xref: /freebsd/sys/vm/vm_pageout.c (revision 3642298923e528d795e3a30ec165d2b469e28b40)
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/resourcevar.h>
89 #include <sys/sched.h>
90 #include <sys/signalvar.h>
91 #include <sys/vnode.h>
92 #include <sys/vmmeter.h>
93 #include <sys/sx.h>
94 #include <sys/sysctl.h>
95 
96 #include <vm/vm.h>
97 #include <vm/vm_param.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_map.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/swap_pager.h>
104 #include <vm/vm_extern.h>
105 #include <vm/uma.h>
106 
107 #include <machine/mutex.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_pmap_collect(void);
117 static void vm_pageout_scan(int pass);
118 
119 struct proc *pageproc;
120 
121 static struct kproc_desc page_kp = {
122 	"pagedaemon",
123 	vm_pageout,
124 	&pageproc
125 };
126 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &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 #endif
150 static int vm_max_launder = 32;
151 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
152 static int vm_pageout_full_stats_interval = 0;
153 static int vm_pageout_algorithm=0;
154 static int defer_swap_pageouts=0;
155 static int disable_swap_pageouts=0;
156 
157 #if defined(NO_SWAPPING)
158 static int vm_swap_enabled=0;
159 static int vm_swap_idle_enabled=0;
160 #else
161 static int vm_swap_enabled=1;
162 static int vm_swap_idle_enabled=0;
163 #endif
164 
165 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
166 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
167 
168 SYSCTL_INT(_vm, OID_AUTO, max_launder,
169 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
170 
171 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
172 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
173 
174 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
175 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
176 
177 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
178 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
179 
180 #if defined(NO_SWAPPING)
181 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
182 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
183 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
184 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
185 #else
186 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
187 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
188 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
189 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
190 #endif
191 
192 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
193 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
194 
195 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
196 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
197 
198 static int pageout_lock_miss;
199 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
200 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
201 
202 #define VM_PAGEOUT_PAGE_COUNT 16
203 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
204 
205 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
206 
207 #if !defined(NO_SWAPPING)
208 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
209 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
210 static void vm_req_vmdaemon(void);
211 #endif
212 static void vm_pageout_page_stats(void);
213 
214 /*
215  * vm_pageout_fallback_object_lock:
216  *
217  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
218  * known to have failed and page queue must be either PQ_ACTIVE or
219  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
220  * while locking the vm object.  Use marker page to detect page queue
221  * changes and maintain notion of next page on page queue.  Return
222  * TRUE if no changes were detected, FALSE otherwise.  vm object is
223  * locked on return.
224  *
225  * This function depends on both the lock portion of struct vm_object
226  * and normal struct vm_page being type stable.
227  */
228 static boolean_t
229 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
230 {
231 	struct vm_page marker;
232 	boolean_t unchanged;
233 	u_short queue;
234 	vm_object_t object;
235 
236 	/*
237 	 * Initialize our marker
238 	 */
239 	bzero(&marker, sizeof(marker));
240 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
241 	marker.queue = m->queue;
242 	marker.wire_count = 1;
243 
244 	queue = m->queue;
245 	object = m->object;
246 
247 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
248 			   m, &marker, pageq);
249 	vm_page_unlock_queues();
250 	VM_OBJECT_LOCK(object);
251 	vm_page_lock_queues();
252 
253 	/* Page queue might have changed. */
254 	*next = TAILQ_NEXT(&marker, pageq);
255 	unchanged = (m->queue == queue &&
256 		     m->object == object &&
257 		     &marker == TAILQ_NEXT(m, pageq));
258 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
259 		     &marker, pageq);
260 	return (unchanged);
261 }
262 
263 /*
264  * vm_pageout_clean:
265  *
266  * Clean the page and remove it from the laundry.
267  *
268  * We set the busy bit to cause potential page faults on this page to
269  * block.  Note the careful timing, however, the busy bit isn't set till
270  * late and we cannot do anything that will mess with the page.
271  */
272 static int
273 vm_pageout_clean(m)
274 	vm_page_t m;
275 {
276 	vm_object_t object;
277 	vm_page_t mc[2*vm_pageout_page_count];
278 	int pageout_count;
279 	int ib, is, page_base;
280 	vm_pindex_t pindex = m->pindex;
281 
282 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
283 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
284 
285 	/*
286 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
287 	 * with the new swapper, but we could have serious problems paging
288 	 * out other object types if there is insufficient memory.
289 	 *
290 	 * Unfortunately, checking free memory here is far too late, so the
291 	 * check has been moved up a procedural level.
292 	 */
293 
294 	/*
295 	 * Don't mess with the page if it's busy, held, or special
296 	 */
297 	if ((m->hold_count != 0) ||
298 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
299 		return 0;
300 	}
301 
302 	mc[vm_pageout_page_count] = m;
303 	pageout_count = 1;
304 	page_base = vm_pageout_page_count;
305 	ib = 1;
306 	is = 1;
307 
308 	/*
309 	 * Scan object for clusterable pages.
310 	 *
311 	 * We can cluster ONLY if: ->> the page is NOT
312 	 * clean, wired, busy, held, or mapped into a
313 	 * buffer, and one of the following:
314 	 * 1) The page is inactive, or a seldom used
315 	 *    active page.
316 	 * -or-
317 	 * 2) we force the issue.
318 	 *
319 	 * During heavy mmap/modification loads the pageout
320 	 * daemon can really fragment the underlying file
321 	 * due to flushing pages out of order and not trying
322 	 * align the clusters (which leave sporatic out-of-order
323 	 * holes).  To solve this problem we do the reverse scan
324 	 * first and attempt to align our cluster, then do a
325 	 * forward scan if room remains.
326 	 */
327 	object = m->object;
328 more:
329 	while (ib && pageout_count < vm_pageout_page_count) {
330 		vm_page_t p;
331 
332 		if (ib > pindex) {
333 			ib = 0;
334 			break;
335 		}
336 
337 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
338 			ib = 0;
339 			break;
340 		}
341 		if (((p->queue - p->pc) == PQ_CACHE) ||
342 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
343 			ib = 0;
344 			break;
345 		}
346 		vm_page_test_dirty(p);
347 		if ((p->dirty & p->valid) == 0 ||
348 		    p->queue != PQ_INACTIVE ||
349 		    p->wire_count != 0 ||	/* may be held by buf cache */
350 		    p->hold_count != 0) {	/* may be undergoing I/O */
351 			ib = 0;
352 			break;
353 		}
354 		mc[--page_base] = p;
355 		++pageout_count;
356 		++ib;
357 		/*
358 		 * alignment boundry, stop here and switch directions.  Do
359 		 * not clear ib.
360 		 */
361 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
362 			break;
363 	}
364 
365 	while (pageout_count < vm_pageout_page_count &&
366 	    pindex + is < object->size) {
367 		vm_page_t p;
368 
369 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
370 			break;
371 		if (((p->queue - p->pc) == PQ_CACHE) ||
372 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
373 			break;
374 		}
375 		vm_page_test_dirty(p);
376 		if ((p->dirty & p->valid) == 0 ||
377 		    p->queue != PQ_INACTIVE ||
378 		    p->wire_count != 0 ||	/* may be held by buf cache */
379 		    p->hold_count != 0) {	/* may be undergoing I/O */
380 			break;
381 		}
382 		mc[page_base + pageout_count] = p;
383 		++pageout_count;
384 		++is;
385 	}
386 
387 	/*
388 	 * If we exhausted our forward scan, continue with the reverse scan
389 	 * when possible, even past a page boundry.  This catches boundry
390 	 * conditions.
391 	 */
392 	if (ib && pageout_count < vm_pageout_page_count)
393 		goto more;
394 
395 	/*
396 	 * we allow reads during pageouts...
397 	 */
398 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
399 }
400 
401 /*
402  * vm_pageout_flush() - launder the given pages
403  *
404  *	The given pages are laundered.  Note that we setup for the start of
405  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
406  *	reference count all in here rather then in the parent.  If we want
407  *	the parent to do more sophisticated things we may have to change
408  *	the ordering.
409  */
410 int
411 vm_pageout_flush(vm_page_t *mc, int count, int flags)
412 {
413 	vm_object_t object = mc[0]->object;
414 	int pageout_status[count];
415 	int numpagedout = 0;
416 	int i;
417 
418 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
419 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
420 	/*
421 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
422 	 * mark the pages read-only.
423 	 *
424 	 * We do not have to fixup the clean/dirty bits here... we can
425 	 * allow the pager to do it after the I/O completes.
426 	 *
427 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
428 	 * edge case with file fragments.
429 	 */
430 	for (i = 0; i < count; i++) {
431 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
432 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
433 			mc[i], i, count));
434 		vm_page_io_start(mc[i]);
435 		pmap_page_protect(mc[i], VM_PROT_READ);
436 	}
437 	vm_page_unlock_queues();
438 	vm_object_pip_add(object, count);
439 
440 	vm_pager_put_pages(object, mc, count,
441 	    (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
442 	    pageout_status);
443 
444 	vm_page_lock_queues();
445 	for (i = 0; i < count; i++) {
446 		vm_page_t mt = mc[i];
447 
448 		KASSERT((mt->flags & PG_WRITEABLE) == 0,
449 		    ("vm_pageout_flush: page %p is not write protected", mt));
450 		switch (pageout_status[i]) {
451 		case VM_PAGER_OK:
452 		case VM_PAGER_PEND:
453 			numpagedout++;
454 			break;
455 		case VM_PAGER_BAD:
456 			/*
457 			 * Page outside of range of object. Right now we
458 			 * essentially lose the changes by pretending it
459 			 * worked.
460 			 */
461 			pmap_clear_modify(mt);
462 			vm_page_undirty(mt);
463 			break;
464 		case VM_PAGER_ERROR:
465 		case VM_PAGER_FAIL:
466 			/*
467 			 * If page couldn't be paged out, then reactivate the
468 			 * page so it doesn't clog the inactive list.  (We
469 			 * will try paging out it again later).
470 			 */
471 			vm_page_activate(mt);
472 			break;
473 		case VM_PAGER_AGAIN:
474 			break;
475 		}
476 
477 		/*
478 		 * If the operation is still going, leave the page busy to
479 		 * block all other accesses. Also, leave the paging in
480 		 * progress indicator set so that we don't attempt an object
481 		 * collapse.
482 		 */
483 		if (pageout_status[i] != VM_PAGER_PEND) {
484 			vm_object_pip_wakeup(object);
485 			vm_page_io_finish(mt);
486 			if (vm_page_count_severe())
487 				vm_page_try_to_cache(mt);
488 		}
489 	}
490 	return numpagedout;
491 }
492 
493 #if !defined(NO_SWAPPING)
494 /*
495  *	vm_pageout_object_deactivate_pages
496  *
497  *	deactivate enough pages to satisfy the inactive target
498  *	requirements or if vm_page_proc_limit is set, then
499  *	deactivate all of the pages in the object and its
500  *	backing_objects.
501  *
502  *	The object and map must be locked.
503  */
504 static void
505 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
506 	pmap_t pmap;
507 	vm_object_t first_object;
508 	long desired;
509 {
510 	vm_object_t backing_object, object;
511 	vm_page_t p, next;
512 	int actcount, rcount, remove_mode;
513 
514 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
515 	if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
516 		return;
517 	for (object = first_object;; object = backing_object) {
518 		if (pmap_resident_count(pmap) <= desired)
519 			goto unlock_return;
520 		if (object->paging_in_progress)
521 			goto unlock_return;
522 
523 		remove_mode = 0;
524 		if (object->shadow_count > 1)
525 			remove_mode = 1;
526 		/*
527 		 * scan the objects entire memory queue
528 		 */
529 		rcount = object->resident_page_count;
530 		p = TAILQ_FIRST(&object->memq);
531 		vm_page_lock_queues();
532 		while (p && (rcount-- > 0)) {
533 			if (pmap_resident_count(pmap) <= desired) {
534 				vm_page_unlock_queues();
535 				goto unlock_return;
536 			}
537 			next = TAILQ_NEXT(p, listq);
538 			cnt.v_pdpages++;
539 			if (p->wire_count != 0 ||
540 			    p->hold_count != 0 ||
541 			    p->busy != 0 ||
542 			    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
543 			    !pmap_page_exists_quick(pmap, p)) {
544 				p = next;
545 				continue;
546 			}
547 			actcount = pmap_ts_referenced(p);
548 			if (actcount) {
549 				vm_page_flag_set(p, PG_REFERENCED);
550 			} else if (p->flags & PG_REFERENCED) {
551 				actcount = 1;
552 			}
553 			if ((p->queue != PQ_ACTIVE) &&
554 				(p->flags & PG_REFERENCED)) {
555 				vm_page_activate(p);
556 				p->act_count += actcount;
557 				vm_page_flag_clear(p, PG_REFERENCED);
558 			} else if (p->queue == PQ_ACTIVE) {
559 				if ((p->flags & PG_REFERENCED) == 0) {
560 					p->act_count -= min(p->act_count, ACT_DECLINE);
561 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
562 						pmap_remove_all(p);
563 						vm_page_deactivate(p);
564 					} else {
565 						vm_pageq_requeue(p);
566 					}
567 				} else {
568 					vm_page_activate(p);
569 					vm_page_flag_clear(p, PG_REFERENCED);
570 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
571 						p->act_count += ACT_ADVANCE;
572 					vm_pageq_requeue(p);
573 				}
574 			} else if (p->queue == PQ_INACTIVE) {
575 				pmap_remove_all(p);
576 			}
577 			p = next;
578 		}
579 		vm_page_unlock_queues();
580 		if ((backing_object = object->backing_object) == NULL)
581 			goto unlock_return;
582 		VM_OBJECT_LOCK(backing_object);
583 		if (object != first_object)
584 			VM_OBJECT_UNLOCK(object);
585 	}
586 unlock_return:
587 	if (object != first_object)
588 		VM_OBJECT_UNLOCK(object);
589 }
590 
591 /*
592  * deactivate some number of pages in a map, try to do it fairly, but
593  * that is really hard to do.
594  */
595 static void
596 vm_pageout_map_deactivate_pages(map, desired)
597 	vm_map_t map;
598 	long desired;
599 {
600 	vm_map_entry_t tmpe;
601 	vm_object_t obj, bigobj;
602 	int nothingwired;
603 
604 	if (!vm_map_trylock(map))
605 		return;
606 
607 	bigobj = NULL;
608 	nothingwired = TRUE;
609 
610 	/*
611 	 * first, search out the biggest object, and try to free pages from
612 	 * that.
613 	 */
614 	tmpe = map->header.next;
615 	while (tmpe != &map->header) {
616 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
617 			obj = tmpe->object.vm_object;
618 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
619 				if (obj->shadow_count <= 1 &&
620 				    (bigobj == NULL ||
621 				     bigobj->resident_page_count < obj->resident_page_count)) {
622 					if (bigobj != NULL)
623 						VM_OBJECT_UNLOCK(bigobj);
624 					bigobj = obj;
625 				} else
626 					VM_OBJECT_UNLOCK(obj);
627 			}
628 		}
629 		if (tmpe->wired_count > 0)
630 			nothingwired = FALSE;
631 		tmpe = tmpe->next;
632 	}
633 
634 	if (bigobj != NULL) {
635 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
636 		VM_OBJECT_UNLOCK(bigobj);
637 	}
638 	/*
639 	 * Next, hunt around for other pages to deactivate.  We actually
640 	 * do this search sort of wrong -- .text first is not the best idea.
641 	 */
642 	tmpe = map->header.next;
643 	while (tmpe != &map->header) {
644 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
645 			break;
646 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
647 			obj = tmpe->object.vm_object;
648 			if (obj != NULL) {
649 				VM_OBJECT_LOCK(obj);
650 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
651 				VM_OBJECT_UNLOCK(obj);
652 			}
653 		}
654 		tmpe = tmpe->next;
655 	}
656 
657 	/*
658 	 * Remove all mappings if a process is swapped out, this will free page
659 	 * table pages.
660 	 */
661 	if (desired == 0 && nothingwired) {
662 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
663 		    vm_map_max(map));
664 	}
665 	vm_map_unlock(map);
666 }
667 #endif		/* !defined(NO_SWAPPING) */
668 
669 /*
670  * This routine is very drastic, but can save the system
671  * in a pinch.
672  */
673 static void
674 vm_pageout_pmap_collect(void)
675 {
676 	int i;
677 	vm_page_t m;
678 	static int warningdone;
679 
680 	if (pmap_pagedaemon_waken == 0)
681 		return;
682 	if (warningdone < 5) {
683 		printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
684 		warningdone++;
685 	}
686 	vm_page_lock_queues();
687 	for (i = 0; i < vm_page_array_size; i++) {
688 		m = &vm_page_array[i];
689 		if (m->wire_count || m->hold_count || m->busy ||
690 		    (m->flags & (PG_BUSY | PG_UNMANAGED)))
691 			continue;
692 		pmap_remove_all(m);
693 	}
694 	vm_page_unlock_queues();
695 	pmap_pagedaemon_waken = 0;
696 }
697 
698 /*
699  *	vm_pageout_scan does the dirty work for the pageout daemon.
700  */
701 static void
702 vm_pageout_scan(int pass)
703 {
704 	vm_page_t m, next;
705 	struct vm_page marker;
706 	int page_shortage, maxscan, pcount;
707 	int addl_page_shortage, addl_page_shortage_init;
708 	struct proc *p, *bigproc;
709 	struct thread *td;
710 	vm_offset_t size, bigsize;
711 	vm_object_t object;
712 	int actcount, cache_cur, cache_first_failure;
713 	static int cache_last_free;
714 	int vnodes_skipped = 0;
715 	int maxlaunder;
716 
717 	mtx_lock(&Giant);
718 	/*
719 	 * Decrease registered cache sizes.
720 	 */
721 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
722 	/*
723 	 * We do this explicitly after the caches have been drained above.
724 	 */
725 	uma_reclaim();
726 	/*
727 	 * Do whatever cleanup that the pmap code can.
728 	 */
729 	vm_pageout_pmap_collect();
730 
731 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
732 
733 	/*
734 	 * Calculate the number of pages we want to either free or move
735 	 * to the cache.
736 	 */
737 	page_shortage = vm_paging_target() + addl_page_shortage_init;
738 
739 	/*
740 	 * Initialize our marker
741 	 */
742 	bzero(&marker, sizeof(marker));
743 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
744 	marker.queue = PQ_INACTIVE;
745 	marker.wire_count = 1;
746 
747 	/*
748 	 * Start scanning the inactive queue for pages we can move to the
749 	 * cache or free.  The scan will stop when the target is reached or
750 	 * we have scanned the entire inactive queue.  Note that m->act_count
751 	 * is not used to form decisions for the inactive queue, only for the
752 	 * active queue.
753 	 *
754 	 * maxlaunder limits the number of dirty pages we flush per scan.
755 	 * For most systems a smaller value (16 or 32) is more robust under
756 	 * extreme memory and disk pressure because any unnecessary writes
757 	 * to disk can result in extreme performance degredation.  However,
758 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
759 	 * used) will die horribly with limited laundering.  If the pageout
760 	 * daemon cannot clean enough pages in the first pass, we let it go
761 	 * all out in succeeding passes.
762 	 */
763 	if ((maxlaunder = vm_max_launder) <= 1)
764 		maxlaunder = 1;
765 	if (pass)
766 		maxlaunder = 10000;
767 	vm_page_lock_queues();
768 rescan0:
769 	addl_page_shortage = addl_page_shortage_init;
770 	maxscan = cnt.v_inactive_count;
771 
772 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
773 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
774 	     m = next) {
775 
776 		cnt.v_pdpages++;
777 
778 		if (m->queue != PQ_INACTIVE) {
779 			goto rescan0;
780 		}
781 
782 		next = TAILQ_NEXT(m, pageq);
783 		object = m->object;
784 
785 		/*
786 		 * skip marker pages
787 		 */
788 		if (m->flags & PG_MARKER)
789 			continue;
790 
791 		/*
792 		 * A held page may be undergoing I/O, so skip it.
793 		 */
794 		if (m->hold_count) {
795 			vm_pageq_requeue(m);
796 			addl_page_shortage++;
797 			continue;
798 		}
799 		/*
800 		 * Don't mess with busy pages, keep in the front of the
801 		 * queue, most likely are being paged out.
802 		 */
803 		if (!VM_OBJECT_TRYLOCK(object) &&
804 		    (!vm_pageout_fallback_object_lock(m, &next) ||
805 		     m->hold_count != 0)) {
806 			VM_OBJECT_UNLOCK(object);
807 			addl_page_shortage++;
808 			continue;
809 		}
810 		if (m->busy || (m->flags & PG_BUSY)) {
811 			VM_OBJECT_UNLOCK(object);
812 			addl_page_shortage++;
813 			continue;
814 		}
815 
816 		/*
817 		 * If the object is not being used, we ignore previous
818 		 * references.
819 		 */
820 		if (object->ref_count == 0) {
821 			vm_page_flag_clear(m, PG_REFERENCED);
822 			pmap_clear_reference(m);
823 
824 		/*
825 		 * Otherwise, if the page has been referenced while in the
826 		 * inactive queue, we bump the "activation count" upwards,
827 		 * making it less likely that the page will be added back to
828 		 * the inactive queue prematurely again.  Here we check the
829 		 * page tables (or emulated bits, if any), given the upper
830 		 * level VM system not knowing anything about existing
831 		 * references.
832 		 */
833 		} else if (((m->flags & PG_REFERENCED) == 0) &&
834 			(actcount = pmap_ts_referenced(m))) {
835 			vm_page_activate(m);
836 			VM_OBJECT_UNLOCK(object);
837 			m->act_count += (actcount + ACT_ADVANCE);
838 			continue;
839 		}
840 
841 		/*
842 		 * If the upper level VM system knows about any page
843 		 * references, we activate the page.  We also set the
844 		 * "activation count" higher than normal so that we will less
845 		 * likely place pages back onto the inactive queue again.
846 		 */
847 		if ((m->flags & PG_REFERENCED) != 0) {
848 			vm_page_flag_clear(m, PG_REFERENCED);
849 			actcount = pmap_ts_referenced(m);
850 			vm_page_activate(m);
851 			VM_OBJECT_UNLOCK(object);
852 			m->act_count += (actcount + ACT_ADVANCE + 1);
853 			continue;
854 		}
855 
856 		/*
857 		 * If the upper level VM system doesn't know anything about
858 		 * the page being dirty, we have to check for it again.  As
859 		 * far as the VM code knows, any partially dirty pages are
860 		 * fully dirty.
861 		 */
862 		if (m->dirty == 0 && !pmap_is_modified(m)) {
863 			/*
864 			 * Avoid a race condition: Unless write access is
865 			 * removed from the page, another processor could
866 			 * modify it before all access is removed by the call
867 			 * to vm_page_cache() below.  If vm_page_cache() finds
868 			 * that the page has been modified when it removes all
869 			 * access, it panics because it cannot cache dirty
870 			 * pages.  In principle, we could eliminate just write
871 			 * access here rather than all access.  In the expected
872 			 * case, when there are no last instant modifications
873 			 * to the page, removing all access will be cheaper
874 			 * overall.
875 			 */
876 			if ((m->flags & PG_WRITEABLE) != 0)
877 				pmap_remove_all(m);
878 		} else {
879 			vm_page_dirty(m);
880 		}
881 
882 		if (m->valid == 0) {
883 			/*
884 			 * Invalid pages can be easily freed
885 			 */
886 			pmap_remove_all(m);
887 			vm_page_free(m);
888 			cnt.v_dfree++;
889 			--page_shortage;
890 		} else if (m->dirty == 0) {
891 			/*
892 			 * Clean pages can be placed onto the cache queue.
893 			 * This effectively frees them.
894 			 */
895 			vm_page_cache(m);
896 			--page_shortage;
897 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
898 			/*
899 			 * Dirty pages need to be paged out, but flushing
900 			 * a page is extremely expensive verses freeing
901 			 * a clean page.  Rather then artificially limiting
902 			 * the number of pages we can flush, we instead give
903 			 * dirty pages extra priority on the inactive queue
904 			 * by forcing them to be cycled through the queue
905 			 * twice before being flushed, after which the
906 			 * (now clean) page will cycle through once more
907 			 * before being freed.  This significantly extends
908 			 * the thrash point for a heavily loaded machine.
909 			 */
910 			vm_page_flag_set(m, PG_WINATCFLS);
911 			vm_pageq_requeue(m);
912 		} else if (maxlaunder > 0) {
913 			/*
914 			 * We always want to try to flush some dirty pages if
915 			 * we encounter them, to keep the system stable.
916 			 * Normally this number is small, but under extreme
917 			 * pressure where there are insufficient clean pages
918 			 * on the inactive queue, we may have to go all out.
919 			 */
920 			int swap_pageouts_ok;
921 			struct vnode *vp = NULL;
922 			struct mount *mp;
923 
924 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
925 				swap_pageouts_ok = 1;
926 			} else {
927 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
928 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
929 				vm_page_count_min());
930 
931 			}
932 
933 			/*
934 			 * We don't bother paging objects that are "dead".
935 			 * Those objects are in a "rundown" state.
936 			 */
937 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
938 				VM_OBJECT_UNLOCK(object);
939 				vm_pageq_requeue(m);
940 				continue;
941 			}
942 
943 			/*
944 			 * The object is already known NOT to be dead.   It
945 			 * is possible for the vget() to block the whole
946 			 * pageout daemon, but the new low-memory handling
947 			 * code should prevent it.
948 			 *
949 			 * The previous code skipped locked vnodes and, worse,
950 			 * reordered pages in the queue.  This results in
951 			 * completely non-deterministic operation and, on a
952 			 * busy system, can lead to extremely non-optimal
953 			 * pageouts.  For example, it can cause clean pages
954 			 * to be freed and dirty pages to be moved to the end
955 			 * of the queue.  Since dirty pages are also moved to
956 			 * the end of the queue once-cleaned, this gives
957 			 * way too large a weighting to defering the freeing
958 			 * of dirty pages.
959 			 *
960 			 * We can't wait forever for the vnode lock, we might
961 			 * deadlock due to a vn_read() getting stuck in
962 			 * vm_wait while holding this vnode.  We skip the
963 			 * vnode if we can't get it in a reasonable amount
964 			 * of time.
965 			 */
966 			if (object->type == OBJT_VNODE) {
967 				vp = object->handle;
968 				mp = NULL;
969 				if (vp->v_type == VREG)
970 					vn_start_write(vp, &mp, V_NOWAIT);
971 				vm_page_unlock_queues();
972 				VI_LOCK(vp);
973 				VM_OBJECT_UNLOCK(object);
974 				if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK |
975 				    LK_TIMELOCK, curthread)) {
976 					VM_OBJECT_LOCK(object);
977 					vm_page_lock_queues();
978 					++pageout_lock_miss;
979 					vn_finished_write(mp);
980 					if (object->flags & OBJ_MIGHTBEDIRTY)
981 						vnodes_skipped++;
982 					VM_OBJECT_UNLOCK(object);
983 					continue;
984 				}
985 				VM_OBJECT_LOCK(object);
986 				vm_page_lock_queues();
987 				/*
988 				 * The page might have been moved to another
989 				 * queue during potential blocking in vget()
990 				 * above.  The page might have been freed and
991 				 * reused for another vnode.  The object might
992 				 * have been reused for another vnode.
993 				 */
994 				if (m->queue != PQ_INACTIVE ||
995 				    m->object != object ||
996 				    object->handle != vp) {
997 					if (object->flags & OBJ_MIGHTBEDIRTY)
998 						vnodes_skipped++;
999 					goto unlock_and_continue;
1000 				}
1001 
1002 				/*
1003 				 * The page may have been busied during the
1004 				 * blocking in vput();  We don't move the
1005 				 * page back onto the end of the queue so that
1006 				 * statistics are more correct if we don't.
1007 				 */
1008 				if (m->busy || (m->flags & PG_BUSY)) {
1009 					goto unlock_and_continue;
1010 				}
1011 
1012 				/*
1013 				 * If the page has become held it might
1014 				 * be undergoing I/O, so skip it
1015 				 */
1016 				if (m->hold_count) {
1017 					vm_pageq_requeue(m);
1018 					if (object->flags & OBJ_MIGHTBEDIRTY)
1019 						vnodes_skipped++;
1020 					goto unlock_and_continue;
1021 				}
1022 			}
1023 
1024 			/*
1025 			 * If a page is dirty, then it is either being washed
1026 			 * (but not yet cleaned) or it is still in the
1027 			 * laundry.  If it is still in the laundry, then we
1028 			 * start the cleaning operation.
1029 			 *
1030 			 * This operation may cluster, invalidating the 'next'
1031 			 * pointer.  To prevent an inordinate number of
1032 			 * restarts we use our marker to remember our place.
1033 			 *
1034 			 * decrement page_shortage on success to account for
1035 			 * the (future) cleaned page.  Otherwise we could wind
1036 			 * up laundering or cleaning too many pages.
1037 			 */
1038 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1039 			if (vm_pageout_clean(m) != 0) {
1040 				--page_shortage;
1041 				--maxlaunder;
1042 			}
1043 			next = TAILQ_NEXT(&marker, pageq);
1044 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1045 unlock_and_continue:
1046 			VM_OBJECT_UNLOCK(object);
1047 			if (vp) {
1048 				vm_page_unlock_queues();
1049 				vput(vp);
1050 				vn_finished_write(mp);
1051 				vm_page_lock_queues();
1052 			}
1053 			continue;
1054 		}
1055 		VM_OBJECT_UNLOCK(object);
1056 	}
1057 
1058 	/*
1059 	 * Compute the number of pages we want to try to move from the
1060 	 * active queue to the inactive queue.
1061 	 */
1062 	page_shortage = vm_paging_target() +
1063 		cnt.v_inactive_target - cnt.v_inactive_count;
1064 	page_shortage += addl_page_shortage;
1065 
1066 	/*
1067 	 * Scan the active queue for things we can deactivate. We nominally
1068 	 * track the per-page activity counter and use it to locate
1069 	 * deactivation candidates.
1070 	 */
1071 	pcount = cnt.v_active_count;
1072 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1073 
1074 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1075 
1076 		KASSERT(m->queue == PQ_ACTIVE,
1077 		    ("vm_pageout_scan: page %p isn't active", m));
1078 
1079 		next = TAILQ_NEXT(m, pageq);
1080 		object = m->object;
1081 		if ((m->flags & PG_MARKER) != 0) {
1082 			m = next;
1083 			continue;
1084 		}
1085 		if (!VM_OBJECT_TRYLOCK(object) &&
1086 		    !vm_pageout_fallback_object_lock(m, &next)) {
1087 			VM_OBJECT_UNLOCK(object);
1088 			m = next;
1089 			continue;
1090 		}
1091 
1092 		/*
1093 		 * Don't deactivate pages that are busy.
1094 		 */
1095 		if ((m->busy != 0) ||
1096 		    (m->flags & PG_BUSY) ||
1097 		    (m->hold_count != 0)) {
1098 			VM_OBJECT_UNLOCK(object);
1099 			vm_pageq_requeue(m);
1100 			m = next;
1101 			continue;
1102 		}
1103 
1104 		/*
1105 		 * The count for pagedaemon pages is done after checking the
1106 		 * page for eligibility...
1107 		 */
1108 		cnt.v_pdpages++;
1109 
1110 		/*
1111 		 * Check to see "how much" the page has been used.
1112 		 */
1113 		actcount = 0;
1114 		if (object->ref_count != 0) {
1115 			if (m->flags & PG_REFERENCED) {
1116 				actcount += 1;
1117 			}
1118 			actcount += pmap_ts_referenced(m);
1119 			if (actcount) {
1120 				m->act_count += ACT_ADVANCE + actcount;
1121 				if (m->act_count > ACT_MAX)
1122 					m->act_count = ACT_MAX;
1123 			}
1124 		}
1125 
1126 		/*
1127 		 * Since we have "tested" this bit, we need to clear it now.
1128 		 */
1129 		vm_page_flag_clear(m, PG_REFERENCED);
1130 
1131 		/*
1132 		 * Only if an object is currently being used, do we use the
1133 		 * page activation count stats.
1134 		 */
1135 		if (actcount && (object->ref_count != 0)) {
1136 			vm_pageq_requeue(m);
1137 		} else {
1138 			m->act_count -= min(m->act_count, ACT_DECLINE);
1139 			if (vm_pageout_algorithm ||
1140 			    object->ref_count == 0 ||
1141 			    m->act_count == 0) {
1142 				page_shortage--;
1143 				if (object->ref_count == 0) {
1144 					pmap_remove_all(m);
1145 					if (m->dirty == 0)
1146 						vm_page_cache(m);
1147 					else
1148 						vm_page_deactivate(m);
1149 				} else {
1150 					vm_page_deactivate(m);
1151 				}
1152 			} else {
1153 				vm_pageq_requeue(m);
1154 			}
1155 		}
1156 		VM_OBJECT_UNLOCK(object);
1157 		m = next;
1158 	}
1159 
1160 	/*
1161 	 * We try to maintain some *really* free pages, this allows interrupt
1162 	 * code to be guaranteed space.  Since both cache and free queues
1163 	 * are considered basically 'free', moving pages from cache to free
1164 	 * does not effect other calculations.
1165 	 */
1166 	cache_cur = cache_last_free;
1167 	cache_first_failure = -1;
1168 	while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur =
1169 	    (cache_cur + PQ_PRIME2) & PQ_L2_MASK) != cache_first_failure) {
1170 		TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl,
1171 		    pageq) {
1172 			KASSERT(m->dirty == 0,
1173 			    ("Found dirty cache page %p", m));
1174 			KASSERT(!pmap_page_is_mapped(m),
1175 			    ("Found mapped cache page %p", m));
1176 			KASSERT((m->flags & PG_UNMANAGED) == 0,
1177 			    ("Found unmanaged cache page %p", m));
1178 			KASSERT(m->wire_count == 0,
1179 			    ("Found wired cache page %p", m));
1180 			if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1181 			    m->object)) {
1182 				KASSERT((m->flags & PG_BUSY) == 0 &&
1183 				    m->busy == 0, ("Found busy cache page %p",
1184 				    m));
1185 				vm_page_free(m);
1186 				VM_OBJECT_UNLOCK(object);
1187 				cnt.v_dfree++;
1188 				cache_last_free = cache_cur;
1189 				cache_first_failure = -1;
1190 				break;
1191 			}
1192 		}
1193 		if (m == NULL && cache_first_failure == -1)
1194 			cache_first_failure = cache_cur;
1195 	}
1196 	vm_page_unlock_queues();
1197 #if !defined(NO_SWAPPING)
1198 	/*
1199 	 * Idle process swapout -- run once per second.
1200 	 */
1201 	if (vm_swap_idle_enabled) {
1202 		static long lsec;
1203 		if (time_second != lsec) {
1204 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1205 			vm_req_vmdaemon();
1206 			lsec = time_second;
1207 		}
1208 	}
1209 #endif
1210 
1211 	/*
1212 	 * If we didn't get enough free pages, and we have skipped a vnode
1213 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1214 	 * if we did not get enough free pages.
1215 	 */
1216 	if (vm_paging_target() > 0) {
1217 		if (vnodes_skipped && vm_page_count_min())
1218 			(void) speedup_syncer();
1219 #if !defined(NO_SWAPPING)
1220 		if (vm_swap_enabled && vm_page_count_target()) {
1221 			vm_req_vmdaemon();
1222 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1223 		}
1224 #endif
1225 	}
1226 
1227 	/*
1228 	 * If we are critically low on one of RAM or swap and low on
1229 	 * the other, kill the largest process.  However, we avoid
1230 	 * doing this on the first pass in order to give ourselves a
1231 	 * chance to flush out dirty vnode-backed pages and to allow
1232 	 * active pages to be moved to the inactive queue and reclaimed.
1233 	 *
1234 	 * We keep the process bigproc locked once we find it to keep anyone
1235 	 * from messing with it; however, there is a possibility of
1236 	 * deadlock if process B is bigproc and one of it's child processes
1237 	 * attempts to propagate a signal to B while we are waiting for A's
1238 	 * lock while walking this list.  To avoid this, we don't block on
1239 	 * the process lock but just skip a process if it is already locked.
1240 	 */
1241 	if (pass != 0 &&
1242 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1243 	     (swap_pager_full && vm_paging_target() > 0))) {
1244 		bigproc = NULL;
1245 		bigsize = 0;
1246 		sx_slock(&allproc_lock);
1247 		FOREACH_PROC_IN_SYSTEM(p) {
1248 			int breakout;
1249 
1250 			if (PROC_TRYLOCK(p) == 0)
1251 				continue;
1252 			/*
1253 			 * If this is a system or protected process, skip it.
1254 			 */
1255 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1256 			    (p->p_flag & P_PROTECTED) ||
1257 			    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1258 				PROC_UNLOCK(p);
1259 				continue;
1260 			}
1261 			/*
1262 			 * If the process is in a non-running type state,
1263 			 * don't touch it.  Check all the threads individually.
1264 			 */
1265 			mtx_lock_spin(&sched_lock);
1266 			breakout = 0;
1267 			FOREACH_THREAD_IN_PROC(p, td) {
1268 				if (!TD_ON_RUNQ(td) &&
1269 				    !TD_IS_RUNNING(td) &&
1270 				    !TD_IS_SLEEPING(td)) {
1271 					breakout = 1;
1272 					break;
1273 				}
1274 			}
1275 			if (breakout) {
1276 				mtx_unlock_spin(&sched_lock);
1277 				PROC_UNLOCK(p);
1278 				continue;
1279 			}
1280 			mtx_unlock_spin(&sched_lock);
1281 			/*
1282 			 * get the process size
1283 			 */
1284 			if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1285 				PROC_UNLOCK(p);
1286 				continue;
1287 			}
1288 			size = vmspace_swap_count(p->p_vmspace);
1289 			vm_map_unlock_read(&p->p_vmspace->vm_map);
1290 			size += vmspace_resident_count(p->p_vmspace);
1291 			/*
1292 			 * if the this process is bigger than the biggest one
1293 			 * remember it.
1294 			 */
1295 			if (size > bigsize) {
1296 				if (bigproc != NULL)
1297 					PROC_UNLOCK(bigproc);
1298 				bigproc = p;
1299 				bigsize = size;
1300 			} else
1301 				PROC_UNLOCK(p);
1302 		}
1303 		sx_sunlock(&allproc_lock);
1304 		if (bigproc != NULL) {
1305 			killproc(bigproc, "out of swap space");
1306 			mtx_lock_spin(&sched_lock);
1307 			sched_nice(bigproc, PRIO_MIN);
1308 			mtx_unlock_spin(&sched_lock);
1309 			PROC_UNLOCK(bigproc);
1310 			wakeup(&cnt.v_free_count);
1311 		}
1312 	}
1313 	mtx_unlock(&Giant);
1314 }
1315 
1316 /*
1317  * This routine tries to maintain the pseudo LRU active queue,
1318  * so that during long periods of time where there is no paging,
1319  * that some statistic accumulation still occurs.  This code
1320  * helps the situation where paging just starts to occur.
1321  */
1322 static void
1323 vm_pageout_page_stats()
1324 {
1325 	vm_object_t object;
1326 	vm_page_t m,next;
1327 	int pcount,tpcount;		/* Number of pages to check */
1328 	static int fullintervalcount = 0;
1329 	int page_shortage;
1330 
1331 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1332 	page_shortage =
1333 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1334 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1335 
1336 	if (page_shortage <= 0)
1337 		return;
1338 
1339 	pcount = cnt.v_active_count;
1340 	fullintervalcount += vm_pageout_stats_interval;
1341 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1342 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1343 		if (pcount > tpcount)
1344 			pcount = tpcount;
1345 	} else {
1346 		fullintervalcount = 0;
1347 	}
1348 
1349 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1350 	while ((m != NULL) && (pcount-- > 0)) {
1351 		int actcount;
1352 
1353 		KASSERT(m->queue == PQ_ACTIVE,
1354 		    ("vm_pageout_page_stats: page %p isn't active", m));
1355 
1356 		next = TAILQ_NEXT(m, pageq);
1357 		object = m->object;
1358 
1359 		if ((m->flags & PG_MARKER) != 0) {
1360 			m = next;
1361 			continue;
1362 		}
1363 		if (!VM_OBJECT_TRYLOCK(object) &&
1364 		    !vm_pageout_fallback_object_lock(m, &next)) {
1365 			VM_OBJECT_UNLOCK(object);
1366 			m = next;
1367 			continue;
1368 		}
1369 
1370 		/*
1371 		 * Don't deactivate pages that are busy.
1372 		 */
1373 		if ((m->busy != 0) ||
1374 		    (m->flags & PG_BUSY) ||
1375 		    (m->hold_count != 0)) {
1376 			VM_OBJECT_UNLOCK(object);
1377 			vm_pageq_requeue(m);
1378 			m = next;
1379 			continue;
1380 		}
1381 
1382 		actcount = 0;
1383 		if (m->flags & PG_REFERENCED) {
1384 			vm_page_flag_clear(m, PG_REFERENCED);
1385 			actcount += 1;
1386 		}
1387 
1388 		actcount += pmap_ts_referenced(m);
1389 		if (actcount) {
1390 			m->act_count += ACT_ADVANCE + actcount;
1391 			if (m->act_count > ACT_MAX)
1392 				m->act_count = ACT_MAX;
1393 			vm_pageq_requeue(m);
1394 		} else {
1395 			if (m->act_count == 0) {
1396 				/*
1397 				 * We turn off page access, so that we have
1398 				 * more accurate RSS stats.  We don't do this
1399 				 * in the normal page deactivation when the
1400 				 * system is loaded VM wise, because the
1401 				 * cost of the large number of page protect
1402 				 * operations would be higher than the value
1403 				 * of doing the operation.
1404 				 */
1405 				pmap_remove_all(m);
1406 				vm_page_deactivate(m);
1407 			} else {
1408 				m->act_count -= min(m->act_count, ACT_DECLINE);
1409 				vm_pageq_requeue(m);
1410 			}
1411 		}
1412 		VM_OBJECT_UNLOCK(object);
1413 		m = next;
1414 	}
1415 }
1416 
1417 /*
1418  *	vm_pageout is the high level pageout daemon.
1419  */
1420 static void
1421 vm_pageout()
1422 {
1423 	int error, pass;
1424 
1425 	/*
1426 	 * Initialize some paging parameters.
1427 	 */
1428 	cnt.v_interrupt_free_min = 2;
1429 	if (cnt.v_page_count < 2000)
1430 		vm_pageout_page_count = 8;
1431 
1432 	/*
1433 	 * v_free_reserved needs to include enough for the largest
1434 	 * swap pager structures plus enough for any pv_entry structs
1435 	 * when paging.
1436 	 */
1437 	if (cnt.v_page_count > 1024)
1438 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1439 	else
1440 		cnt.v_free_min = 4;
1441 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1442 	    cnt.v_interrupt_free_min;
1443 	cnt.v_free_reserved = vm_pageout_page_count +
1444 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_L2_SIZE;
1445 	cnt.v_free_severe = cnt.v_free_min / 2;
1446 	cnt.v_free_min += cnt.v_free_reserved;
1447 	cnt.v_free_severe += cnt.v_free_reserved;
1448 
1449 	/*
1450 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1451 	 * that these are more a measure of the VM cache queue hysteresis
1452 	 * then the VM free queue.  Specifically, v_free_target is the
1453 	 * high water mark (free+cache pages).
1454 	 *
1455 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1456 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1457 	 * be big enough to handle memory needs while the pageout daemon
1458 	 * is signalled and run to free more pages.
1459 	 */
1460 	if (cnt.v_free_count > 6144)
1461 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1462 	else
1463 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1464 
1465 	if (cnt.v_free_count > 2048) {
1466 		cnt.v_cache_min = cnt.v_free_target;
1467 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1468 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1469 	} else {
1470 		cnt.v_cache_min = 0;
1471 		cnt.v_cache_max = 0;
1472 		cnt.v_inactive_target = cnt.v_free_count / 4;
1473 	}
1474 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1475 		cnt.v_inactive_target = cnt.v_free_count / 3;
1476 
1477 	/* XXX does not really belong here */
1478 	if (vm_page_max_wired == 0)
1479 		vm_page_max_wired = cnt.v_free_count / 3;
1480 
1481 	if (vm_pageout_stats_max == 0)
1482 		vm_pageout_stats_max = cnt.v_free_target;
1483 
1484 	/*
1485 	 * Set interval in seconds for stats scan.
1486 	 */
1487 	if (vm_pageout_stats_interval == 0)
1488 		vm_pageout_stats_interval = 5;
1489 	if (vm_pageout_full_stats_interval == 0)
1490 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1491 
1492 	swap_pager_swap_init();
1493 	pass = 0;
1494 	/*
1495 	 * The pageout daemon is never done, so loop forever.
1496 	 */
1497 	while (TRUE) {
1498 		vm_page_lock_queues();
1499 		/*
1500 		 * If we have enough free memory, wakeup waiters.  Do
1501 		 * not clear vm_pages_needed until we reach our target,
1502 		 * otherwise we may be woken up over and over again and
1503 		 * waste a lot of cpu.
1504 		 */
1505 		if (vm_pages_needed && !vm_page_count_min()) {
1506 			if (!vm_paging_needed())
1507 				vm_pages_needed = 0;
1508 			wakeup(&cnt.v_free_count);
1509 		}
1510 		if (vm_pages_needed) {
1511 			/*
1512 			 * Still not done, take a second pass without waiting
1513 			 * (unlimited dirty cleaning), otherwise sleep a bit
1514 			 * and try again.
1515 			 */
1516 			++pass;
1517 			if (pass > 1)
1518 				msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1519 				       "psleep", hz/2);
1520 		} else {
1521 			/*
1522 			 * Good enough, sleep & handle stats.  Prime the pass
1523 			 * for the next run.
1524 			 */
1525 			if (pass > 1)
1526 				pass = 1;
1527 			else
1528 				pass = 0;
1529 			error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1530 				    "psleep", vm_pageout_stats_interval * hz);
1531 			if (error && !vm_pages_needed) {
1532 				pass = 0;
1533 				vm_pageout_page_stats();
1534 				vm_page_unlock_queues();
1535 				continue;
1536 			}
1537 		}
1538 		if (vm_pages_needed)
1539 			cnt.v_pdwakeups++;
1540 		vm_page_unlock_queues();
1541 		vm_pageout_scan(pass);
1542 	}
1543 }
1544 
1545 /*
1546  * Unless the page queue lock is held by the caller, this function
1547  * should be regarded as advisory.  Specifically, the caller should
1548  * not msleep() on &cnt.v_free_count following this function unless
1549  * the page queue lock is held until the msleep() is performed.
1550  */
1551 void
1552 pagedaemon_wakeup()
1553 {
1554 
1555 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1556 		vm_pages_needed = 1;
1557 		wakeup(&vm_pages_needed);
1558 	}
1559 }
1560 
1561 #if !defined(NO_SWAPPING)
1562 static void
1563 vm_req_vmdaemon()
1564 {
1565 	static int lastrun = 0;
1566 
1567 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1568 		wakeup(&vm_daemon_needed);
1569 		lastrun = ticks;
1570 	}
1571 }
1572 
1573 static void
1574 vm_daemon()
1575 {
1576 	struct rlimit rsslim;
1577 	struct proc *p;
1578 	struct thread *td;
1579 	int breakout;
1580 
1581 	mtx_lock(&Giant);
1582 	while (TRUE) {
1583 		tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1584 		if (vm_pageout_req_swapout) {
1585 			swapout_procs(vm_pageout_req_swapout);
1586 			vm_pageout_req_swapout = 0;
1587 		}
1588 		/*
1589 		 * scan the processes for exceeding their rlimits or if
1590 		 * process is swapped out -- deactivate pages
1591 		 */
1592 		sx_slock(&allproc_lock);
1593 		LIST_FOREACH(p, &allproc, p_list) {
1594 			vm_pindex_t limit, size;
1595 
1596 			/*
1597 			 * if this is a system process or if we have already
1598 			 * looked at this process, skip it.
1599 			 */
1600 			PROC_LOCK(p);
1601 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1602 				PROC_UNLOCK(p);
1603 				continue;
1604 			}
1605 			/*
1606 			 * if the process is in a non-running type state,
1607 			 * don't touch it.
1608 			 */
1609 			mtx_lock_spin(&sched_lock);
1610 			breakout = 0;
1611 			FOREACH_THREAD_IN_PROC(p, td) {
1612 				if (!TD_ON_RUNQ(td) &&
1613 				    !TD_IS_RUNNING(td) &&
1614 				    !TD_IS_SLEEPING(td)) {
1615 					breakout = 1;
1616 					break;
1617 				}
1618 			}
1619 			mtx_unlock_spin(&sched_lock);
1620 			if (breakout) {
1621 				PROC_UNLOCK(p);
1622 				continue;
1623 			}
1624 			/*
1625 			 * get a limit
1626 			 */
1627 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1628 			limit = OFF_TO_IDX(
1629 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1630 
1631 			/*
1632 			 * let processes that are swapped out really be
1633 			 * swapped out set the limit to nothing (will force a
1634 			 * swap-out.)
1635 			 */
1636 			if ((p->p_sflag & PS_INMEM) == 0)
1637 				limit = 0;	/* XXX */
1638 			PROC_UNLOCK(p);
1639 
1640 			size = vmspace_resident_count(p->p_vmspace);
1641 			if (limit >= 0 && size >= limit) {
1642 				vm_pageout_map_deactivate_pages(
1643 				    &p->p_vmspace->vm_map, limit);
1644 			}
1645 		}
1646 		sx_sunlock(&allproc_lock);
1647 	}
1648 }
1649 #endif			/* !defined(NO_SWAPPING) */
1650