1550 tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1551 }
1552 } else {
1553 pass = 0;
1554 vm_pages_needed = 0;
1555 }
984263bc
MD
1556 }
1557}
1558
20479584
MD
1559/*
1560 * Called after allocating a page out of the cache or free queue
1561 * to possibly wake the pagedaemon up to replentish our supply.
1562 *
1563 * We try to generate some hysteresis by waking the pagedaemon up
1564 * when our free+cache pages go below the severe level. The pagedaemon
1565 * tries to get the count back up to at least the minimum, and through
1566 * to the target level if possible.
1567 *
1568 * If the pagedaemon is already active bump vm_pages_needed as a hint
1569 * that there are even more requests pending.
1570 */
984263bc 1571void
57e43348 1572pagedaemon_wakeup(void)
984263bc 1573{
20479584
MD
1574 if (vm_page_count_severe() && curthread != pagethread) {
1575 if (vm_pages_needed == 0) {
1576 vm_pages_needed = 1;
1577 wakeup(&vm_pages_needed);
1578 } else if (vm_page_count_min(0)) {
1579 ++vm_pages_needed;
1580 }
984263bc
MD
1581 }
1582}
1583
1584#if !defined(NO_SWAPPING)
1585static void
57e43348 1586vm_req_vmdaemon(void)
984263bc
MD
1587{
1588 static int lastrun = 0;
1589
1590 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1591 wakeup(&vm_daemon_needed);
1592 lastrun = ticks;
1593 }
1594}
1595
8fa76237
MD
1596static int vm_daemon_callback(struct proc *p, void *data __unused);
1597
984263bc 1598static void
57e43348 1599vm_daemon(void)
984263bc 1600{
984263bc 1601 while (TRUE) {
377d4740 1602 tsleep(&vm_daemon_needed, 0, "psleep", 0);
984263bc
MD
1603 if (vm_pageout_req_swapout) {
1604 swapout_procs(vm_pageout_req_swapout);
1605 vm_pageout_req_swapout = 0;
1606 }
1607 /*
1608 * scan the processes for exceeding their rlimits or if
1609 * process is swapped out -- deactivate pages
1610 */
8fa76237
MD
1611 allproc_scan(vm_daemon_callback, NULL);
1612 }
1613}
984263bc 1614
8fa76237
MD
1615static int
1616vm_daemon_callback(struct proc *p, void *data __unused)
1617{
1618 vm_pindex_t limit, size;
984263bc 1619
8fa76237
MD
1620 /*
1621 * if this is a system process or if we have already
1622 * looked at this process, skip it.
1623 */
1624 if (p->p_flag & (P_SYSTEM | P_WEXIT))
1625 return (0);
984263bc 1626
8fa76237
MD
1627 /*
1628 * if the process is in a non-running type state,
1629 * don't touch it.
1630 */
164b8401 1631 if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
8fa76237 1632 return (0);
984263bc 1633
8fa76237
MD
1634 /*
1635 * get a limit
1636 */
1637 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1638 p->p_rlimit[RLIMIT_RSS].rlim_max));
1639
1640 /*
1641 * let processes that are swapped out really be
1642 * swapped out. Set the limit to nothing to get as
1643 * many pages out to swap as possible.
1644 */
1645 if (p->p_flag & P_SWAPPEDOUT)
1646 limit = 0;
1647
1648 size = vmspace_resident_count(p->p_vmspace);
1649 if (limit >= 0 && size >= limit) {
1650 vm_pageout_map_deactivate_pages(
1651 &p->p_vmspace->vm_map, limit);
984263bc 1652 }
8fa76237 1653 return (0);
984263bc 1654}
8fa76237 1655
984263bc 1656#endif