xref: /freebsd/sys/vm/vm_pageout.c (revision aa0a1e58f0189b0fde359a8bda032887e72057fa)
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 	object = m->object;
330 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
331 
332 	/*
333 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
334 	 * with the new swapper, but we could have serious problems paging
335 	 * out other object types if there is insufficient memory.
336 	 *
337 	 * Unfortunately, checking free memory here is far too late, so the
338 	 * check has been moved up a procedural level.
339 	 */
340 
341 	/*
342 	 * Can't clean the page if it's busy or held.
343 	 */
344 	KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
345 	    ("vm_pageout_clean: page %p is busy", m));
346 	KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
347 	vm_page_unlock(m);
348 
349 	mc[vm_pageout_page_count] = pb = ps = m;
350 	pageout_count = 1;
351 	page_base = vm_pageout_page_count;
352 	ib = 1;
353 	is = 1;
354 
355 	/*
356 	 * Scan object for clusterable pages.
357 	 *
358 	 * We can cluster ONLY if: ->> the page is NOT
359 	 * clean, wired, busy, held, or mapped into a
360 	 * buffer, and one of the following:
361 	 * 1) The page is inactive, or a seldom used
362 	 *    active page.
363 	 * -or-
364 	 * 2) we force the issue.
365 	 *
366 	 * During heavy mmap/modification loads the pageout
367 	 * daemon can really fragment the underlying file
368 	 * due to flushing pages out of order and not trying
369 	 * align the clusters (which leave sporatic out-of-order
370 	 * holes).  To solve this problem we do the reverse scan
371 	 * first and attempt to align our cluster, then do a
372 	 * forward scan if room remains.
373 	 */
374 more:
375 	while (ib && pageout_count < vm_pageout_page_count) {
376 		vm_page_t p;
377 
378 		if (ib > pindex) {
379 			ib = 0;
380 			break;
381 		}
382 
383 		if ((p = vm_page_prev(pb)) == NULL ||
384 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
385 			ib = 0;
386 			break;
387 		}
388 		vm_page_lock(p);
389 		vm_page_test_dirty(p);
390 		if (p->dirty == 0 ||
391 		    p->queue != PQ_INACTIVE ||
392 		    p->hold_count != 0) {	/* may be undergoing I/O */
393 			vm_page_unlock(p);
394 			ib = 0;
395 			break;
396 		}
397 		vm_page_unlock(p);
398 		mc[--page_base] = pb = p;
399 		++pageout_count;
400 		++ib;
401 		/*
402 		 * alignment boundry, stop here and switch directions.  Do
403 		 * not clear ib.
404 		 */
405 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
406 			break;
407 	}
408 
409 	while (pageout_count < vm_pageout_page_count &&
410 	    pindex + is < object->size) {
411 		vm_page_t p;
412 
413 		if ((p = vm_page_next(ps)) == NULL ||
414 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
415 			break;
416 		vm_page_lock(p);
417 		vm_page_test_dirty(p);
418 		if (p->dirty == 0 ||
419 		    p->queue != PQ_INACTIVE ||
420 		    p->hold_count != 0) {	/* may be undergoing I/O */
421 			vm_page_unlock(p);
422 			break;
423 		}
424 		vm_page_unlock(p);
425 		mc[page_base + pageout_count] = ps = p;
426 		++pageout_count;
427 		++is;
428 	}
429 
430 	/*
431 	 * If we exhausted our forward scan, continue with the reverse scan
432 	 * when possible, even past a page boundry.  This catches boundry
433 	 * conditions.
434 	 */
435 	if (ib && pageout_count < vm_pageout_page_count)
436 		goto more;
437 
438 	/*
439 	 * we allow reads during pageouts...
440 	 */
441 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL));
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  *	Returned runlen is the count of pages between mreq and first
454  *	page after mreq with status VM_PAGER_AGAIN.
455  */
456 int
457 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen)
458 {
459 	vm_object_t object = mc[0]->object;
460 	int pageout_status[count];
461 	int numpagedout = 0;
462 	int i, runlen;
463 
464 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
465 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
466 
467 	/*
468 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
469 	 * mark the pages read-only.
470 	 *
471 	 * We do not have to fixup the clean/dirty bits here... we can
472 	 * allow the pager to do it after the I/O completes.
473 	 *
474 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
475 	 * edge case with file fragments.
476 	 */
477 	for (i = 0; i < count; i++) {
478 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
479 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
480 			mc[i], i, count));
481 		vm_page_io_start(mc[i]);
482 		pmap_remove_write(mc[i]);
483 	}
484 	vm_object_pip_add(object, count);
485 
486 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
487 
488 	runlen = count - mreq;
489 	for (i = 0; i < count; i++) {
490 		vm_page_t mt = mc[i];
491 
492 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
493 		    (mt->flags & PG_WRITEABLE) == 0,
494 		    ("vm_pageout_flush: page %p is not write protected", mt));
495 		switch (pageout_status[i]) {
496 		case VM_PAGER_OK:
497 		case VM_PAGER_PEND:
498 			numpagedout++;
499 			break;
500 		case VM_PAGER_BAD:
501 			/*
502 			 * Page outside of range of object. Right now we
503 			 * essentially lose the changes by pretending it
504 			 * worked.
505 			 */
506 			vm_page_undirty(mt);
507 			break;
508 		case VM_PAGER_ERROR:
509 		case VM_PAGER_FAIL:
510 			/*
511 			 * If page couldn't be paged out, then reactivate the
512 			 * page so it doesn't clog the inactive list.  (We
513 			 * will try paging out it again later).
514 			 */
515 			vm_page_lock(mt);
516 			vm_page_activate(mt);
517 			vm_page_unlock(mt);
518 			break;
519 		case VM_PAGER_AGAIN:
520 			if (i >= mreq && i - mreq < runlen)
521 				runlen = i - mreq;
522 			break;
523 		}
524 
525 		/*
526 		 * If the operation is still going, leave the page busy to
527 		 * block all other accesses. Also, leave the paging in
528 		 * progress indicator set so that we don't attempt an object
529 		 * collapse.
530 		 */
531 		if (pageout_status[i] != VM_PAGER_PEND) {
532 			vm_object_pip_wakeup(object);
533 			vm_page_io_finish(mt);
534 			if (vm_page_count_severe()) {
535 				vm_page_lock(mt);
536 				vm_page_try_to_cache(mt);
537 				vm_page_unlock(mt);
538 			}
539 		}
540 	}
541 	if (prunlen != NULL)
542 		*prunlen = runlen;
543 	return (numpagedout);
544 }
545 
546 #if !defined(NO_SWAPPING)
547 /*
548  *	vm_pageout_object_deactivate_pages
549  *
550  *	Deactivate enough pages to satisfy the inactive target
551  *	requirements.
552  *
553  *	The object and map must be locked.
554  */
555 static void
556 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
557     long desired)
558 {
559 	vm_object_t backing_object, object;
560 	vm_page_t p;
561 	int actcount, remove_mode;
562 
563 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
564 	if (first_object->type == OBJT_DEVICE ||
565 	    first_object->type == OBJT_SG)
566 		return;
567 	for (object = first_object;; object = backing_object) {
568 		if (pmap_resident_count(pmap) <= desired)
569 			goto unlock_return;
570 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
571 		if (object->type == OBJT_PHYS || object->paging_in_progress)
572 			goto unlock_return;
573 
574 		remove_mode = 0;
575 		if (object->shadow_count > 1)
576 			remove_mode = 1;
577 		/*
578 		 * Scan the object's entire memory queue.
579 		 */
580 		TAILQ_FOREACH(p, &object->memq, listq) {
581 			if (pmap_resident_count(pmap) <= desired)
582 				goto unlock_return;
583 			if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
584 				continue;
585 			PCPU_INC(cnt.v_pdpages);
586 			vm_page_lock(p);
587 			if (p->wire_count != 0 || p->hold_count != 0 ||
588 			    !pmap_page_exists_quick(pmap, p)) {
589 				vm_page_unlock(p);
590 				continue;
591 			}
592 			actcount = pmap_ts_referenced(p);
593 			if ((p->flags & PG_REFERENCED) != 0) {
594 				if (actcount == 0)
595 					actcount = 1;
596 				vm_page_lock_queues();
597 				vm_page_flag_clear(p, PG_REFERENCED);
598 				vm_page_unlock_queues();
599 			}
600 			if (p->queue != PQ_ACTIVE && actcount != 0) {
601 				vm_page_activate(p);
602 				p->act_count += actcount;
603 			} else if (p->queue == PQ_ACTIVE) {
604 				if (actcount == 0) {
605 					p->act_count -= min(p->act_count,
606 					    ACT_DECLINE);
607 					if (!remove_mode &&
608 					    (vm_pageout_algorithm ||
609 					    p->act_count == 0)) {
610 						pmap_remove_all(p);
611 						vm_page_deactivate(p);
612 					} else {
613 						vm_page_lock_queues();
614 						vm_page_requeue(p);
615 						vm_page_unlock_queues();
616 					}
617 				} else {
618 					vm_page_activate(p);
619 					if (p->act_count < ACT_MAX -
620 					    ACT_ADVANCE)
621 						p->act_count += ACT_ADVANCE;
622 					vm_page_lock_queues();
623 					vm_page_requeue(p);
624 					vm_page_unlock_queues();
625 				}
626 			} else if (p->queue == PQ_INACTIVE)
627 				pmap_remove_all(p);
628 			vm_page_unlock(p);
629 		}
630 		if ((backing_object = object->backing_object) == NULL)
631 			goto unlock_return;
632 		VM_OBJECT_LOCK(backing_object);
633 		if (object != first_object)
634 			VM_OBJECT_UNLOCK(object);
635 	}
636 unlock_return:
637 	if (object != first_object)
638 		VM_OBJECT_UNLOCK(object);
639 }
640 
641 /*
642  * deactivate some number of pages in a map, try to do it fairly, but
643  * that is really hard to do.
644  */
645 static void
646 vm_pageout_map_deactivate_pages(map, desired)
647 	vm_map_t map;
648 	long desired;
649 {
650 	vm_map_entry_t tmpe;
651 	vm_object_t obj, bigobj;
652 	int nothingwired;
653 
654 	if (!vm_map_trylock(map))
655 		return;
656 
657 	bigobj = NULL;
658 	nothingwired = TRUE;
659 
660 	/*
661 	 * first, search out the biggest object, and try to free pages from
662 	 * that.
663 	 */
664 	tmpe = map->header.next;
665 	while (tmpe != &map->header) {
666 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
667 			obj = tmpe->object.vm_object;
668 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
669 				if (obj->shadow_count <= 1 &&
670 				    (bigobj == NULL ||
671 				     bigobj->resident_page_count < obj->resident_page_count)) {
672 					if (bigobj != NULL)
673 						VM_OBJECT_UNLOCK(bigobj);
674 					bigobj = obj;
675 				} else
676 					VM_OBJECT_UNLOCK(obj);
677 			}
678 		}
679 		if (tmpe->wired_count > 0)
680 			nothingwired = FALSE;
681 		tmpe = tmpe->next;
682 	}
683 
684 	if (bigobj != NULL) {
685 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
686 		VM_OBJECT_UNLOCK(bigobj);
687 	}
688 	/*
689 	 * Next, hunt around for other pages to deactivate.  We actually
690 	 * do this search sort of wrong -- .text first is not the best idea.
691 	 */
692 	tmpe = map->header.next;
693 	while (tmpe != &map->header) {
694 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
695 			break;
696 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
697 			obj = tmpe->object.vm_object;
698 			if (obj != NULL) {
699 				VM_OBJECT_LOCK(obj);
700 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
701 				VM_OBJECT_UNLOCK(obj);
702 			}
703 		}
704 		tmpe = tmpe->next;
705 	}
706 
707 	/*
708 	 * Remove all mappings if a process is swapped out, this will free page
709 	 * table pages.
710 	 */
711 	if (desired == 0 && nothingwired) {
712 		tmpe = map->header.next;
713 		while (tmpe != &map->header) {
714 			pmap_remove(vm_map_pmap(map), tmpe->start, tmpe->end);
715 			tmpe = tmpe->next;
716 		}
717 	}
718 	vm_map_unlock(map);
719 }
720 #endif		/* !defined(NO_SWAPPING) */
721 
722 /*
723  *	vm_pageout_scan does the dirty work for the pageout daemon.
724  */
725 static void
726 vm_pageout_scan(int pass)
727 {
728 	vm_page_t m, next;
729 	struct vm_page marker;
730 	int page_shortage, maxscan, pcount;
731 	int addl_page_shortage, addl_page_shortage_init;
732 	vm_object_t object;
733 	int actcount;
734 	int vnodes_skipped = 0;
735 	int maxlaunder;
736 
737 	/*
738 	 * Decrease registered cache sizes.
739 	 */
740 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
741 	/*
742 	 * We do this explicitly after the caches have been drained above.
743 	 */
744 	uma_reclaim();
745 
746 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
747 
748 	/*
749 	 * Calculate the number of pages we want to either free or move
750 	 * to the cache.
751 	 */
752 	page_shortage = vm_paging_target() + addl_page_shortage_init;
753 
754 	vm_pageout_init_marker(&marker, PQ_INACTIVE);
755 
756 	/*
757 	 * Start scanning the inactive queue for pages we can move to the
758 	 * cache or free.  The scan will stop when the target is reached or
759 	 * we have scanned the entire inactive queue.  Note that m->act_count
760 	 * is not used to form decisions for the inactive queue, only for the
761 	 * active queue.
762 	 *
763 	 * maxlaunder limits the number of dirty pages we flush per scan.
764 	 * For most systems a smaller value (16 or 32) is more robust under
765 	 * extreme memory and disk pressure because any unnecessary writes
766 	 * to disk can result in extreme performance degredation.  However,
767 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
768 	 * used) will die horribly with limited laundering.  If the pageout
769 	 * daemon cannot clean enough pages in the first pass, we let it go
770 	 * all out in succeeding passes.
771 	 */
772 	if ((maxlaunder = vm_max_launder) <= 1)
773 		maxlaunder = 1;
774 	if (pass)
775 		maxlaunder = 10000;
776 	vm_page_lock_queues();
777 rescan0:
778 	addl_page_shortage = addl_page_shortage_init;
779 	maxscan = cnt.v_inactive_count;
780 
781 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
782 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
783 	     m = next) {
784 
785 		cnt.v_pdpages++;
786 
787 		if (m->queue != PQ_INACTIVE)
788 			goto rescan0;
789 
790 		next = TAILQ_NEXT(m, pageq);
791 
792 		/*
793 		 * skip marker pages
794 		 */
795 		if (m->flags & PG_MARKER)
796 			continue;
797 
798 		/*
799 		 * Lock the page.
800 		 */
801 		if (!vm_pageout_page_lock(m, &next)) {
802 			vm_page_unlock(m);
803 			addl_page_shortage++;
804 			continue;
805 		}
806 
807 		/*
808 		 * A held page may be undergoing I/O, so skip it.
809 		 */
810 		if (m->hold_count) {
811 			vm_page_unlock(m);
812 			vm_page_requeue(m);
813 			addl_page_shortage++;
814 			continue;
815 		}
816 
817 		/*
818 		 * Don't mess with busy pages, keep in the front of the
819 		 * queue, most likely are being paged out.
820 		 */
821 		object = m->object;
822 		if (!VM_OBJECT_TRYLOCK(object) &&
823 		    (!vm_pageout_fallback_object_lock(m, &next) ||
824 			m->hold_count != 0)) {
825 			VM_OBJECT_UNLOCK(object);
826 			vm_page_unlock(m);
827 			addl_page_shortage++;
828 			continue;
829 		}
830 		if (m->busy || (m->oflags & VPO_BUSY)) {
831 			vm_page_unlock(m);
832 			VM_OBJECT_UNLOCK(object);
833 			addl_page_shortage++;
834 			continue;
835 		}
836 
837 		/*
838 		 * If the object is not being used, we ignore previous
839 		 * references.
840 		 */
841 		if (object->ref_count == 0) {
842 			vm_page_flag_clear(m, PG_REFERENCED);
843 			KASSERT(!pmap_page_is_mapped(m),
844 			    ("vm_pageout_scan: page %p is mapped", m));
845 
846 		/*
847 		 * Otherwise, if the page has been referenced while in the
848 		 * inactive queue, we bump the "activation count" upwards,
849 		 * making it less likely that the page will be added back to
850 		 * the inactive queue prematurely again.  Here we check the
851 		 * page tables (or emulated bits, if any), given the upper
852 		 * level VM system not knowing anything about existing
853 		 * references.
854 		 */
855 		} else if (((m->flags & PG_REFERENCED) == 0) &&
856 			(actcount = pmap_ts_referenced(m))) {
857 			vm_page_activate(m);
858 			vm_page_unlock(m);
859 			m->act_count += actcount + ACT_ADVANCE;
860 			VM_OBJECT_UNLOCK(object);
861 			continue;
862 		}
863 
864 		/*
865 		 * If the upper level VM system knows about any page
866 		 * references, we activate the page.  We also set the
867 		 * "activation count" higher than normal so that we will less
868 		 * likely place pages back onto the inactive queue again.
869 		 */
870 		if ((m->flags & PG_REFERENCED) != 0) {
871 			vm_page_flag_clear(m, PG_REFERENCED);
872 			actcount = pmap_ts_referenced(m);
873 			vm_page_activate(m);
874 			vm_page_unlock(m);
875 			m->act_count += actcount + ACT_ADVANCE + 1;
876 			VM_OBJECT_UNLOCK(object);
877 			continue;
878 		}
879 
880 		/*
881 		 * If the upper level VM system does not believe that the page
882 		 * is fully dirty, but it is mapped for write access, then we
883 		 * consult the pmap to see if the page's dirty status should
884 		 * be updated.
885 		 */
886 		if (m->dirty != VM_PAGE_BITS_ALL &&
887 		    (m->flags & PG_WRITEABLE) != 0) {
888 			/*
889 			 * Avoid a race condition: Unless write access is
890 			 * removed from the page, another processor could
891 			 * modify it before all access is removed by the call
892 			 * to vm_page_cache() below.  If vm_page_cache() finds
893 			 * that the page has been modified when it removes all
894 			 * access, it panics because it cannot cache dirty
895 			 * pages.  In principle, we could eliminate just write
896 			 * access here rather than all access.  In the expected
897 			 * case, when there are no last instant modifications
898 			 * to the page, removing all access will be cheaper
899 			 * overall.
900 			 */
901 			if (pmap_is_modified(m))
902 				vm_page_dirty(m);
903 			else if (m->dirty == 0)
904 				pmap_remove_all(m);
905 		}
906 
907 		if (m->valid == 0) {
908 			/*
909 			 * Invalid pages can be easily freed
910 			 */
911 			vm_page_free(m);
912 			cnt.v_dfree++;
913 			--page_shortage;
914 		} else if (m->dirty == 0) {
915 			/*
916 			 * Clean pages can be placed onto the cache queue.
917 			 * This effectively frees them.
918 			 */
919 			vm_page_cache(m);
920 			--page_shortage;
921 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
922 			/*
923 			 * Dirty pages need to be paged out, but flushing
924 			 * a page is extremely expensive verses freeing
925 			 * a clean page.  Rather then artificially limiting
926 			 * the number of pages we can flush, we instead give
927 			 * dirty pages extra priority on the inactive queue
928 			 * by forcing them to be cycled through the queue
929 			 * twice before being flushed, after which the
930 			 * (now clean) page will cycle through once more
931 			 * before being freed.  This significantly extends
932 			 * the thrash point for a heavily loaded machine.
933 			 */
934 			vm_page_flag_set(m, PG_WINATCFLS);
935 			vm_page_requeue(m);
936 		} else if (maxlaunder > 0) {
937 			/*
938 			 * We always want to try to flush some dirty pages if
939 			 * we encounter them, to keep the system stable.
940 			 * Normally this number is small, but under extreme
941 			 * pressure where there are insufficient clean pages
942 			 * on the inactive queue, we may have to go all out.
943 			 */
944 			int swap_pageouts_ok, vfslocked = 0;
945 			struct vnode *vp = NULL;
946 			struct mount *mp = NULL;
947 
948 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
949 				swap_pageouts_ok = 1;
950 			} else {
951 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
952 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
953 				vm_page_count_min());
954 
955 			}
956 
957 			/*
958 			 * We don't bother paging objects that are "dead".
959 			 * Those objects are in a "rundown" state.
960 			 */
961 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
962 				vm_page_unlock(m);
963 				VM_OBJECT_UNLOCK(object);
964 				vm_page_requeue(m);
965 				continue;
966 			}
967 
968 			/*
969 			 * Following operations may unlock
970 			 * vm_page_queue_mtx, invalidating the 'next'
971 			 * pointer.  To prevent an inordinate number
972 			 * of restarts we use our marker to remember
973 			 * our place.
974 			 *
975 			 */
976 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
977 					   m, &marker, pageq);
978 			/*
979 			 * The object is already known NOT to be dead.   It
980 			 * is possible for the vget() to block the whole
981 			 * pageout daemon, but the new low-memory handling
982 			 * code should prevent it.
983 			 *
984 			 * The previous code skipped locked vnodes and, worse,
985 			 * reordered pages in the queue.  This results in
986 			 * completely non-deterministic operation and, on a
987 			 * busy system, can lead to extremely non-optimal
988 			 * pageouts.  For example, it can cause clean pages
989 			 * to be freed and dirty pages to be moved to the end
990 			 * of the queue.  Since dirty pages are also moved to
991 			 * the end of the queue once-cleaned, this gives
992 			 * way too large a weighting to defering the freeing
993 			 * of dirty pages.
994 			 *
995 			 * We can't wait forever for the vnode lock, we might
996 			 * deadlock due to a vn_read() getting stuck in
997 			 * vm_wait while holding this vnode.  We skip the
998 			 * vnode if we can't get it in a reasonable amount
999 			 * of time.
1000 			 */
1001 			if (object->type == OBJT_VNODE) {
1002 				vm_page_unlock_queues();
1003 				vm_page_unlock(m);
1004 				vp = object->handle;
1005 				if (vp->v_type == VREG &&
1006 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1007 					mp = NULL;
1008 					++pageout_lock_miss;
1009 					if (object->flags & OBJ_MIGHTBEDIRTY)
1010 						vnodes_skipped++;
1011 					vm_page_lock_queues();
1012 					goto unlock_and_continue;
1013 				}
1014 				KASSERT(mp != NULL,
1015 				    ("vp %p with NULL v_mount", vp));
1016 				vm_object_reference_locked(object);
1017 				VM_OBJECT_UNLOCK(object);
1018 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1019 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1020 				    curthread)) {
1021 					VM_OBJECT_LOCK(object);
1022 					vm_page_lock_queues();
1023 					++pageout_lock_miss;
1024 					if (object->flags & OBJ_MIGHTBEDIRTY)
1025 						vnodes_skipped++;
1026 					vp = NULL;
1027 					goto unlock_and_continue;
1028 				}
1029 				VM_OBJECT_LOCK(object);
1030 				vm_page_lock(m);
1031 				vm_page_lock_queues();
1032 				/*
1033 				 * The page might have been moved to another
1034 				 * queue during potential blocking in vget()
1035 				 * above.  The page might have been freed and
1036 				 * reused for another vnode.
1037 				 */
1038 				if (m->queue != PQ_INACTIVE ||
1039 				    m->object != object ||
1040 				    TAILQ_NEXT(m, pageq) != &marker) {
1041 					vm_page_unlock(m);
1042 					if (object->flags & OBJ_MIGHTBEDIRTY)
1043 						vnodes_skipped++;
1044 					goto unlock_and_continue;
1045 				}
1046 
1047 				/*
1048 				 * The page may have been busied during the
1049 				 * blocking in vget().  We don't move the
1050 				 * page back onto the end of the queue so that
1051 				 * statistics are more correct if we don't.
1052 				 */
1053 				if (m->busy || (m->oflags & VPO_BUSY)) {
1054 					vm_page_unlock(m);
1055 					goto unlock_and_continue;
1056 				}
1057 
1058 				/*
1059 				 * If the page has become held it might
1060 				 * be undergoing I/O, so skip it
1061 				 */
1062 				if (m->hold_count) {
1063 					vm_page_unlock(m);
1064 					vm_page_requeue(m);
1065 					if (object->flags & OBJ_MIGHTBEDIRTY)
1066 						vnodes_skipped++;
1067 					goto unlock_and_continue;
1068 				}
1069 			}
1070 
1071 			/*
1072 			 * If a page is dirty, then it is either being washed
1073 			 * (but not yet cleaned) or it is still in the
1074 			 * laundry.  If it is still in the laundry, then we
1075 			 * start the cleaning operation.
1076 			 *
1077 			 * decrement page_shortage on success to account for
1078 			 * the (future) cleaned page.  Otherwise we could wind
1079 			 * up laundering or cleaning too many pages.
1080 			 */
1081 			vm_page_unlock_queues();
1082 			if (vm_pageout_clean(m) != 0) {
1083 				--page_shortage;
1084 				--maxlaunder;
1085 			}
1086 			vm_page_lock_queues();
1087 unlock_and_continue:
1088 			vm_page_lock_assert(m, MA_NOTOWNED);
1089 			VM_OBJECT_UNLOCK(object);
1090 			if (mp != NULL) {
1091 				vm_page_unlock_queues();
1092 				if (vp != NULL)
1093 					vput(vp);
1094 				VFS_UNLOCK_GIANT(vfslocked);
1095 				vm_object_deallocate(object);
1096 				vn_finished_write(mp);
1097 				vm_page_lock_queues();
1098 			}
1099 			next = TAILQ_NEXT(&marker, pageq);
1100 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1101 				     &marker, pageq);
1102 			vm_page_lock_assert(m, MA_NOTOWNED);
1103 			continue;
1104 		}
1105 		vm_page_unlock(m);
1106 		VM_OBJECT_UNLOCK(object);
1107 	}
1108 
1109 	/*
1110 	 * Compute the number of pages we want to try to move from the
1111 	 * active queue to the inactive queue.
1112 	 */
1113 	page_shortage = vm_paging_target() +
1114 		cnt.v_inactive_target - cnt.v_inactive_count;
1115 	page_shortage += addl_page_shortage;
1116 
1117 	/*
1118 	 * Scan the active queue for things we can deactivate. We nominally
1119 	 * track the per-page activity counter and use it to locate
1120 	 * deactivation candidates.
1121 	 */
1122 	pcount = cnt.v_active_count;
1123 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1124 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1125 
1126 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1127 
1128 		KASSERT(m->queue == PQ_ACTIVE,
1129 		    ("vm_pageout_scan: page %p isn't active", m));
1130 
1131 		next = TAILQ_NEXT(m, pageq);
1132 		if ((m->flags & PG_MARKER) != 0) {
1133 			m = next;
1134 			continue;
1135 		}
1136 		if (!vm_pageout_page_lock(m, &next)) {
1137 			vm_page_unlock(m);
1138 			m = next;
1139 			continue;
1140 		}
1141 		object = m->object;
1142 		if (!VM_OBJECT_TRYLOCK(object) &&
1143 		    !vm_pageout_fallback_object_lock(m, &next)) {
1144 			VM_OBJECT_UNLOCK(object);
1145 			vm_page_unlock(m);
1146 			m = next;
1147 			continue;
1148 		}
1149 
1150 		/*
1151 		 * Don't deactivate pages that are busy.
1152 		 */
1153 		if ((m->busy != 0) ||
1154 		    (m->oflags & VPO_BUSY) ||
1155 		    (m->hold_count != 0)) {
1156 			vm_page_unlock(m);
1157 			VM_OBJECT_UNLOCK(object);
1158 			vm_page_requeue(m);
1159 			m = next;
1160 			continue;
1161 		}
1162 
1163 		/*
1164 		 * The count for pagedaemon pages is done after checking the
1165 		 * page for eligibility...
1166 		 */
1167 		cnt.v_pdpages++;
1168 
1169 		/*
1170 		 * Check to see "how much" the page has been used.
1171 		 */
1172 		actcount = 0;
1173 		if (object->ref_count != 0) {
1174 			if (m->flags & PG_REFERENCED) {
1175 				actcount += 1;
1176 			}
1177 			actcount += pmap_ts_referenced(m);
1178 			if (actcount) {
1179 				m->act_count += ACT_ADVANCE + actcount;
1180 				if (m->act_count > ACT_MAX)
1181 					m->act_count = ACT_MAX;
1182 			}
1183 		}
1184 
1185 		/*
1186 		 * Since we have "tested" this bit, we need to clear it now.
1187 		 */
1188 		vm_page_flag_clear(m, PG_REFERENCED);
1189 
1190 		/*
1191 		 * Only if an object is currently being used, do we use the
1192 		 * page activation count stats.
1193 		 */
1194 		if (actcount && (object->ref_count != 0)) {
1195 			vm_page_requeue(m);
1196 		} else {
1197 			m->act_count -= min(m->act_count, ACT_DECLINE);
1198 			if (vm_pageout_algorithm ||
1199 			    object->ref_count == 0 ||
1200 			    m->act_count == 0) {
1201 				page_shortage--;
1202 				if (object->ref_count == 0) {
1203 					KASSERT(!pmap_page_is_mapped(m),
1204 				    ("vm_pageout_scan: page %p is mapped", m));
1205 					if (m->dirty == 0)
1206 						vm_page_cache(m);
1207 					else
1208 						vm_page_deactivate(m);
1209 				} else {
1210 					vm_page_deactivate(m);
1211 				}
1212 			} else {
1213 				vm_page_requeue(m);
1214 			}
1215 		}
1216 		vm_page_unlock(m);
1217 		VM_OBJECT_UNLOCK(object);
1218 		m = next;
1219 	}
1220 	vm_page_unlock_queues();
1221 #if !defined(NO_SWAPPING)
1222 	/*
1223 	 * Idle process swapout -- run once per second.
1224 	 */
1225 	if (vm_swap_idle_enabled) {
1226 		static long lsec;
1227 		if (time_second != lsec) {
1228 			vm_req_vmdaemon(VM_SWAP_IDLE);
1229 			lsec = time_second;
1230 		}
1231 	}
1232 #endif
1233 
1234 	/*
1235 	 * If we didn't get enough free pages, and we have skipped a vnode
1236 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1237 	 * if we did not get enough free pages.
1238 	 */
1239 	if (vm_paging_target() > 0) {
1240 		if (vnodes_skipped && vm_page_count_min())
1241 			(void) speedup_syncer();
1242 #if !defined(NO_SWAPPING)
1243 		if (vm_swap_enabled && vm_page_count_target())
1244 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1245 #endif
1246 	}
1247 
1248 	/*
1249 	 * If we are critically low on one of RAM or swap and low on
1250 	 * the other, kill the largest process.  However, we avoid
1251 	 * doing this on the first pass in order to give ourselves a
1252 	 * chance to flush out dirty vnode-backed pages and to allow
1253 	 * active pages to be moved to the inactive queue and reclaimed.
1254 	 */
1255 	if (pass != 0 &&
1256 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1257 	     (swap_pager_full && vm_paging_target() > 0)))
1258 		vm_pageout_oom(VM_OOM_MEM);
1259 }
1260 
1261 
1262 void
1263 vm_pageout_oom(int shortage)
1264 {
1265 	struct proc *p, *bigproc;
1266 	vm_offset_t size, bigsize;
1267 	struct thread *td;
1268 	struct vmspace *vm;
1269 
1270 	/*
1271 	 * We keep the process bigproc locked once we find it to keep anyone
1272 	 * from messing with it; however, there is a possibility of
1273 	 * deadlock if process B is bigproc and one of it's child processes
1274 	 * attempts to propagate a signal to B while we are waiting for A's
1275 	 * lock while walking this list.  To avoid this, we don't block on
1276 	 * the process lock but just skip a process if it is already locked.
1277 	 */
1278 	bigproc = NULL;
1279 	bigsize = 0;
1280 	sx_slock(&allproc_lock);
1281 	FOREACH_PROC_IN_SYSTEM(p) {
1282 		int breakout;
1283 
1284 		if (p->p_state != PRS_NORMAL)
1285 			continue;
1286 		if (PROC_TRYLOCK(p) == 0)
1287 			continue;
1288 		/*
1289 		 * If this is a system, protected or killed process, skip it.
1290 		 */
1291 		if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1292 		    (p->p_pid == 1) || P_KILLED(p) ||
1293 		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1294 			PROC_UNLOCK(p);
1295 			continue;
1296 		}
1297 		/*
1298 		 * If the process is in a non-running type state,
1299 		 * don't touch it.  Check all the threads individually.
1300 		 */
1301 		breakout = 0;
1302 		FOREACH_THREAD_IN_PROC(p, td) {
1303 			thread_lock(td);
1304 			if (!TD_ON_RUNQ(td) &&
1305 			    !TD_IS_RUNNING(td) &&
1306 			    !TD_IS_SLEEPING(td)) {
1307 				thread_unlock(td);
1308 				breakout = 1;
1309 				break;
1310 			}
1311 			thread_unlock(td);
1312 		}
1313 		if (breakout) {
1314 			PROC_UNLOCK(p);
1315 			continue;
1316 		}
1317 		/*
1318 		 * get the process size
1319 		 */
1320 		vm = vmspace_acquire_ref(p);
1321 		if (vm == NULL) {
1322 			PROC_UNLOCK(p);
1323 			continue;
1324 		}
1325 		if (!vm_map_trylock_read(&vm->vm_map)) {
1326 			vmspace_free(vm);
1327 			PROC_UNLOCK(p);
1328 			continue;
1329 		}
1330 		size = vmspace_swap_count(vm);
1331 		vm_map_unlock_read(&vm->vm_map);
1332 		if (shortage == VM_OOM_MEM)
1333 			size += vmspace_resident_count(vm);
1334 		vmspace_free(vm);
1335 		/*
1336 		 * if the this process is bigger than the biggest one
1337 		 * remember it.
1338 		 */
1339 		if (size > bigsize) {
1340 			if (bigproc != NULL)
1341 				PROC_UNLOCK(bigproc);
1342 			bigproc = p;
1343 			bigsize = size;
1344 		} else
1345 			PROC_UNLOCK(p);
1346 	}
1347 	sx_sunlock(&allproc_lock);
1348 	if (bigproc != NULL) {
1349 		killproc(bigproc, "out of swap space");
1350 		sched_nice(bigproc, PRIO_MIN);
1351 		PROC_UNLOCK(bigproc);
1352 		wakeup(&cnt.v_free_count);
1353 	}
1354 }
1355 
1356 /*
1357  * This routine tries to maintain the pseudo LRU active queue,
1358  * so that during long periods of time where there is no paging,
1359  * that some statistic accumulation still occurs.  This code
1360  * helps the situation where paging just starts to occur.
1361  */
1362 static void
1363 vm_pageout_page_stats()
1364 {
1365 	vm_object_t object;
1366 	vm_page_t m,next;
1367 	int pcount,tpcount;		/* Number of pages to check */
1368 	static int fullintervalcount = 0;
1369 	int page_shortage;
1370 
1371 	page_shortage =
1372 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1373 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1374 
1375 	if (page_shortage <= 0)
1376 		return;
1377 
1378 	vm_page_lock_queues();
1379 	pcount = cnt.v_active_count;
1380 	fullintervalcount += vm_pageout_stats_interval;
1381 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1382 		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1383 		    cnt.v_page_count;
1384 		if (pcount > tpcount)
1385 			pcount = tpcount;
1386 	} else {
1387 		fullintervalcount = 0;
1388 	}
1389 
1390 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1391 	while ((m != NULL) && (pcount-- > 0)) {
1392 		int actcount;
1393 
1394 		KASSERT(m->queue == PQ_ACTIVE,
1395 		    ("vm_pageout_page_stats: page %p isn't active", m));
1396 
1397 		next = TAILQ_NEXT(m, pageq);
1398 		if ((m->flags & PG_MARKER) != 0) {
1399 			m = next;
1400 			continue;
1401 		}
1402 		vm_page_lock_assert(m, MA_NOTOWNED);
1403 		if (!vm_pageout_page_lock(m, &next)) {
1404 			vm_page_unlock(m);
1405 			m = next;
1406 			continue;
1407 		}
1408 		object = m->object;
1409 		if (!VM_OBJECT_TRYLOCK(object) &&
1410 		    !vm_pageout_fallback_object_lock(m, &next)) {
1411 			VM_OBJECT_UNLOCK(object);
1412 			vm_page_unlock(m);
1413 			m = next;
1414 			continue;
1415 		}
1416 
1417 		/*
1418 		 * Don't deactivate pages that are busy.
1419 		 */
1420 		if ((m->busy != 0) ||
1421 		    (m->oflags & VPO_BUSY) ||
1422 		    (m->hold_count != 0)) {
1423 			vm_page_unlock(m);
1424 			VM_OBJECT_UNLOCK(object);
1425 			vm_page_requeue(m);
1426 			m = next;
1427 			continue;
1428 		}
1429 
1430 		actcount = 0;
1431 		if (m->flags & PG_REFERENCED) {
1432 			vm_page_flag_clear(m, PG_REFERENCED);
1433 			actcount += 1;
1434 		}
1435 
1436 		actcount += pmap_ts_referenced(m);
1437 		if (actcount) {
1438 			m->act_count += ACT_ADVANCE + actcount;
1439 			if (m->act_count > ACT_MAX)
1440 				m->act_count = ACT_MAX;
1441 			vm_page_requeue(m);
1442 		} else {
1443 			if (m->act_count == 0) {
1444 				/*
1445 				 * We turn off page access, so that we have
1446 				 * more accurate RSS stats.  We don't do this
1447 				 * in the normal page deactivation when the
1448 				 * system is loaded VM wise, because the
1449 				 * cost of the large number of page protect
1450 				 * operations would be higher than the value
1451 				 * of doing the operation.
1452 				 */
1453 				pmap_remove_all(m);
1454 				vm_page_deactivate(m);
1455 			} else {
1456 				m->act_count -= min(m->act_count, ACT_DECLINE);
1457 				vm_page_requeue(m);
1458 			}
1459 		}
1460 		vm_page_unlock(m);
1461 		VM_OBJECT_UNLOCK(object);
1462 		m = next;
1463 	}
1464 	vm_page_unlock_queues();
1465 }
1466 
1467 /*
1468  *	vm_pageout is the high level pageout daemon.
1469  */
1470 static void
1471 vm_pageout()
1472 {
1473 	int error, pass;
1474 
1475 	/*
1476 	 * Initialize some paging parameters.
1477 	 */
1478 	cnt.v_interrupt_free_min = 2;
1479 	if (cnt.v_page_count < 2000)
1480 		vm_pageout_page_count = 8;
1481 
1482 	/*
1483 	 * v_free_reserved needs to include enough for the largest
1484 	 * swap pager structures plus enough for any pv_entry structs
1485 	 * when paging.
1486 	 */
1487 	if (cnt.v_page_count > 1024)
1488 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1489 	else
1490 		cnt.v_free_min = 4;
1491 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1492 	    cnt.v_interrupt_free_min;
1493 	cnt.v_free_reserved = vm_pageout_page_count +
1494 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1495 	cnt.v_free_severe = cnt.v_free_min / 2;
1496 	cnt.v_free_min += cnt.v_free_reserved;
1497 	cnt.v_free_severe += cnt.v_free_reserved;
1498 
1499 	/*
1500 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1501 	 * that these are more a measure of the VM cache queue hysteresis
1502 	 * then the VM free queue.  Specifically, v_free_target is the
1503 	 * high water mark (free+cache pages).
1504 	 *
1505 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1506 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1507 	 * be big enough to handle memory needs while the pageout daemon
1508 	 * is signalled and run to free more pages.
1509 	 */
1510 	if (cnt.v_free_count > 6144)
1511 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1512 	else
1513 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1514 
1515 	if (cnt.v_free_count > 2048) {
1516 		cnt.v_cache_min = cnt.v_free_target;
1517 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1518 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1519 	} else {
1520 		cnt.v_cache_min = 0;
1521 		cnt.v_cache_max = 0;
1522 		cnt.v_inactive_target = cnt.v_free_count / 4;
1523 	}
1524 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1525 		cnt.v_inactive_target = cnt.v_free_count / 3;
1526 
1527 	/* XXX does not really belong here */
1528 	if (vm_page_max_wired == 0)
1529 		vm_page_max_wired = cnt.v_free_count / 3;
1530 
1531 	if (vm_pageout_stats_max == 0)
1532 		vm_pageout_stats_max = cnt.v_free_target;
1533 
1534 	/*
1535 	 * Set interval in seconds for stats scan.
1536 	 */
1537 	if (vm_pageout_stats_interval == 0)
1538 		vm_pageout_stats_interval = 5;
1539 	if (vm_pageout_full_stats_interval == 0)
1540 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1541 
1542 	swap_pager_swap_init();
1543 	pass = 0;
1544 	/*
1545 	 * The pageout daemon is never done, so loop forever.
1546 	 */
1547 	while (TRUE) {
1548 		/*
1549 		 * If we have enough free memory, wakeup waiters.  Do
1550 		 * not clear vm_pages_needed until we reach our target,
1551 		 * otherwise we may be woken up over and over again and
1552 		 * waste a lot of cpu.
1553 		 */
1554 		mtx_lock(&vm_page_queue_free_mtx);
1555 		if (vm_pages_needed && !vm_page_count_min()) {
1556 			if (!vm_paging_needed())
1557 				vm_pages_needed = 0;
1558 			wakeup(&cnt.v_free_count);
1559 		}
1560 		if (vm_pages_needed) {
1561 			/*
1562 			 * Still not done, take a second pass without waiting
1563 			 * (unlimited dirty cleaning), otherwise sleep a bit
1564 			 * and try again.
1565 			 */
1566 			++pass;
1567 			if (pass > 1)
1568 				msleep(&vm_pages_needed,
1569 				    &vm_page_queue_free_mtx, PVM, "psleep",
1570 				    hz / 2);
1571 		} else {
1572 			/*
1573 			 * Good enough, sleep & handle stats.  Prime the pass
1574 			 * for the next run.
1575 			 */
1576 			if (pass > 1)
1577 				pass = 1;
1578 			else
1579 				pass = 0;
1580 			error = msleep(&vm_pages_needed,
1581 			    &vm_page_queue_free_mtx, PVM, "psleep",
1582 			    vm_pageout_stats_interval * hz);
1583 			if (error && !vm_pages_needed) {
1584 				mtx_unlock(&vm_page_queue_free_mtx);
1585 				pass = 0;
1586 				vm_pageout_page_stats();
1587 				continue;
1588 			}
1589 		}
1590 		if (vm_pages_needed)
1591 			cnt.v_pdwakeups++;
1592 		mtx_unlock(&vm_page_queue_free_mtx);
1593 		vm_pageout_scan(pass);
1594 	}
1595 }
1596 
1597 /*
1598  * Unless the free page queue lock is held by the caller, this function
1599  * should be regarded as advisory.  Specifically, the caller should
1600  * not msleep() on &cnt.v_free_count following this function unless
1601  * the free page queue lock is held until the msleep() is performed.
1602  */
1603 void
1604 pagedaemon_wakeup()
1605 {
1606 
1607 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1608 		vm_pages_needed = 1;
1609 		wakeup(&vm_pages_needed);
1610 	}
1611 }
1612 
1613 #if !defined(NO_SWAPPING)
1614 static void
1615 vm_req_vmdaemon(int req)
1616 {
1617 	static int lastrun = 0;
1618 
1619 	mtx_lock(&vm_daemon_mtx);
1620 	vm_pageout_req_swapout |= req;
1621 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1622 		wakeup(&vm_daemon_needed);
1623 		lastrun = ticks;
1624 	}
1625 	mtx_unlock(&vm_daemon_mtx);
1626 }
1627 
1628 static void
1629 vm_daemon()
1630 {
1631 	struct rlimit rsslim;
1632 	struct proc *p;
1633 	struct thread *td;
1634 	struct vmspace *vm;
1635 	int breakout, swapout_flags;
1636 
1637 	while (TRUE) {
1638 		mtx_lock(&vm_daemon_mtx);
1639 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1640 		swapout_flags = vm_pageout_req_swapout;
1641 		vm_pageout_req_swapout = 0;
1642 		mtx_unlock(&vm_daemon_mtx);
1643 		if (swapout_flags)
1644 			swapout_procs(swapout_flags);
1645 
1646 		/*
1647 		 * scan the processes for exceeding their rlimits or if
1648 		 * process is swapped out -- deactivate pages
1649 		 */
1650 		sx_slock(&allproc_lock);
1651 		FOREACH_PROC_IN_SYSTEM(p) {
1652 			vm_pindex_t limit, size;
1653 
1654 			if (p->p_state != PRS_NORMAL)
1655 				continue;
1656 			/*
1657 			 * if this is a system process or if we have already
1658 			 * looked at this process, skip it.
1659 			 */
1660 			PROC_LOCK(p);
1661 			if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1662 				PROC_UNLOCK(p);
1663 				continue;
1664 			}
1665 			/*
1666 			 * if the process is in a non-running type state,
1667 			 * don't touch it.
1668 			 */
1669 			breakout = 0;
1670 			FOREACH_THREAD_IN_PROC(p, td) {
1671 				thread_lock(td);
1672 				if (!TD_ON_RUNQ(td) &&
1673 				    !TD_IS_RUNNING(td) &&
1674 				    !TD_IS_SLEEPING(td)) {
1675 					thread_unlock(td);
1676 					breakout = 1;
1677 					break;
1678 				}
1679 				thread_unlock(td);
1680 			}
1681 			if (breakout) {
1682 				PROC_UNLOCK(p);
1683 				continue;
1684 			}
1685 			/*
1686 			 * get a limit
1687 			 */
1688 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1689 			limit = OFF_TO_IDX(
1690 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1691 
1692 			/*
1693 			 * let processes that are swapped out really be
1694 			 * swapped out set the limit to nothing (will force a
1695 			 * swap-out.)
1696 			 */
1697 			if ((p->p_flag & P_INMEM) == 0)
1698 				limit = 0;	/* XXX */
1699 			vm = vmspace_acquire_ref(p);
1700 			PROC_UNLOCK(p);
1701 			if (vm == NULL)
1702 				continue;
1703 
1704 			size = vmspace_resident_count(vm);
1705 			if (limit >= 0 && size >= limit) {
1706 				vm_pageout_map_deactivate_pages(
1707 				    &vm->vm_map, limit);
1708 			}
1709 			vmspace_free(vm);
1710 		}
1711 		sx_sunlock(&allproc_lock);
1712 	}
1713 }
1714 #endif			/* !defined(NO_SWAPPING) */
1715