* Remove a procedural layer in the scheduler clock code by having
[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 $
dcc99b62 69 * $DragonFly: src/sys/vm/vm_pageout.c,v 1.14 2005/06/27 18:38:00 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);
109static void vm_pageout_scan (int pass);
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
198#define VM_PAGEOUT_PAGE_COUNT 16
199int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
200
201int vm_page_max_wired; /* XXX max # of wired pages system-wide */
202
203#if !defined(NO_SWAPPING)
1388df65
RG
204typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
205static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
984263bc 206static freeer_fcn_t vm_pageout_object_deactivate_pages;
1388df65 207static void vm_req_vmdaemon (void);
984263bc
MD
208#endif
209static void vm_pageout_page_stats(void);
210
211/*
212 * vm_pageout_clean:
213 *
06ecca5a
MD
214 * Clean the page and remove it from the laundry. The page must not be
215 * busy on-call.
984263bc
MD
216 *
217 * We set the busy bit to cause potential page faults on this page to
218 * block. Note the careful timing, however, the busy bit isn't set till
219 * late and we cannot do anything that will mess with the page.
220 */
221
222static int
57e43348 223vm_pageout_clean(vm_page_t m)
984263bc 224{
5f910b2f 225 vm_object_t object;
984263bc
MD
226 vm_page_t mc[2*vm_pageout_page_count];
227 int pageout_count;
228 int ib, is, page_base;
229 vm_pindex_t pindex = m->pindex;
230
231 object = m->object;
232
233 /*
234 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
235 * with the new swapper, but we could have serious problems paging
236 * out other object types if there is insufficient memory.
237 *
238 * Unfortunately, checking free memory here is far too late, so the
239 * check has been moved up a procedural level.
240 */
241
242 /*
243 * Don't mess with the page if it's busy, held, or special
244 */
245 if ((m->hold_count != 0) ||
246 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
247 return 0;
248 }
249
250 mc[vm_pageout_page_count] = m;
251 pageout_count = 1;
252 page_base = vm_pageout_page_count;
253 ib = 1;
254 is = 1;
255
256 /*
257 * Scan object for clusterable pages.
258 *
259 * We can cluster ONLY if: ->> the page is NOT
260 * clean, wired, busy, held, or mapped into a
261 * buffer, and one of the following:
262 * 1) The page is inactive, or a seldom used
263 * active page.
264 * -or-
265 * 2) we force the issue.
266 *
267 * During heavy mmap/modification loads the pageout
268 * daemon can really fragment the underlying file
269 * due to flushing pages out of order and not trying
270 * align the clusters (which leave sporatic out-of-order
271 * holes). To solve this problem we do the reverse scan
272 * first and attempt to align our cluster, then do a
273 * forward scan if room remains.
274 */
275
276more:
277 while (ib && pageout_count < vm_pageout_page_count) {
278 vm_page_t p;
279
280 if (ib > pindex) {
281 ib = 0;
282 break;
283 }
284
285 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
286 ib = 0;
287 break;
288 }
289 if (((p->queue - p->pc) == PQ_CACHE) ||
290 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
291 ib = 0;
292 break;
293 }
294 vm_page_test_dirty(p);
295 if ((p->dirty & p->valid) == 0 ||
296 p->queue != PQ_INACTIVE ||
297 p->wire_count != 0 || /* may be held by buf cache */
298 p->hold_count != 0) { /* may be undergoing I/O */
299 ib = 0;
300 break;
301 }
302 mc[--page_base] = p;
303 ++pageout_count;
304 ++ib;
305 /*
306 * alignment boundry, stop here and switch directions. Do
307 * not clear ib.
308 */
309 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
310 break;
311 }
312
313 while (pageout_count < vm_pageout_page_count &&
314 pindex + is < object->size) {
315 vm_page_t p;
316
317 if ((p = vm_page_lookup(object, pindex + is)) == NULL)
318 break;
319 if (((p->queue - p->pc) == PQ_CACHE) ||
320 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
321 break;
322 }
323 vm_page_test_dirty(p);
324 if ((p->dirty & p->valid) == 0 ||
325 p->queue != PQ_INACTIVE ||
326 p->wire_count != 0 || /* may be held by buf cache */
327 p->hold_count != 0) { /* may be undergoing I/O */
328 break;
329 }
330 mc[page_base + pageout_count] = p;
331 ++pageout_count;
332 ++is;
333 }
334
335 /*
336 * If we exhausted our forward scan, continue with the reverse scan
337 * when possible, even past a page boundry. This catches boundry
338 * conditions.
339 */
340 if (ib && pageout_count < vm_pageout_page_count)
341 goto more;
342
343 /*
344 * we allow reads during pageouts...
345 */
346 return vm_pageout_flush(&mc[page_base], pageout_count, 0);
347}
348
349/*
350 * vm_pageout_flush() - launder the given pages
351 *
352 * The given pages are laundered. Note that we setup for the start of
353 * I/O ( i.e. busy the page ), mark it read-only, and bump the object
354 * reference count all in here rather then in the parent. If we want
355 * the parent to do more sophisticated things we may have to change
356 * the ordering.
357 */
358
359int
57e43348 360vm_pageout_flush(vm_page_t *mc, int count, int flags)
984263bc 361{
5f910b2f 362 vm_object_t object;
984263bc
MD
363 int pageout_status[count];
364 int numpagedout = 0;
365 int i;
366
367 /*
368 * Initiate I/O. Bump the vm_page_t->busy counter and
369 * mark the pages read-only.
370 *
371 * We do not have to fixup the clean/dirty bits here... we can
372 * allow the pager to do it after the I/O completes.
373 */
374
375 for (i = 0; i < count; i++) {
376 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
377 vm_page_io_start(mc[i]);
378 vm_page_protect(mc[i], VM_PROT_READ);
379 }
380
381 object = mc[0]->object;
382 vm_object_pip_add(object, count);
383
384 vm_pager_put_pages(object, mc, count,
385 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
386 pageout_status);
387
388 for (i = 0; i < count; i++) {
389 vm_page_t mt = mc[i];
390
391 switch (pageout_status[i]) {
392 case VM_PAGER_OK:
393 numpagedout++;
394 break;
395 case VM_PAGER_PEND:
396 numpagedout++;
397 break;
398 case VM_PAGER_BAD:
399 /*
400 * Page outside of range of object. Right now we
401 * essentially lose the changes by pretending it
402 * worked.
403 */
404 pmap_clear_modify(mt);
405 vm_page_undirty(mt);
406 break;
407 case VM_PAGER_ERROR:
408 case VM_PAGER_FAIL:
409 /*
410 * If page couldn't be paged out, then reactivate the
411 * page so it doesn't clog the inactive list. (We
412 * will try paging out it again later).
413 */
414 vm_page_activate(mt);
415 break;
416 case VM_PAGER_AGAIN:
417 break;
418 }
419
420 /*
421 * If the operation is still going, leave the page busy to
422 * block all other accesses. Also, leave the paging in
423 * progress indicator set so that we don't attempt an object
424 * collapse.
425 */
426 if (pageout_status[i] != VM_PAGER_PEND) {
427 vm_object_pip_wakeup(object);
428 vm_page_io_finish(mt);
429 if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
430 vm_page_protect(mt, VM_PROT_READ);
431 }
432 }
433 return numpagedout;
434}
435
436#if !defined(NO_SWAPPING)
437/*
438 * vm_pageout_object_deactivate_pages
439 *
440 * deactivate enough pages to satisfy the inactive target
441 * requirements or if vm_page_proc_limit is set, then
442 * deactivate all of the pages in the object and its
443 * backing_objects.
444 *
445 * The object and map must be locked.
446 */
447static void
57e43348 448vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
06ecca5a 449 vm_pindex_t desired, int map_remove_only)
984263bc 450{
5f910b2f 451 vm_page_t p, next;
984263bc
MD
452 int rcount;
453 int remove_mode;
984263bc
MD
454
455 if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
456 return;
457
458 while (object) {
459 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
460 return;
461 if (object->paging_in_progress)
462 return;
463
464 remove_mode = map_remove_only;
465 if (object->shadow_count > 1)
466 remove_mode = 1;
06ecca5a
MD
467
468 /*
469 * scan the objects entire memory queue. spl protection is
470 * required to avoid an interrupt unbusy/free race against
471 * our busy check.
472 */
5fd012e0 473 crit_enter();
984263bc
MD
474 rcount = object->resident_page_count;
475 p = TAILQ_FIRST(&object->memq);
06ecca5a 476
984263bc
MD
477 while (p && (rcount-- > 0)) {
478 int actcount;
06ecca5a 479 if (pmap_resident_count(vm_map_pmap(map)) <= desired) {
5fd012e0 480 crit_exit();
984263bc 481 return;
06ecca5a 482 }
984263bc 483 next = TAILQ_NEXT(p, listq);
12e4aaff 484 mycpu->gd_cnt.v_pdpages++;
984263bc
MD
485 if (p->wire_count != 0 ||
486 p->hold_count != 0 ||
487 p->busy != 0 ||
488 (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
489 !pmap_page_exists_quick(vm_map_pmap(map), p)) {
490 p = next;
491 continue;
492 }
493
494 actcount = pmap_ts_referenced(p);
495 if (actcount) {
496 vm_page_flag_set(p, PG_REFERENCED);
497 } else if (p->flags & PG_REFERENCED) {
498 actcount = 1;
499 }
500
501 if ((p->queue != PQ_ACTIVE) &&
502 (p->flags & PG_REFERENCED)) {
503 vm_page_activate(p);
504 p->act_count += actcount;
505 vm_page_flag_clear(p, PG_REFERENCED);
506 } else if (p->queue == PQ_ACTIVE) {
507 if ((p->flags & PG_REFERENCED) == 0) {
508 p->act_count -= min(p->act_count, ACT_DECLINE);
509 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
510 vm_page_protect(p, VM_PROT_NONE);
511 vm_page_deactivate(p);
512 } else {
984263bc
MD
513 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
514 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
984263bc
MD
515 }
516 } else {
517 vm_page_activate(p);
518 vm_page_flag_clear(p, PG_REFERENCED);
519 if (p->act_count < (ACT_MAX - ACT_ADVANCE))
520 p->act_count += ACT_ADVANCE;
984263bc
MD
521 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
522 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
984263bc
MD
523 }
524 } else if (p->queue == PQ_INACTIVE) {
525 vm_page_protect(p, VM_PROT_NONE);
526 }
527 p = next;
528 }
5fd012e0 529 crit_exit();
984263bc
MD
530 object = object->backing_object;
531 }
984263bc
MD
532}
533
534/*
535 * deactivate some number of pages in a map, try to do it fairly, but
536 * that is really hard to do.
537 */
538static void
57e43348 539vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
984263bc
MD
540{
541 vm_map_entry_t tmpe;
542 vm_object_t obj, bigobj;
543 int nothingwired;
544
41a01a4d 545 if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, NULL, curthread)) {
984263bc
MD
546 return;
547 }
548
549 bigobj = NULL;
550 nothingwired = TRUE;
551
552 /*
553 * first, search out the biggest object, and try to free pages from
554 * that.
555 */
556 tmpe = map->header.next;
557 while (tmpe != &map->header) {
558 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
559 obj = tmpe->object.vm_object;
560 if ((obj != NULL) && (obj->shadow_count <= 1) &&
561 ((bigobj == NULL) ||
562 (bigobj->resident_page_count < obj->resident_page_count))) {
563 bigobj = obj;
564 }
565 }
566 if (tmpe->wired_count > 0)
567 nothingwired = FALSE;
568 tmpe = tmpe->next;
569 }
570
571 if (bigobj)
572 vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
573
574 /*
575 * Next, hunt around for other pages to deactivate. We actually
576 * do this search sort of wrong -- .text first is not the best idea.
577 */
578 tmpe = map->header.next;
579 while (tmpe != &map->header) {
580 if (pmap_resident_count(vm_map_pmap(map)) <= desired)
581 break;
582 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
583 obj = tmpe->object.vm_object;
584 if (obj)
585 vm_pageout_object_deactivate_pages(map, obj, desired, 0);
586 }
587 tmpe = tmpe->next;
588 };
589
590 /*
591 * Remove all mappings if a process is swapped out, this will free page
592 * table pages.
593 */
594 if (desired == 0 && nothingwired)
595 pmap_remove(vm_map_pmap(map),
596 VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
597 vm_map_unlock(map);
984263bc
MD
598}
599#endif
600
601/*
602 * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore
603 * to vnode deadlocks. We only do it for OBJT_DEFAULT and OBJT_SWAP objects
604 * which we know can be trivially freed.
605 */
606
607void
608vm_pageout_page_free(vm_page_t m) {
609 vm_object_t object = m->object;
610 int type = object->type;
611
612 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
613 vm_object_reference(object);
614 vm_page_busy(m);
615 vm_page_protect(m, VM_PROT_NONE);
616 vm_page_free(m);
617 if (type == OBJT_SWAP || type == OBJT_DEFAULT)
618 vm_object_deallocate(object);
619}
620
621/*
622 * vm_pageout_scan does the dirty work for the pageout daemon.
623 */
624static void
625vm_pageout_scan(int pass)
626{
627 vm_page_t m, next;
628 struct vm_page marker;
629 int page_shortage, maxscan, pcount;
630 int addl_page_shortage, addl_page_shortage_init;
631 struct proc *p, *bigproc;
632 vm_offset_t size, bigsize;
633 vm_object_t object;
634 int actcount;
635 int vnodes_skipped = 0;
636 int maxlaunder;
984263bc
MD
637
638 /*
639 * Do whatever cleanup that the pmap code can.
640 */
641 pmap_collect();
642
643 addl_page_shortage_init = vm_pageout_deficit;
644 vm_pageout_deficit = 0;
645
646 /*
647 * Calculate the number of pages we want to either free or move
648 * to the cache.
649 */
650 page_shortage = vm_paging_target() + addl_page_shortage_init;
651
652 /*
653 * Initialize our marker
654 */
655 bzero(&marker, sizeof(marker));
656 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
657 marker.queue = PQ_INACTIVE;
658 marker.wire_count = 1;
659
660 /*
661 * Start scanning the inactive queue for pages we can move to the
662 * cache or free. The scan will stop when the target is reached or
663 * we have scanned the entire inactive queue. Note that m->act_count
664 * is not used to form decisions for the inactive queue, only for the
665 * active queue.
666 *
667 * maxlaunder limits the number of dirty pages we flush per scan.
668 * For most systems a smaller value (16 or 32) is more robust under
669 * extreme memory and disk pressure because any unnecessary writes
670 * to disk can result in extreme performance degredation. However,
671 * systems with excessive dirty pages (especially when MAP_NOSYNC is
672 * used) will die horribly with limited laundering. If the pageout
673 * daemon cannot clean enough pages in the first pass, we let it go
674 * all out in succeeding passes.
675 */
676 if ((maxlaunder = vm_max_launder) <= 1)
677 maxlaunder = 1;
678 if (pass)
679 maxlaunder = 10000;
680
06ecca5a 681 /*
5fd012e0
MD
682 * We will generally be in a critical section throughout the
683 * scan, but we can release it temporarily when we are sitting on a
684 * non-busy page without fear. this is required to prevent an
685 * interrupt from unbusying or freeing a page prior to our busy
686 * check, leaving us on the wrong queue or checking the wrong
687 * page.
06ecca5a 688 */
5fd012e0 689 crit_enter();
984263bc
MD
690rescan0:
691 addl_page_shortage = addl_page_shortage_init;
12e4aaff 692 maxscan = vmstats.v_inactive_count;
984263bc
MD
693 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
694 m != NULL && maxscan-- > 0 && page_shortage > 0;
06ecca5a
MD
695 m = next
696 ) {
12e4aaff 697 mycpu->gd_cnt.v_pdpages++;
984263bc 698
06ecca5a
MD
699 /*
700 * Give interrupts a chance
701 */
5fd012e0
MD
702 crit_exit();
703 crit_enter();
984263bc 704
06ecca5a
MD
705 /*
706 * It's easier for some of the conditions below to just loop
707 * and catch queue changes here rather then check everywhere
708 * else.
709 */
710 if (m->queue != PQ_INACTIVE)
711 goto rescan0;
984263bc
MD
712 next = TAILQ_NEXT(m, pageq);
713
714 /*
715 * skip marker pages
716 */
717 if (m->flags & PG_MARKER)
718 continue;
719
720 /*
721 * A held page may be undergoing I/O, so skip it.
722 */
723 if (m->hold_count) {
984263bc
MD
724 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
725 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
726 addl_page_shortage++;
727 continue;
728 }
06ecca5a 729
984263bc
MD
730 /*
731 * Dont mess with busy pages, keep in the front of the
732 * queue, most likely are being paged out.
733 */
734 if (m->busy || (m->flags & PG_BUSY)) {
735 addl_page_shortage++;
736 continue;
737 }
738
984263bc 739 if (m->object->ref_count == 0) {
06ecca5a
MD
740 /*
741 * If the object is not being used, we ignore previous
742 * references.
743 */
984263bc
MD
744 vm_page_flag_clear(m, PG_REFERENCED);
745 pmap_clear_reference(m);
746
984263bc 747 } else if (((m->flags & PG_REFERENCED) == 0) &&
06ecca5a
MD
748 (actcount = pmap_ts_referenced(m))) {
749 /*
750 * Otherwise, if the page has been referenced while
751 * in the inactive queue, we bump the "activation
752 * count" upwards, making it less likely that the
753 * page will be added back to the inactive queue
754 * prematurely again. Here we check the page tables
755 * (or emulated bits, if any), given the upper level
756 * VM system not knowing anything about existing
757 * references.
758 */
984263bc
MD
759 vm_page_activate(m);
760 m->act_count += (actcount + ACT_ADVANCE);
761 continue;
762 }
763
764 /*
765 * If the upper level VM system knows about any page
766 * references, we activate the page. We also set the
767 * "activation count" higher than normal so that we will less
768 * likely place pages back onto the inactive queue again.
769 */
770 if ((m->flags & PG_REFERENCED) != 0) {
771 vm_page_flag_clear(m, PG_REFERENCED);
772 actcount = pmap_ts_referenced(m);
773 vm_page_activate(m);
774 m->act_count += (actcount + ACT_ADVANCE + 1);
775 continue;
776 }
777
778 /*
779 * If the upper level VM system doesn't know anything about
780 * the page being dirty, we have to check for it again. As
781 * far as the VM code knows, any partially dirty pages are
782 * fully dirty.
41a01a4d
MD
783 *
784 * Pages marked PG_WRITEABLE may be mapped into the user
785 * address space of a process running on another cpu. A
786 * user process (without holding the MP lock) running on
787 * another cpu may be able to touch the page while we are
788 * trying to remove it. To prevent this from occuring we
789 * must call pmap_remove_all() or otherwise make the page
790 * read-only. If the race occured pmap_remove_all() is
791 * responsible for setting m->dirty.
984263bc
MD
792 */
793 if (m->dirty == 0) {
794 vm_page_test_dirty(m);
41a01a4d
MD
795#if 0
796 if (m->dirty == 0 && (m->flags & PG_WRITEABLE) != 0)
797 pmap_remove_all(m);
798#endif
984263bc
MD
799 } else {
800 vm_page_dirty(m);
801 }
802
984263bc 803 if (m->valid == 0) {
41a01a4d
MD
804 /*
805 * Invalid pages can be easily freed
806 */
984263bc 807 vm_pageout_page_free(m);
12e4aaff 808 mycpu->gd_cnt.v_dfree++;
984263bc 809 --page_shortage;
984263bc
MD
810 } else if (m->dirty == 0) {
811 /*
41a01a4d
MD
812 * Clean pages can be placed onto the cache queue.
813 * This effectively frees them.
984263bc
MD
814 */
815 vm_page_cache(m);
816 --page_shortage;
817 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
818 /*
819 * Dirty pages need to be paged out, but flushing
820 * a page is extremely expensive verses freeing
821 * a clean page. Rather then artificially limiting
822 * the number of pages we can flush, we instead give
823 * dirty pages extra priority on the inactive queue
824 * by forcing them to be cycled through the queue
825 * twice before being flushed, after which the
826 * (now clean) page will cycle through once more
827 * before being freed. This significantly extends
828 * the thrash point for a heavily loaded machine.
829 */
984263bc
MD
830 vm_page_flag_set(m, PG_WINATCFLS);
831 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
832 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
833 } else if (maxlaunder > 0) {
834 /*
835 * We always want to try to flush some dirty pages if
836 * we encounter them, to keep the system stable.
837 * Normally this number is small, but under extreme
838 * pressure where there are insufficient clean pages
839 * on the inactive queue, we may have to go all out.
840 */
841 int swap_pageouts_ok;
842 struct vnode *vp = NULL;
843
844 object = m->object;
845
846 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
847 swap_pageouts_ok = 1;
848 } else {
849 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
850 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
851 vm_page_count_min());
852
853 }
854
855 /*
856 * We don't bother paging objects that are "dead".
857 * Those objects are in a "rundown" state.
858 */
859 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
984263bc
MD
860 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
861 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
862 continue;
863 }
864
865 /*
866 * The object is already known NOT to be dead. It
867 * is possible for the vget() to block the whole
868 * pageout daemon, but the new low-memory handling
869 * code should prevent it.
870 *
871 * The previous code skipped locked vnodes and, worse,
872 * reordered pages in the queue. This results in
873 * completely non-deterministic operation because,
874 * quite often, a vm_fault has initiated an I/O and
875 * is holding a locked vnode at just the point where
876 * the pageout daemon is woken up.
877 *
878 * We can't wait forever for the vnode lock, we might
879 * deadlock due to a vn_read() getting stuck in
880 * vm_wait while holding this vnode. We skip the
881 * vnode if we can't get it in a reasonable amount
882 * of time.
883 */
884
885 if (object->type == OBJT_VNODE) {
886 vp = object->handle;
887
5fd012e0 888 if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK, curthread)) {
984263bc
MD
889 ++pageout_lock_miss;
890 if (object->flags & OBJ_MIGHTBEDIRTY)
891 vnodes_skipped++;
892 continue;
893 }
894
895 /*
896 * The page might have been moved to another
897 * queue during potential blocking in vget()
898 * above. The page might have been freed and
899 * reused for another vnode. The object might
900 * have been reused for another vnode.
901 */
902 if (m->queue != PQ_INACTIVE ||
903 m->object != object ||
904 object->handle != vp) {
905 if (object->flags & OBJ_MIGHTBEDIRTY)
906 vnodes_skipped++;
907 vput(vp);
908 continue;
909 }
910
911 /*
912 * The page may have been busied during the
913 * blocking in vput(); We don't move the
914 * page back onto the end of the queue so that
915 * statistics are more correct if we don't.
916 */
917 if (m->busy || (m->flags & PG_BUSY)) {
918 vput(vp);
919 continue;
920 }
921
922 /*
923 * If the page has become held it might
924 * be undergoing I/O, so skip it
925 */
926 if (m->hold_count) {
984263bc
MD
927 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
928 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
984263bc
MD
929 if (object->flags & OBJ_MIGHTBEDIRTY)
930 vnodes_skipped++;
931 vput(vp);
932 continue;
933 }
934 }
935
936 /*
937 * If a page is dirty, then it is either being washed
938 * (but not yet cleaned) or it is still in the
939 * laundry. If it is still in the laundry, then we
940 * start the cleaning operation.
941 *
942 * This operation may cluster, invalidating the 'next'
943 * pointer. To prevent an inordinate number of
944 * restarts we use our marker to remember our place.
945 *
946 * decrement page_shortage on success to account for
947 * the (future) cleaned page. Otherwise we could wind
948 * up laundering or cleaning too many pages.
949 */
984263bc 950 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
984263bc
MD
951 if (vm_pageout_clean(m) != 0) {
952 --page_shortage;
953 --maxlaunder;
954 }
984263bc
MD
955 next = TAILQ_NEXT(&marker, pageq);
956 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
984263bc
MD
957 if (vp != NULL)
958 vput(vp);
959 }
960 }
961
962 /*
963 * Compute the number of pages we want to try to move from the
964 * active queue to the inactive queue.
965 */
966 page_shortage = vm_paging_target() +
12e4aaff 967 vmstats.v_inactive_target - vmstats.v_inactive_count;
984263bc
MD
968 page_shortage += addl_page_shortage;
969
970 /*
971 * Scan the active queue for things we can deactivate. We nominally
972 * track the per-page activity counter and use it to locate
973 * deactivation candidates.
06ecca5a 974 *
5fd012e0 975 * NOTE: we are still in a critical section.
984263bc 976 */
12e4aaff 977 pcount = vmstats.v_active_count;
984263bc
MD
978 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
979
980 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
06ecca5a
MD
981 /*
982 * Give interrupts a chance.
983 */
5fd012e0
MD
984 crit_exit();
985 crit_enter();
984263bc
MD
986
987 /*
06ecca5a 988 * If the page was ripped out from under us, just stop.
984263bc 989 */
06ecca5a 990 if (m->queue != PQ_ACTIVE)
984263bc 991 break;
984263bc 992 next = TAILQ_NEXT(m, pageq);
06ecca5a 993
984263bc
MD
994 /*
995 * Don't deactivate pages that are busy.
996 */
997 if ((m->busy != 0) ||
998 (m->flags & PG_BUSY) ||
999 (m->hold_count != 0)) {
984263bc
MD
1000 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1001 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1002 m = next;
1003 continue;
1004 }
1005
1006 /*
1007 * The count for pagedaemon pages is done after checking the
1008 * page for eligibility...
1009 */
12e4aaff 1010 mycpu->gd_cnt.v_pdpages++;
984263bc
MD
1011
1012 /*
1013 * Check to see "how much" the page has been used.
1014 */
1015 actcount = 0;
1016 if (m->object->ref_count != 0) {
1017 if (m->flags & PG_REFERENCED) {
1018 actcount += 1;
1019 }
1020 actcount += pmap_ts_referenced(m);
1021 if (actcount) {
1022 m->act_count += ACT_ADVANCE + actcount;
1023 if (m->act_count > ACT_MAX)
1024 m->act_count = ACT_MAX;
1025 }
1026 }
1027
1028 /*
1029 * Since we have "tested" this bit, we need to clear it now.
1030 */
1031 vm_page_flag_clear(m, PG_REFERENCED);
1032
1033 /*
1034 * Only if an object is currently being used, do we use the
1035 * page activation count stats.
1036 */
1037 if (actcount && (m->object->ref_count != 0)) {
984263bc
MD
1038 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1039 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1040 } else {
1041 m->act_count -= min(m->act_count, ACT_DECLINE);
1042 if (vm_pageout_algorithm ||
1043 m->object->ref_count == 0 ||
1044 m->act_count == 0) {
1045 page_shortage--;
1046 if (m->object->ref_count == 0) {
1047 vm_page_protect(m, VM_PROT_NONE);
1048 if (m->dirty == 0)
1049 vm_page_cache(m);
1050 else
1051 vm_page_deactivate(m);
1052 } else {
1053 vm_page_deactivate(m);
1054 }
1055 } else {
984263bc
MD
1056 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1057 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1058 }
1059 }
1060 m = next;
1061 }
1062
984263bc
MD
1063 /*
1064 * We try to maintain some *really* free pages, this allows interrupt
1065 * code to be guaranteed space. Since both cache and free queues
1066 * are considered basically 'free', moving pages from cache to free
1067 * does not effect other calculations.
06ecca5a 1068 *
5fd012e0 1069 * NOTE: we are still in a critical section.
984263bc
MD
1070 */
1071
12e4aaff 1072 while (vmstats.v_free_count < vmstats.v_free_reserved) {
984263bc
MD
1073 static int cache_rover = 0;
1074 m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1075 if (!m)
1076 break;
1077 if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1078 m->busy ||
1079 m->hold_count ||
1080 m->wire_count) {
1081#ifdef INVARIANTS
1082 printf("Warning: busy page %p found in cache\n", m);
1083#endif
1084 vm_page_deactivate(m);
1085 continue;
1086 }
1087 cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1088 vm_pageout_page_free(m);
12e4aaff 1089 mycpu->gd_cnt.v_dfree++;
984263bc 1090 }
06ecca5a 1091
5fd012e0 1092 crit_exit();
984263bc
MD
1093
1094#if !defined(NO_SWAPPING)
1095 /*
1096 * Idle process swapout -- run once per second.
1097 */
1098 if (vm_swap_idle_enabled) {
1099 static long lsec;
1100 if (time_second != lsec) {
1101 vm_pageout_req_swapout |= VM_SWAP_IDLE;
1102 vm_req_vmdaemon();
1103 lsec = time_second;
1104 }
1105 }
1106#endif
1107
1108 /*
1109 * If we didn't get enough free pages, and we have skipped a vnode
1110 * in a writeable object, wakeup the sync daemon. And kick swapout
1111 * if we did not get enough free pages.
1112 */
1113 if (vm_paging_target() > 0) {
1114 if (vnodes_skipped && vm_page_count_min())
1115 (void) speedup_syncer();
1116#if !defined(NO_SWAPPING)
1117 if (vm_swap_enabled && vm_page_count_target()) {
1118 vm_req_vmdaemon();
1119 vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1120 }
1121#endif
1122 }
1123
1124 /*
1125 * If we are out of swap and were not able to reach our paging
1126 * target, kill the largest process.
1127 */
1128 if ((vm_swap_size < 64 && vm_page_count_min()) ||
1129 (swap_pager_full && vm_paging_target() > 0)) {
1130#if 0
1131 if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) {
1132#endif
1133 bigproc = NULL;
1134 bigsize = 0;
1135 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1136 /*
1137 * if this is a system process, skip it
1138 */
1139 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1140 ((p->p_pid < 48) && (vm_swap_size != 0))) {
1141 continue;
1142 }
1143 /*
1144 * if the process is in a non-running type state,
1145 * don't touch it.
1146 */
1147 if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1148 continue;
1149 }
1150 /*
1151 * get the process size
1152 */
1153 size = vmspace_resident_count(p->p_vmspace) +
1154 vmspace_swap_count(p->p_vmspace);
1155 /*
1156 * if the this process is bigger than the biggest one
1157 * remember it.
1158 */
1159 if (size > bigsize) {
1160 bigproc = p;
1161 bigsize = size;
1162 }
1163 }
1164 if (bigproc != NULL) {
1165 killproc(bigproc, "out of swap space");
984263bc 1166 bigproc->p_nice = PRIO_MIN;
a77ac49d 1167 bigproc->p_usched->resetpriority(bigproc);
12e4aaff 1168 wakeup(&vmstats.v_free_count);
984263bc
MD
1169 }
1170 }
1171}
1172
1173/*
1174 * This routine tries to maintain the pseudo LRU active queue,
1175 * so that during long periods of time where there is no paging,
1176 * that some statistic accumulation still occurs. This code
1177 * helps the situation where paging just starts to occur.
1178 */
1179static void
57e43348 1180vm_pageout_page_stats(void)
984263bc 1181{
984263bc
MD
1182 vm_page_t m,next;
1183 int pcount,tpcount; /* Number of pages to check */
1184 static int fullintervalcount = 0;
1185 int page_shortage;
984263bc
MD
1186
1187 page_shortage =
12e4aaff
MD
1188 (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1189 (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
984263bc
MD
1190
1191 if (page_shortage <= 0)
1192 return;
1193
5fd012e0 1194 crit_enter();
984263bc 1195
12e4aaff 1196 pcount = vmstats.v_active_count;
984263bc
MD
1197 fullintervalcount += vm_pageout_stats_interval;
1198 if (fullintervalcount < vm_pageout_full_stats_interval) {
12e4aaff 1199 tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
984263bc
MD
1200 if (pcount > tpcount)
1201 pcount = tpcount;
1202 } else {
1203 fullintervalcount = 0;
1204 }
1205
1206 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1207 while ((m != NULL) && (pcount-- > 0)) {
1208 int actcount;
1209
1210 if (m->queue != PQ_ACTIVE) {
1211 break;
1212 }
1213
1214 next = TAILQ_NEXT(m, pageq);
1215 /*
1216 * Don't deactivate pages that are busy.
1217 */
1218 if ((m->busy != 0) ||
1219 (m->flags & PG_BUSY) ||
1220 (m->hold_count != 0)) {
984263bc
MD
1221 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1222 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1223 m = next;
1224 continue;
1225 }
1226
1227 actcount = 0;
1228 if (m->flags & PG_REFERENCED) {
1229 vm_page_flag_clear(m, PG_REFERENCED);
1230 actcount += 1;
1231 }
1232
1233 actcount += pmap_ts_referenced(m);
1234 if (actcount) {
1235 m->act_count += ACT_ADVANCE + actcount;
1236 if (m->act_count > ACT_MAX)
1237 m->act_count = ACT_MAX;
984263bc
MD
1238 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1239 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1240 } else {
1241 if (m->act_count == 0) {
1242 /*
1243 * We turn off page access, so that we have
1244 * more accurate RSS stats. We don't do this
1245 * in the normal page deactivation when the
1246 * system is loaded VM wise, because the
1247 * cost of the large number of page protect
1248 * operations would be higher than the value
1249 * of doing the operation.
1250 */
1251 vm_page_protect(m, VM_PROT_NONE);
1252 vm_page_deactivate(m);
1253 } else {
1254 m->act_count -= min(m->act_count, ACT_DECLINE);
984263bc
MD
1255 TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1256 TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
984263bc
MD
1257 }
1258 }
1259
1260 m = next;
1261 }
5fd012e0 1262 crit_exit();
984263bc
MD
1263}
1264
1265static int
57e43348 1266vm_pageout_free_page_calc(vm_size_t count)
984263bc 1267{
12e4aaff 1268 if (count < vmstats.v_page_count)
984263bc
MD
1269 return 0;
1270 /*
1271 * free_reserved needs to include enough for the largest swap pager
1272 * structures plus enough for any pv_entry structs when paging.
1273 */
12e4aaff
MD
1274 if (vmstats.v_page_count > 1024)
1275 vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
984263bc 1276 else
12e4aaff
MD
1277 vmstats.v_free_min = 4;
1278 vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1279 vmstats.v_interrupt_free_min;
1280 vmstats.v_free_reserved = vm_pageout_page_count +
1281 vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1282 vmstats.v_free_severe = vmstats.v_free_min / 2;
1283 vmstats.v_free_min += vmstats.v_free_reserved;
1284 vmstats.v_free_severe += vmstats.v_free_reserved;
984263bc
MD
1285 return 1;
1286}
1287
1288
1289/*
1290 * vm_pageout is the high level pageout daemon.
1291 */
1292static void
57e43348 1293vm_pageout(void)
984263bc
MD
1294{
1295 int pass;
1296
1297 /*
1298 * Initialize some paging parameters.
1299 */
1300
12e4aaff
MD
1301 vmstats.v_interrupt_free_min = 2;
1302 if (vmstats.v_page_count < 2000)
984263bc
MD
1303 vm_pageout_page_count = 8;
1304
12e4aaff 1305 vm_pageout_free_page_calc(vmstats.v_page_count);
984263bc
MD
1306 /*
1307 * v_free_target and v_cache_min control pageout hysteresis. Note
1308 * that these are more a measure of the VM cache queue hysteresis
1309 * then the VM free queue. Specifically, v_free_target is the
1310 * high water mark (free+cache pages).
1311 *
1312 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1313 * low water mark, while v_free_min is the stop. v_cache_min must
1314 * be big enough to handle memory needs while the pageout daemon
1315 * is signalled and run to free more pages.
1316 */
12e4aaff
MD
1317 if (vmstats.v_free_count > 6144)
1318 vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
984263bc 1319 else
12e4aaff 1320 vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
984263bc 1321
12e4aaff
MD
1322 if (vmstats.v_free_count > 2048) {
1323 vmstats.v_cache_min = vmstats.v_free_target;
1324 vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1325 vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2;
984263bc 1326 } else {
12e4aaff
MD
1327 vmstats.v_cache_min = 0;
1328 vmstats.v_cache_max = 0;
1329 vmstats.v_inactive_target = vmstats.v_free_count / 4;
984263bc 1330 }
12e4aaff
MD
1331 if (vmstats.v_inactive_target > vmstats.v_free_count / 3)
1332 vmstats.v_inactive_target = vmstats.v_free_count / 3;
984263bc
MD
1333
1334 /* XXX does not really belong here */
1335 if (vm_page_max_wired == 0)
12e4aaff 1336 vm_page_max_wired = vmstats.v_free_count / 3;
984263bc
MD
1337
1338 if (vm_pageout_stats_max == 0)
12e4aaff 1339 vm_pageout_stats_max = vmstats.v_free_target;
984263bc
MD
1340
1341 /*
1342 * Set interval in seconds for stats scan.
1343 */
1344 if (vm_pageout_stats_interval == 0)
1345 vm_pageout_stats_interval = 5;
1346 if (vm_pageout_full_stats_interval == 0)
1347 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1348
1349
1350 /*
1351 * Set maximum free per pass
1352 */
1353 if (vm_pageout_stats_free_max == 0)
1354 vm_pageout_stats_free_max = 5;
1355
1356 swap_pager_swap_init();
1357 pass = 0;
1358 /*
1359 * The pageout daemon is never done, so loop forever.
1360 */
1361 while (TRUE) {
1362 int error;
984263bc
MD
1363
1364 /*
1365 * If we have enough free memory, wakeup waiters. Do
1366 * not clear vm_pages_needed until we reach our target,
1367 * otherwise we may be woken up over and over again and
1368 * waste a lot of cpu.
1369 */
5fd012e0 1370 crit_enter();
984263bc
MD
1371 if (vm_pages_needed && !vm_page_count_min()) {
1372 if (vm_paging_needed() <= 0)
1373 vm_pages_needed = 0;
12e4aaff 1374 wakeup(&vmstats.v_free_count);
984263bc
MD
1375 }
1376 if (vm_pages_needed) {
1377 /*
1378 * Still not done, take a second pass without waiting
1379 * (unlimited dirty cleaning), otherwise sleep a bit
1380 * and try again.
1381 */
1382 ++pass;
1383 if (pass > 1)
377d4740 1384 tsleep(&vm_pages_needed, 0, "psleep", hz/2);
984263bc
MD
1385 } else {
1386 /*
1387 * Good enough, sleep & handle stats. Prime the pass
1388 * for the next run.
1389 */
1390 if (pass > 1)
1391 pass = 1;
1392 else
1393 pass = 0;
1394 error = tsleep(&vm_pages_needed,
377d4740 1395 0, "psleep", vm_pageout_stats_interval * hz);
984263bc 1396 if (error && !vm_pages_needed) {
5fd012e0 1397 crit_exit();
984263bc
MD
1398 pass = 0;
1399 vm_pageout_page_stats();
1400 continue;
1401 }
1402 }
1403
1404 if (vm_pages_needed)
12e4aaff 1405 mycpu->gd_cnt.v_pdwakeups++;
5fd012e0 1406 crit_exit();
984263bc
MD
1407 vm_pageout_scan(pass);
1408 vm_pageout_deficit = 0;
1409 }
1410}
1411
1412void
57e43348 1413pagedaemon_wakeup(void)
984263bc 1414{
bc6dffab 1415 if (!vm_pages_needed && curthread != pagethread) {
984263bc
MD
1416 vm_pages_needed++;
1417 wakeup(&vm_pages_needed);
1418 }
1419}
1420
1421#if !defined(NO_SWAPPING)
1422static void
57e43348 1423vm_req_vmdaemon(void)
984263bc
MD
1424{
1425 static int lastrun = 0;
1426
1427 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1428 wakeup(&vm_daemon_needed);
1429 lastrun = ticks;
1430 }
1431}
1432
1433static void
57e43348 1434vm_daemon(void)
984263bc
MD
1435{
1436 struct proc *p;
1437
1438 while (TRUE) {
377d4740 1439 tsleep(&vm_daemon_needed, 0, "psleep", 0);
984263bc
MD
1440 if (vm_pageout_req_swapout) {
1441 swapout_procs(vm_pageout_req_swapout);
1442 vm_pageout_req_swapout = 0;
1443 }
1444 /*
1445 * scan the processes for exceeding their rlimits or if
1446 * process is swapped out -- deactivate pages
1447 */
1448
1449 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1450 vm_pindex_t limit, size;
1451
1452 /*
1453 * if this is a system process or if we have already
1454 * looked at this process, skip it.
1455 */
1456 if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1457 continue;
1458 }
1459 /*
1460 * if the process is in a non-running type state,
1461 * don't touch it.
1462 */
1463 if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1464 continue;
1465 }
1466 /*
1467 * get a limit
1468 */
1469 limit = OFF_TO_IDX(
1470 qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1471 p->p_rlimit[RLIMIT_RSS].rlim_max));
1472
1473 /*
1474 * let processes that are swapped out really be
1475 * swapped out set the limit to nothing (will force a
1476 * swap-out.)
1477 */
1478 if ((p->p_flag & P_INMEM) == 0)
1479 limit = 0; /* XXX */
1480
1481 size = vmspace_resident_count(p->p_vmspace);
1482 if (limit >= 0 && size >= limit) {
1483 vm_pageout_map_deactivate_pages(
1484 &p->p_vmspace->vm_map, limit);
1485 }
1486 }
1487 }
1488}
1489#endif