xref: /freebsd/sys/vm/vm_pageout.c (revision 6486b015fc84e96725fef22b0e3363351399ae83)
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/racct.h>
90 #include <sys/resourcevar.h>
91 #include <sys/sched.h>
92 #include <sys/signalvar.h>
93 #include <sys/vnode.h>
94 #include <sys/vmmeter.h>
95 #include <sys/sx.h>
96 #include <sys/sysctl.h>
97 
98 #include <vm/vm.h>
99 #include <vm/vm_param.h>
100 #include <vm/vm_object.h>
101 #include <vm/vm_page.h>
102 #include <vm/vm_map.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_pager.h>
105 #include <vm/swap_pager.h>
106 #include <vm/vm_extern.h>
107 #include <vm/uma.h>
108 
109 /*
110  * System initialization
111  */
112 
113 /* the kernel process "vm_pageout"*/
114 static void vm_pageout(void);
115 static int vm_pageout_clean(vm_page_t);
116 static void vm_pageout_scan(int pass);
117 
118 struct proc *pageproc;
119 
120 static struct kproc_desc page_kp = {
121 	"pagedaemon",
122 	vm_pageout,
123 	&pageproc
124 };
125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
126     &page_kp);
127 
128 #if !defined(NO_SWAPPING)
129 /* the kernel process "vm_daemon"*/
130 static void vm_daemon(void);
131 static struct	proc *vmproc;
132 
133 static struct kproc_desc vm_kp = {
134 	"vmdaemon",
135 	vm_daemon,
136 	&vmproc
137 };
138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
139 #endif
140 
141 
142 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
143 int vm_pageout_deficit;		/* Estimated number of pages deficit */
144 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
145 
146 #if !defined(NO_SWAPPING)
147 static int vm_pageout_req_swapout;	/* XXX */
148 static int vm_daemon_needed;
149 static struct mtx vm_daemon_mtx;
150 /* Allow for use by vm_pageout before vm_daemon is initialized. */
151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
152 #endif
153 static int vm_max_launder = 32;
154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
155 static int vm_pageout_full_stats_interval = 0;
156 static int vm_pageout_algorithm=0;
157 static int defer_swap_pageouts=0;
158 static int disable_swap_pageouts=0;
159 
160 #if defined(NO_SWAPPING)
161 static int vm_swap_enabled=0;
162 static int vm_swap_idle_enabled=0;
163 #else
164 static int vm_swap_enabled=1;
165 static int vm_swap_idle_enabled=0;
166 #endif
167 
168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
169 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
170 
171 SYSCTL_INT(_vm, OID_AUTO, max_launder,
172 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
173 
174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
175 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
176 
177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
178 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
179 
180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
181 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
182 
183 #if defined(NO_SWAPPING)
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 #else
189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
190 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
192 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
193 #endif
194 
195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
196 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
197 
198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
199 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
200 
201 static int pageout_lock_miss;
202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
203 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
204 
205 #define VM_PAGEOUT_PAGE_COUNT 16
206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
207 
208 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
209 SYSCTL_INT(_vm, OID_AUTO, max_wired,
210 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
211 
212 #if !defined(NO_SWAPPING)
213 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
214 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
215 static void vm_req_vmdaemon(int req);
216 #endif
217 static void vm_pageout_page_stats(void);
218 
219 /*
220  * Initialize a dummy page for marking the caller's place in the specified
221  * paging queue.  In principle, this function only needs to set the flag
222  * PG_MARKER.  Nonetheless, it sets the flag VPO_BUSY and initializes the hold
223  * count to one as safety precautions.
224  */
225 static void
226 vm_pageout_init_marker(vm_page_t marker, u_short queue)
227 {
228 
229 	bzero(marker, sizeof(*marker));
230 	marker->flags = PG_MARKER;
231 	marker->oflags = VPO_BUSY;
232 	marker->queue = queue;
233 	marker->hold_count = 1;
234 }
235 
236 /*
237  * vm_pageout_fallback_object_lock:
238  *
239  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
240  * known to have failed and page queue must be either PQ_ACTIVE or
241  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
242  * while locking the vm object.  Use marker page to detect page queue
243  * changes and maintain notion of next page on page queue.  Return
244  * TRUE if no changes were detected, FALSE otherwise.  vm object is
245  * locked on return.
246  *
247  * This function depends on both the lock portion of struct vm_object
248  * and normal struct vm_page being type stable.
249  */
250 boolean_t
251 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
252 {
253 	struct vm_page marker;
254 	boolean_t unchanged;
255 	u_short queue;
256 	vm_object_t object;
257 
258 	queue = m->queue;
259 	vm_pageout_init_marker(&marker, queue);
260 	object = m->object;
261 
262 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
263 			   m, &marker, pageq);
264 	vm_page_unlock_queues();
265 	vm_page_unlock(m);
266 	VM_OBJECT_LOCK(object);
267 	vm_page_lock(m);
268 	vm_page_lock_queues();
269 
270 	/* Page queue might have changed. */
271 	*next = TAILQ_NEXT(&marker, pageq);
272 	unchanged = (m->queue == queue &&
273 		     m->object == object &&
274 		     &marker == TAILQ_NEXT(m, pageq));
275 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
276 		     &marker, pageq);
277 	return (unchanged);
278 }
279 
280 /*
281  * Lock the page while holding the page queue lock.  Use marker page
282  * to detect page queue changes and maintain notion of next page on
283  * page queue.  Return TRUE if no changes were detected, FALSE
284  * otherwise.  The page is locked on return. The page queue lock might
285  * be dropped and reacquired.
286  *
287  * This function depends on normal struct vm_page being type stable.
288  */
289 boolean_t
290 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
291 {
292 	struct vm_page marker;
293 	boolean_t unchanged;
294 	u_short queue;
295 
296 	vm_page_lock_assert(m, MA_NOTOWNED);
297 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
298 
299 	if (vm_page_trylock(m))
300 		return (TRUE);
301 
302 	queue = m->queue;
303 	vm_pageout_init_marker(&marker, queue);
304 
305 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, m, &marker, pageq);
306 	vm_page_unlock_queues();
307 	vm_page_lock(m);
308 	vm_page_lock_queues();
309 
310 	/* Page queue might have changed. */
311 	*next = TAILQ_NEXT(&marker, pageq);
312 	unchanged = (m->queue == queue && &marker == TAILQ_NEXT(m, pageq));
313 	TAILQ_REMOVE(&vm_page_queues[queue].pl, &marker, pageq);
314 	return (unchanged);
315 }
316 
317 /*
318  * vm_pageout_clean:
319  *
320  * Clean the page and remove it from the laundry.
321  *
322  * We set the busy bit to cause potential page faults on this page to
323  * block.  Note the careful timing, however, the busy bit isn't set till
324  * late and we cannot do anything that will mess with the page.
325  */
326 static int
327 vm_pageout_clean(vm_page_t m)
328 {
329 	vm_object_t object;
330 	vm_page_t mc[2*vm_pageout_page_count], pb, ps;
331 	int pageout_count;
332 	int ib, is, page_base;
333 	vm_pindex_t pindex = m->pindex;
334 
335 	vm_page_lock_assert(m, MA_OWNED);
336 	object = m->object;
337 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
338 
339 	/*
340 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
341 	 * with the new swapper, but we could have serious problems paging
342 	 * out other object types if there is insufficient memory.
343 	 *
344 	 * Unfortunately, checking free memory here is far too late, so the
345 	 * check has been moved up a procedural level.
346 	 */
347 
348 	/*
349 	 * Can't clean the page if it's busy or held.
350 	 */
351 	KASSERT(m->busy == 0 && (m->oflags & VPO_BUSY) == 0,
352 	    ("vm_pageout_clean: page %p is busy", m));
353 	KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
354 	vm_page_unlock(m);
355 
356 	mc[vm_pageout_page_count] = pb = ps = m;
357 	pageout_count = 1;
358 	page_base = vm_pageout_page_count;
359 	ib = 1;
360 	is = 1;
361 
362 	/*
363 	 * Scan object for clusterable pages.
364 	 *
365 	 * We can cluster ONLY if: ->> the page is NOT
366 	 * clean, wired, busy, held, or mapped into a
367 	 * buffer, and one of the following:
368 	 * 1) The page is inactive, or a seldom used
369 	 *    active page.
370 	 * -or-
371 	 * 2) we force the issue.
372 	 *
373 	 * During heavy mmap/modification loads the pageout
374 	 * daemon can really fragment the underlying file
375 	 * due to flushing pages out of order and not trying
376 	 * align the clusters (which leave sporatic out-of-order
377 	 * holes).  To solve this problem we do the reverse scan
378 	 * first and attempt to align our cluster, then do a
379 	 * forward scan if room remains.
380 	 */
381 more:
382 	while (ib && pageout_count < vm_pageout_page_count) {
383 		vm_page_t p;
384 
385 		if (ib > pindex) {
386 			ib = 0;
387 			break;
388 		}
389 
390 		if ((p = vm_page_prev(pb)) == NULL ||
391 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0) {
392 			ib = 0;
393 			break;
394 		}
395 		vm_page_lock(p);
396 		vm_page_test_dirty(p);
397 		if (p->dirty == 0 ||
398 		    p->queue != PQ_INACTIVE ||
399 		    p->hold_count != 0) {	/* may be undergoing I/O */
400 			vm_page_unlock(p);
401 			ib = 0;
402 			break;
403 		}
404 		vm_page_unlock(p);
405 		mc[--page_base] = pb = p;
406 		++pageout_count;
407 		++ib;
408 		/*
409 		 * alignment boundry, stop here and switch directions.  Do
410 		 * not clear ib.
411 		 */
412 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
413 			break;
414 	}
415 
416 	while (pageout_count < vm_pageout_page_count &&
417 	    pindex + is < object->size) {
418 		vm_page_t p;
419 
420 		if ((p = vm_page_next(ps)) == NULL ||
421 		    (p->oflags & VPO_BUSY) != 0 || p->busy != 0)
422 			break;
423 		vm_page_lock(p);
424 		vm_page_test_dirty(p);
425 		if (p->dirty == 0 ||
426 		    p->queue != PQ_INACTIVE ||
427 		    p->hold_count != 0) {	/* may be undergoing I/O */
428 			vm_page_unlock(p);
429 			break;
430 		}
431 		vm_page_unlock(p);
432 		mc[page_base + pageout_count] = ps = p;
433 		++pageout_count;
434 		++is;
435 	}
436 
437 	/*
438 	 * If we exhausted our forward scan, continue with the reverse scan
439 	 * when possible, even past a page boundry.  This catches boundry
440 	 * conditions.
441 	 */
442 	if (ib && pageout_count < vm_pageout_page_count)
443 		goto more;
444 
445 	/*
446 	 * we allow reads during pageouts...
447 	 */
448 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
449 	    NULL));
450 }
451 
452 /*
453  * vm_pageout_flush() - launder the given pages
454  *
455  *	The given pages are laundered.  Note that we setup for the start of
456  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
457  *	reference count all in here rather then in the parent.  If we want
458  *	the parent to do more sophisticated things we may have to change
459  *	the ordering.
460  *
461  *	Returned runlen is the count of pages between mreq and first
462  *	page after mreq with status VM_PAGER_AGAIN.
463  *	*eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
464  *	for any page in runlen set.
465  */
466 int
467 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
468     boolean_t *eio)
469 {
470 	vm_object_t object = mc[0]->object;
471 	int pageout_status[count];
472 	int numpagedout = 0;
473 	int i, runlen;
474 
475 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
476 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
477 
478 	/*
479 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
480 	 * mark the pages read-only.
481 	 *
482 	 * We do not have to fixup the clean/dirty bits here... we can
483 	 * allow the pager to do it after the I/O completes.
484 	 *
485 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
486 	 * edge case with file fragments.
487 	 */
488 	for (i = 0; i < count; i++) {
489 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
490 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
491 			mc[i], i, count));
492 		vm_page_io_start(mc[i]);
493 		pmap_remove_write(mc[i]);
494 	}
495 	vm_object_pip_add(object, count);
496 
497 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
498 
499 	runlen = count - mreq;
500 	if (eio != NULL)
501 		*eio = FALSE;
502 	for (i = 0; i < count; i++) {
503 		vm_page_t mt = mc[i];
504 
505 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
506 		    (mt->aflags & PGA_WRITEABLE) == 0,
507 		    ("vm_pageout_flush: page %p is not write protected", mt));
508 		switch (pageout_status[i]) {
509 		case VM_PAGER_OK:
510 		case VM_PAGER_PEND:
511 			numpagedout++;
512 			break;
513 		case VM_PAGER_BAD:
514 			/*
515 			 * Page outside of range of object. Right now we
516 			 * essentially lose the changes by pretending it
517 			 * worked.
518 			 */
519 			vm_page_undirty(mt);
520 			break;
521 		case VM_PAGER_ERROR:
522 		case VM_PAGER_FAIL:
523 			/*
524 			 * If page couldn't be paged out, then reactivate the
525 			 * page so it doesn't clog the inactive list.  (We
526 			 * will try paging out it again later).
527 			 */
528 			vm_page_lock(mt);
529 			vm_page_activate(mt);
530 			vm_page_unlock(mt);
531 			if (eio != NULL && i >= mreq && i - mreq < runlen)
532 				*eio = TRUE;
533 			break;
534 		case VM_PAGER_AGAIN:
535 			if (i >= mreq && i - mreq < runlen)
536 				runlen = i - mreq;
537 			break;
538 		}
539 
540 		/*
541 		 * If the operation is still going, leave the page busy to
542 		 * block all other accesses. Also, leave the paging in
543 		 * progress indicator set so that we don't attempt an object
544 		 * collapse.
545 		 */
546 		if (pageout_status[i] != VM_PAGER_PEND) {
547 			vm_object_pip_wakeup(object);
548 			vm_page_io_finish(mt);
549 			if (vm_page_count_severe()) {
550 				vm_page_lock(mt);
551 				vm_page_try_to_cache(mt);
552 				vm_page_unlock(mt);
553 			}
554 		}
555 	}
556 	if (prunlen != NULL)
557 		*prunlen = runlen;
558 	return (numpagedout);
559 }
560 
561 #if !defined(NO_SWAPPING)
562 /*
563  *	vm_pageout_object_deactivate_pages
564  *
565  *	Deactivate enough pages to satisfy the inactive target
566  *	requirements.
567  *
568  *	The object and map must be locked.
569  */
570 static void
571 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
572     long desired)
573 {
574 	vm_object_t backing_object, object;
575 	vm_page_t p;
576 	int actcount, remove_mode;
577 
578 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
579 	if (first_object->type == OBJT_DEVICE ||
580 	    first_object->type == OBJT_SG)
581 		return;
582 	for (object = first_object;; object = backing_object) {
583 		if (pmap_resident_count(pmap) <= desired)
584 			goto unlock_return;
585 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
586 		if (object->type == OBJT_PHYS || object->paging_in_progress)
587 			goto unlock_return;
588 
589 		remove_mode = 0;
590 		if (object->shadow_count > 1)
591 			remove_mode = 1;
592 		/*
593 		 * Scan the object's entire memory queue.
594 		 */
595 		TAILQ_FOREACH(p, &object->memq, listq) {
596 			if (pmap_resident_count(pmap) <= desired)
597 				goto unlock_return;
598 			if ((p->oflags & VPO_BUSY) != 0 || p->busy != 0)
599 				continue;
600 			PCPU_INC(cnt.v_pdpages);
601 			vm_page_lock(p);
602 			if (p->wire_count != 0 || p->hold_count != 0 ||
603 			    !pmap_page_exists_quick(pmap, p)) {
604 				vm_page_unlock(p);
605 				continue;
606 			}
607 			actcount = pmap_ts_referenced(p);
608 			if ((p->aflags & PGA_REFERENCED) != 0) {
609 				if (actcount == 0)
610 					actcount = 1;
611 				vm_page_aflag_clear(p, PGA_REFERENCED);
612 			}
613 			if (p->queue != PQ_ACTIVE && actcount != 0) {
614 				vm_page_activate(p);
615 				p->act_count += actcount;
616 			} else if (p->queue == PQ_ACTIVE) {
617 				if (actcount == 0) {
618 					p->act_count -= min(p->act_count,
619 					    ACT_DECLINE);
620 					if (!remove_mode &&
621 					    (vm_pageout_algorithm ||
622 					    p->act_count == 0)) {
623 						pmap_remove_all(p);
624 						vm_page_deactivate(p);
625 					} else {
626 						vm_page_lock_queues();
627 						vm_page_requeue(p);
628 						vm_page_unlock_queues();
629 					}
630 				} else {
631 					vm_page_activate(p);
632 					if (p->act_count < ACT_MAX -
633 					    ACT_ADVANCE)
634 						p->act_count += ACT_ADVANCE;
635 					vm_page_lock_queues();
636 					vm_page_requeue(p);
637 					vm_page_unlock_queues();
638 				}
639 			} else if (p->queue == PQ_INACTIVE)
640 				pmap_remove_all(p);
641 			vm_page_unlock(p);
642 		}
643 		if ((backing_object = object->backing_object) == NULL)
644 			goto unlock_return;
645 		VM_OBJECT_LOCK(backing_object);
646 		if (object != first_object)
647 			VM_OBJECT_UNLOCK(object);
648 	}
649 unlock_return:
650 	if (object != first_object)
651 		VM_OBJECT_UNLOCK(object);
652 }
653 
654 /*
655  * deactivate some number of pages in a map, try to do it fairly, but
656  * that is really hard to do.
657  */
658 static void
659 vm_pageout_map_deactivate_pages(map, desired)
660 	vm_map_t map;
661 	long desired;
662 {
663 	vm_map_entry_t tmpe;
664 	vm_object_t obj, bigobj;
665 	int nothingwired;
666 
667 	if (!vm_map_trylock(map))
668 		return;
669 
670 	bigobj = NULL;
671 	nothingwired = TRUE;
672 
673 	/*
674 	 * first, search out the biggest object, and try to free pages from
675 	 * that.
676 	 */
677 	tmpe = map->header.next;
678 	while (tmpe != &map->header) {
679 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
680 			obj = tmpe->object.vm_object;
681 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
682 				if (obj->shadow_count <= 1 &&
683 				    (bigobj == NULL ||
684 				     bigobj->resident_page_count < obj->resident_page_count)) {
685 					if (bigobj != NULL)
686 						VM_OBJECT_UNLOCK(bigobj);
687 					bigobj = obj;
688 				} else
689 					VM_OBJECT_UNLOCK(obj);
690 			}
691 		}
692 		if (tmpe->wired_count > 0)
693 			nothingwired = FALSE;
694 		tmpe = tmpe->next;
695 	}
696 
697 	if (bigobj != NULL) {
698 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
699 		VM_OBJECT_UNLOCK(bigobj);
700 	}
701 	/*
702 	 * Next, hunt around for other pages to deactivate.  We actually
703 	 * do this search sort of wrong -- .text first is not the best idea.
704 	 */
705 	tmpe = map->header.next;
706 	while (tmpe != &map->header) {
707 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
708 			break;
709 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
710 			obj = tmpe->object.vm_object;
711 			if (obj != NULL) {
712 				VM_OBJECT_LOCK(obj);
713 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
714 				VM_OBJECT_UNLOCK(obj);
715 			}
716 		}
717 		tmpe = tmpe->next;
718 	}
719 
720 	/*
721 	 * Remove all mappings if a process is swapped out, this will free page
722 	 * table pages.
723 	 */
724 	if (desired == 0 && nothingwired) {
725 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
726 		    vm_map_max(map));
727 	}
728 	vm_map_unlock(map);
729 }
730 #endif		/* !defined(NO_SWAPPING) */
731 
732 /*
733  *	vm_pageout_scan does the dirty work for the pageout daemon.
734  */
735 static void
736 vm_pageout_scan(int pass)
737 {
738 	vm_page_t m, next;
739 	struct vm_page marker;
740 	int page_shortage, maxscan, pcount;
741 	int addl_page_shortage, addl_page_shortage_init;
742 	vm_object_t object;
743 	int actcount;
744 	int vnodes_skipped = 0;
745 	int maxlaunder;
746 
747 	/*
748 	 * Decrease registered cache sizes.
749 	 */
750 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
751 	/*
752 	 * We do this explicitly after the caches have been drained above.
753 	 */
754 	uma_reclaim();
755 
756 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
757 
758 	/*
759 	 * Calculate the number of pages we want to either free or move
760 	 * to the cache.
761 	 */
762 	page_shortage = vm_paging_target() + addl_page_shortage_init;
763 
764 	vm_pageout_init_marker(&marker, PQ_INACTIVE);
765 
766 	/*
767 	 * Start scanning the inactive queue for pages we can move to the
768 	 * cache or free.  The scan will stop when the target is reached or
769 	 * we have scanned the entire inactive queue.  Note that m->act_count
770 	 * is not used to form decisions for the inactive queue, only for the
771 	 * active queue.
772 	 *
773 	 * maxlaunder limits the number of dirty pages we flush per scan.
774 	 * For most systems a smaller value (16 or 32) is more robust under
775 	 * extreme memory and disk pressure because any unnecessary writes
776 	 * to disk can result in extreme performance degredation.  However,
777 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
778 	 * used) will die horribly with limited laundering.  If the pageout
779 	 * daemon cannot clean enough pages in the first pass, we let it go
780 	 * all out in succeeding passes.
781 	 */
782 	if ((maxlaunder = vm_max_launder) <= 1)
783 		maxlaunder = 1;
784 	if (pass)
785 		maxlaunder = 10000;
786 	vm_page_lock_queues();
787 rescan0:
788 	addl_page_shortage = addl_page_shortage_init;
789 	maxscan = cnt.v_inactive_count;
790 
791 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
792 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
793 	     m = next) {
794 
795 		cnt.v_pdpages++;
796 
797 		if (m->queue != PQ_INACTIVE)
798 			goto rescan0;
799 
800 		next = TAILQ_NEXT(m, pageq);
801 
802 		/*
803 		 * skip marker pages
804 		 */
805 		if (m->flags & PG_MARKER)
806 			continue;
807 
808 		/*
809 		 * Lock the page.
810 		 */
811 		if (!vm_pageout_page_lock(m, &next)) {
812 			vm_page_unlock(m);
813 			addl_page_shortage++;
814 			continue;
815 		}
816 
817 		/*
818 		 * A held page may be undergoing I/O, so skip it.
819 		 */
820 		if (m->hold_count) {
821 			vm_page_unlock(m);
822 			vm_page_requeue(m);
823 			addl_page_shortage++;
824 			continue;
825 		}
826 
827 		/*
828 		 * Don't mess with busy pages, keep in the front of the
829 		 * queue, most likely are being paged out.
830 		 */
831 		object = m->object;
832 		if (!VM_OBJECT_TRYLOCK(object) &&
833 		    (!vm_pageout_fallback_object_lock(m, &next) ||
834 			m->hold_count != 0)) {
835 			VM_OBJECT_UNLOCK(object);
836 			vm_page_unlock(m);
837 			addl_page_shortage++;
838 			continue;
839 		}
840 		if (m->busy || (m->oflags & VPO_BUSY)) {
841 			vm_page_unlock(m);
842 			VM_OBJECT_UNLOCK(object);
843 			addl_page_shortage++;
844 			continue;
845 		}
846 
847 		/*
848 		 * If the object is not being used, we ignore previous
849 		 * references.
850 		 */
851 		if (object->ref_count == 0) {
852 			vm_page_aflag_clear(m, PGA_REFERENCED);
853 			KASSERT(!pmap_page_is_mapped(m),
854 			    ("vm_pageout_scan: page %p is mapped", m));
855 
856 		/*
857 		 * Otherwise, if the page has been referenced while in the
858 		 * inactive queue, we bump the "activation count" upwards,
859 		 * making it less likely that the page will be added back to
860 		 * the inactive queue prematurely again.  Here we check the
861 		 * page tables (or emulated bits, if any), given the upper
862 		 * level VM system not knowing anything about existing
863 		 * references.
864 		 */
865 		} else if (((m->aflags & PGA_REFERENCED) == 0) &&
866 			(actcount = pmap_ts_referenced(m))) {
867 			vm_page_activate(m);
868 			vm_page_unlock(m);
869 			m->act_count += actcount + ACT_ADVANCE;
870 			VM_OBJECT_UNLOCK(object);
871 			continue;
872 		}
873 
874 		/*
875 		 * If the upper level VM system knows about any page
876 		 * references, we activate the page.  We also set the
877 		 * "activation count" higher than normal so that we will less
878 		 * likely place pages back onto the inactive queue again.
879 		 */
880 		if ((m->aflags & PGA_REFERENCED) != 0) {
881 			vm_page_aflag_clear(m, PGA_REFERENCED);
882 			actcount = pmap_ts_referenced(m);
883 			vm_page_activate(m);
884 			vm_page_unlock(m);
885 			m->act_count += actcount + ACT_ADVANCE + 1;
886 			VM_OBJECT_UNLOCK(object);
887 			continue;
888 		}
889 
890 		/*
891 		 * If the upper level VM system does not believe that the page
892 		 * is fully dirty, but it is mapped for write access, then we
893 		 * consult the pmap to see if the page's dirty status should
894 		 * be updated.
895 		 */
896 		if (m->dirty != VM_PAGE_BITS_ALL &&
897 		    (m->aflags & PGA_WRITEABLE) != 0) {
898 			/*
899 			 * Avoid a race condition: Unless write access is
900 			 * removed from the page, another processor could
901 			 * modify it before all access is removed by the call
902 			 * to vm_page_cache() below.  If vm_page_cache() finds
903 			 * that the page has been modified when it removes all
904 			 * access, it panics because it cannot cache dirty
905 			 * pages.  In principle, we could eliminate just write
906 			 * access here rather than all access.  In the expected
907 			 * case, when there are no last instant modifications
908 			 * to the page, removing all access will be cheaper
909 			 * overall.
910 			 */
911 			if (pmap_is_modified(m))
912 				vm_page_dirty(m);
913 			else if (m->dirty == 0)
914 				pmap_remove_all(m);
915 		}
916 
917 		if (m->valid == 0) {
918 			/*
919 			 * Invalid pages can be easily freed
920 			 */
921 			vm_page_free(m);
922 			cnt.v_dfree++;
923 			--page_shortage;
924 		} else if (m->dirty == 0) {
925 			/*
926 			 * Clean pages can be placed onto the cache queue.
927 			 * This effectively frees them.
928 			 */
929 			vm_page_cache(m);
930 			--page_shortage;
931 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
932 			/*
933 			 * Dirty pages need to be paged out, but flushing
934 			 * a page is extremely expensive verses freeing
935 			 * a clean page.  Rather then artificially limiting
936 			 * the number of pages we can flush, we instead give
937 			 * dirty pages extra priority on the inactive queue
938 			 * by forcing them to be cycled through the queue
939 			 * twice before being flushed, after which the
940 			 * (now clean) page will cycle through once more
941 			 * before being freed.  This significantly extends
942 			 * the thrash point for a heavily loaded machine.
943 			 */
944 			m->flags |= PG_WINATCFLS;
945 			vm_page_requeue(m);
946 		} else if (maxlaunder > 0) {
947 			/*
948 			 * We always want to try to flush some dirty pages if
949 			 * we encounter them, to keep the system stable.
950 			 * Normally this number is small, but under extreme
951 			 * pressure where there are insufficient clean pages
952 			 * on the inactive queue, we may have to go all out.
953 			 */
954 			int swap_pageouts_ok, vfslocked = 0;
955 			struct vnode *vp = NULL;
956 			struct mount *mp = NULL;
957 
958 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
959 				swap_pageouts_ok = 1;
960 			} else {
961 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
962 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
963 				vm_page_count_min());
964 
965 			}
966 
967 			/*
968 			 * We don't bother paging objects that are "dead".
969 			 * Those objects are in a "rundown" state.
970 			 */
971 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
972 				vm_page_unlock(m);
973 				VM_OBJECT_UNLOCK(object);
974 				vm_page_requeue(m);
975 				continue;
976 			}
977 
978 			/*
979 			 * Following operations may unlock
980 			 * vm_page_queue_mtx, invalidating the 'next'
981 			 * pointer.  To prevent an inordinate number
982 			 * of restarts we use our marker to remember
983 			 * our place.
984 			 *
985 			 */
986 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
987 					   m, &marker, pageq);
988 			/*
989 			 * The object is already known NOT to be dead.   It
990 			 * is possible for the vget() to block the whole
991 			 * pageout daemon, but the new low-memory handling
992 			 * code should prevent it.
993 			 *
994 			 * The previous code skipped locked vnodes and, worse,
995 			 * reordered pages in the queue.  This results in
996 			 * completely non-deterministic operation and, on a
997 			 * busy system, can lead to extremely non-optimal
998 			 * pageouts.  For example, it can cause clean pages
999 			 * to be freed and dirty pages to be moved to the end
1000 			 * of the queue.  Since dirty pages are also moved to
1001 			 * the end of the queue once-cleaned, this gives
1002 			 * way too large a weighting to defering the freeing
1003 			 * of dirty pages.
1004 			 *
1005 			 * We can't wait forever for the vnode lock, we might
1006 			 * deadlock due to a vn_read() getting stuck in
1007 			 * vm_wait while holding this vnode.  We skip the
1008 			 * vnode if we can't get it in a reasonable amount
1009 			 * of time.
1010 			 */
1011 			if (object->type == OBJT_VNODE) {
1012 				vm_page_unlock_queues();
1013 				vm_page_unlock(m);
1014 				vp = object->handle;
1015 				if (vp->v_type == VREG &&
1016 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1017 					mp = NULL;
1018 					++pageout_lock_miss;
1019 					if (object->flags & OBJ_MIGHTBEDIRTY)
1020 						vnodes_skipped++;
1021 					vm_page_lock_queues();
1022 					goto unlock_and_continue;
1023 				}
1024 				KASSERT(mp != NULL,
1025 				    ("vp %p with NULL v_mount", vp));
1026 				vm_object_reference_locked(object);
1027 				VM_OBJECT_UNLOCK(object);
1028 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1029 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
1030 				    curthread)) {
1031 					VM_OBJECT_LOCK(object);
1032 					vm_page_lock_queues();
1033 					++pageout_lock_miss;
1034 					if (object->flags & OBJ_MIGHTBEDIRTY)
1035 						vnodes_skipped++;
1036 					vp = NULL;
1037 					goto unlock_and_continue;
1038 				}
1039 				VM_OBJECT_LOCK(object);
1040 				vm_page_lock(m);
1041 				vm_page_lock_queues();
1042 				/*
1043 				 * The page might have been moved to another
1044 				 * queue during potential blocking in vget()
1045 				 * above.  The page might have been freed and
1046 				 * reused for another vnode.
1047 				 */
1048 				if (m->queue != PQ_INACTIVE ||
1049 				    m->object != object ||
1050 				    TAILQ_NEXT(m, pageq) != &marker) {
1051 					vm_page_unlock(m);
1052 					if (object->flags & OBJ_MIGHTBEDIRTY)
1053 						vnodes_skipped++;
1054 					goto unlock_and_continue;
1055 				}
1056 
1057 				/*
1058 				 * The page may have been busied during the
1059 				 * blocking in vget().  We don't move the
1060 				 * page back onto the end of the queue so that
1061 				 * statistics are more correct if we don't.
1062 				 */
1063 				if (m->busy || (m->oflags & VPO_BUSY)) {
1064 					vm_page_unlock(m);
1065 					goto unlock_and_continue;
1066 				}
1067 
1068 				/*
1069 				 * If the page has become held it might
1070 				 * be undergoing I/O, so skip it
1071 				 */
1072 				if (m->hold_count) {
1073 					vm_page_unlock(m);
1074 					vm_page_requeue(m);
1075 					if (object->flags & OBJ_MIGHTBEDIRTY)
1076 						vnodes_skipped++;
1077 					goto unlock_and_continue;
1078 				}
1079 			}
1080 
1081 			/*
1082 			 * If a page is dirty, then it is either being washed
1083 			 * (but not yet cleaned) or it is still in the
1084 			 * laundry.  If it is still in the laundry, then we
1085 			 * start the cleaning operation.
1086 			 *
1087 			 * decrement page_shortage on success to account for
1088 			 * the (future) cleaned page.  Otherwise we could wind
1089 			 * up laundering or cleaning too many pages.
1090 			 */
1091 			vm_page_unlock_queues();
1092 			if (vm_pageout_clean(m) != 0) {
1093 				--page_shortage;
1094 				--maxlaunder;
1095 			}
1096 			vm_page_lock_queues();
1097 unlock_and_continue:
1098 			vm_page_lock_assert(m, MA_NOTOWNED);
1099 			VM_OBJECT_UNLOCK(object);
1100 			if (mp != NULL) {
1101 				vm_page_unlock_queues();
1102 				if (vp != NULL)
1103 					vput(vp);
1104 				VFS_UNLOCK_GIANT(vfslocked);
1105 				vm_object_deallocate(object);
1106 				vn_finished_write(mp);
1107 				vm_page_lock_queues();
1108 			}
1109 			next = TAILQ_NEXT(&marker, pageq);
1110 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1111 				     &marker, pageq);
1112 			vm_page_lock_assert(m, MA_NOTOWNED);
1113 			continue;
1114 		}
1115 		vm_page_unlock(m);
1116 		VM_OBJECT_UNLOCK(object);
1117 	}
1118 
1119 	/*
1120 	 * Compute the number of pages we want to try to move from the
1121 	 * active queue to the inactive queue.
1122 	 */
1123 	page_shortage = vm_paging_target() +
1124 		cnt.v_inactive_target - cnt.v_inactive_count;
1125 	page_shortage += addl_page_shortage;
1126 
1127 	/*
1128 	 * Scan the active queue for things we can deactivate. We nominally
1129 	 * track the per-page activity counter and use it to locate
1130 	 * deactivation candidates.
1131 	 */
1132 	pcount = cnt.v_active_count;
1133 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1134 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1135 
1136 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1137 
1138 		KASSERT(m->queue == PQ_ACTIVE,
1139 		    ("vm_pageout_scan: page %p isn't active", m));
1140 
1141 		next = TAILQ_NEXT(m, pageq);
1142 		if ((m->flags & PG_MARKER) != 0) {
1143 			m = next;
1144 			continue;
1145 		}
1146 		if (!vm_pageout_page_lock(m, &next)) {
1147 			vm_page_unlock(m);
1148 			m = next;
1149 			continue;
1150 		}
1151 		object = m->object;
1152 		if (!VM_OBJECT_TRYLOCK(object) &&
1153 		    !vm_pageout_fallback_object_lock(m, &next)) {
1154 			VM_OBJECT_UNLOCK(object);
1155 			vm_page_unlock(m);
1156 			m = next;
1157 			continue;
1158 		}
1159 
1160 		/*
1161 		 * Don't deactivate pages that are busy.
1162 		 */
1163 		if ((m->busy != 0) ||
1164 		    (m->oflags & VPO_BUSY) ||
1165 		    (m->hold_count != 0)) {
1166 			vm_page_unlock(m);
1167 			VM_OBJECT_UNLOCK(object);
1168 			vm_page_requeue(m);
1169 			m = next;
1170 			continue;
1171 		}
1172 
1173 		/*
1174 		 * The count for pagedaemon pages is done after checking the
1175 		 * page for eligibility...
1176 		 */
1177 		cnt.v_pdpages++;
1178 
1179 		/*
1180 		 * Check to see "how much" the page has been used.
1181 		 */
1182 		actcount = 0;
1183 		if (object->ref_count != 0) {
1184 			if (m->aflags & PGA_REFERENCED) {
1185 				actcount += 1;
1186 			}
1187 			actcount += pmap_ts_referenced(m);
1188 			if (actcount) {
1189 				m->act_count += ACT_ADVANCE + actcount;
1190 				if (m->act_count > ACT_MAX)
1191 					m->act_count = ACT_MAX;
1192 			}
1193 		}
1194 
1195 		/*
1196 		 * Since we have "tested" this bit, we need to clear it now.
1197 		 */
1198 		vm_page_aflag_clear(m, PGA_REFERENCED);
1199 
1200 		/*
1201 		 * Only if an object is currently being used, do we use the
1202 		 * page activation count stats.
1203 		 */
1204 		if (actcount && (object->ref_count != 0)) {
1205 			vm_page_requeue(m);
1206 		} else {
1207 			m->act_count -= min(m->act_count, ACT_DECLINE);
1208 			if (vm_pageout_algorithm ||
1209 			    object->ref_count == 0 ||
1210 			    m->act_count == 0) {
1211 				page_shortage--;
1212 				if (object->ref_count == 0) {
1213 					KASSERT(!pmap_page_is_mapped(m),
1214 				    ("vm_pageout_scan: page %p is mapped", m));
1215 					if (m->dirty == 0)
1216 						vm_page_cache(m);
1217 					else
1218 						vm_page_deactivate(m);
1219 				} else {
1220 					vm_page_deactivate(m);
1221 				}
1222 			} else {
1223 				vm_page_requeue(m);
1224 			}
1225 		}
1226 		vm_page_unlock(m);
1227 		VM_OBJECT_UNLOCK(object);
1228 		m = next;
1229 	}
1230 	vm_page_unlock_queues();
1231 #if !defined(NO_SWAPPING)
1232 	/*
1233 	 * Idle process swapout -- run once per second.
1234 	 */
1235 	if (vm_swap_idle_enabled) {
1236 		static long lsec;
1237 		if (time_second != lsec) {
1238 			vm_req_vmdaemon(VM_SWAP_IDLE);
1239 			lsec = time_second;
1240 		}
1241 	}
1242 #endif
1243 
1244 	/*
1245 	 * If we didn't get enough free pages, and we have skipped a vnode
1246 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1247 	 * if we did not get enough free pages.
1248 	 */
1249 	if (vm_paging_target() > 0) {
1250 		if (vnodes_skipped && vm_page_count_min())
1251 			(void) speedup_syncer();
1252 #if !defined(NO_SWAPPING)
1253 		if (vm_swap_enabled && vm_page_count_target())
1254 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1255 #endif
1256 	}
1257 
1258 	/*
1259 	 * If we are critically low on one of RAM or swap and low on
1260 	 * the other, kill the largest process.  However, we avoid
1261 	 * doing this on the first pass in order to give ourselves a
1262 	 * chance to flush out dirty vnode-backed pages and to allow
1263 	 * active pages to be moved to the inactive queue and reclaimed.
1264 	 */
1265 	if (pass != 0 &&
1266 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1267 	     (swap_pager_full && vm_paging_target() > 0)))
1268 		vm_pageout_oom(VM_OOM_MEM);
1269 }
1270 
1271 
1272 void
1273 vm_pageout_oom(int shortage)
1274 {
1275 	struct proc *p, *bigproc;
1276 	vm_offset_t size, bigsize;
1277 	struct thread *td;
1278 	struct vmspace *vm;
1279 
1280 	/*
1281 	 * We keep the process bigproc locked once we find it to keep anyone
1282 	 * from messing with it; however, there is a possibility of
1283 	 * deadlock if process B is bigproc and one of it's child processes
1284 	 * attempts to propagate a signal to B while we are waiting for A's
1285 	 * lock while walking this list.  To avoid this, we don't block on
1286 	 * the process lock but just skip a process if it is already locked.
1287 	 */
1288 	bigproc = NULL;
1289 	bigsize = 0;
1290 	sx_slock(&allproc_lock);
1291 	FOREACH_PROC_IN_SYSTEM(p) {
1292 		int breakout;
1293 
1294 		if (PROC_TRYLOCK(p) == 0)
1295 			continue;
1296 		/*
1297 		 * If this is a system, protected or killed process, skip it.
1298 		 */
1299 		if (p->p_state != PRS_NORMAL ||
1300 		    (p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1301 		    (p->p_pid == 1) || P_KILLED(p) ||
1302 		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1303 			PROC_UNLOCK(p);
1304 			continue;
1305 		}
1306 		/*
1307 		 * If the process is in a non-running type state,
1308 		 * don't touch it.  Check all the threads individually.
1309 		 */
1310 		breakout = 0;
1311 		FOREACH_THREAD_IN_PROC(p, td) {
1312 			thread_lock(td);
1313 			if (!TD_ON_RUNQ(td) &&
1314 			    !TD_IS_RUNNING(td) &&
1315 			    !TD_IS_SLEEPING(td) &&
1316 			    !TD_IS_SUSPENDED(td)) {
1317 				thread_unlock(td);
1318 				breakout = 1;
1319 				break;
1320 			}
1321 			thread_unlock(td);
1322 		}
1323 		if (breakout) {
1324 			PROC_UNLOCK(p);
1325 			continue;
1326 		}
1327 		/*
1328 		 * get the process size
1329 		 */
1330 		vm = vmspace_acquire_ref(p);
1331 		if (vm == NULL) {
1332 			PROC_UNLOCK(p);
1333 			continue;
1334 		}
1335 		if (!vm_map_trylock_read(&vm->vm_map)) {
1336 			vmspace_free(vm);
1337 			PROC_UNLOCK(p);
1338 			continue;
1339 		}
1340 		size = vmspace_swap_count(vm);
1341 		vm_map_unlock_read(&vm->vm_map);
1342 		if (shortage == VM_OOM_MEM)
1343 			size += vmspace_resident_count(vm);
1344 		vmspace_free(vm);
1345 		/*
1346 		 * if the this process is bigger than the biggest one
1347 		 * remember it.
1348 		 */
1349 		if (size > bigsize) {
1350 			if (bigproc != NULL)
1351 				PROC_UNLOCK(bigproc);
1352 			bigproc = p;
1353 			bigsize = size;
1354 		} else
1355 			PROC_UNLOCK(p);
1356 	}
1357 	sx_sunlock(&allproc_lock);
1358 	if (bigproc != NULL) {
1359 		killproc(bigproc, "out of swap space");
1360 		sched_nice(bigproc, PRIO_MIN);
1361 		PROC_UNLOCK(bigproc);
1362 		wakeup(&cnt.v_free_count);
1363 	}
1364 }
1365 
1366 /*
1367  * This routine tries to maintain the pseudo LRU active queue,
1368  * so that during long periods of time where there is no paging,
1369  * that some statistic accumulation still occurs.  This code
1370  * helps the situation where paging just starts to occur.
1371  */
1372 static void
1373 vm_pageout_page_stats()
1374 {
1375 	vm_object_t object;
1376 	vm_page_t m,next;
1377 	int pcount,tpcount;		/* Number of pages to check */
1378 	static int fullintervalcount = 0;
1379 	int page_shortage;
1380 
1381 	page_shortage =
1382 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1383 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1384 
1385 	if (page_shortage <= 0)
1386 		return;
1387 
1388 	vm_page_lock_queues();
1389 	pcount = cnt.v_active_count;
1390 	fullintervalcount += vm_pageout_stats_interval;
1391 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1392 		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1393 		    cnt.v_page_count;
1394 		if (pcount > tpcount)
1395 			pcount = tpcount;
1396 	} else {
1397 		fullintervalcount = 0;
1398 	}
1399 
1400 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1401 	while ((m != NULL) && (pcount-- > 0)) {
1402 		int actcount;
1403 
1404 		KASSERT(m->queue == PQ_ACTIVE,
1405 		    ("vm_pageout_page_stats: page %p isn't active", m));
1406 
1407 		next = TAILQ_NEXT(m, pageq);
1408 		if ((m->flags & PG_MARKER) != 0) {
1409 			m = next;
1410 			continue;
1411 		}
1412 		vm_page_lock_assert(m, MA_NOTOWNED);
1413 		if (!vm_pageout_page_lock(m, &next)) {
1414 			vm_page_unlock(m);
1415 			m = next;
1416 			continue;
1417 		}
1418 		object = m->object;
1419 		if (!VM_OBJECT_TRYLOCK(object) &&
1420 		    !vm_pageout_fallback_object_lock(m, &next)) {
1421 			VM_OBJECT_UNLOCK(object);
1422 			vm_page_unlock(m);
1423 			m = next;
1424 			continue;
1425 		}
1426 
1427 		/*
1428 		 * Don't deactivate pages that are busy.
1429 		 */
1430 		if ((m->busy != 0) ||
1431 		    (m->oflags & VPO_BUSY) ||
1432 		    (m->hold_count != 0)) {
1433 			vm_page_unlock(m);
1434 			VM_OBJECT_UNLOCK(object);
1435 			vm_page_requeue(m);
1436 			m = next;
1437 			continue;
1438 		}
1439 
1440 		actcount = 0;
1441 		if (m->aflags & PGA_REFERENCED) {
1442 			vm_page_aflag_clear(m, PGA_REFERENCED);
1443 			actcount += 1;
1444 		}
1445 
1446 		actcount += pmap_ts_referenced(m);
1447 		if (actcount) {
1448 			m->act_count += ACT_ADVANCE + actcount;
1449 			if (m->act_count > ACT_MAX)
1450 				m->act_count = ACT_MAX;
1451 			vm_page_requeue(m);
1452 		} else {
1453 			if (m->act_count == 0) {
1454 				/*
1455 				 * We turn off page access, so that we have
1456 				 * more accurate RSS stats.  We don't do this
1457 				 * in the normal page deactivation when the
1458 				 * system is loaded VM wise, because the
1459 				 * cost of the large number of page protect
1460 				 * operations would be higher than the value
1461 				 * of doing the operation.
1462 				 */
1463 				pmap_remove_all(m);
1464 				vm_page_deactivate(m);
1465 			} else {
1466 				m->act_count -= min(m->act_count, ACT_DECLINE);
1467 				vm_page_requeue(m);
1468 			}
1469 		}
1470 		vm_page_unlock(m);
1471 		VM_OBJECT_UNLOCK(object);
1472 		m = next;
1473 	}
1474 	vm_page_unlock_queues();
1475 }
1476 
1477 /*
1478  *	vm_pageout is the high level pageout daemon.
1479  */
1480 static void
1481 vm_pageout()
1482 {
1483 	int error, pass;
1484 
1485 	/*
1486 	 * Initialize some paging parameters.
1487 	 */
1488 	cnt.v_interrupt_free_min = 2;
1489 	if (cnt.v_page_count < 2000)
1490 		vm_pageout_page_count = 8;
1491 
1492 	/*
1493 	 * v_free_reserved needs to include enough for the largest
1494 	 * swap pager structures plus enough for any pv_entry structs
1495 	 * when paging.
1496 	 */
1497 	if (cnt.v_page_count > 1024)
1498 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1499 	else
1500 		cnt.v_free_min = 4;
1501 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1502 	    cnt.v_interrupt_free_min;
1503 	cnt.v_free_reserved = vm_pageout_page_count +
1504 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1505 	cnt.v_free_severe = cnt.v_free_min / 2;
1506 	cnt.v_free_min += cnt.v_free_reserved;
1507 	cnt.v_free_severe += cnt.v_free_reserved;
1508 
1509 	/*
1510 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1511 	 * that these are more a measure of the VM cache queue hysteresis
1512 	 * then the VM free queue.  Specifically, v_free_target is the
1513 	 * high water mark (free+cache pages).
1514 	 *
1515 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1516 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1517 	 * be big enough to handle memory needs while the pageout daemon
1518 	 * is signalled and run to free more pages.
1519 	 */
1520 	if (cnt.v_free_count > 6144)
1521 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1522 	else
1523 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1524 
1525 	if (cnt.v_free_count > 2048) {
1526 		cnt.v_cache_min = cnt.v_free_target;
1527 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1528 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1529 	} else {
1530 		cnt.v_cache_min = 0;
1531 		cnt.v_cache_max = 0;
1532 		cnt.v_inactive_target = cnt.v_free_count / 4;
1533 	}
1534 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1535 		cnt.v_inactive_target = cnt.v_free_count / 3;
1536 
1537 	/* XXX does not really belong here */
1538 	if (vm_page_max_wired == 0)
1539 		vm_page_max_wired = cnt.v_free_count / 3;
1540 
1541 	if (vm_pageout_stats_max == 0)
1542 		vm_pageout_stats_max = cnt.v_free_target;
1543 
1544 	/*
1545 	 * Set interval in seconds for stats scan.
1546 	 */
1547 	if (vm_pageout_stats_interval == 0)
1548 		vm_pageout_stats_interval = 5;
1549 	if (vm_pageout_full_stats_interval == 0)
1550 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1551 
1552 	swap_pager_swap_init();
1553 	pass = 0;
1554 	/*
1555 	 * The pageout daemon is never done, so loop forever.
1556 	 */
1557 	while (TRUE) {
1558 		/*
1559 		 * If we have enough free memory, wakeup waiters.  Do
1560 		 * not clear vm_pages_needed until we reach our target,
1561 		 * otherwise we may be woken up over and over again and
1562 		 * waste a lot of cpu.
1563 		 */
1564 		mtx_lock(&vm_page_queue_free_mtx);
1565 		if (vm_pages_needed && !vm_page_count_min()) {
1566 			if (!vm_paging_needed())
1567 				vm_pages_needed = 0;
1568 			wakeup(&cnt.v_free_count);
1569 		}
1570 		if (vm_pages_needed) {
1571 			/*
1572 			 * Still not done, take a second pass without waiting
1573 			 * (unlimited dirty cleaning), otherwise sleep a bit
1574 			 * and try again.
1575 			 */
1576 			++pass;
1577 			if (pass > 1)
1578 				msleep(&vm_pages_needed,
1579 				    &vm_page_queue_free_mtx, PVM, "psleep",
1580 				    hz / 2);
1581 		} else {
1582 			/*
1583 			 * Good enough, sleep & handle stats.  Prime the pass
1584 			 * for the next run.
1585 			 */
1586 			if (pass > 1)
1587 				pass = 1;
1588 			else
1589 				pass = 0;
1590 			error = msleep(&vm_pages_needed,
1591 			    &vm_page_queue_free_mtx, PVM, "psleep",
1592 			    vm_pageout_stats_interval * hz);
1593 			if (error && !vm_pages_needed) {
1594 				mtx_unlock(&vm_page_queue_free_mtx);
1595 				pass = 0;
1596 				vm_pageout_page_stats();
1597 				continue;
1598 			}
1599 		}
1600 		if (vm_pages_needed)
1601 			cnt.v_pdwakeups++;
1602 		mtx_unlock(&vm_page_queue_free_mtx);
1603 		vm_pageout_scan(pass);
1604 	}
1605 }
1606 
1607 /*
1608  * Unless the free page queue lock is held by the caller, this function
1609  * should be regarded as advisory.  Specifically, the caller should
1610  * not msleep() on &cnt.v_free_count following this function unless
1611  * the free page queue lock is held until the msleep() is performed.
1612  */
1613 void
1614 pagedaemon_wakeup()
1615 {
1616 
1617 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1618 		vm_pages_needed = 1;
1619 		wakeup(&vm_pages_needed);
1620 	}
1621 }
1622 
1623 #if !defined(NO_SWAPPING)
1624 static void
1625 vm_req_vmdaemon(int req)
1626 {
1627 	static int lastrun = 0;
1628 
1629 	mtx_lock(&vm_daemon_mtx);
1630 	vm_pageout_req_swapout |= req;
1631 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1632 		wakeup(&vm_daemon_needed);
1633 		lastrun = ticks;
1634 	}
1635 	mtx_unlock(&vm_daemon_mtx);
1636 }
1637 
1638 static void
1639 vm_daemon()
1640 {
1641 	struct rlimit rsslim;
1642 	struct proc *p;
1643 	struct thread *td;
1644 	struct vmspace *vm;
1645 	int breakout, swapout_flags, tryagain, attempts;
1646 #ifdef RACCT
1647 	uint64_t rsize, ravailable;
1648 #endif
1649 
1650 	while (TRUE) {
1651 		mtx_lock(&vm_daemon_mtx);
1652 #ifdef RACCT
1653 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", hz);
1654 #else
1655 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1656 #endif
1657 		swapout_flags = vm_pageout_req_swapout;
1658 		vm_pageout_req_swapout = 0;
1659 		mtx_unlock(&vm_daemon_mtx);
1660 		if (swapout_flags)
1661 			swapout_procs(swapout_flags);
1662 
1663 		/*
1664 		 * scan the processes for exceeding their rlimits or if
1665 		 * process is swapped out -- deactivate pages
1666 		 */
1667 		tryagain = 0;
1668 		attempts = 0;
1669 again:
1670 		attempts++;
1671 		sx_slock(&allproc_lock);
1672 		FOREACH_PROC_IN_SYSTEM(p) {
1673 			vm_pindex_t limit, size;
1674 
1675 			/*
1676 			 * if this is a system process or if we have already
1677 			 * looked at this process, skip it.
1678 			 */
1679 			PROC_LOCK(p);
1680 			if (p->p_state != PRS_NORMAL ||
1681 			    p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1682 				PROC_UNLOCK(p);
1683 				continue;
1684 			}
1685 			/*
1686 			 * if the process is in a non-running type state,
1687 			 * don't touch it.
1688 			 */
1689 			breakout = 0;
1690 			FOREACH_THREAD_IN_PROC(p, td) {
1691 				thread_lock(td);
1692 				if (!TD_ON_RUNQ(td) &&
1693 				    !TD_IS_RUNNING(td) &&
1694 				    !TD_IS_SLEEPING(td) &&
1695 				    !TD_IS_SUSPENDED(td)) {
1696 					thread_unlock(td);
1697 					breakout = 1;
1698 					break;
1699 				}
1700 				thread_unlock(td);
1701 			}
1702 			if (breakout) {
1703 				PROC_UNLOCK(p);
1704 				continue;
1705 			}
1706 			/*
1707 			 * get a limit
1708 			 */
1709 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1710 			limit = OFF_TO_IDX(
1711 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1712 
1713 			/*
1714 			 * let processes that are swapped out really be
1715 			 * swapped out set the limit to nothing (will force a
1716 			 * swap-out.)
1717 			 */
1718 			if ((p->p_flag & P_INMEM) == 0)
1719 				limit = 0;	/* XXX */
1720 			vm = vmspace_acquire_ref(p);
1721 			PROC_UNLOCK(p);
1722 			if (vm == NULL)
1723 				continue;
1724 
1725 			size = vmspace_resident_count(vm);
1726 			if (limit >= 0 && size >= limit) {
1727 				vm_pageout_map_deactivate_pages(
1728 				    &vm->vm_map, limit);
1729 			}
1730 #ifdef RACCT
1731 			rsize = IDX_TO_OFF(size);
1732 			PROC_LOCK(p);
1733 			racct_set(p, RACCT_RSS, rsize);
1734 			ravailable = racct_get_available(p, RACCT_RSS);
1735 			PROC_UNLOCK(p);
1736 			if (rsize > ravailable) {
1737 				/*
1738 				 * Don't be overly aggressive; this might be
1739 				 * an innocent process, and the limit could've
1740 				 * been exceeded by some memory hog.  Don't
1741 				 * try to deactivate more than 1/4th of process'
1742 				 * resident set size.
1743 				 */
1744 				if (attempts <= 8) {
1745 					if (ravailable < rsize - (rsize / 4))
1746 						ravailable = rsize - (rsize / 4);
1747 				}
1748 				vm_pageout_map_deactivate_pages(
1749 				    &vm->vm_map, OFF_TO_IDX(ravailable));
1750 				/* Update RSS usage after paging out. */
1751 				size = vmspace_resident_count(vm);
1752 				rsize = IDX_TO_OFF(size);
1753 				PROC_LOCK(p);
1754 				racct_set(p, RACCT_RSS, rsize);
1755 				PROC_UNLOCK(p);
1756 				if (rsize > ravailable)
1757 					tryagain = 1;
1758 			}
1759 #endif
1760 			vmspace_free(vm);
1761 		}
1762 		sx_sunlock(&allproc_lock);
1763 		if (tryagain != 0 && attempts <= 10)
1764 			goto again;
1765 	}
1766 }
1767 #endif			/* !defined(NO_SWAPPING) */
1768