xref: /freebsd/sys/vm/vm_pageout.c (revision b2db760808f74bb53c232900091c9da801ebbfcc)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
77 
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
94 #include <sys/sx.h>
95 #include <sys/sysctl.h>
96 
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/uma.h>
107 
108 /*
109  * System initialization
110  */
111 
112 /* the kernel process "vm_pageout"*/
113 static void vm_pageout(void);
114 static int vm_pageout_clean(vm_page_t);
115 static void vm_pageout_scan(int pass);
116 
117 struct proc *pageproc;
118 
119 static struct kproc_desc page_kp = {
120 	"pagedaemon",
121 	vm_pageout,
122 	&pageproc
123 };
124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
125     &page_kp);
126 
127 #if !defined(NO_SWAPPING)
128 /* the kernel process "vm_daemon"*/
129 static void vm_daemon(void);
130 static struct	proc *vmproc;
131 
132 static struct kproc_desc vm_kp = {
133 	"vmdaemon",
134 	vm_daemon,
135 	&vmproc
136 };
137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
138 #endif
139 
140 
141 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
142 int vm_pageout_deficit;		/* Estimated number of pages deficit */
143 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
144 
145 #if !defined(NO_SWAPPING)
146 static int vm_pageout_req_swapout;	/* XXX */
147 static int vm_daemon_needed;
148 static struct mtx vm_daemon_mtx;
149 /* Allow for use by vm_pageout before vm_daemon is initialized. */
150 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
151 #endif
152 static int vm_max_launder = 32;
153 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
154 static int vm_pageout_full_stats_interval = 0;
155 static int vm_pageout_algorithm=0;
156 static int defer_swap_pageouts=0;
157 static int disable_swap_pageouts=0;
158 
159 #if defined(NO_SWAPPING)
160 static int vm_swap_enabled=0;
161 static int vm_swap_idle_enabled=0;
162 #else
163 static int vm_swap_enabled=1;
164 static int vm_swap_idle_enabled=0;
165 #endif
166 
167 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
168 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
169 
170 SYSCTL_INT(_vm, OID_AUTO, max_launder,
171 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
172 
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
174 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
175 
176 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
177 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
178 
179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
180 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
181 
182 #if defined(NO_SWAPPING)
183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
184 	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
186 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
187 #else
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
192 #endif
193 
194 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
195 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
196 
197 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
198 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
199 
200 static int pageout_lock_miss;
201 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
202 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
203 
204 #define VM_PAGEOUT_PAGE_COUNT 16
205 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
206 
207 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
208 SYSCTL_INT(_vm, OID_AUTO, max_wired,
209 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
210 
211 #if !defined(NO_SWAPPING)
212 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
213 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
214 static void vm_req_vmdaemon(int req);
215 #endif
216 static void vm_pageout_page_stats(void);
217 
218 static void
219 vm_pageout_init_marker(vm_page_t marker, u_short queue)
220 {
221 
222 	bzero(marker, sizeof(*marker));
223 	marker->flags = PG_FICTITIOUS | PG_MARKER;
224 	marker->oflags = VPO_BUSY;
225 	marker->queue = queue;
226 	marker->wire_count = 1;
227 }
228 
229 /*
230  * vm_pageout_fallback_object_lock:
231  *
232  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
233  * known to have failed and page queue must be either PQ_ACTIVE or
234  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
235  * while locking the vm object.  Use marker page to detect page queue
236  * changes and maintain notion of next page on page queue.  Return
237  * TRUE if no changes were detected, FALSE otherwise.  vm object is
238  * locked on return.
239  *
240  * This function depends on both the lock portion of struct vm_object
241  * and normal struct vm_page being type stable.
242  */
243 boolean_t
244 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
245 {
246 	struct vm_page marker;
247 	boolean_t unchanged;
248 	u_short queue;
249 	vm_object_t object;
250 
251 	queue = m->queue;
252 	vm_pageout_init_marker(&marker, queue);
253 	object = m->object;
254 
255 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
256 			   m, &marker, pageq);
257 	vm_page_unlock_queues();
258 	vm_page_unlock(m);
259 	VM_OBJECT_LOCK(object);
260 	vm_page_lock(m);
261 	vm_page_lock_queues();
262 
263 	/* Page queue might have changed. */
264 	*next = TAILQ_NEXT(&marker, pageq);
265 	unchanged = (m->queue == queue &&
266 		     m->object == object &&
267 		     &marker == TAILQ_NEXT(m, pageq));
268 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
269 		     &marker, pageq);
270 	return (unchanged);
271 }
272 
273 /*
274  * Lock the page while holding the page queue lock.  Use marker page
275  * to detect page queue changes and maintain notion of next page on
276  * page queue.  Return TRUE if no changes were detected, FALSE
277  * otherwise.  The page is locked on return. The page queue lock might
278  * be dropped and reacquired.
279  *
280  * This function depends on normal struct vm_page being type stable.
281  */
282 boolean_t
283 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
284 {
285 	struct vm_page marker;
286 	boolean_t unchanged;
287 	u_short queue;
288 
289 	vm_page_lock_assert(m, MA_NOTOWNED);
290 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
291 
292 	if (vm_page_trylock(m))
293 		return (TRUE);
294 
295 	queue = m->queue;
296 	vm_pageout_init_marker(&marker, queue);
297 
298 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq);
299 	vm_page_unlock_queues();
300 	vm_page_lock(m);
301 	vm_page_lock_queues();
302 
303 	/* Page queue might have changed. */
304 	*next = TAILQ_NEXT(&marker, pageq);
305 	unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
306 	TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq);
307 	return (unchanged);
308 }
309 
310 /*
311  * vm_pageout_clean:
312  *
313  * Clean the page and remove it from the laundry.
314  *
315  * We set the busy bit to cause potential page faults on this page to
316  * block.  Note the careful timing, however, the busy bit isn't set till
317  * late and we cannot do anything that will mess with the page.
318  */
319 static int
320 vm_pageout_clean(vm_page_t m)
321 {
322 	vm_object_t object;
323 	vm_page_t mc[2*vm_pageout_page_count], pb, ps;
324 	int pageout_count;
325 	int ib, is, page_base;
326 	vm_pindex_t pindex = m->pindex;
327 
328 	vm_page_lock_assert(m, MA_OWNED);
329 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
330 
331 	/*
332 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
333 	 * with the new swapper, but we could have serious problems paging
334 	 * out other object types if there is insufficient memory.
335 	 *
336 	 * Unfortunately, checking free memory here is far too late, so the
337 	 * check has been moved up a procedural level.
338 	 */
339 
340 	/*
341 	 * Can't clean the page if it's busy or held.
342 	 */
343 	KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
344 	    ("vm_pageout_clean: page %p is busy", m));
345 	KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
346 
347 	mc[vm_pageout_page_count] = pb = ps = m;
348 	pageout_count = 1;
349 	page_base = vm_pageout_page_count;
350 	ib = 1;
351 	is = 1;
352 
353 	/*
354 	 * Scan object for clusterable pages.
355 	 *
356 	 * We can cluster ONLY if: ->> the page is NOT
357 	 * clean, wired, busy, held, or mapped into a
358 	 * buffer, and one of the following:
359 	 * 1) The page is inactive, or a seldom used
360 	 *    active page.
361 	 * -or-
362 	 * 2) we force the issue.
363 	 *
364 	 * During heavy mmap/modification loads the pageout
365 	 * daemon can really fragment the underlying file
366 	 * due to flushing pages out of order and not trying
367 	 * align the clusters (which leave sporatic out-of-order
368 	 * holes).  To solve this problem we do the reverse scan
369 	 * first and attempt to align our cluster, then do a
370 	 * forward scan if room remains.
371 	 */
372 	object = m->object;
373 more:
374 	while (ib && pageout_count < vm_pageout_page_count) {
375 		vm_page_t p;
376 
377 		if (ib > pindex) {
378 			ib = 0;
379 			break;
380 		}
381 
382 		if ((p = vm_page_prev(pb)) == NULL ||
383 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
384 			ib = 0;
385 			break;
386 		}
387 		vm_page_lock(p);
388 		vm_page_test_dirty(p);
389 		if (p->dirty == 0 ||
390 		    p->queue != PQ_INACTIVE ||
391 		    p->hold_count != 0) {	/* may be undergoing I/O */
392 			vm_page_unlock(p);
393 			ib = 0;
394 			break;
395 		}
396 		vm_page_unlock(p);
397 		mc[--page_base] = pb = p;
398 		++pageout_count;
399 		++ib;
400 		/*
401 		 * alignment boundry, stop here and switch directions.  Do
402 		 * not clear ib.
403 		 */
404 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
405 			break;
406 	}
407 
408 	while (pageout_count < vm_pageout_page_count &&
409 	    pindex + is < object->size) {
410 		vm_page_t p;
411 
412 		if ((p = vm_page_next(ps)) == NULL ||
413 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
414 			break;
415 		vm_page_lock(p);
416 		vm_page_test_dirty(p);
417 		if (p->dirty == 0 ||
418 		    p->queue != PQ_INACTIVE ||
419 		    p->hold_count != 0) {	/* may be undergoing I/O */
420 			vm_page_unlock(p);
421 			break;
422 		}
423 		vm_page_unlock(p);
424 		mc[page_base + pageout_count] = ps = p;
425 		++pageout_count;
426 		++is;
427 	}
428 
429 	/*
430 	 * If we exhausted our forward scan, continue with the reverse scan
431 	 * when possible, even past a page boundry.  This catches boundry
432 	 * conditions.
433 	 */
434 	if (ib && pageout_count < vm_pageout_page_count)
435 		goto more;
436 
437 	vm_page_unlock(m);
438 	/*
439 	 * we allow reads during pageouts...
440 	 */
441 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
442 }
443 
444 /*
445  * vm_pageout_flush() - launder the given pages
446  *
447  *	The given pages are laundered.  Note that we setup for the start of
448  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
449  *	reference count all in here rather then in the parent.  If we want
450  *	the parent to do more sophisticated things we may have to change
451  *	the ordering.
452  */
453 int
454 vm_pageout_flush(vm_page_t *mc, int count, int flags)
455 {
456 	vm_object_t object = mc[0]->object;
457 	int pageout_status[count];
458 	int numpagedout = 0;
459 	int i;
460 
461 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
462 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
463 
464 	/*
465 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
466 	 * mark the pages read-only.
467 	 *
468 	 * We do not have to fixup the clean/dirty bits here... we can
469 	 * allow the pager to do it after the I/O completes.
470 	 *
471 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
472 	 * edge case with file fragments.
473 	 */
474 	for (i = 0; i < count; i++) {
475 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
476 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
477 			mc[i], i, count));
478 		vm_page_io_start(mc[i]);
479 		pmap_remove_write(mc[i]);
480 	}
481 	vm_object_pip_add(object, count);
482 
483 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
484 
485 	for (i = 0; i < count; i++) {
486 		vm_page_t mt = mc[i];
487 
488 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
489 		    (mt->flags & PG_WRITEABLE) == 0,
490 		    ("vm_pageout_flush: page %p is not write protected", mt));
491 		switch (pageout_status[i]) {
492 		case VM_PAGER_OK:
493 		case VM_PAGER_PEND:
494 			numpagedout++;
495 			break;
496 		case VM_PAGER_BAD:
497 			/*
498 			 * Page outside of range of object. Right now we
499 			 * essentially lose the changes by pretending it
500 			 * worked.
501 			 */
502 			vm_page_undirty(mt);
503 			break;
504 		case VM_PAGER_ERROR:
505 		case VM_PAGER_FAIL:
506 			/*
507 			 * If page couldn't be paged out, then reactivate the
508 			 * page so it doesn't clog the inactive list.  (We
509 			 * will try paging out it again later).
510 			 */
511 			vm_page_lock(mt);
512 			vm_page_activate(mt);
513 			vm_page_unlock(mt);
514 			break;
515 		case VM_PAGER_AGAIN:
516 			break;
517 		}
518 
519 		/*
520 		 * If the operation is still going, leave the page busy to
521 		 * block all other accesses. Also, leave the paging in
522 		 * progress indicator set so that we don't attempt an object
523 		 * collapse.
524 		 */
525 		if (pageout_status[i] != VM_PAGER_PEND) {
526 			vm_object_pip_wakeup(object);
527 			vm_page_io_finish(mt);
528 			if (vm_page_count_severe()) {
529 				vm_page_lock(mt);
530 				vm_page_try_to_cache(mt);
531 				vm_page_unlock(mt);
532 			}
533 		}
534 	}
535 	return (numpagedout);
536 }
537 
538 #if !defined(NO_SWAPPING)
539 /*
540  *	vm_pageout_object_deactivate_pages
541  *
542  *	Deactivate enough pages to satisfy the inactive target
543  *	requirements.
544  *
545  *	The object and map must be locked.
546  */
547 static void
548 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
549     long desired)
550 {
551 	vm_object_t backing_object, object;
552 	vm_page_t p;
553 	int actcount, remove_mode;
554 
555 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
556 	if (first_object->type == OBJT_DEVICE ||
557 	    first_object->type == OBJT_SG)
558 		return;
559 	for (object = first_object;; object = backing_object) {
560 		if (pmap_resident_count(pmap) <= desired)
561 			goto unlock_return;
562 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
563 		if (object->type == OBJT_PHYS || object->paging_in_progress)
564 			goto unlock_return;
565 
566 		remove_mode = 0;
567 		if (object->shadow_count > 1)
568 			remove_mode = 1;
569 		/*
570 		 * Scan the object's entire memory queue.
571 		 */
572 		TAILQ_FOREACH(p, &object->memq, listq) {
573 			if (pmap_resident_count(pmap) <= desired)
574 				goto unlock_return;
575 			if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
576 				continue;
577 			PCPU_INC(cnt.v_pdpages);
578 			vm_page_lock(p);
579 			if (p->wire_count != 0 || p->hold_count != 0 ||
580 			    !pmap_page_exists_quick(pmap, p)) {
581 				vm_page_unlock(p);
582 				continue;
583 			}
584 			actcount = pmap_ts_referenced(p);
585 			if ((p->flags & PG_REFERENCED) != 0) {
586 				if (actcount == 0)
587 					actcount = 1;
588 				vm_page_lock_queues();
589 				vm_page_flag_clear(p, PG_REFERENCED);
590 				vm_page_unlock_queues();
591 			}
592 			if (p->queue != PQ_ACTIVE && actcount != 0) {
593 				vm_page_activate(p);
594 				p->act_count += actcount;
595 			} else if (p->queue == PQ_ACTIVE) {
596 				if (actcount == 0) {
597 					p->act_count -= min(p->act_count,
598 					    ACT_DECLINE);
599 					if (!remove_mode &&
600 					    (vm_pageout_algorithm ||
601 					    p->act_count == 0)) {
602 						pmap_remove_all(p);
603 						vm_page_deactivate(p);
604 					} else {
605 						vm_page_lock_queues();
606 						vm_page_requeue(p);
607 						vm_page_unlock_queues();
608 					}
609 				} else {
610 					vm_page_activate(p);
611 					if (p->act_count < ACT_MAX -
612 					    ACT_ADVANCE)
613 						p->act_count += ACT_ADVANCE;
614 					vm_page_lock_queues();
615 					vm_page_requeue(p);
616 					vm_page_unlock_queues();
617 				}
618 			} else if (p->queue == PQ_INACTIVE)
619 				pmap_remove_all(p);
620 			vm_page_unlock(p);
621 		}
622 		if ((backing_object = object->backing_object) == NULL)
623 			goto unlock_return;
624 		VM_OBJECT_LOCK(backing_object);
625 		if (object != first_object)
626 			VM_OBJECT_UNLOCK(object);
627 	}
628 unlock_return:
629 	if (object != first_object)
630 		VM_OBJECT_UNLOCK(object);
631 }
632 
633 /*
634  * deactivate some number of pages in a map, try to do it fairly, but
635  * that is really hard to do.
636  */
637 static void
638 vm_pageout_map_deactivate_pages(map, desired)
639 	vm_map_t map;
640 	long desired;
641 {
642 	vm_map_entry_t tmpe;
643 	vm_object_t obj, bigobj;
644 	int nothingwired;
645 
646 	if (!vm_map_trylock(map))
647 		return;
648 
649 	bigobj = NULL;
650 	nothingwired = TRUE;
651 
652 	/*
653 	 * first, search out the biggest object, and try to free pages from
654 	 * that.
655 	 */
656 	tmpe = map->header.next;
657 	while (tmpe != &map->header) {
658 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
659 			obj = tmpe->object.vm_object;
660 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
661 				if (obj->shadow_count <= 1 &&
662 				    (bigobj == NULL ||
663 				     bigobj->resident_page_count < obj->resident_page_count)) {
664 					if (bigobj != NULL)
665 						VM_OBJECT_UNLOCK(bigobj);
666 					bigobj = obj;
667 				} else
668 					VM_OBJECT_UNLOCK(obj);
669 			}
670 		}
671 		if (tmpe->wired_count > 0)
672 			nothingwired = FALSE;
673 		tmpe = tmpe->next;
674 	}
675 
676 	if (bigobj != NULL) {
677 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
678 		VM_OBJECT_UNLOCK(bigobj);
679 	}
680 	/*
681 	 * Next, hunt around for other pages to deactivate.  We actually
682 	 * do this search sort of wrong -- .text first is not the best idea.
683 	 */
684 	tmpe = map->header.next;
685 	while (tmpe != &map->header) {
686 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
687 			break;
688 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
689 			obj = tmpe->object.vm_object;
690 			if (obj != NULL) {
691 				VM_OBJECT_LOCK(obj);
692 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
693 				VM_OBJECT_UNLOCK(obj);
694 			}
695 		}
696 		tmpe = tmpe->next;
697 	}
698 
699 	/*
700 	 * Remove all mappings if a process is swapped out, this will free page
701 	 * table pages.
702 	 */
703 	if (desired == 0 && nothingwired) {
704 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
705 		    vm_map_max(map));
706 	}
707 	vm_map_unlock(map);
708 }
709 #endif		/* !defined(NO_SWAPPING) */
710 
711 /*
712  *	vm_pageout_scan does the dirty work for the pageout daemon.
713  */
714 static void
715 vm_pageout_scan(int pass)
716 {
717 	vm_page_t m, next;
718 	struct vm_page marker;
719 	int page_shortage, maxscan, pcount;
720 	int addl_page_shortage, addl_page_shortage_init;
721 	vm_object_t object;
722 	int actcount;
723 	int vnodes_skipped = 0;
724 	int maxlaunder;
725 
726 	/*
727 	 * Decrease registered cache sizes.
728 	 */
729 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
730 	/*
731 	 * We do this explicitly after the caches have been drained above.
732 	 */
733 	uma_reclaim();
734 
735 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
736 
737 	/*
738 	 * Calculate the number of pages we want to either free or move
739 	 * to the cache.
740 	 */
741 	page_shortage = vm_paging_target() + addl_page_shortage_init;
742 
743 	vm_pageout_init_marker(&marker, PQ_INACTIVE);
744 
745 	/*
746 	 * Start scanning the inactive queue for pages we can move to the
747 	 * cache or free.  The scan will stop when the target is reached or
748 	 * we have scanned the entire inactive queue.  Note that m->act_count
749 	 * is not used to form decisions for the inactive queue, only for the
750 	 * active queue.
751 	 *
752 	 * maxlaunder limits the number of dirty pages we flush per scan.
753 	 * For most systems a smaller value (16 or 32) is more robust under
754 	 * extreme memory and disk pressure because any unnecessary writes
755 	 * to disk can result in extreme performance degredation.  However,
756 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
757 	 * used) will die horribly with limited laundering.  If the pageout
758 	 * daemon cannot clean enough pages in the first pass, we let it go
759 	 * all out in succeeding passes.
760 	 */
761 	if ((maxlaunder = vm_max_launder) <= 1)
762 		maxlaunder = 1;
763 	if (pass)
764 		maxlaunder = 10000;
765 	vm_page_lock_queues();
766 rescan0:
767 	addl_page_shortage = addl_page_shortage_init;
768 	maxscan = cnt.v_inactive_count;
769 
770 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
771 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
772 	     m = next) {
773 
774 		cnt.v_pdpages++;
775 
776 		if (m->queue != PQ_INACTIVE)
777 			goto rescan0;
778 
779 		next = TAILQ_NEXT(m, pageq);
780 
781 		/*
782 		 * skip marker pages
783 		 */
784 		if (m->flags & PG_MARKER)
785 			continue;
786 
787 		/*
788 		 * Lock the page.
789 		 */
790 		if (!vm_pageout_page_lock(m, &next)) {
791 			vm_page_unlock(m);
792 			addl_page_shortage++;
793 			continue;
794 		}
795 
796 		/*
797 		 * A held page may be undergoing I/O, so skip it.
798 		 */
799 		if (m->hold_count) {
800 			vm_page_unlock(m);
801 			vm_page_requeue(m);
802 			addl_page_shortage++;
803 			continue;
804 		}
805 
806 		/*
807 		 * Don't mess with busy pages, keep in the front of the
808 		 * queue, most likely are being paged out.
809 		 */
810 		object = m->object;
811 		if (!VM_OBJECT_TRYLOCK(object) &&
812 		    (!vm_pageout_fallback_object_lock(m, &next) ||
813 			m->hold_count != 0)) {
814 			VM_OBJECT_UNLOCK(object);
815 			vm_page_unlock(m);
816 			addl_page_shortage++;
817 			continue;
818 		}
819 		if (m->busy || (m->oflags & VPO_BUSY)) {
820 			vm_page_unlock(m);
821 			VM_OBJECT_UNLOCK(object);
822 			addl_page_shortage++;
823 			continue;
824 		}
825 
826 		/*
827 		 * If the object is not being used, we ignore previous
828 		 * references.
829 		 */
830 		if (object->ref_count == 0) {
831 			vm_page_flag_clear(m, PG_REFERENCED);
832 			KASSERT(!pmap_page_is_mapped(m),
833 			    ("vm_pageout_scan: page %p is mapped", m));
834 
835 		/*
836 		 * Otherwise, if the page has been referenced while in the
837 		 * inactive queue, we bump the "activation count" upwards,
838 		 * making it less likely that the page will be added back to
839 		 * the inactive queue prematurely again.  Here we check the
840 		 * page tables (or emulated bits, if any), given the upper
841 		 * level VM system not knowing anything about existing
842 		 * references.
843 		 */
844 		} else if (((m->flags & PG_REFERENCED) == 0) &&
845 			(actcount = pmap_ts_referenced(m))) {
846 			vm_page_activate(m);
847 			VM_OBJECT_UNLOCK(object);
848 			m->act_count += (actcount + ACT_ADVANCE);
849 			vm_page_unlock(m);
850 			continue;
851 		}
852 
853 		/*
854 		 * If the upper level VM system knows about any page
855 		 * references, we activate the page.  We also set the
856 		 * "activation count" higher than normal so that we will less
857 		 * likely place pages back onto the inactive queue again.
858 		 */
859 		if ((m->flags & PG_REFERENCED) != 0) {
860 			vm_page_flag_clear(m, PG_REFERENCED);
861 			actcount = pmap_ts_referenced(m);
862 			vm_page_activate(m);
863 			VM_OBJECT_UNLOCK(object);
864 			m->act_count += (actcount + ACT_ADVANCE + 1);
865 			vm_page_unlock(m);
866 			continue;
867 		}
868 
869 		/*
870 		 * If the upper level VM system does not believe that the page
871 		 * is fully dirty, but it is mapped for write access, then we
872 		 * consult the pmap to see if the page's dirty status should
873 		 * be updated.
874 		 */
875 		if (m->dirty != VM_PAGE_BITS_ALL &&
876 		    (m->flags & PG_WRITEABLE) != 0) {
877 			/*
878 			 * Avoid a race condition: Unless write access is
879 			 * removed from the page, another processor could
880 			 * modify it before all access is removed by the call
881 			 * to vm_page_cache() below.  If vm_page_cache() finds
882 			 * that the page has been modified when it removes all
883 			 * access, it panics because it cannot cache dirty
884 			 * pages.  In principle, we could eliminate just write
885 			 * access here rather than all access.  In the expected
886 			 * case, when there are no last instant modifications
887 			 * to the page, removing all access will be cheaper
888 			 * overall.
889 			 */
890 			if (pmap_is_modified(m))
891 				vm_page_dirty(m);
892 			else if (m->dirty == 0)
893 				pmap_remove_all(m);
894 		}
895 
896 		if (m->valid == 0) {
897 			/*
898 			 * Invalid pages can be easily freed
899 			 */
900 			vm_page_free(m);
901 			cnt.v_dfree++;
902 			--page_shortage;
903 		} else if (m->dirty == 0) {
904 			/*
905 			 * Clean pages can be placed onto the cache queue.
906 			 * This effectively frees them.
907 			 */
908 			vm_page_cache(m);
909 			--page_shortage;
910 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
911 			/*
912 			 * Dirty pages need to be paged out, but flushing
913 			 * a page is extremely expensive verses freeing
914 			 * a clean page.  Rather then artificially limiting
915 			 * the number of pages we can flush, we instead give
916 			 * dirty pages extra priority on the inactive queue
917 			 * by forcing them to be cycled through the queue
918 			 * twice before being flushed, after which the
919 			 * (now clean) page will cycle through once more
920 			 * before being freed.  This significantly extends
921 			 * the thrash point for a heavily loaded machine.
922 			 */
923 			vm_page_flag_set(m, PG_WINATCFLS);
924 			vm_page_requeue(m);
925 		} else if (maxlaunder > 0) {
926 			/*
927 			 * We always want to try to flush some dirty pages if
928 			 * we encounter them, to keep the system stable.
929 			 * Normally this number is small, but under extreme
930 			 * pressure where there are insufficient clean pages
931 			 * on the inactive queue, we may have to go all out.
932 			 */
933 			int swap_pageouts_ok, vfslocked = 0;
934 			struct vnode *vp = NULL;
935 			struct mount *mp = NULL;
936 
937 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
938 				swap_pageouts_ok = 1;
939 			} else {
940 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
941 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
942 				vm_page_count_min());
943 
944 			}
945 
946 			/*
947 			 * We don't bother paging objects that are "dead".
948 			 * Those objects are in a "rundown" state.
949 			 */
950 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
951 				vm_page_unlock(m);
952 				VM_OBJECT_UNLOCK(object);
953 				vm_page_requeue(m);
954 				continue;
955 			}
956 
957 			/*
958 			 * Following operations may unlock
959 			 * vm_page_queue_mtx, invalidating the 'next'
960 			 * pointer.  To prevent an inordinate number
961 			 * of restarts we use our marker to remember
962 			 * our place.
963 			 *
964 			 */
965 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
966 					   m, &marker, pageq);
967 			/*
968 			 * The object is already known NOT to be dead.   It
969 			 * is possible for the vget() to block the whole
970 			 * pageout daemon, but the new low-memory handling
971 			 * code should prevent it.
972 			 *
973 			 * The previous code skipped locked vnodes and, worse,
974 			 * reordered pages in the queue.  This results in
975 			 * completely non-deterministic operation and, on a
976 			 * busy system, can lead to extremely non-optimal
977 			 * pageouts.  For example, it can cause clean pages
978 			 * to be freed and dirty pages to be moved to the end
979 			 * of the queue.  Since dirty pages are also moved to
980 			 * the end of the queue once-cleaned, this gives
981 			 * way too large a weighting to defering the freeing
982 			 * of dirty pages.
983 			 *
984 			 * We can't wait forever for the vnode lock, we might
985 			 * deadlock due to a vn_read() getting stuck in
986 			 * vm_wait while holding this vnode.  We skip the
987 			 * vnode if we can't get it in a reasonable amount
988 			 * of time.
989 			 */
990 			if (object->type == OBJT_VNODE) {
991 				vm_page_unlock_queues();
992 				vm_page_unlock(m);
993 				vp = object->handle;
994 				if (vp->v_type == VREG &&
995 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
996 					mp = NULL;
997 					++pageout_lock_miss;
998 					if (object->flags & OBJ_MIGHTBEDIRTY)
999 						vnodes_skipped++;
1000 					vm_page_lock_queues();
1001 					goto unlock_and_continue;
1002 				}
1003 				KASSERT(mp != NULL,
1004 				    ("vp %p with NULL v_mount", vp));
1005 				vm_object_reference_locked(object);
1006 				VM_OBJECT_UNLOCK(object);
1007 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1008 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1009 				    curthread)) {
1010 					VM_OBJECT_LOCK(object);
1011 					vm_page_lock_queues();
1012 					++pageout_lock_miss;
1013 					if (object->flags & OBJ_MIGHTBEDIRTY)
1014 						vnodes_skipped++;
1015 					vp = NULL;
1016 					goto unlock_and_continue;
1017 				}
1018 				VM_OBJECT_LOCK(object);
1019 				vm_page_lock(m);
1020 				vm_page_lock_queues();
1021 				/*
1022 				 * The page might have been moved to another
1023 				 * queue during potential blocking in vget()
1024 				 * above.  The page might have been freed and
1025 				 * reused for another vnode.
1026 				 */
1027 				if (m->queue != PQ_INACTIVE ||
1028 				    m->object != object ||
1029 				    TAILQ_NEXT(m, pageq) != &marker) {
1030 					vm_page_unlock(m);
1031 					if (object->flags & OBJ_MIGHTBEDIRTY)
1032 						vnodes_skipped++;
1033 					goto unlock_and_continue;
1034 				}
1035 
1036 				/*
1037 				 * The page may have been busied during the
1038 				 * blocking in vget().  We don't move the
1039 				 * page back onto the end of the queue so that
1040 				 * statistics are more correct if we don't.
1041 				 */
1042 				if (m->busy || (m->oflags & VPO_BUSY)) {
1043 					vm_page_unlock(m);
1044 					goto unlock_and_continue;
1045 				}
1046 
1047 				/*
1048 				 * If the page has become held it might
1049 				 * be undergoing I/O, so skip it
1050 				 */
1051 				if (m->hold_count) {
1052 					vm_page_unlock(m);
1053 					vm_page_requeue(m);
1054 					if (object->flags & OBJ_MIGHTBEDIRTY)
1055 						vnodes_skipped++;
1056 					goto unlock_and_continue;
1057 				}
1058 			}
1059 
1060 			/*
1061 			 * If a page is dirty, then it is either being washed
1062 			 * (but not yet cleaned) or it is still in the
1063 			 * laundry.  If it is still in the laundry, then we
1064 			 * start the cleaning operation.
1065 			 *
1066 			 * decrement page_shortage on success to account for
1067 			 * the (future) cleaned page.  Otherwise we could wind
1068 			 * up laundering or cleaning too many pages.
1069 			 */
1070 			vm_page_unlock_queues();
1071 			if (vm_pageout_clean(m) != 0) {
1072 				--page_shortage;
1073 				--maxlaunder;
1074 			}
1075 			vm_page_lock_queues();
1076 unlock_and_continue:
1077 			vm_page_lock_assert(m, MA_NOTOWNED);
1078 			VM_OBJECT_UNLOCK(object);
1079 			if (mp != NULL) {
1080 				vm_page_unlock_queues();
1081 				if (vp != NULL)
1082 					vput(vp);
1083 				VFS_UNLOCK_GIANT(vfslocked);
1084 				vm_object_deallocate(object);
1085 				vn_finished_write(mp);
1086 				vm_page_lock_queues();
1087 			}
1088 			next = TAILQ_NEXT(&marker, pageq);
1089 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1090 				     &marker, pageq);
1091 			vm_page_lock_assert(m, MA_NOTOWNED);
1092 			continue;
1093 		}
1094 		vm_page_unlock(m);
1095 		VM_OBJECT_UNLOCK(object);
1096 	}
1097 
1098 	/*
1099 	 * Compute the number of pages we want to try to move from the
1100 	 * active queue to the inactive queue.
1101 	 */
1102 	page_shortage = vm_paging_target() +
1103 		cnt.v_inactive_target - cnt.v_inactive_count;
1104 	page_shortage += addl_page_shortage;
1105 
1106 	/*
1107 	 * Scan the active queue for things we can deactivate. We nominally
1108 	 * track the per-page activity counter and use it to locate
1109 	 * deactivation candidates.
1110 	 */
1111 	pcount = cnt.v_active_count;
1112 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1113 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1114 
1115 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1116 
1117 		KASSERT(m->queue == PQ_ACTIVE,
1118 		    ("vm_pageout_scan: page %p isn't active", m));
1119 
1120 		next = TAILQ_NEXT(m, pageq);
1121 		if ((m->flags & PG_MARKER) != 0) {
1122 			m = next;
1123 			continue;
1124 		}
1125 		if (!vm_pageout_page_lock(m, &next)) {
1126 			vm_page_unlock(m);
1127 			m = next;
1128 			continue;
1129 		}
1130 		object = m->object;
1131 		if (!VM_OBJECT_TRYLOCK(object) &&
1132 		    !vm_pageout_fallback_object_lock(m, &next)) {
1133 			VM_OBJECT_UNLOCK(object);
1134 			vm_page_unlock(m);
1135 			m = next;
1136 			continue;
1137 		}
1138 
1139 		/*
1140 		 * Don't deactivate pages that are busy.
1141 		 */
1142 		if ((m->busy != 0) ||
1143 		    (m->oflags & VPO_BUSY) ||
1144 		    (m->hold_count != 0)) {
1145 			vm_page_unlock(m);
1146 			VM_OBJECT_UNLOCK(object);
1147 			vm_page_requeue(m);
1148 			m = next;
1149 			continue;
1150 		}
1151 
1152 		/*
1153 		 * The count for pagedaemon pages is done after checking the
1154 		 * page for eligibility...
1155 		 */
1156 		cnt.v_pdpages++;
1157 
1158 		/*
1159 		 * Check to see "how much" the page has been used.
1160 		 */
1161 		actcount = 0;
1162 		if (object->ref_count != 0) {
1163 			if (m->flags & PG_REFERENCED) {
1164 				actcount += 1;
1165 			}
1166 			actcount += pmap_ts_referenced(m);
1167 			if (actcount) {
1168 				m->act_count += ACT_ADVANCE + actcount;
1169 				if (m->act_count > ACT_MAX)
1170 					m->act_count = ACT_MAX;
1171 			}
1172 		}
1173 
1174 		/*
1175 		 * Since we have "tested" this bit, we need to clear it now.
1176 		 */
1177 		vm_page_flag_clear(m, PG_REFERENCED);
1178 
1179 		/*
1180 		 * Only if an object is currently being used, do we use the
1181 		 * page activation count stats.
1182 		 */
1183 		if (actcount && (object->ref_count != 0)) {
1184 			vm_page_requeue(m);
1185 		} else {
1186 			m->act_count -= min(m->act_count, ACT_DECLINE);
1187 			if (vm_pageout_algorithm ||
1188 			    object->ref_count == 0 ||
1189 			    m->act_count == 0) {
1190 				page_shortage--;
1191 				if (object->ref_count == 0) {
1192 					KASSERT(!pmap_page_is_mapped(m),
1193 				    ("vm_pageout_scan: page %p is mapped", m));
1194 					if (m->dirty == 0)
1195 						vm_page_cache(m);
1196 					else
1197 						vm_page_deactivate(m);
1198 				} else {
1199 					vm_page_deactivate(m);
1200 				}
1201 			} else {
1202 				vm_page_requeue(m);
1203 			}
1204 		}
1205 		vm_page_unlock(m);
1206 		VM_OBJECT_UNLOCK(object);
1207 		m = next;
1208 	}
1209 	vm_page_unlock_queues();
1210 #if !defined(NO_SWAPPING)
1211 	/*
1212 	 * Idle process swapout -- run once per second.
1213 	 */
1214 	if (vm_swap_idle_enabled) {
1215 		static long lsec;
1216 		if (time_second != lsec) {
1217 			vm_req_vmdaemon(VM_SWAP_IDLE);
1218 			lsec = time_second;
1219 		}
1220 	}
1221 #endif
1222 
1223 	/*
1224 	 * If we didn't get enough free pages, and we have skipped a vnode
1225 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1226 	 * if we did not get enough free pages.
1227 	 */
1228 	if (vm_paging_target() > 0) {
1229 		if (vnodes_skipped && vm_page_count_min())
1230 			(void) speedup_syncer();
1231 #if !defined(NO_SWAPPING)
1232 		if (vm_swap_enabled && vm_page_count_target())
1233 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1234 #endif
1235 	}
1236 
1237 	/*
1238 	 * If we are critically low on one of RAM or swap and low on
1239 	 * the other, kill the largest process.  However, we avoid
1240 	 * doing this on the first pass in order to give ourselves a
1241 	 * chance to flush out dirty vnode-backed pages and to allow
1242 	 * active pages to be moved to the inactive queue and reclaimed.
1243 	 */
1244 	if (pass != 0 &&
1245 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1246 	     (swap_pager_full && vm_paging_target() > 0)))
1247 		vm_pageout_oom(VM_OOM_MEM);
1248 }
1249 
1250 
1251 void
1252 vm_pageout_oom(int shortage)
1253 {
1254 	struct proc *p, *bigproc;
1255 	vm_offset_t size, bigsize;
1256 	struct thread *td;
1257 	struct vmspace *vm;
1258 
1259 	/*
1260 	 * We keep the process bigproc locked once we find it to keep anyone
1261 	 * from messing with it; however, there is a possibility of
1262 	 * deadlock if process B is bigproc and one of it's child processes
1263 	 * attempts to propagate a signal to B while we are waiting for A's
1264 	 * lock while walking this list.  To avoid this, we don't block on
1265 	 * the process lock but just skip a process if it is already locked.
1266 	 */
1267 	bigproc = NULL;
1268 	bigsize = 0;
1269 	sx_slock(&allproc_lock);
1270 	FOREACH_PROC_IN_SYSTEM(p) {
1271 		int breakout;
1272 
1273 		if (PROC_TRYLOCK(p) == 0)
1274 			continue;
1275 		/*
1276 		 * If this is a system, protected or killed process, skip it.
1277 		 */
1278 		if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1279 		    (p->p_pid == 1) || P_KILLED(p) ||
1280 		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1281 			PROC_UNLOCK(p);
1282 			continue;
1283 		}
1284 		/*
1285 		 * If the process is in a non-running type state,
1286 		 * don't touch it.  Check all the threads individually.
1287 		 */
1288 		breakout = 0;
1289 		FOREACH_THREAD_IN_PROC(p, td) {
1290 			thread_lock(td);
1291 			if (!TD_ON_RUNQ(td) &&
1292 			    !TD_IS_RUNNING(td) &&
1293 			    !TD_IS_SLEEPING(td)) {
1294 				thread_unlock(td);
1295 				breakout = 1;
1296 				break;
1297 			}
1298 			thread_unlock(td);
1299 		}
1300 		if (breakout) {
1301 			PROC_UNLOCK(p);
1302 			continue;
1303 		}
1304 		/*
1305 		 * get the process size
1306 		 */
1307 		vm = vmspace_acquire_ref(p);
1308 		if (vm == NULL) {
1309 			PROC_UNLOCK(p);
1310 			continue;
1311 		}
1312 		if (!vm_map_trylock_read(&vm->vm_map)) {
1313 			vmspace_free(vm);
1314 			PROC_UNLOCK(p);
1315 			continue;
1316 		}
1317 		size = vmspace_swap_count(vm);
1318 		vm_map_unlock_read(&vm->vm_map);
1319 		if (shortage == VM_OOM_MEM)
1320 			size += vmspace_resident_count(vm);
1321 		vmspace_free(vm);
1322 		/*
1323 		 * if the this process is bigger than the biggest one
1324 		 * remember it.
1325 		 */
1326 		if (size > bigsize) {
1327 			if (bigproc != NULL)
1328 				PROC_UNLOCK(bigproc);
1329 			bigproc = p;
1330 			bigsize = size;
1331 		} else
1332 			PROC_UNLOCK(p);
1333 	}
1334 	sx_sunlock(&allproc_lock);
1335 	if (bigproc != NULL) {
1336 		killproc(bigproc, "out of swap space");
1337 		sched_nice(bigproc, PRIO_MIN);
1338 		PROC_UNLOCK(bigproc);
1339 		wakeup(&cnt.v_free_count);
1340 	}
1341 }
1342 
1343 /*
1344  * This routine tries to maintain the pseudo LRU active queue,
1345  * so that during long periods of time where there is no paging,
1346  * that some statistic accumulation still occurs.  This code
1347  * helps the situation where paging just starts to occur.
1348  */
1349 static void
1350 vm_pageout_page_stats()
1351 {
1352 	vm_object_t object;
1353 	vm_page_t m,next;
1354 	int pcount,tpcount;		/* Number of pages to check */
1355 	static int fullintervalcount = 0;
1356 	int page_shortage;
1357 
1358 	page_shortage =
1359 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1360 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1361 
1362 	if (page_shortage <= 0)
1363 		return;
1364 
1365 	vm_page_lock_queues();
1366 	pcount = cnt.v_active_count;
1367 	fullintervalcount += vm_pageout_stats_interval;
1368 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1369 		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1370 		    cnt.v_page_count;
1371 		if (pcount > tpcount)
1372 			pcount = tpcount;
1373 	} else {
1374 		fullintervalcount = 0;
1375 	}
1376 
1377 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1378 	while ((m != NULL) && (pcount-- > 0)) {
1379 		int actcount;
1380 
1381 		KASSERT(m->queue == PQ_ACTIVE,
1382 		    ("vm_pageout_page_stats: page %p isn't active", m));
1383 
1384 		next = TAILQ_NEXT(m, pageq);
1385 		if ((m->flags & PG_MARKER) != 0) {
1386 			m = next;
1387 			continue;
1388 		}
1389 		vm_page_lock_assert(m, MA_NOTOWNED);
1390 		if (!vm_pageout_page_lock(m, &next)) {
1391 			vm_page_unlock(m);
1392 			m = next;
1393 			continue;
1394 		}
1395 		object = m->object;
1396 		if (!VM_OBJECT_TRYLOCK(object) &&
1397 		    !vm_pageout_fallback_object_lock(m, &next)) {
1398 			VM_OBJECT_UNLOCK(object);
1399 			vm_page_unlock(m);
1400 			m = next;
1401 			continue;
1402 		}
1403 
1404 		/*
1405 		 * Don't deactivate pages that are busy.
1406 		 */
1407 		if ((m->busy != 0) ||
1408 		    (m->oflags & VPO_BUSY) ||
1409 		    (m->hold_count != 0)) {
1410 			vm_page_unlock(m);
1411 			VM_OBJECT_UNLOCK(object);
1412 			vm_page_requeue(m);
1413 			m = next;
1414 			continue;
1415 		}
1416 
1417 		actcount = 0;
1418 		if (m->flags & PG_REFERENCED) {
1419 			vm_page_flag_clear(m, PG_REFERENCED);
1420 			actcount += 1;
1421 		}
1422 
1423 		actcount += pmap_ts_referenced(m);
1424 		if (actcount) {
1425 			m->act_count += ACT_ADVANCE + actcount;
1426 			if (m->act_count > ACT_MAX)
1427 				m->act_count = ACT_MAX;
1428 			vm_page_requeue(m);
1429 		} else {
1430 			if (m->act_count == 0) {
1431 				/*
1432 				 * We turn off page access, so that we have
1433 				 * more accurate RSS stats.  We don't do this
1434 				 * in the normal page deactivation when the
1435 				 * system is loaded VM wise, because the
1436 				 * cost of the large number of page protect
1437 				 * operations would be higher than the value
1438 				 * of doing the operation.
1439 				 */
1440 				pmap_remove_all(m);
1441 				vm_page_deactivate(m);
1442 			} else {
1443 				m->act_count -= min(m->act_count, ACT_DECLINE);
1444 				vm_page_requeue(m);
1445 			}
1446 		}
1447 		vm_page_unlock(m);
1448 		VM_OBJECT_UNLOCK(object);
1449 		m = next;
1450 	}
1451 	vm_page_unlock_queues();
1452 }
1453 
1454 /*
1455  *	vm_pageout is the high level pageout daemon.
1456  */
1457 static void
1458 vm_pageout()
1459 {
1460 	int error, pass;
1461 
1462 	/*
1463 	 * Initialize some paging parameters.
1464 	 */
1465 	cnt.v_interrupt_free_min = 2;
1466 	if (cnt.v_page_count < 2000)
1467 		vm_pageout_page_count = 8;
1468 
1469 	/*
1470 	 * v_free_reserved needs to include enough for the largest
1471 	 * swap pager structures plus enough for any pv_entry structs
1472 	 * when paging.
1473 	 */
1474 	if (cnt.v_page_count > 1024)
1475 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1476 	else
1477 		cnt.v_free_min = 4;
1478 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1479 	    cnt.v_interrupt_free_min;
1480 	cnt.v_free_reserved = vm_pageout_page_count +
1481 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1482 	cnt.v_free_severe = cnt.v_free_min / 2;
1483 	cnt.v_free_min += cnt.v_free_reserved;
1484 	cnt.v_free_severe += cnt.v_free_reserved;
1485 
1486 	/*
1487 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1488 	 * that these are more a measure of the VM cache queue hysteresis
1489 	 * then the VM free queue.  Specifically, v_free_target is the
1490 	 * high water mark (free+cache pages).
1491 	 *
1492 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1493 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1494 	 * be big enough to handle memory needs while the pageout daemon
1495 	 * is signalled and run to free more pages.
1496 	 */
1497 	if (cnt.v_free_count > 6144)
1498 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1499 	else
1500 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1501 
1502 	if (cnt.v_free_count > 2048) {
1503 		cnt.v_cache_min = cnt.v_free_target;
1504 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1505 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1506 	} else {
1507 		cnt.v_cache_min = 0;
1508 		cnt.v_cache_max = 0;
1509 		cnt.v_inactive_target = cnt.v_free_count / 4;
1510 	}
1511 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1512 		cnt.v_inactive_target = cnt.v_free_count / 3;
1513 
1514 	/* XXX does not really belong here */
1515 	if (vm_page_max_wired == 0)
1516 		vm_page_max_wired = cnt.v_free_count / 3;
1517 
1518 	if (vm_pageout_stats_max == 0)
1519 		vm_pageout_stats_max = cnt.v_free_target;
1520 
1521 	/*
1522 	 * Set interval in seconds for stats scan.
1523 	 */
1524 	if (vm_pageout_stats_interval == 0)
1525 		vm_pageout_stats_interval = 5;
1526 	if (vm_pageout_full_stats_interval == 0)
1527 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1528 
1529 	swap_pager_swap_init();
1530 	pass = 0;
1531 	/*
1532 	 * The pageout daemon is never done, so loop forever.
1533 	 */
1534 	while (TRUE) {
1535 		/*
1536 		 * If we have enough free memory, wakeup waiters.  Do
1537 		 * not clear vm_pages_needed until we reach our target,
1538 		 * otherwise we may be woken up over and over again and
1539 		 * waste a lot of cpu.
1540 		 */
1541 		mtx_lock(&vm_page_queue_free_mtx);
1542 		if (vm_pages_needed && !vm_page_count_min()) {
1543 			if (!vm_paging_needed())
1544 				vm_pages_needed = 0;
1545 			wakeup(&cnt.v_free_count);
1546 		}
1547 		if (vm_pages_needed) {
1548 			/*
1549 			 * Still not done, take a second pass without waiting
1550 			 * (unlimited dirty cleaning), otherwise sleep a bit
1551 			 * and try again.
1552 			 */
1553 			++pass;
1554 			if (pass > 1)
1555 				msleep(&vm_pages_needed,
1556 				    &vm_page_queue_free_mtx, PVM, "psleep",
1557 				    hz / 2);
1558 		} else {
1559 			/*
1560 			 * Good enough, sleep & handle stats.  Prime the pass
1561 			 * for the next run.
1562 			 */
1563 			if (pass > 1)
1564 				pass = 1;
1565 			else
1566 				pass = 0;
1567 			error = msleep(&vm_pages_needed,
1568 			    &vm_page_queue_free_mtx, PVM, "psleep",
1569 			    vm_pageout_stats_interval * hz);
1570 			if (error && !vm_pages_needed) {
1571 				mtx_unlock(&vm_page_queue_free_mtx);
1572 				pass = 0;
1573 				vm_pageout_page_stats();
1574 				continue;
1575 			}
1576 		}
1577 		if (vm_pages_needed)
1578 			cnt.v_pdwakeups++;
1579 		mtx_unlock(&vm_page_queue_free_mtx);
1580 		vm_pageout_scan(pass);
1581 	}
1582 }
1583 
1584 /*
1585  * Unless the free page queue lock is held by the caller, this function
1586  * should be regarded as advisory.  Specifically, the caller should
1587  * not msleep() on &cnt.v_free_count following this function unless
1588  * the free page queue lock is held until the msleep() is performed.
1589  */
1590 void
1591 pagedaemon_wakeup()
1592 {
1593 
1594 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1595 		vm_pages_needed = 1;
1596 		wakeup(&vm_pages_needed);
1597 	}
1598 }
1599 
1600 #if !defined(NO_SWAPPING)
1601 static void
1602 vm_req_vmdaemon(int req)
1603 {
1604 	static int lastrun = 0;
1605 
1606 	mtx_lock(&vm_daemon_mtx);
1607 	vm_pageout_req_swapout |= req;
1608 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1609 		wakeup(&vm_daemon_needed);
1610 		lastrun = ticks;
1611 	}
1612 	mtx_unlock(&vm_daemon_mtx);
1613 }
1614 
1615 static void
1616 vm_daemon()
1617 {
1618 	struct rlimit rsslim;
1619 	struct proc *p;
1620 	struct thread *td;
1621 	struct vmspace *vm;
1622 	int breakout, swapout_flags;
1623 
1624 	while (TRUE) {
1625 		mtx_lock(&vm_daemon_mtx);
1626 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1627 		swapout_flags = vm_pageout_req_swapout;
1628 		vm_pageout_req_swapout = 0;
1629 		mtx_unlock(&vm_daemon_mtx);
1630 		if (swapout_flags)
1631 			swapout_procs(swapout_flags);
1632 
1633 		/*
1634 		 * scan the processes for exceeding their rlimits or if
1635 		 * process is swapped out -- deactivate pages
1636 		 */
1637 		sx_slock(&allproc_lock);
1638 		FOREACH_PROC_IN_SYSTEM(p) {
1639 			vm_pindex_t limit, size;
1640 
1641 			/*
1642 			 * if this is a system process or if we have already
1643 			 * looked at this process, skip it.
1644 			 */
1645 			PROC_LOCK(p);
1646 			if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1647 				PROC_UNLOCK(p);
1648 				continue;
1649 			}
1650 			/*
1651 			 * if the process is in a non-running type state,
1652 			 * don't touch it.
1653 			 */
1654 			breakout = 0;
1655 			FOREACH_THREAD_IN_PROC(p, td) {
1656 				thread_lock(td);
1657 				if (!TD_ON_RUNQ(td) &&
1658 				    !TD_IS_RUNNING(td) &&
1659 				    !TD_IS_SLEEPING(td)) {
1660 					thread_unlock(td);
1661 					breakout = 1;
1662 					break;
1663 				}
1664 				thread_unlock(td);
1665 			}
1666 			if (breakout) {
1667 				PROC_UNLOCK(p);
1668 				continue;
1669 			}
1670 			/*
1671 			 * get a limit
1672 			 */
1673 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1674 			limit = OFF_TO_IDX(
1675 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1676 
1677 			/*
1678 			 * let processes that are swapped out really be
1679 			 * swapped out set the limit to nothing (will force a
1680 			 * swap-out.)
1681 			 */
1682 			if ((p->p_flag & P_INMEM) == 0)
1683 				limit = 0;	/* XXX */
1684 			vm = vmspace_acquire_ref(p);
1685 			PROC_UNLOCK(p);
1686 			if (vm == NULL)
1687 				continue;
1688 
1689 			size = vmspace_resident_count(vm);
1690 			if (limit >= 0 && size >= limit) {
1691 				vm_pageout_map_deactivate_pages(
1692 				    &vm->vm_map, limit);
1693 			}
1694 			vmspace_free(vm);
1695 		}
1696 		sx_sunlock(&allproc_lock);
1697 	}
1698 }
1699 #endif			/* !defined(NO_SWAPPING) */
1700