xref: /freebsd/sys/vm/vm_pageout.c (revision 262e143bd46171a6415a5b28af260a5efa2a3db8)
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_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, &page_kp)
126 
127 #if !defined(NO_SWAPPING)
128 /* the kernel process "vm_daemon"*/
129 static void vm_daemon(void);
130 static struct	proc *vmproc;
131 
132 static struct kproc_desc vm_kp = {
133 	"vmdaemon",
134 	vm_daemon,
135 	&vmproc
136 };
137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
138 #endif
139 
140 
141 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
142 int vm_pageout_deficit;		/* Estimated number of pages deficit */
143 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
144 
145 #if !defined(NO_SWAPPING)
146 static int vm_pageout_req_swapout;	/* XXX */
147 static int vm_daemon_needed;
148 #endif
149 static int vm_max_launder = 32;
150 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
151 static int vm_pageout_full_stats_interval = 0;
152 static int vm_pageout_algorithm=0;
153 static int defer_swap_pageouts=0;
154 static int disable_swap_pageouts=0;
155 
156 #if defined(NO_SWAPPING)
157 static int vm_swap_enabled=0;
158 static int vm_swap_idle_enabled=0;
159 #else
160 static int vm_swap_enabled=1;
161 static int vm_swap_idle_enabled=0;
162 #endif
163 
164 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
165 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
166 
167 SYSCTL_INT(_vm, OID_AUTO, max_launder,
168 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
169 
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
171 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
172 
173 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
174 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
175 
176 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
177 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
178 
179 #if defined(NO_SWAPPING)
180 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
181 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
182 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
183 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
184 #else
185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
186 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
188 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
189 #endif
190 
191 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
192 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
193 
194 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
195 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
196 
197 static int pageout_lock_miss;
198 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
199 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
200 
201 #define VM_PAGEOUT_PAGE_COUNT 16
202 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
203 
204 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
205 
206 #if !defined(NO_SWAPPING)
207 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
208 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
209 static void vm_req_vmdaemon(void);
210 #endif
211 static void vm_pageout_page_stats(void);
212 
213 /*
214  * vm_pageout_fallback_object_lock:
215  *
216  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
217  * known to have failed and page queue must be either PQ_ACTIVE or
218  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
219  * while locking the vm object.  Use marker page to detect page queue
220  * changes and maintain notion of next page on page queue.  Return
221  * TRUE if no changes were detected, FALSE otherwise.  vm object is
222  * locked on return.
223  *
224  * This function depends on both the lock portion of struct vm_object
225  * and normal struct vm_page being type stable.
226  */
227 static boolean_t
228 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
229 {
230 	struct vm_page marker;
231 	boolean_t unchanged;
232 	u_short queue;
233 	vm_object_t object;
234 
235 	/*
236 	 * Initialize our marker
237 	 */
238 	bzero(&marker, sizeof(marker));
239 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
240 	marker.queue = m->queue;
241 	marker.wire_count = 1;
242 
243 	queue = m->queue;
244 	object = m->object;
245 
246 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
247 			   m, &marker, pageq);
248 	vm_page_unlock_queues();
249 	VM_OBJECT_LOCK(object);
250 	vm_page_lock_queues();
251 
252 	/* Page queue might have changed. */
253 	*next = TAILQ_NEXT(&marker, pageq);
254 	unchanged = (m->queue == queue &&
255 		     m->object == object &&
256 		     &marker == TAILQ_NEXT(m, pageq));
257 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
258 		     &marker, pageq);
259 	return (unchanged);
260 }
261 
262 /*
263  * vm_pageout_clean:
264  *
265  * Clean the page and remove it from the laundry.
266  *
267  * We set the busy bit to cause potential page faults on this page to
268  * block.  Note the careful timing, however, the busy bit isn't set till
269  * late and we cannot do anything that will mess with the page.
270  */
271 static int
272 vm_pageout_clean(m)
273 	vm_page_t m;
274 {
275 	vm_object_t object;
276 	vm_page_t mc[2*vm_pageout_page_count];
277 	int pageout_count;
278 	int ib, is, page_base;
279 	vm_pindex_t pindex = m->pindex;
280 
281 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
282 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
283 
284 	/*
285 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
286 	 * with the new swapper, but we could have serious problems paging
287 	 * out other object types if there is insufficient memory.
288 	 *
289 	 * Unfortunately, checking free memory here is far too late, so the
290 	 * check has been moved up a procedural level.
291 	 */
292 
293 	/*
294 	 * Don't mess with the page if it's busy, held, or special
295 	 */
296 	if ((m->hold_count != 0) ||
297 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
298 		return 0;
299 	}
300 
301 	mc[vm_pageout_page_count] = m;
302 	pageout_count = 1;
303 	page_base = vm_pageout_page_count;
304 	ib = 1;
305 	is = 1;
306 
307 	/*
308 	 * Scan object for clusterable pages.
309 	 *
310 	 * We can cluster ONLY if: ->> the page is NOT
311 	 * clean, wired, busy, held, or mapped into a
312 	 * buffer, and one of the following:
313 	 * 1) The page is inactive, or a seldom used
314 	 *    active page.
315 	 * -or-
316 	 * 2) we force the issue.
317 	 *
318 	 * During heavy mmap/modification loads the pageout
319 	 * daemon can really fragment the underlying file
320 	 * due to flushing pages out of order and not trying
321 	 * align the clusters (which leave sporatic out-of-order
322 	 * holes).  To solve this problem we do the reverse scan
323 	 * first and attempt to align our cluster, then do a
324 	 * forward scan if room remains.
325 	 */
326 	object = m->object;
327 more:
328 	while (ib && pageout_count < vm_pageout_page_count) {
329 		vm_page_t p;
330 
331 		if (ib > pindex) {
332 			ib = 0;
333 			break;
334 		}
335 
336 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
337 			ib = 0;
338 			break;
339 		}
340 		if (((p->queue - p->pc) == PQ_CACHE) ||
341 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
342 			ib = 0;
343 			break;
344 		}
345 		vm_page_test_dirty(p);
346 		if ((p->dirty & p->valid) == 0 ||
347 		    p->queue != PQ_INACTIVE ||
348 		    p->wire_count != 0 ||	/* may be held by buf cache */
349 		    p->hold_count != 0) {	/* may be undergoing I/O */
350 			ib = 0;
351 			break;
352 		}
353 		mc[--page_base] = p;
354 		++pageout_count;
355 		++ib;
356 		/*
357 		 * alignment boundry, stop here and switch directions.  Do
358 		 * not clear ib.
359 		 */
360 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
361 			break;
362 	}
363 
364 	while (pageout_count < vm_pageout_page_count &&
365 	    pindex + is < object->size) {
366 		vm_page_t p;
367 
368 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
369 			break;
370 		if (((p->queue - p->pc) == PQ_CACHE) ||
371 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
372 			break;
373 		}
374 		vm_page_test_dirty(p);
375 		if ((p->dirty & p->valid) == 0 ||
376 		    p->queue != PQ_INACTIVE ||
377 		    p->wire_count != 0 ||	/* may be held by buf cache */
378 		    p->hold_count != 0) {	/* may be undergoing I/O */
379 			break;
380 		}
381 		mc[page_base + pageout_count] = p;
382 		++pageout_count;
383 		++is;
384 	}
385 
386 	/*
387 	 * If we exhausted our forward scan, continue with the reverse scan
388 	 * when possible, even past a page boundry.  This catches boundry
389 	 * conditions.
390 	 */
391 	if (ib && pageout_count < vm_pageout_page_count)
392 		goto more;
393 
394 	/*
395 	 * we allow reads during pageouts...
396 	 */
397 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
398 }
399 
400 /*
401  * vm_pageout_flush() - launder the given pages
402  *
403  *	The given pages are laundered.  Note that we setup for the start of
404  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
405  *	reference count all in here rather then in the parent.  If we want
406  *	the parent to do more sophisticated things we may have to change
407  *	the ordering.
408  */
409 int
410 vm_pageout_flush(vm_page_t *mc, int count, int flags)
411 {
412 	vm_object_t object = mc[0]->object;
413 	int pageout_status[count];
414 	int numpagedout = 0;
415 	int i;
416 
417 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
418 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
419 	/*
420 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
421 	 * mark the pages read-only.
422 	 *
423 	 * We do not have to fixup the clean/dirty bits here... we can
424 	 * allow the pager to do it after the I/O completes.
425 	 *
426 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
427 	 * edge case with file fragments.
428 	 */
429 	for (i = 0; i < count; i++) {
430 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
431 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
432 			mc[i], i, count));
433 		vm_page_io_start(mc[i]);
434 		pmap_page_protect(mc[i], VM_PROT_READ);
435 	}
436 	vm_page_unlock_queues();
437 	vm_object_pip_add(object, count);
438 
439 	vm_pager_put_pages(object, mc, count,
440 	    (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
441 	    pageout_status);
442 
443 	vm_page_lock_queues();
444 	for (i = 0; i < count; i++) {
445 		vm_page_t mt = mc[i];
446 
447 		KASSERT((mt->flags & PG_WRITEABLE) == 0,
448 		    ("vm_pageout_flush: page %p is not write protected", mt));
449 		switch (pageout_status[i]) {
450 		case VM_PAGER_OK:
451 		case VM_PAGER_PEND:
452 			numpagedout++;
453 			break;
454 		case VM_PAGER_BAD:
455 			/*
456 			 * Page outside of range of object. Right now we
457 			 * essentially lose the changes by pretending it
458 			 * worked.
459 			 */
460 			pmap_clear_modify(mt);
461 			vm_page_undirty(mt);
462 			break;
463 		case VM_PAGER_ERROR:
464 		case VM_PAGER_FAIL:
465 			/*
466 			 * If page couldn't be paged out, then reactivate the
467 			 * page so it doesn't clog the inactive list.  (We
468 			 * will try paging out it again later).
469 			 */
470 			vm_page_activate(mt);
471 			break;
472 		case VM_PAGER_AGAIN:
473 			break;
474 		}
475 
476 		/*
477 		 * If the operation is still going, leave the page busy to
478 		 * block all other accesses. Also, leave the paging in
479 		 * progress indicator set so that we don't attempt an object
480 		 * collapse.
481 		 */
482 		if (pageout_status[i] != VM_PAGER_PEND) {
483 			vm_object_pip_wakeup(object);
484 			vm_page_io_finish(mt);
485 			if (vm_page_count_severe())
486 				vm_page_try_to_cache(mt);
487 		}
488 	}
489 	return numpagedout;
490 }
491 
492 #if !defined(NO_SWAPPING)
493 /*
494  *	vm_pageout_object_deactivate_pages
495  *
496  *	deactivate enough pages to satisfy the inactive target
497  *	requirements or if vm_page_proc_limit is set, then
498  *	deactivate all of the pages in the object and its
499  *	backing_objects.
500  *
501  *	The object and map must be locked.
502  */
503 static void
504 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
505 	pmap_t pmap;
506 	vm_object_t first_object;
507 	long desired;
508 {
509 	vm_object_t backing_object, object;
510 	vm_page_t p, next;
511 	int actcount, rcount, remove_mode;
512 
513 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
514 	if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
515 		return;
516 	for (object = first_object;; object = backing_object) {
517 		if (pmap_resident_count(pmap) <= desired)
518 			goto unlock_return;
519 		if (object->paging_in_progress)
520 			goto unlock_return;
521 
522 		remove_mode = 0;
523 		if (object->shadow_count > 1)
524 			remove_mode = 1;
525 		/*
526 		 * scan the objects entire memory queue
527 		 */
528 		rcount = object->resident_page_count;
529 		p = TAILQ_FIRST(&object->memq);
530 		vm_page_lock_queues();
531 		while (p && (rcount-- > 0)) {
532 			if (pmap_resident_count(pmap) <= desired) {
533 				vm_page_unlock_queues();
534 				goto unlock_return;
535 			}
536 			next = TAILQ_NEXT(p, listq);
537 			cnt.v_pdpages++;
538 			if (p->wire_count != 0 ||
539 			    p->hold_count != 0 ||
540 			    p->busy != 0 ||
541 			    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
542 			    !pmap_page_exists_quick(pmap, p)) {
543 				p = next;
544 				continue;
545 			}
546 			actcount = pmap_ts_referenced(p);
547 			if (actcount) {
548 				vm_page_flag_set(p, PG_REFERENCED);
549 			} else if (p->flags & PG_REFERENCED) {
550 				actcount = 1;
551 			}
552 			if ((p->queue != PQ_ACTIVE) &&
553 				(p->flags & PG_REFERENCED)) {
554 				vm_page_activate(p);
555 				p->act_count += actcount;
556 				vm_page_flag_clear(p, PG_REFERENCED);
557 			} else if (p->queue == PQ_ACTIVE) {
558 				if ((p->flags & PG_REFERENCED) == 0) {
559 					p->act_count -= min(p->act_count, ACT_DECLINE);
560 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
561 						pmap_remove_all(p);
562 						vm_page_deactivate(p);
563 					} else {
564 						vm_pageq_requeue(p);
565 					}
566 				} else {
567 					vm_page_activate(p);
568 					vm_page_flag_clear(p, PG_REFERENCED);
569 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
570 						p->act_count += ACT_ADVANCE;
571 					vm_pageq_requeue(p);
572 				}
573 			} else if (p->queue == PQ_INACTIVE) {
574 				pmap_remove_all(p);
575 			}
576 			p = next;
577 		}
578 		vm_page_unlock_queues();
579 		if ((backing_object = object->backing_object) == NULL)
580 			goto unlock_return;
581 		VM_OBJECT_LOCK(backing_object);
582 		if (object != first_object)
583 			VM_OBJECT_UNLOCK(object);
584 	}
585 unlock_return:
586 	if (object != first_object)
587 		VM_OBJECT_UNLOCK(object);
588 }
589 
590 /*
591  * deactivate some number of pages in a map, try to do it fairly, but
592  * that is really hard to do.
593  */
594 static void
595 vm_pageout_map_deactivate_pages(map, desired)
596 	vm_map_t map;
597 	long desired;
598 {
599 	vm_map_entry_t tmpe;
600 	vm_object_t obj, bigobj;
601 	int nothingwired;
602 
603 	if (!vm_map_trylock(map))
604 		return;
605 
606 	bigobj = NULL;
607 	nothingwired = TRUE;
608 
609 	/*
610 	 * first, search out the biggest object, and try to free pages from
611 	 * that.
612 	 */
613 	tmpe = map->header.next;
614 	while (tmpe != &map->header) {
615 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
616 			obj = tmpe->object.vm_object;
617 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
618 				if (obj->shadow_count <= 1 &&
619 				    (bigobj == NULL ||
620 				     bigobj->resident_page_count < obj->resident_page_count)) {
621 					if (bigobj != NULL)
622 						VM_OBJECT_UNLOCK(bigobj);
623 					bigobj = obj;
624 				} else
625 					VM_OBJECT_UNLOCK(obj);
626 			}
627 		}
628 		if (tmpe->wired_count > 0)
629 			nothingwired = FALSE;
630 		tmpe = tmpe->next;
631 	}
632 
633 	if (bigobj != NULL) {
634 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
635 		VM_OBJECT_UNLOCK(bigobj);
636 	}
637 	/*
638 	 * Next, hunt around for other pages to deactivate.  We actually
639 	 * do this search sort of wrong -- .text first is not the best idea.
640 	 */
641 	tmpe = map->header.next;
642 	while (tmpe != &map->header) {
643 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
644 			break;
645 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
646 			obj = tmpe->object.vm_object;
647 			if (obj != NULL) {
648 				VM_OBJECT_LOCK(obj);
649 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
650 				VM_OBJECT_UNLOCK(obj);
651 			}
652 		}
653 		tmpe = tmpe->next;
654 	}
655 
656 	/*
657 	 * Remove all mappings if a process is swapped out, this will free page
658 	 * table pages.
659 	 */
660 	if (desired == 0 && nothingwired) {
661 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
662 		    vm_map_max(map));
663 	}
664 	vm_map_unlock(map);
665 }
666 #endif		/* !defined(NO_SWAPPING) */
667 
668 /*
669  *	vm_pageout_scan does the dirty work for the pageout daemon.
670  */
671 static void
672 vm_pageout_scan(int pass)
673 {
674 	vm_page_t m, next;
675 	struct vm_page marker;
676 	int page_shortage, maxscan, pcount;
677 	int addl_page_shortage, addl_page_shortage_init;
678 	struct proc *p, *bigproc;
679 	struct thread *td;
680 	vm_offset_t size, bigsize;
681 	vm_object_t object;
682 	int actcount, cache_cur, cache_first_failure;
683 	static int cache_last_free;
684 	int vnodes_skipped = 0;
685 	int maxlaunder;
686 
687 	mtx_lock(&Giant);
688 	/*
689 	 * Decrease registered cache sizes.
690 	 */
691 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
692 	/*
693 	 * We do this explicitly after the caches have been drained above.
694 	 */
695 	uma_reclaim();
696 
697 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
698 
699 	/*
700 	 * Calculate the number of pages we want to either free or move
701 	 * to the cache.
702 	 */
703 	page_shortage = vm_paging_target() + addl_page_shortage_init;
704 
705 	/*
706 	 * Initialize our marker
707 	 */
708 	bzero(&marker, sizeof(marker));
709 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
710 	marker.queue = PQ_INACTIVE;
711 	marker.wire_count = 1;
712 
713 	/*
714 	 * Start scanning the inactive queue for pages we can move to the
715 	 * cache or free.  The scan will stop when the target is reached or
716 	 * we have scanned the entire inactive queue.  Note that m->act_count
717 	 * is not used to form decisions for the inactive queue, only for the
718 	 * active queue.
719 	 *
720 	 * maxlaunder limits the number of dirty pages we flush per scan.
721 	 * For most systems a smaller value (16 or 32) is more robust under
722 	 * extreme memory and disk pressure because any unnecessary writes
723 	 * to disk can result in extreme performance degredation.  However,
724 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
725 	 * used) will die horribly with limited laundering.  If the pageout
726 	 * daemon cannot clean enough pages in the first pass, we let it go
727 	 * all out in succeeding passes.
728 	 */
729 	if ((maxlaunder = vm_max_launder) <= 1)
730 		maxlaunder = 1;
731 	if (pass)
732 		maxlaunder = 10000;
733 	vm_page_lock_queues();
734 rescan0:
735 	addl_page_shortage = addl_page_shortage_init;
736 	maxscan = cnt.v_inactive_count;
737 
738 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
739 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
740 	     m = next) {
741 
742 		cnt.v_pdpages++;
743 
744 		if (m->queue != PQ_INACTIVE) {
745 			goto rescan0;
746 		}
747 
748 		next = TAILQ_NEXT(m, pageq);
749 		object = m->object;
750 
751 		/*
752 		 * skip marker pages
753 		 */
754 		if (m->flags & PG_MARKER)
755 			continue;
756 
757 		/*
758 		 * A held page may be undergoing I/O, so skip it.
759 		 */
760 		if (m->hold_count) {
761 			vm_pageq_requeue(m);
762 			addl_page_shortage++;
763 			continue;
764 		}
765 		/*
766 		 * Don't mess with busy pages, keep in the front of the
767 		 * queue, most likely are being paged out.
768 		 */
769 		if (!VM_OBJECT_TRYLOCK(object) &&
770 		    (!vm_pageout_fallback_object_lock(m, &next) ||
771 		     m->hold_count != 0)) {
772 			VM_OBJECT_UNLOCK(object);
773 			addl_page_shortage++;
774 			continue;
775 		}
776 		if (m->busy || (m->flags & PG_BUSY)) {
777 			VM_OBJECT_UNLOCK(object);
778 			addl_page_shortage++;
779 			continue;
780 		}
781 
782 		/*
783 		 * If the object is not being used, we ignore previous
784 		 * references.
785 		 */
786 		if (object->ref_count == 0) {
787 			vm_page_flag_clear(m, PG_REFERENCED);
788 			pmap_clear_reference(m);
789 
790 		/*
791 		 * Otherwise, if the page has been referenced while in the
792 		 * inactive queue, we bump the "activation count" upwards,
793 		 * making it less likely that the page will be added back to
794 		 * the inactive queue prematurely again.  Here we check the
795 		 * page tables (or emulated bits, if any), given the upper
796 		 * level VM system not knowing anything about existing
797 		 * references.
798 		 */
799 		} else if (((m->flags & PG_REFERENCED) == 0) &&
800 			(actcount = pmap_ts_referenced(m))) {
801 			vm_page_activate(m);
802 			VM_OBJECT_UNLOCK(object);
803 			m->act_count += (actcount + ACT_ADVANCE);
804 			continue;
805 		}
806 
807 		/*
808 		 * If the upper level VM system knows about any page
809 		 * references, we activate the page.  We also set the
810 		 * "activation count" higher than normal so that we will less
811 		 * likely place pages back onto the inactive queue again.
812 		 */
813 		if ((m->flags & PG_REFERENCED) != 0) {
814 			vm_page_flag_clear(m, PG_REFERENCED);
815 			actcount = pmap_ts_referenced(m);
816 			vm_page_activate(m);
817 			VM_OBJECT_UNLOCK(object);
818 			m->act_count += (actcount + ACT_ADVANCE + 1);
819 			continue;
820 		}
821 
822 		/*
823 		 * If the upper level VM system doesn't know anything about
824 		 * the page being dirty, we have to check for it again.  As
825 		 * far as the VM code knows, any partially dirty pages are
826 		 * fully dirty.
827 		 */
828 		if (m->dirty == 0 && !pmap_is_modified(m)) {
829 			/*
830 			 * Avoid a race condition: Unless write access is
831 			 * removed from the page, another processor could
832 			 * modify it before all access is removed by the call
833 			 * to vm_page_cache() below.  If vm_page_cache() finds
834 			 * that the page has been modified when it removes all
835 			 * access, it panics because it cannot cache dirty
836 			 * pages.  In principle, we could eliminate just write
837 			 * access here rather than all access.  In the expected
838 			 * case, when there are no last instant modifications
839 			 * to the page, removing all access will be cheaper
840 			 * overall.
841 			 */
842 			if ((m->flags & PG_WRITEABLE) != 0)
843 				pmap_remove_all(m);
844 		} else {
845 			vm_page_dirty(m);
846 		}
847 
848 		if (m->valid == 0) {
849 			/*
850 			 * Invalid pages can be easily freed
851 			 */
852 			pmap_remove_all(m);
853 			vm_page_free(m);
854 			cnt.v_dfree++;
855 			--page_shortage;
856 		} else if (m->dirty == 0) {
857 			/*
858 			 * Clean pages can be placed onto the cache queue.
859 			 * This effectively frees them.
860 			 */
861 			vm_page_cache(m);
862 			--page_shortage;
863 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
864 			/*
865 			 * Dirty pages need to be paged out, but flushing
866 			 * a page is extremely expensive verses freeing
867 			 * a clean page.  Rather then artificially limiting
868 			 * the number of pages we can flush, we instead give
869 			 * dirty pages extra priority on the inactive queue
870 			 * by forcing them to be cycled through the queue
871 			 * twice before being flushed, after which the
872 			 * (now clean) page will cycle through once more
873 			 * before being freed.  This significantly extends
874 			 * the thrash point for a heavily loaded machine.
875 			 */
876 			vm_page_flag_set(m, PG_WINATCFLS);
877 			vm_pageq_requeue(m);
878 		} else if (maxlaunder > 0) {
879 			/*
880 			 * We always want to try to flush some dirty pages if
881 			 * we encounter them, to keep the system stable.
882 			 * Normally this number is small, but under extreme
883 			 * pressure where there are insufficient clean pages
884 			 * on the inactive queue, we may have to go all out.
885 			 */
886 			int swap_pageouts_ok;
887 			struct vnode *vp = NULL;
888 			struct mount *mp;
889 
890 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
891 				swap_pageouts_ok = 1;
892 			} else {
893 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
894 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
895 				vm_page_count_min());
896 
897 			}
898 
899 			/*
900 			 * We don't bother paging objects that are "dead".
901 			 * Those objects are in a "rundown" state.
902 			 */
903 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
904 				VM_OBJECT_UNLOCK(object);
905 				vm_pageq_requeue(m);
906 				continue;
907 			}
908 
909 			/*
910 			 * The object is already known NOT to be dead.   It
911 			 * is possible for the vget() to block the whole
912 			 * pageout daemon, but the new low-memory handling
913 			 * code should prevent it.
914 			 *
915 			 * The previous code skipped locked vnodes and, worse,
916 			 * reordered pages in the queue.  This results in
917 			 * completely non-deterministic operation and, on a
918 			 * busy system, can lead to extremely non-optimal
919 			 * pageouts.  For example, it can cause clean pages
920 			 * to be freed and dirty pages to be moved to the end
921 			 * of the queue.  Since dirty pages are also moved to
922 			 * the end of the queue once-cleaned, this gives
923 			 * way too large a weighting to defering the freeing
924 			 * of dirty pages.
925 			 *
926 			 * We can't wait forever for the vnode lock, we might
927 			 * deadlock due to a vn_read() getting stuck in
928 			 * vm_wait while holding this vnode.  We skip the
929 			 * vnode if we can't get it in a reasonable amount
930 			 * of time.
931 			 */
932 			if (object->type == OBJT_VNODE) {
933 				vp = object->handle;
934 				mp = NULL;
935 				if (vp->v_type == VREG)
936 					vn_start_write(vp, &mp, V_NOWAIT);
937 				vm_page_unlock_queues();
938 				VI_LOCK(vp);
939 				VM_OBJECT_UNLOCK(object);
940 				if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK |
941 				    LK_TIMELOCK, curthread)) {
942 					VM_OBJECT_LOCK(object);
943 					vm_page_lock_queues();
944 					++pageout_lock_miss;
945 					vn_finished_write(mp);
946 					if (object->flags & OBJ_MIGHTBEDIRTY)
947 						vnodes_skipped++;
948 					VM_OBJECT_UNLOCK(object);
949 					continue;
950 				}
951 				VM_OBJECT_LOCK(object);
952 				vm_page_lock_queues();
953 				/*
954 				 * The page might have been moved to another
955 				 * queue during potential blocking in vget()
956 				 * above.  The page might have been freed and
957 				 * reused for another vnode.  The object might
958 				 * have been reused for another vnode.
959 				 */
960 				if (m->queue != PQ_INACTIVE ||
961 				    m->object != object ||
962 				    object->handle != vp) {
963 					if (object->flags & OBJ_MIGHTBEDIRTY)
964 						vnodes_skipped++;
965 					goto unlock_and_continue;
966 				}
967 
968 				/*
969 				 * The page may have been busied during the
970 				 * blocking in vput();  We don't move the
971 				 * page back onto the end of the queue so that
972 				 * statistics are more correct if we don't.
973 				 */
974 				if (m->busy || (m->flags & PG_BUSY)) {
975 					goto unlock_and_continue;
976 				}
977 
978 				/*
979 				 * If the page has become held it might
980 				 * be undergoing I/O, so skip it
981 				 */
982 				if (m->hold_count) {
983 					vm_pageq_requeue(m);
984 					if (object->flags & OBJ_MIGHTBEDIRTY)
985 						vnodes_skipped++;
986 					goto unlock_and_continue;
987 				}
988 			}
989 
990 			/*
991 			 * If a page is dirty, then it is either being washed
992 			 * (but not yet cleaned) or it is still in the
993 			 * laundry.  If it is still in the laundry, then we
994 			 * start the cleaning operation.
995 			 *
996 			 * This operation may cluster, invalidating the 'next'
997 			 * pointer.  To prevent an inordinate number of
998 			 * restarts we use our marker to remember our place.
999 			 *
1000 			 * decrement page_shortage on success to account for
1001 			 * the (future) cleaned page.  Otherwise we could wind
1002 			 * up laundering or cleaning too many pages.
1003 			 */
1004 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1005 			if (vm_pageout_clean(m) != 0) {
1006 				--page_shortage;
1007 				--maxlaunder;
1008 			}
1009 			next = TAILQ_NEXT(&marker, pageq);
1010 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1011 unlock_and_continue:
1012 			VM_OBJECT_UNLOCK(object);
1013 			if (vp) {
1014 				vm_page_unlock_queues();
1015 				vput(vp);
1016 				vn_finished_write(mp);
1017 				vm_page_lock_queues();
1018 			}
1019 			continue;
1020 		}
1021 		VM_OBJECT_UNLOCK(object);
1022 	}
1023 
1024 	/*
1025 	 * Compute the number of pages we want to try to move from the
1026 	 * active queue to the inactive queue.
1027 	 */
1028 	page_shortage = vm_paging_target() +
1029 		cnt.v_inactive_target - cnt.v_inactive_count;
1030 	page_shortage += addl_page_shortage;
1031 
1032 	/*
1033 	 * Scan the active queue for things we can deactivate. We nominally
1034 	 * track the per-page activity counter and use it to locate
1035 	 * deactivation candidates.
1036 	 */
1037 	pcount = cnt.v_active_count;
1038 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1039 
1040 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1041 
1042 		KASSERT(m->queue == PQ_ACTIVE,
1043 		    ("vm_pageout_scan: page %p isn't active", m));
1044 
1045 		next = TAILQ_NEXT(m, pageq);
1046 		object = m->object;
1047 		if ((m->flags & PG_MARKER) != 0) {
1048 			m = next;
1049 			continue;
1050 		}
1051 		if (!VM_OBJECT_TRYLOCK(object) &&
1052 		    !vm_pageout_fallback_object_lock(m, &next)) {
1053 			VM_OBJECT_UNLOCK(object);
1054 			m = next;
1055 			continue;
1056 		}
1057 
1058 		/*
1059 		 * Don't deactivate pages that are busy.
1060 		 */
1061 		if ((m->busy != 0) ||
1062 		    (m->flags & PG_BUSY) ||
1063 		    (m->hold_count != 0)) {
1064 			VM_OBJECT_UNLOCK(object);
1065 			vm_pageq_requeue(m);
1066 			m = next;
1067 			continue;
1068 		}
1069 
1070 		/*
1071 		 * The count for pagedaemon pages is done after checking the
1072 		 * page for eligibility...
1073 		 */
1074 		cnt.v_pdpages++;
1075 
1076 		/*
1077 		 * Check to see "how much" the page has been used.
1078 		 */
1079 		actcount = 0;
1080 		if (object->ref_count != 0) {
1081 			if (m->flags & PG_REFERENCED) {
1082 				actcount += 1;
1083 			}
1084 			actcount += pmap_ts_referenced(m);
1085 			if (actcount) {
1086 				m->act_count += ACT_ADVANCE + actcount;
1087 				if (m->act_count > ACT_MAX)
1088 					m->act_count = ACT_MAX;
1089 			}
1090 		}
1091 
1092 		/*
1093 		 * Since we have "tested" this bit, we need to clear it now.
1094 		 */
1095 		vm_page_flag_clear(m, PG_REFERENCED);
1096 
1097 		/*
1098 		 * Only if an object is currently being used, do we use the
1099 		 * page activation count stats.
1100 		 */
1101 		if (actcount && (object->ref_count != 0)) {
1102 			vm_pageq_requeue(m);
1103 		} else {
1104 			m->act_count -= min(m->act_count, ACT_DECLINE);
1105 			if (vm_pageout_algorithm ||
1106 			    object->ref_count == 0 ||
1107 			    m->act_count == 0) {
1108 				page_shortage--;
1109 				if (object->ref_count == 0) {
1110 					pmap_remove_all(m);
1111 					if (m->dirty == 0)
1112 						vm_page_cache(m);
1113 					else
1114 						vm_page_deactivate(m);
1115 				} else {
1116 					vm_page_deactivate(m);
1117 				}
1118 			} else {
1119 				vm_pageq_requeue(m);
1120 			}
1121 		}
1122 		VM_OBJECT_UNLOCK(object);
1123 		m = next;
1124 	}
1125 
1126 	/*
1127 	 * We try to maintain some *really* free pages, this allows interrupt
1128 	 * code to be guaranteed space.  Since both cache and free queues
1129 	 * are considered basically 'free', moving pages from cache to free
1130 	 * does not effect other calculations.
1131 	 */
1132 	cache_cur = cache_last_free;
1133 	cache_first_failure = -1;
1134 	while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur =
1135 	    (cache_cur + PQ_PRIME2) & PQ_L2_MASK) != cache_first_failure) {
1136 		TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl,
1137 		    pageq) {
1138 			KASSERT(m->dirty == 0,
1139 			    ("Found dirty cache page %p", m));
1140 			KASSERT(!pmap_page_is_mapped(m),
1141 			    ("Found mapped cache page %p", m));
1142 			KASSERT((m->flags & PG_UNMANAGED) == 0,
1143 			    ("Found unmanaged cache page %p", m));
1144 			KASSERT(m->wire_count == 0,
1145 			    ("Found wired cache page %p", m));
1146 			if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1147 			    m->object)) {
1148 				KASSERT((m->flags & PG_BUSY) == 0 &&
1149 				    m->busy == 0, ("Found busy cache page %p",
1150 				    m));
1151 				vm_page_free(m);
1152 				VM_OBJECT_UNLOCK(object);
1153 				cnt.v_dfree++;
1154 				cache_last_free = cache_cur;
1155 				cache_first_failure = -1;
1156 				break;
1157 			}
1158 		}
1159 		if (m == NULL && cache_first_failure == -1)
1160 			cache_first_failure = cache_cur;
1161 	}
1162 	vm_page_unlock_queues();
1163 #if !defined(NO_SWAPPING)
1164 	/*
1165 	 * Idle process swapout -- run once per second.
1166 	 */
1167 	if (vm_swap_idle_enabled) {
1168 		static long lsec;
1169 		if (time_second != lsec) {
1170 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1171 			vm_req_vmdaemon();
1172 			lsec = time_second;
1173 		}
1174 	}
1175 #endif
1176 
1177 	/*
1178 	 * If we didn't get enough free pages, and we have skipped a vnode
1179 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1180 	 * if we did not get enough free pages.
1181 	 */
1182 	if (vm_paging_target() > 0) {
1183 		if (vnodes_skipped && vm_page_count_min())
1184 			(void) speedup_syncer();
1185 #if !defined(NO_SWAPPING)
1186 		if (vm_swap_enabled && vm_page_count_target()) {
1187 			vm_req_vmdaemon();
1188 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1189 		}
1190 #endif
1191 	}
1192 
1193 	/*
1194 	 * If we are critically low on one of RAM or swap and low on
1195 	 * the other, kill the largest process.  However, we avoid
1196 	 * doing this on the first pass in order to give ourselves a
1197 	 * chance to flush out dirty vnode-backed pages and to allow
1198 	 * active pages to be moved to the inactive queue and reclaimed.
1199 	 *
1200 	 * We keep the process bigproc locked once we find it to keep anyone
1201 	 * from messing with it; however, there is a possibility of
1202 	 * deadlock if process B is bigproc and one of it's child processes
1203 	 * attempts to propagate a signal to B while we are waiting for A's
1204 	 * lock while walking this list.  To avoid this, we don't block on
1205 	 * the process lock but just skip a process if it is already locked.
1206 	 */
1207 	if (pass != 0 &&
1208 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1209 	     (swap_pager_full && vm_paging_target() > 0))) {
1210 		bigproc = NULL;
1211 		bigsize = 0;
1212 		sx_slock(&allproc_lock);
1213 		FOREACH_PROC_IN_SYSTEM(p) {
1214 			int breakout;
1215 
1216 			if (PROC_TRYLOCK(p) == 0)
1217 				continue;
1218 			/*
1219 			 * If this is a system or protected process, skip it.
1220 			 */
1221 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1222 			    (p->p_flag & P_PROTECTED) ||
1223 			    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1224 				PROC_UNLOCK(p);
1225 				continue;
1226 			}
1227 			/*
1228 			 * If the process is in a non-running type state,
1229 			 * don't touch it.  Check all the threads individually.
1230 			 */
1231 			mtx_lock_spin(&sched_lock);
1232 			breakout = 0;
1233 			FOREACH_THREAD_IN_PROC(p, td) {
1234 				if (!TD_ON_RUNQ(td) &&
1235 				    !TD_IS_RUNNING(td) &&
1236 				    !TD_IS_SLEEPING(td)) {
1237 					breakout = 1;
1238 					break;
1239 				}
1240 			}
1241 			if (breakout) {
1242 				mtx_unlock_spin(&sched_lock);
1243 				PROC_UNLOCK(p);
1244 				continue;
1245 			}
1246 			mtx_unlock_spin(&sched_lock);
1247 			/*
1248 			 * get the process size
1249 			 */
1250 			if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1251 				PROC_UNLOCK(p);
1252 				continue;
1253 			}
1254 			size = vmspace_swap_count(p->p_vmspace);
1255 			vm_map_unlock_read(&p->p_vmspace->vm_map);
1256 			size += vmspace_resident_count(p->p_vmspace);
1257 			/*
1258 			 * if the this process is bigger than the biggest one
1259 			 * remember it.
1260 			 */
1261 			if (size > bigsize) {
1262 				if (bigproc != NULL)
1263 					PROC_UNLOCK(bigproc);
1264 				bigproc = p;
1265 				bigsize = size;
1266 			} else
1267 				PROC_UNLOCK(p);
1268 		}
1269 		sx_sunlock(&allproc_lock);
1270 		if (bigproc != NULL) {
1271 			killproc(bigproc, "out of swap space");
1272 			mtx_lock_spin(&sched_lock);
1273 			sched_nice(bigproc, PRIO_MIN);
1274 			mtx_unlock_spin(&sched_lock);
1275 			PROC_UNLOCK(bigproc);
1276 			wakeup(&cnt.v_free_count);
1277 		}
1278 	}
1279 	mtx_unlock(&Giant);
1280 }
1281 
1282 /*
1283  * This routine tries to maintain the pseudo LRU active queue,
1284  * so that during long periods of time where there is no paging,
1285  * that some statistic accumulation still occurs.  This code
1286  * helps the situation where paging just starts to occur.
1287  */
1288 static void
1289 vm_pageout_page_stats()
1290 {
1291 	vm_object_t object;
1292 	vm_page_t m,next;
1293 	int pcount,tpcount;		/* Number of pages to check */
1294 	static int fullintervalcount = 0;
1295 	int page_shortage;
1296 
1297 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1298 	page_shortage =
1299 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1300 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1301 
1302 	if (page_shortage <= 0)
1303 		return;
1304 
1305 	pcount = cnt.v_active_count;
1306 	fullintervalcount += vm_pageout_stats_interval;
1307 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1308 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1309 		if (pcount > tpcount)
1310 			pcount = tpcount;
1311 	} else {
1312 		fullintervalcount = 0;
1313 	}
1314 
1315 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1316 	while ((m != NULL) && (pcount-- > 0)) {
1317 		int actcount;
1318 
1319 		KASSERT(m->queue == PQ_ACTIVE,
1320 		    ("vm_pageout_page_stats: page %p isn't active", m));
1321 
1322 		next = TAILQ_NEXT(m, pageq);
1323 		object = m->object;
1324 
1325 		if ((m->flags & PG_MARKER) != 0) {
1326 			m = next;
1327 			continue;
1328 		}
1329 		if (!VM_OBJECT_TRYLOCK(object) &&
1330 		    !vm_pageout_fallback_object_lock(m, &next)) {
1331 			VM_OBJECT_UNLOCK(object);
1332 			m = next;
1333 			continue;
1334 		}
1335 
1336 		/*
1337 		 * Don't deactivate pages that are busy.
1338 		 */
1339 		if ((m->busy != 0) ||
1340 		    (m->flags & PG_BUSY) ||
1341 		    (m->hold_count != 0)) {
1342 			VM_OBJECT_UNLOCK(object);
1343 			vm_pageq_requeue(m);
1344 			m = next;
1345 			continue;
1346 		}
1347 
1348 		actcount = 0;
1349 		if (m->flags & PG_REFERENCED) {
1350 			vm_page_flag_clear(m, PG_REFERENCED);
1351 			actcount += 1;
1352 		}
1353 
1354 		actcount += pmap_ts_referenced(m);
1355 		if (actcount) {
1356 			m->act_count += ACT_ADVANCE + actcount;
1357 			if (m->act_count > ACT_MAX)
1358 				m->act_count = ACT_MAX;
1359 			vm_pageq_requeue(m);
1360 		} else {
1361 			if (m->act_count == 0) {
1362 				/*
1363 				 * We turn off page access, so that we have
1364 				 * more accurate RSS stats.  We don't do this
1365 				 * in the normal page deactivation when the
1366 				 * system is loaded VM wise, because the
1367 				 * cost of the large number of page protect
1368 				 * operations would be higher than the value
1369 				 * of doing the operation.
1370 				 */
1371 				pmap_remove_all(m);
1372 				vm_page_deactivate(m);
1373 			} else {
1374 				m->act_count -= min(m->act_count, ACT_DECLINE);
1375 				vm_pageq_requeue(m);
1376 			}
1377 		}
1378 		VM_OBJECT_UNLOCK(object);
1379 		m = next;
1380 	}
1381 }
1382 
1383 /*
1384  *	vm_pageout is the high level pageout daemon.
1385  */
1386 static void
1387 vm_pageout()
1388 {
1389 	int error, pass;
1390 
1391 	/*
1392 	 * Initialize some paging parameters.
1393 	 */
1394 	cnt.v_interrupt_free_min = 2;
1395 	if (cnt.v_page_count < 2000)
1396 		vm_pageout_page_count = 8;
1397 
1398 	/*
1399 	 * v_free_reserved needs to include enough for the largest
1400 	 * swap pager structures plus enough for any pv_entry structs
1401 	 * when paging.
1402 	 */
1403 	if (cnt.v_page_count > 1024)
1404 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1405 	else
1406 		cnt.v_free_min = 4;
1407 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1408 	    cnt.v_interrupt_free_min;
1409 	cnt.v_free_reserved = vm_pageout_page_count +
1410 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_L2_SIZE;
1411 	cnt.v_free_severe = cnt.v_free_min / 2;
1412 	cnt.v_free_min += cnt.v_free_reserved;
1413 	cnt.v_free_severe += cnt.v_free_reserved;
1414 
1415 	/*
1416 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1417 	 * that these are more a measure of the VM cache queue hysteresis
1418 	 * then the VM free queue.  Specifically, v_free_target is the
1419 	 * high water mark (free+cache pages).
1420 	 *
1421 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1422 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1423 	 * be big enough to handle memory needs while the pageout daemon
1424 	 * is signalled and run to free more pages.
1425 	 */
1426 	if (cnt.v_free_count > 6144)
1427 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1428 	else
1429 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1430 
1431 	if (cnt.v_free_count > 2048) {
1432 		cnt.v_cache_min = cnt.v_free_target;
1433 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1434 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1435 	} else {
1436 		cnt.v_cache_min = 0;
1437 		cnt.v_cache_max = 0;
1438 		cnt.v_inactive_target = cnt.v_free_count / 4;
1439 	}
1440 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1441 		cnt.v_inactive_target = cnt.v_free_count / 3;
1442 
1443 	/* XXX does not really belong here */
1444 	if (vm_page_max_wired == 0)
1445 		vm_page_max_wired = cnt.v_free_count / 3;
1446 
1447 	if (vm_pageout_stats_max == 0)
1448 		vm_pageout_stats_max = cnt.v_free_target;
1449 
1450 	/*
1451 	 * Set interval in seconds for stats scan.
1452 	 */
1453 	if (vm_pageout_stats_interval == 0)
1454 		vm_pageout_stats_interval = 5;
1455 	if (vm_pageout_full_stats_interval == 0)
1456 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1457 
1458 	swap_pager_swap_init();
1459 	pass = 0;
1460 	/*
1461 	 * The pageout daemon is never done, so loop forever.
1462 	 */
1463 	while (TRUE) {
1464 		vm_page_lock_queues();
1465 		/*
1466 		 * If we have enough free memory, wakeup waiters.  Do
1467 		 * not clear vm_pages_needed until we reach our target,
1468 		 * otherwise we may be woken up over and over again and
1469 		 * waste a lot of cpu.
1470 		 */
1471 		if (vm_pages_needed && !vm_page_count_min()) {
1472 			if (!vm_paging_needed())
1473 				vm_pages_needed = 0;
1474 			wakeup(&cnt.v_free_count);
1475 		}
1476 		if (vm_pages_needed) {
1477 			/*
1478 			 * Still not done, take a second pass without waiting
1479 			 * (unlimited dirty cleaning), otherwise sleep a bit
1480 			 * and try again.
1481 			 */
1482 			++pass;
1483 			if (pass > 1)
1484 				msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1485 				       "psleep", hz/2);
1486 		} else {
1487 			/*
1488 			 * Good enough, sleep & handle stats.  Prime the pass
1489 			 * for the next run.
1490 			 */
1491 			if (pass > 1)
1492 				pass = 1;
1493 			else
1494 				pass = 0;
1495 			error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1496 				    "psleep", vm_pageout_stats_interval * hz);
1497 			if (error && !vm_pages_needed) {
1498 				pass = 0;
1499 				vm_pageout_page_stats();
1500 				vm_page_unlock_queues();
1501 				continue;
1502 			}
1503 		}
1504 		if (vm_pages_needed)
1505 			cnt.v_pdwakeups++;
1506 		vm_page_unlock_queues();
1507 		vm_pageout_scan(pass);
1508 	}
1509 }
1510 
1511 /*
1512  * Unless the page queue lock is held by the caller, this function
1513  * should be regarded as advisory.  Specifically, the caller should
1514  * not msleep() on &cnt.v_free_count following this function unless
1515  * the page queue lock is held until the msleep() is performed.
1516  */
1517 void
1518 pagedaemon_wakeup()
1519 {
1520 
1521 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1522 		vm_pages_needed = 1;
1523 		wakeup(&vm_pages_needed);
1524 	}
1525 }
1526 
1527 #if !defined(NO_SWAPPING)
1528 static void
1529 vm_req_vmdaemon()
1530 {
1531 	static int lastrun = 0;
1532 
1533 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1534 		wakeup(&vm_daemon_needed);
1535 		lastrun = ticks;
1536 	}
1537 }
1538 
1539 static void
1540 vm_daemon()
1541 {
1542 	struct rlimit rsslim;
1543 	struct proc *p;
1544 	struct thread *td;
1545 	int breakout;
1546 
1547 	mtx_lock(&Giant);
1548 	while (TRUE) {
1549 		tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1550 		if (vm_pageout_req_swapout) {
1551 			swapout_procs(vm_pageout_req_swapout);
1552 			vm_pageout_req_swapout = 0;
1553 		}
1554 		/*
1555 		 * scan the processes for exceeding their rlimits or if
1556 		 * process is swapped out -- deactivate pages
1557 		 */
1558 		sx_slock(&allproc_lock);
1559 		LIST_FOREACH(p, &allproc, p_list) {
1560 			vm_pindex_t limit, size;
1561 
1562 			/*
1563 			 * if this is a system process or if we have already
1564 			 * looked at this process, skip it.
1565 			 */
1566 			PROC_LOCK(p);
1567 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1568 				PROC_UNLOCK(p);
1569 				continue;
1570 			}
1571 			/*
1572 			 * if the process is in a non-running type state,
1573 			 * don't touch it.
1574 			 */
1575 			mtx_lock_spin(&sched_lock);
1576 			breakout = 0;
1577 			FOREACH_THREAD_IN_PROC(p, td) {
1578 				if (!TD_ON_RUNQ(td) &&
1579 				    !TD_IS_RUNNING(td) &&
1580 				    !TD_IS_SLEEPING(td)) {
1581 					breakout = 1;
1582 					break;
1583 				}
1584 			}
1585 			mtx_unlock_spin(&sched_lock);
1586 			if (breakout) {
1587 				PROC_UNLOCK(p);
1588 				continue;
1589 			}
1590 			/*
1591 			 * get a limit
1592 			 */
1593 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1594 			limit = OFF_TO_IDX(
1595 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1596 
1597 			/*
1598 			 * let processes that are swapped out really be
1599 			 * swapped out set the limit to nothing (will force a
1600 			 * swap-out.)
1601 			 */
1602 			if ((p->p_sflag & PS_INMEM) == 0)
1603 				limit = 0;	/* XXX */
1604 			PROC_UNLOCK(p);
1605 
1606 			size = vmspace_resident_count(p->p_vmspace);
1607 			if (limit >= 0 && size >= limit) {
1608 				vm_pageout_map_deactivate_pages(
1609 				    &p->p_vmspace->vm_map, limit);
1610 			}
1611 		}
1612 		sx_sunlock(&allproc_lock);
1613 	}
1614 }
1615 #endif			/* !defined(NO_SWAPPING) */
1616