xref: /freebsd/sys/vm/vm_pageout.c (revision 1e413cf93298b5b97441a21d9a50fdcd0ee9945e)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
77 
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
94 #include <sys/sx.h>
95 #include <sys/sysctl.h>
96 
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/uma.h>
107 
108 #include <machine/mutex.h>
109 
110 /*
111  * System initialization
112  */
113 
114 /* the kernel process "vm_pageout"*/
115 static void vm_pageout(void);
116 static int vm_pageout_clean(vm_page_t);
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 static struct mtx vm_daemon_mtx;
150 /* Allow for use by vm_pageout before vm_daemon is initialized. */
151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
152 #endif
153 static int vm_max_launder = 32;
154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
155 static int vm_pageout_full_stats_interval = 0;
156 static int vm_pageout_algorithm=0;
157 static int defer_swap_pageouts=0;
158 static int disable_swap_pageouts=0;
159 
160 #if defined(NO_SWAPPING)
161 static int vm_swap_enabled=0;
162 static int vm_swap_idle_enabled=0;
163 #else
164 static int vm_swap_enabled=1;
165 static int vm_swap_idle_enabled=0;
166 #endif
167 
168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
169 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
170 
171 SYSCTL_INT(_vm, OID_AUTO, max_launder,
172 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
173 
174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
175 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
176 
177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
178 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
179 
180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
181 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
182 
183 #if defined(NO_SWAPPING)
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
188 #else
189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
190 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
192 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
193 #endif
194 
195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
196 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
197 
198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
199 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
200 
201 static int pageout_lock_miss;
202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
203 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
204 
205 #define VM_PAGEOUT_PAGE_COUNT 16
206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
207 
208 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
209 SYSCTL_INT(_vm, OID_AUTO, max_wired,
210 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
211 
212 #if !defined(NO_SWAPPING)
213 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
214 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
215 static void vm_req_vmdaemon(int req);
216 #endif
217 static void vm_pageout_page_stats(void);
218 
219 /*
220  * vm_pageout_fallback_object_lock:
221  *
222  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
223  * known to have failed and page queue must be either PQ_ACTIVE or
224  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
225  * while locking the vm object.  Use marker page to detect page queue
226  * changes and maintain notion of next page on page queue.  Return
227  * TRUE if no changes were detected, FALSE otherwise.  vm object is
228  * locked on return.
229  *
230  * This function depends on both the lock portion of struct vm_object
231  * and normal struct vm_page being type stable.
232  */
233 boolean_t
234 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
235 {
236 	struct vm_page marker;
237 	boolean_t unchanged;
238 	u_short queue;
239 	vm_object_t object;
240 
241 	/*
242 	 * Initialize our marker
243 	 */
244 	bzero(&marker, sizeof(marker));
245 	marker.flags = PG_FICTITIOUS | PG_MARKER;
246 	marker.oflags = VPO_BUSY;
247 	marker.queue = m->queue;
248 	marker.wire_count = 1;
249 
250 	queue = m->queue;
251 	object = m->object;
252 
253 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
254 			   m, &marker, pageq);
255 	vm_page_unlock_queues();
256 	VM_OBJECT_LOCK(object);
257 	vm_page_lock_queues();
258 
259 	/* Page queue might have changed. */
260 	*next = TAILQ_NEXT(&marker, pageq);
261 	unchanged = (m->queue == queue &&
262 		     m->object == object &&
263 		     &marker == TAILQ_NEXT(m, pageq));
264 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
265 		     &marker, pageq);
266 	return (unchanged);
267 }
268 
269 /*
270  * vm_pageout_clean:
271  *
272  * Clean the page and remove it from the laundry.
273  *
274  * We set the busy bit to cause potential page faults on this page to
275  * block.  Note the careful timing, however, the busy bit isn't set till
276  * late and we cannot do anything that will mess with the page.
277  */
278 static int
279 vm_pageout_clean(m)
280 	vm_page_t m;
281 {
282 	vm_object_t object;
283 	vm_page_t mc[2*vm_pageout_page_count];
284 	int pageout_count;
285 	int ib, is, page_base;
286 	vm_pindex_t pindex = m->pindex;
287 
288 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
289 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
290 
291 	/*
292 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
293 	 * with the new swapper, but we could have serious problems paging
294 	 * out other object types if there is insufficient memory.
295 	 *
296 	 * Unfortunately, checking free memory here is far too late, so the
297 	 * check has been moved up a procedural level.
298 	 */
299 
300 	/*
301 	 * Can't clean the page if it's busy or held.
302 	 */
303 	if ((m->hold_count != 0) ||
304 	    ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
305 		return 0;
306 	}
307 
308 	mc[vm_pageout_page_count] = m;
309 	pageout_count = 1;
310 	page_base = vm_pageout_page_count;
311 	ib = 1;
312 	is = 1;
313 
314 	/*
315 	 * Scan object for clusterable pages.
316 	 *
317 	 * We can cluster ONLY if: ->> the page is NOT
318 	 * clean, wired, busy, held, or mapped into a
319 	 * buffer, and one of the following:
320 	 * 1) The page is inactive, or a seldom used
321 	 *    active page.
322 	 * -or-
323 	 * 2) we force the issue.
324 	 *
325 	 * During heavy mmap/modification loads the pageout
326 	 * daemon can really fragment the underlying file
327 	 * due to flushing pages out of order and not trying
328 	 * align the clusters (which leave sporatic out-of-order
329 	 * holes).  To solve this problem we do the reverse scan
330 	 * first and attempt to align our cluster, then do a
331 	 * forward scan if room remains.
332 	 */
333 	object = m->object;
334 more:
335 	while (ib && pageout_count < vm_pageout_page_count) {
336 		vm_page_t p;
337 
338 		if (ib > pindex) {
339 			ib = 0;
340 			break;
341 		}
342 
343 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
344 			ib = 0;
345 			break;
346 		}
347 		if ((p->oflags & VPO_BUSY) || p->busy) {
348 			ib = 0;
349 			break;
350 		}
351 		vm_page_test_dirty(p);
352 		if ((p->dirty & p->valid) == 0 ||
353 		    p->queue != PQ_INACTIVE ||
354 		    p->wire_count != 0 ||	/* may be held by buf cache */
355 		    p->hold_count != 0) {	/* may be undergoing I/O */
356 			ib = 0;
357 			break;
358 		}
359 		mc[--page_base] = p;
360 		++pageout_count;
361 		++ib;
362 		/*
363 		 * alignment boundry, stop here and switch directions.  Do
364 		 * not clear ib.
365 		 */
366 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
367 			break;
368 	}
369 
370 	while (pageout_count < vm_pageout_page_count &&
371 	    pindex + is < object->size) {
372 		vm_page_t p;
373 
374 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
375 			break;
376 		if ((p->oflags & VPO_BUSY) || p->busy) {
377 			break;
378 		}
379 		vm_page_test_dirty(p);
380 		if ((p->dirty & p->valid) == 0 ||
381 		    p->queue != PQ_INACTIVE ||
382 		    p->wire_count != 0 ||	/* may be held by buf cache */
383 		    p->hold_count != 0) {	/* may be undergoing I/O */
384 			break;
385 		}
386 		mc[page_base + pageout_count] = p;
387 		++pageout_count;
388 		++is;
389 	}
390 
391 	/*
392 	 * If we exhausted our forward scan, continue with the reverse scan
393 	 * when possible, even past a page boundry.  This catches boundry
394 	 * conditions.
395 	 */
396 	if (ib && pageout_count < vm_pageout_page_count)
397 		goto more;
398 
399 	/*
400 	 * we allow reads during pageouts...
401 	 */
402 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
403 }
404 
405 /*
406  * vm_pageout_flush() - launder the given pages
407  *
408  *	The given pages are laundered.  Note that we setup for the start of
409  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
410  *	reference count all in here rather then in the parent.  If we want
411  *	the parent to do more sophisticated things we may have to change
412  *	the ordering.
413  */
414 int
415 vm_pageout_flush(vm_page_t *mc, int count, int flags)
416 {
417 	vm_object_t object = mc[0]->object;
418 	int pageout_status[count];
419 	int numpagedout = 0;
420 	int i;
421 
422 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
423 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
424 	/*
425 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
426 	 * mark the pages read-only.
427 	 *
428 	 * We do not have to fixup the clean/dirty bits here... we can
429 	 * allow the pager to do it after the I/O completes.
430 	 *
431 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
432 	 * edge case with file fragments.
433 	 */
434 	for (i = 0; i < count; i++) {
435 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
436 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
437 			mc[i], i, count));
438 		vm_page_io_start(mc[i]);
439 		pmap_remove_write(mc[i]);
440 	}
441 	vm_page_unlock_queues();
442 	vm_object_pip_add(object, count);
443 
444 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
445 
446 	vm_page_lock_queues();
447 	for (i = 0; i < count; i++) {
448 		vm_page_t mt = mc[i];
449 
450 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
451 		    (mt->flags & PG_WRITEABLE) == 0,
452 		    ("vm_pageout_flush: page %p is not write protected", mt));
453 		switch (pageout_status[i]) {
454 		case VM_PAGER_OK:
455 		case VM_PAGER_PEND:
456 			numpagedout++;
457 			break;
458 		case VM_PAGER_BAD:
459 			/*
460 			 * Page outside of range of object. Right now we
461 			 * essentially lose the changes by pretending it
462 			 * worked.
463 			 */
464 			pmap_clear_modify(mt);
465 			vm_page_undirty(mt);
466 			break;
467 		case VM_PAGER_ERROR:
468 		case VM_PAGER_FAIL:
469 			/*
470 			 * If page couldn't be paged out, then reactivate the
471 			 * page so it doesn't clog the inactive list.  (We
472 			 * will try paging out it again later).
473 			 */
474 			vm_page_activate(mt);
475 			break;
476 		case VM_PAGER_AGAIN:
477 			break;
478 		}
479 
480 		/*
481 		 * If the operation is still going, leave the page busy to
482 		 * block all other accesses. Also, leave the paging in
483 		 * progress indicator set so that we don't attempt an object
484 		 * collapse.
485 		 */
486 		if (pageout_status[i] != VM_PAGER_PEND) {
487 			vm_object_pip_wakeup(object);
488 			vm_page_io_finish(mt);
489 			if (vm_page_count_severe())
490 				vm_page_try_to_cache(mt);
491 		}
492 	}
493 	return numpagedout;
494 }
495 
496 #if !defined(NO_SWAPPING)
497 /*
498  *	vm_pageout_object_deactivate_pages
499  *
500  *	deactivate enough pages to satisfy the inactive target
501  *	requirements or if vm_page_proc_limit is set, then
502  *	deactivate all of the pages in the object and its
503  *	backing_objects.
504  *
505  *	The object and map must be locked.
506  */
507 static void
508 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
509 	pmap_t pmap;
510 	vm_object_t first_object;
511 	long desired;
512 {
513 	vm_object_t backing_object, object;
514 	vm_page_t p, next;
515 	int actcount, rcount, remove_mode;
516 
517 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
518 	if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
519 		return;
520 	for (object = first_object;; object = backing_object) {
521 		if (pmap_resident_count(pmap) <= desired)
522 			goto unlock_return;
523 		if (object->paging_in_progress)
524 			goto unlock_return;
525 
526 		remove_mode = 0;
527 		if (object->shadow_count > 1)
528 			remove_mode = 1;
529 		/*
530 		 * scan the objects entire memory queue
531 		 */
532 		rcount = object->resident_page_count;
533 		p = TAILQ_FIRST(&object->memq);
534 		vm_page_lock_queues();
535 		while (p && (rcount-- > 0)) {
536 			if (pmap_resident_count(pmap) <= desired) {
537 				vm_page_unlock_queues();
538 				goto unlock_return;
539 			}
540 			next = TAILQ_NEXT(p, listq);
541 			cnt.v_pdpages++;
542 			if (p->wire_count != 0 ||
543 			    p->hold_count != 0 ||
544 			    p->busy != 0 ||
545 			    (p->oflags & VPO_BUSY) ||
546 			    (p->flags & PG_UNMANAGED) ||
547 			    !pmap_page_exists_quick(pmap, p)) {
548 				p = next;
549 				continue;
550 			}
551 			actcount = pmap_ts_referenced(p);
552 			if (actcount) {
553 				vm_page_flag_set(p, PG_REFERENCED);
554 			} else if (p->flags & PG_REFERENCED) {
555 				actcount = 1;
556 			}
557 			if ((p->queue != PQ_ACTIVE) &&
558 				(p->flags & PG_REFERENCED)) {
559 				vm_page_activate(p);
560 				p->act_count += actcount;
561 				vm_page_flag_clear(p, PG_REFERENCED);
562 			} else if (p->queue == PQ_ACTIVE) {
563 				if ((p->flags & PG_REFERENCED) == 0) {
564 					p->act_count -= min(p->act_count, ACT_DECLINE);
565 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
566 						pmap_remove_all(p);
567 						vm_page_deactivate(p);
568 					} else {
569 						vm_pageq_requeue(p);
570 					}
571 				} else {
572 					vm_page_activate(p);
573 					vm_page_flag_clear(p, PG_REFERENCED);
574 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
575 						p->act_count += ACT_ADVANCE;
576 					vm_pageq_requeue(p);
577 				}
578 			} else if (p->queue == PQ_INACTIVE) {
579 				pmap_remove_all(p);
580 			}
581 			p = next;
582 		}
583 		vm_page_unlock_queues();
584 		if ((backing_object = object->backing_object) == NULL)
585 			goto unlock_return;
586 		VM_OBJECT_LOCK(backing_object);
587 		if (object != first_object)
588 			VM_OBJECT_UNLOCK(object);
589 	}
590 unlock_return:
591 	if (object != first_object)
592 		VM_OBJECT_UNLOCK(object);
593 }
594 
595 /*
596  * deactivate some number of pages in a map, try to do it fairly, but
597  * that is really hard to do.
598  */
599 static void
600 vm_pageout_map_deactivate_pages(map, desired)
601 	vm_map_t map;
602 	long desired;
603 {
604 	vm_map_entry_t tmpe;
605 	vm_object_t obj, bigobj;
606 	int nothingwired;
607 
608 	if (!vm_map_trylock(map))
609 		return;
610 
611 	bigobj = NULL;
612 	nothingwired = TRUE;
613 
614 	/*
615 	 * first, search out the biggest object, and try to free pages from
616 	 * that.
617 	 */
618 	tmpe = map->header.next;
619 	while (tmpe != &map->header) {
620 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
621 			obj = tmpe->object.vm_object;
622 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
623 				if (obj->shadow_count <= 1 &&
624 				    (bigobj == NULL ||
625 				     bigobj->resident_page_count < obj->resident_page_count)) {
626 					if (bigobj != NULL)
627 						VM_OBJECT_UNLOCK(bigobj);
628 					bigobj = obj;
629 				} else
630 					VM_OBJECT_UNLOCK(obj);
631 			}
632 		}
633 		if (tmpe->wired_count > 0)
634 			nothingwired = FALSE;
635 		tmpe = tmpe->next;
636 	}
637 
638 	if (bigobj != NULL) {
639 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
640 		VM_OBJECT_UNLOCK(bigobj);
641 	}
642 	/*
643 	 * Next, hunt around for other pages to deactivate.  We actually
644 	 * do this search sort of wrong -- .text first is not the best idea.
645 	 */
646 	tmpe = map->header.next;
647 	while (tmpe != &map->header) {
648 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
649 			break;
650 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
651 			obj = tmpe->object.vm_object;
652 			if (obj != NULL) {
653 				VM_OBJECT_LOCK(obj);
654 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
655 				VM_OBJECT_UNLOCK(obj);
656 			}
657 		}
658 		tmpe = tmpe->next;
659 	}
660 
661 	/*
662 	 * Remove all mappings if a process is swapped out, this will free page
663 	 * table pages.
664 	 */
665 	if (desired == 0 && nothingwired) {
666 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
667 		    vm_map_max(map));
668 	}
669 	vm_map_unlock(map);
670 }
671 #endif		/* !defined(NO_SWAPPING) */
672 
673 /*
674  *	vm_pageout_scan does the dirty work for the pageout daemon.
675  */
676 static void
677 vm_pageout_scan(int pass)
678 {
679 	vm_page_t m, next;
680 	struct vm_page marker;
681 	int page_shortage, maxscan, pcount;
682 	int addl_page_shortage, addl_page_shortage_init;
683 	struct proc *p, *bigproc;
684 	struct thread *td;
685 	vm_offset_t size, bigsize;
686 	vm_object_t object;
687 	int actcount;
688 	int vnodes_skipped = 0;
689 	int maxlaunder;
690 
691 	/*
692 	 * Decrease registered cache sizes.
693 	 */
694 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
695 	/*
696 	 * We do this explicitly after the caches have been drained above.
697 	 */
698 	uma_reclaim();
699 
700 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
701 
702 	/*
703 	 * Calculate the number of pages we want to either free or move
704 	 * to the cache.
705 	 */
706 	page_shortage = vm_paging_target() + addl_page_shortage_init;
707 
708 	/*
709 	 * Initialize our marker
710 	 */
711 	bzero(&marker, sizeof(marker));
712 	marker.flags = PG_FICTITIOUS | PG_MARKER;
713 	marker.oflags = VPO_BUSY;
714 	marker.queue = PQ_INACTIVE;
715 	marker.wire_count = 1;
716 
717 	/*
718 	 * Start scanning the inactive queue for pages we can move to the
719 	 * cache or free.  The scan will stop when the target is reached or
720 	 * we have scanned the entire inactive queue.  Note that m->act_count
721 	 * is not used to form decisions for the inactive queue, only for the
722 	 * active queue.
723 	 *
724 	 * maxlaunder limits the number of dirty pages we flush per scan.
725 	 * For most systems a smaller value (16 or 32) is more robust under
726 	 * extreme memory and disk pressure because any unnecessary writes
727 	 * to disk can result in extreme performance degredation.  However,
728 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
729 	 * used) will die horribly with limited laundering.  If the pageout
730 	 * daemon cannot clean enough pages in the first pass, we let it go
731 	 * all out in succeeding passes.
732 	 */
733 	if ((maxlaunder = vm_max_launder) <= 1)
734 		maxlaunder = 1;
735 	if (pass)
736 		maxlaunder = 10000;
737 	vm_page_lock_queues();
738 rescan0:
739 	addl_page_shortage = addl_page_shortage_init;
740 	maxscan = cnt.v_inactive_count;
741 
742 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
743 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
744 	     m = next) {
745 
746 		cnt.v_pdpages++;
747 
748 		if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
749 			goto rescan0;
750 		}
751 
752 		next = TAILQ_NEXT(m, pageq);
753 		object = m->object;
754 
755 		/*
756 		 * skip marker pages
757 		 */
758 		if (m->flags & PG_MARKER)
759 			continue;
760 
761 		/*
762 		 * A held page may be undergoing I/O, so skip it.
763 		 */
764 		if (m->hold_count) {
765 			vm_pageq_requeue(m);
766 			addl_page_shortage++;
767 			continue;
768 		}
769 		/*
770 		 * Don't mess with busy pages, keep in the front of the
771 		 * queue, most likely are being paged out.
772 		 */
773 		if (!VM_OBJECT_TRYLOCK(object) &&
774 		    (!vm_pageout_fallback_object_lock(m, &next) ||
775 		     m->hold_count != 0)) {
776 			VM_OBJECT_UNLOCK(object);
777 			addl_page_shortage++;
778 			continue;
779 		}
780 		if (m->busy || (m->oflags & VPO_BUSY)) {
781 			VM_OBJECT_UNLOCK(object);
782 			addl_page_shortage++;
783 			continue;
784 		}
785 
786 		/*
787 		 * If the object is not being used, we ignore previous
788 		 * references.
789 		 */
790 		if (object->ref_count == 0) {
791 			vm_page_flag_clear(m, PG_REFERENCED);
792 			pmap_clear_reference(m);
793 
794 		/*
795 		 * Otherwise, if the page has been referenced while in the
796 		 * inactive queue, we bump the "activation count" upwards,
797 		 * making it less likely that the page will be added back to
798 		 * the inactive queue prematurely again.  Here we check the
799 		 * page tables (or emulated bits, if any), given the upper
800 		 * level VM system not knowing anything about existing
801 		 * references.
802 		 */
803 		} else if (((m->flags & PG_REFERENCED) == 0) &&
804 			(actcount = pmap_ts_referenced(m))) {
805 			vm_page_activate(m);
806 			VM_OBJECT_UNLOCK(object);
807 			m->act_count += (actcount + ACT_ADVANCE);
808 			continue;
809 		}
810 
811 		/*
812 		 * If the upper level VM system knows about any page
813 		 * references, we activate the page.  We also set the
814 		 * "activation count" higher than normal so that we will less
815 		 * likely place pages back onto the inactive queue again.
816 		 */
817 		if ((m->flags & PG_REFERENCED) != 0) {
818 			vm_page_flag_clear(m, PG_REFERENCED);
819 			actcount = pmap_ts_referenced(m);
820 			vm_page_activate(m);
821 			VM_OBJECT_UNLOCK(object);
822 			m->act_count += (actcount + ACT_ADVANCE + 1);
823 			continue;
824 		}
825 
826 		/*
827 		 * If the upper level VM system doesn't know anything about
828 		 * the page being dirty, we have to check for it again.  As
829 		 * far as the VM code knows, any partially dirty pages are
830 		 * fully dirty.
831 		 */
832 		if (m->dirty == 0 && !pmap_is_modified(m)) {
833 			/*
834 			 * Avoid a race condition: Unless write access is
835 			 * removed from the page, another processor could
836 			 * modify it before all access is removed by the call
837 			 * to vm_page_cache() below.  If vm_page_cache() finds
838 			 * that the page has been modified when it removes all
839 			 * access, it panics because it cannot cache dirty
840 			 * pages.  In principle, we could eliminate just write
841 			 * access here rather than all access.  In the expected
842 			 * case, when there are no last instant modifications
843 			 * to the page, removing all access will be cheaper
844 			 * overall.
845 			 */
846 			if ((m->flags & PG_WRITEABLE) != 0)
847 				pmap_remove_all(m);
848 		} else {
849 			vm_page_dirty(m);
850 		}
851 
852 		if (m->valid == 0) {
853 			/*
854 			 * Invalid pages can be easily freed
855 			 */
856 			vm_page_free(m);
857 			cnt.v_dfree++;
858 			--page_shortage;
859 		} else if (m->dirty == 0) {
860 			/*
861 			 * Clean pages can be placed onto the cache queue.
862 			 * This effectively frees them.
863 			 */
864 			vm_page_cache(m);
865 			--page_shortage;
866 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
867 			/*
868 			 * Dirty pages need to be paged out, but flushing
869 			 * a page is extremely expensive verses freeing
870 			 * a clean page.  Rather then artificially limiting
871 			 * the number of pages we can flush, we instead give
872 			 * dirty pages extra priority on the inactive queue
873 			 * by forcing them to be cycled through the queue
874 			 * twice before being flushed, after which the
875 			 * (now clean) page will cycle through once more
876 			 * before being freed.  This significantly extends
877 			 * the thrash point for a heavily loaded machine.
878 			 */
879 			vm_page_flag_set(m, PG_WINATCFLS);
880 			vm_pageq_requeue(m);
881 		} else if (maxlaunder > 0) {
882 			/*
883 			 * We always want to try to flush some dirty pages if
884 			 * we encounter them, to keep the system stable.
885 			 * Normally this number is small, but under extreme
886 			 * pressure where there are insufficient clean pages
887 			 * on the inactive queue, we may have to go all out.
888 			 */
889 			int swap_pageouts_ok, vfslocked = 0;
890 			struct vnode *vp = NULL;
891 			struct mount *mp = NULL;
892 
893 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
894 				swap_pageouts_ok = 1;
895 			} else {
896 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
897 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
898 				vm_page_count_min());
899 
900 			}
901 
902 			/*
903 			 * We don't bother paging objects that are "dead".
904 			 * Those objects are in a "rundown" state.
905 			 */
906 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
907 				VM_OBJECT_UNLOCK(object);
908 				vm_pageq_requeue(m);
909 				continue;
910 			}
911 
912 			/*
913 			 * Following operations may unlock
914 			 * vm_page_queue_mtx, invalidating the 'next'
915 			 * pointer.  To prevent an inordinate number
916 			 * of restarts we use our marker to remember
917 			 * our place.
918 			 *
919 			 */
920 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
921 					   m, &marker, pageq);
922 			/*
923 			 * The object is already known NOT to be dead.   It
924 			 * is possible for the vget() to block the whole
925 			 * pageout daemon, but the new low-memory handling
926 			 * code should prevent it.
927 			 *
928 			 * The previous code skipped locked vnodes and, worse,
929 			 * reordered pages in the queue.  This results in
930 			 * completely non-deterministic operation and, on a
931 			 * busy system, can lead to extremely non-optimal
932 			 * pageouts.  For example, it can cause clean pages
933 			 * to be freed and dirty pages to be moved to the end
934 			 * of the queue.  Since dirty pages are also moved to
935 			 * the end of the queue once-cleaned, this gives
936 			 * way too large a weighting to defering the freeing
937 			 * of dirty pages.
938 			 *
939 			 * We can't wait forever for the vnode lock, we might
940 			 * deadlock due to a vn_read() getting stuck in
941 			 * vm_wait while holding this vnode.  We skip the
942 			 * vnode if we can't get it in a reasonable amount
943 			 * of time.
944 			 */
945 			if (object->type == OBJT_VNODE) {
946 				vp = object->handle;
947 				if (vp->v_type == VREG &&
948 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
949 					KASSERT(mp == NULL,
950 					    ("vm_pageout_scan: mp != NULL"));
951 					++pageout_lock_miss;
952 					if (object->flags & OBJ_MIGHTBEDIRTY)
953 						vnodes_skipped++;
954 					goto unlock_and_continue;
955 				}
956 				vm_page_unlock_queues();
957 				vm_object_reference_locked(object);
958 				VM_OBJECT_UNLOCK(object);
959 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
960 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
961 				    curthread)) {
962 					VM_OBJECT_LOCK(object);
963 					vm_page_lock_queues();
964 					++pageout_lock_miss;
965 					if (object->flags & OBJ_MIGHTBEDIRTY)
966 						vnodes_skipped++;
967 					vp = NULL;
968 					goto unlock_and_continue;
969 				}
970 				VM_OBJECT_LOCK(object);
971 				vm_page_lock_queues();
972 				/*
973 				 * The page might have been moved to another
974 				 * queue during potential blocking in vget()
975 				 * above.  The page might have been freed and
976 				 * reused for another vnode.
977 				 */
978 				if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
979 				    m->object != object ||
980 				    TAILQ_NEXT(m, pageq) != &marker) {
981 					if (object->flags & OBJ_MIGHTBEDIRTY)
982 						vnodes_skipped++;
983 					goto unlock_and_continue;
984 				}
985 
986 				/*
987 				 * The page may have been busied during the
988 				 * blocking in vget().  We don't move the
989 				 * page back onto the end of the queue so that
990 				 * statistics are more correct if we don't.
991 				 */
992 				if (m->busy || (m->oflags & VPO_BUSY)) {
993 					goto unlock_and_continue;
994 				}
995 
996 				/*
997 				 * If the page has become held it might
998 				 * be undergoing I/O, so skip it
999 				 */
1000 				if (m->hold_count) {
1001 					vm_pageq_requeue(m);
1002 					if (object->flags & OBJ_MIGHTBEDIRTY)
1003 						vnodes_skipped++;
1004 					goto unlock_and_continue;
1005 				}
1006 			}
1007 
1008 			/*
1009 			 * If a page is dirty, then it is either being washed
1010 			 * (but not yet cleaned) or it is still in the
1011 			 * laundry.  If it is still in the laundry, then we
1012 			 * start the cleaning operation.
1013 			 *
1014 			 * decrement page_shortage on success to account for
1015 			 * the (future) cleaned page.  Otherwise we could wind
1016 			 * up laundering or cleaning too many pages.
1017 			 */
1018 			if (vm_pageout_clean(m) != 0) {
1019 				--page_shortage;
1020 				--maxlaunder;
1021 			}
1022 unlock_and_continue:
1023 			VM_OBJECT_UNLOCK(object);
1024 			if (mp != NULL) {
1025 				vm_page_unlock_queues();
1026 				if (vp != NULL)
1027 					vput(vp);
1028 				VFS_UNLOCK_GIANT(vfslocked);
1029 				vm_object_deallocate(object);
1030 				vn_finished_write(mp);
1031 				vm_page_lock_queues();
1032 			}
1033 			next = TAILQ_NEXT(&marker, pageq);
1034 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1035 				     &marker, pageq);
1036 			continue;
1037 		}
1038 		VM_OBJECT_UNLOCK(object);
1039 	}
1040 
1041 	/*
1042 	 * Compute the number of pages we want to try to move from the
1043 	 * active queue to the inactive queue.
1044 	 */
1045 	page_shortage = vm_paging_target() +
1046 		cnt.v_inactive_target - cnt.v_inactive_count;
1047 	page_shortage += addl_page_shortage;
1048 
1049 	/*
1050 	 * Scan the active queue for things we can deactivate. We nominally
1051 	 * track the per-page activity counter and use it to locate
1052 	 * deactivation candidates.
1053 	 */
1054 	pcount = cnt.v_active_count;
1055 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1056 
1057 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1058 
1059 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1060 		    ("vm_pageout_scan: page %p isn't active", m));
1061 
1062 		next = TAILQ_NEXT(m, pageq);
1063 		object = m->object;
1064 		if ((m->flags & PG_MARKER) != 0) {
1065 			m = next;
1066 			continue;
1067 		}
1068 		if (!VM_OBJECT_TRYLOCK(object) &&
1069 		    !vm_pageout_fallback_object_lock(m, &next)) {
1070 			VM_OBJECT_UNLOCK(object);
1071 			m = next;
1072 			continue;
1073 		}
1074 
1075 		/*
1076 		 * Don't deactivate pages that are busy.
1077 		 */
1078 		if ((m->busy != 0) ||
1079 		    (m->oflags & VPO_BUSY) ||
1080 		    (m->hold_count != 0)) {
1081 			VM_OBJECT_UNLOCK(object);
1082 			vm_pageq_requeue(m);
1083 			m = next;
1084 			continue;
1085 		}
1086 
1087 		/*
1088 		 * The count for pagedaemon pages is done after checking the
1089 		 * page for eligibility...
1090 		 */
1091 		cnt.v_pdpages++;
1092 
1093 		/*
1094 		 * Check to see "how much" the page has been used.
1095 		 */
1096 		actcount = 0;
1097 		if (object->ref_count != 0) {
1098 			if (m->flags & PG_REFERENCED) {
1099 				actcount += 1;
1100 			}
1101 			actcount += pmap_ts_referenced(m);
1102 			if (actcount) {
1103 				m->act_count += ACT_ADVANCE + actcount;
1104 				if (m->act_count > ACT_MAX)
1105 					m->act_count = ACT_MAX;
1106 			}
1107 		}
1108 
1109 		/*
1110 		 * Since we have "tested" this bit, we need to clear it now.
1111 		 */
1112 		vm_page_flag_clear(m, PG_REFERENCED);
1113 
1114 		/*
1115 		 * Only if an object is currently being used, do we use the
1116 		 * page activation count stats.
1117 		 */
1118 		if (actcount && (object->ref_count != 0)) {
1119 			vm_pageq_requeue(m);
1120 		} else {
1121 			m->act_count -= min(m->act_count, ACT_DECLINE);
1122 			if (vm_pageout_algorithm ||
1123 			    object->ref_count == 0 ||
1124 			    m->act_count == 0) {
1125 				page_shortage--;
1126 				if (object->ref_count == 0) {
1127 					pmap_remove_all(m);
1128 					if (m->dirty == 0)
1129 						vm_page_cache(m);
1130 					else
1131 						vm_page_deactivate(m);
1132 				} else {
1133 					vm_page_deactivate(m);
1134 				}
1135 			} else {
1136 				vm_pageq_requeue(m);
1137 			}
1138 		}
1139 		VM_OBJECT_UNLOCK(object);
1140 		m = next;
1141 	}
1142 	vm_page_unlock_queues();
1143 #if !defined(NO_SWAPPING)
1144 	/*
1145 	 * Idle process swapout -- run once per second.
1146 	 */
1147 	if (vm_swap_idle_enabled) {
1148 		static long lsec;
1149 		if (time_second != lsec) {
1150 			vm_req_vmdaemon(VM_SWAP_IDLE);
1151 			lsec = time_second;
1152 		}
1153 	}
1154 #endif
1155 
1156 	/*
1157 	 * If we didn't get enough free pages, and we have skipped a vnode
1158 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1159 	 * if we did not get enough free pages.
1160 	 */
1161 	if (vm_paging_target() > 0) {
1162 		if (vnodes_skipped && vm_page_count_min())
1163 			(void) speedup_syncer();
1164 #if !defined(NO_SWAPPING)
1165 		if (vm_swap_enabled && vm_page_count_target())
1166 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1167 #endif
1168 	}
1169 
1170 	/*
1171 	 * If we are critically low on one of RAM or swap and low on
1172 	 * the other, kill the largest process.  However, we avoid
1173 	 * doing this on the first pass in order to give ourselves a
1174 	 * chance to flush out dirty vnode-backed pages and to allow
1175 	 * active pages to be moved to the inactive queue and reclaimed.
1176 	 *
1177 	 * We keep the process bigproc locked once we find it to keep anyone
1178 	 * from messing with it; however, there is a possibility of
1179 	 * deadlock if process B is bigproc and one of it's child processes
1180 	 * attempts to propagate a signal to B while we are waiting for A's
1181 	 * lock while walking this list.  To avoid this, we don't block on
1182 	 * the process lock but just skip a process if it is already locked.
1183 	 */
1184 	if (pass != 0 &&
1185 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1186 	     (swap_pager_full && vm_paging_target() > 0))) {
1187 		bigproc = NULL;
1188 		bigsize = 0;
1189 		sx_slock(&allproc_lock);
1190 		FOREACH_PROC_IN_SYSTEM(p) {
1191 			int breakout;
1192 
1193 			if (PROC_TRYLOCK(p) == 0)
1194 				continue;
1195 			/*
1196 			 * If this is a system or protected process, skip it.
1197 			 */
1198 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1199 			    (p->p_flag & P_PROTECTED) ||
1200 			    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1201 				PROC_UNLOCK(p);
1202 				continue;
1203 			}
1204 			/*
1205 			 * If the process is in a non-running type state,
1206 			 * don't touch it.  Check all the threads individually.
1207 			 */
1208 			PROC_SLOCK(p);
1209 			breakout = 0;
1210 			FOREACH_THREAD_IN_PROC(p, td) {
1211 				thread_lock(td);
1212 				if (!TD_ON_RUNQ(td) &&
1213 				    !TD_IS_RUNNING(td) &&
1214 				    !TD_IS_SLEEPING(td)) {
1215 					thread_unlock(td);
1216 					breakout = 1;
1217 					break;
1218 				}
1219 				thread_unlock(td);
1220 			}
1221 			PROC_SUNLOCK(p);
1222 			if (breakout) {
1223 				PROC_UNLOCK(p);
1224 				continue;
1225 			}
1226 			/*
1227 			 * get the process size
1228 			 */
1229 			if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1230 				PROC_UNLOCK(p);
1231 				continue;
1232 			}
1233 			size = vmspace_swap_count(p->p_vmspace);
1234 			vm_map_unlock_read(&p->p_vmspace->vm_map);
1235 			size += vmspace_resident_count(p->p_vmspace);
1236 			/*
1237 			 * if the this process is bigger than the biggest one
1238 			 * remember it.
1239 			 */
1240 			if (size > bigsize) {
1241 				if (bigproc != NULL)
1242 					PROC_UNLOCK(bigproc);
1243 				bigproc = p;
1244 				bigsize = size;
1245 			} else
1246 				PROC_UNLOCK(p);
1247 		}
1248 		sx_sunlock(&allproc_lock);
1249 		if (bigproc != NULL) {
1250 			killproc(bigproc, "out of swap space");
1251 			PROC_SLOCK(bigproc);
1252 			sched_nice(bigproc, PRIO_MIN);
1253 			PROC_SUNLOCK(bigproc);
1254 			PROC_UNLOCK(bigproc);
1255 			wakeup(&cnt.v_free_count);
1256 		}
1257 	}
1258 }
1259 
1260 /*
1261  * This routine tries to maintain the pseudo LRU active queue,
1262  * so that during long periods of time where there is no paging,
1263  * that some statistic accumulation still occurs.  This code
1264  * helps the situation where paging just starts to occur.
1265  */
1266 static void
1267 vm_pageout_page_stats()
1268 {
1269 	vm_object_t object;
1270 	vm_page_t m,next;
1271 	int pcount,tpcount;		/* Number of pages to check */
1272 	static int fullintervalcount = 0;
1273 	int page_shortage;
1274 
1275 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1276 	page_shortage =
1277 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1278 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1279 
1280 	if (page_shortage <= 0)
1281 		return;
1282 
1283 	pcount = cnt.v_active_count;
1284 	fullintervalcount += vm_pageout_stats_interval;
1285 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1286 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1287 		if (pcount > tpcount)
1288 			pcount = tpcount;
1289 	} else {
1290 		fullintervalcount = 0;
1291 	}
1292 
1293 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1294 	while ((m != NULL) && (pcount-- > 0)) {
1295 		int actcount;
1296 
1297 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1298 		    ("vm_pageout_page_stats: page %p isn't active", m));
1299 
1300 		next = TAILQ_NEXT(m, pageq);
1301 		object = m->object;
1302 
1303 		if ((m->flags & PG_MARKER) != 0) {
1304 			m = next;
1305 			continue;
1306 		}
1307 		if (!VM_OBJECT_TRYLOCK(object) &&
1308 		    !vm_pageout_fallback_object_lock(m, &next)) {
1309 			VM_OBJECT_UNLOCK(object);
1310 			m = next;
1311 			continue;
1312 		}
1313 
1314 		/*
1315 		 * Don't deactivate pages that are busy.
1316 		 */
1317 		if ((m->busy != 0) ||
1318 		    (m->oflags & VPO_BUSY) ||
1319 		    (m->hold_count != 0)) {
1320 			VM_OBJECT_UNLOCK(object);
1321 			vm_pageq_requeue(m);
1322 			m = next;
1323 			continue;
1324 		}
1325 
1326 		actcount = 0;
1327 		if (m->flags & PG_REFERENCED) {
1328 			vm_page_flag_clear(m, PG_REFERENCED);
1329 			actcount += 1;
1330 		}
1331 
1332 		actcount += pmap_ts_referenced(m);
1333 		if (actcount) {
1334 			m->act_count += ACT_ADVANCE + actcount;
1335 			if (m->act_count > ACT_MAX)
1336 				m->act_count = ACT_MAX;
1337 			vm_pageq_requeue(m);
1338 		} else {
1339 			if (m->act_count == 0) {
1340 				/*
1341 				 * We turn off page access, so that we have
1342 				 * more accurate RSS stats.  We don't do this
1343 				 * in the normal page deactivation when the
1344 				 * system is loaded VM wise, because the
1345 				 * cost of the large number of page protect
1346 				 * operations would be higher than the value
1347 				 * of doing the operation.
1348 				 */
1349 				pmap_remove_all(m);
1350 				vm_page_deactivate(m);
1351 			} else {
1352 				m->act_count -= min(m->act_count, ACT_DECLINE);
1353 				vm_pageq_requeue(m);
1354 			}
1355 		}
1356 		VM_OBJECT_UNLOCK(object);
1357 		m = next;
1358 	}
1359 }
1360 
1361 /*
1362  *	vm_pageout is the high level pageout daemon.
1363  */
1364 static void
1365 vm_pageout()
1366 {
1367 	int error, pass;
1368 
1369 	/*
1370 	 * Initialize some paging parameters.
1371 	 */
1372 	cnt.v_interrupt_free_min = 2;
1373 	if (cnt.v_page_count < 2000)
1374 		vm_pageout_page_count = 8;
1375 
1376 	/*
1377 	 * v_free_reserved needs to include enough for the largest
1378 	 * swap pager structures plus enough for any pv_entry structs
1379 	 * when paging.
1380 	 */
1381 	if (cnt.v_page_count > 1024)
1382 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1383 	else
1384 		cnt.v_free_min = 4;
1385 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1386 	    cnt.v_interrupt_free_min;
1387 	cnt.v_free_reserved = vm_pageout_page_count +
1388 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1389 	cnt.v_free_severe = cnt.v_free_min / 2;
1390 	cnt.v_free_min += cnt.v_free_reserved;
1391 	cnt.v_free_severe += cnt.v_free_reserved;
1392 
1393 	/*
1394 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1395 	 * that these are more a measure of the VM cache queue hysteresis
1396 	 * then the VM free queue.  Specifically, v_free_target is the
1397 	 * high water mark (free+cache pages).
1398 	 *
1399 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1400 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1401 	 * be big enough to handle memory needs while the pageout daemon
1402 	 * is signalled and run to free more pages.
1403 	 */
1404 	if (cnt.v_free_count > 6144)
1405 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1406 	else
1407 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1408 
1409 	if (cnt.v_free_count > 2048) {
1410 		cnt.v_cache_min = cnt.v_free_target;
1411 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1412 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1413 	} else {
1414 		cnt.v_cache_min = 0;
1415 		cnt.v_cache_max = 0;
1416 		cnt.v_inactive_target = cnt.v_free_count / 4;
1417 	}
1418 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1419 		cnt.v_inactive_target = cnt.v_free_count / 3;
1420 
1421 	/* XXX does not really belong here */
1422 	if (vm_page_max_wired == 0)
1423 		vm_page_max_wired = cnt.v_free_count / 3;
1424 
1425 	if (vm_pageout_stats_max == 0)
1426 		vm_pageout_stats_max = cnt.v_free_target;
1427 
1428 	/*
1429 	 * Set interval in seconds for stats scan.
1430 	 */
1431 	if (vm_pageout_stats_interval == 0)
1432 		vm_pageout_stats_interval = 5;
1433 	if (vm_pageout_full_stats_interval == 0)
1434 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1435 
1436 	swap_pager_swap_init();
1437 	pass = 0;
1438 	/*
1439 	 * The pageout daemon is never done, so loop forever.
1440 	 */
1441 	while (TRUE) {
1442 		/*
1443 		 * If we have enough free memory, wakeup waiters.  Do
1444 		 * not clear vm_pages_needed until we reach our target,
1445 		 * otherwise we may be woken up over and over again and
1446 		 * waste a lot of cpu.
1447 		 */
1448 		mtx_lock(&vm_page_queue_free_mtx);
1449 		if (vm_pages_needed && !vm_page_count_min()) {
1450 			if (!vm_paging_needed())
1451 				vm_pages_needed = 0;
1452 			wakeup(&cnt.v_free_count);
1453 		}
1454 		if (vm_pages_needed) {
1455 			/*
1456 			 * Still not done, take a second pass without waiting
1457 			 * (unlimited dirty cleaning), otherwise sleep a bit
1458 			 * and try again.
1459 			 */
1460 			++pass;
1461 			if (pass > 1)
1462 				msleep(&vm_pages_needed,
1463 				    &vm_page_queue_free_mtx, PVM, "psleep",
1464 				    hz / 2);
1465 		} else {
1466 			/*
1467 			 * Good enough, sleep & handle stats.  Prime the pass
1468 			 * for the next run.
1469 			 */
1470 			if (pass > 1)
1471 				pass = 1;
1472 			else
1473 				pass = 0;
1474 			error = msleep(&vm_pages_needed,
1475 			    &vm_page_queue_free_mtx, PVM, "psleep",
1476 			    vm_pageout_stats_interval * hz);
1477 			if (error && !vm_pages_needed) {
1478 				mtx_unlock(&vm_page_queue_free_mtx);
1479 				pass = 0;
1480 				vm_page_lock_queues();
1481 				vm_pageout_page_stats();
1482 				vm_page_unlock_queues();
1483 				continue;
1484 			}
1485 		}
1486 		if (vm_pages_needed)
1487 			cnt.v_pdwakeups++;
1488 		mtx_unlock(&vm_page_queue_free_mtx);
1489 		vm_pageout_scan(pass);
1490 	}
1491 }
1492 
1493 /*
1494  * Unless the free page queue lock is held by the caller, this function
1495  * should be regarded as advisory.  Specifically, the caller should
1496  * not msleep() on &cnt.v_free_count following this function unless
1497  * the free page queue lock is held until the msleep() is performed.
1498  */
1499 void
1500 pagedaemon_wakeup()
1501 {
1502 
1503 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1504 		vm_pages_needed = 1;
1505 		wakeup(&vm_pages_needed);
1506 	}
1507 }
1508 
1509 #if !defined(NO_SWAPPING)
1510 static void
1511 vm_req_vmdaemon(int req)
1512 {
1513 	static int lastrun = 0;
1514 
1515 	mtx_lock(&vm_daemon_mtx);
1516 	vm_pageout_req_swapout |= req;
1517 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1518 		wakeup(&vm_daemon_needed);
1519 		lastrun = ticks;
1520 	}
1521 	mtx_unlock(&vm_daemon_mtx);
1522 }
1523 
1524 static void
1525 vm_daemon()
1526 {
1527 	struct rlimit rsslim;
1528 	struct proc *p;
1529 	struct thread *td;
1530 	int breakout, swapout_flags;
1531 
1532 	while (TRUE) {
1533 		mtx_lock(&vm_daemon_mtx);
1534 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1535 		swapout_flags = vm_pageout_req_swapout;
1536 		vm_pageout_req_swapout = 0;
1537 		mtx_unlock(&vm_daemon_mtx);
1538 		if (swapout_flags)
1539 			swapout_procs(swapout_flags);
1540 
1541 		/*
1542 		 * scan the processes for exceeding their rlimits or if
1543 		 * process is swapped out -- deactivate pages
1544 		 */
1545 		sx_slock(&allproc_lock);
1546 		FOREACH_PROC_IN_SYSTEM(p) {
1547 			vm_pindex_t limit, size;
1548 
1549 			/*
1550 			 * if this is a system process or if we have already
1551 			 * looked at this process, skip it.
1552 			 */
1553 			PROC_LOCK(p);
1554 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1555 				PROC_UNLOCK(p);
1556 				continue;
1557 			}
1558 			/*
1559 			 * if the process is in a non-running type state,
1560 			 * don't touch it.
1561 			 */
1562 			PROC_SLOCK(p);
1563 			breakout = 0;
1564 			FOREACH_THREAD_IN_PROC(p, td) {
1565 				thread_lock(td);
1566 				if (!TD_ON_RUNQ(td) &&
1567 				    !TD_IS_RUNNING(td) &&
1568 				    !TD_IS_SLEEPING(td)) {
1569 					thread_unlock(td);
1570 					breakout = 1;
1571 					break;
1572 				}
1573 				thread_unlock(td);
1574 			}
1575 			PROC_SUNLOCK(p);
1576 			if (breakout) {
1577 				PROC_UNLOCK(p);
1578 				continue;
1579 			}
1580 			/*
1581 			 * get a limit
1582 			 */
1583 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1584 			limit = OFF_TO_IDX(
1585 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1586 
1587 			/*
1588 			 * let processes that are swapped out really be
1589 			 * swapped out set the limit to nothing (will force a
1590 			 * swap-out.)
1591 			 */
1592 			if ((p->p_flag & P_INMEM) == 0)
1593 				limit = 0;	/* XXX */
1594 			PROC_UNLOCK(p);
1595 
1596 			size = vmspace_resident_count(p->p_vmspace);
1597 			if (limit >= 0 && size >= limit) {
1598 				vm_pageout_map_deactivate_pages(
1599 				    &p->p_vmspace->vm_map, limit);
1600 			}
1601 		}
1602 		sx_sunlock(&allproc_lock);
1603 	}
1604 }
1605 #endif			/* !defined(NO_SWAPPING) */
1606