xref: /freebsd/sys/vm/vm_pageout.c (revision d8b878873e7aa8df1972cc6a642804b17eb61087)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
77 
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
94 #include <sys/sx.h>
95 #include <sys/sysctl.h>
96 
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/uma.h>
107 
108 /*
109  * System initialization
110  */
111 
112 /* the kernel process "vm_pageout"*/
113 static void vm_pageout(void);
114 static int vm_pageout_clean(vm_page_t);
115 static void vm_pageout_scan(int pass);
116 
117 struct proc *pageproc;
118 
119 static struct kproc_desc page_kp = {
120 	"pagedaemon",
121 	vm_pageout,
122 	&pageproc
123 };
124 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
125     &page_kp);
126 
127 #if !defined(NO_SWAPPING)
128 /* the kernel process "vm_daemon"*/
129 static void vm_daemon(void);
130 static struct	proc *vmproc;
131 
132 static struct kproc_desc vm_kp = {
133 	"vmdaemon",
134 	vm_daemon,
135 	&vmproc
136 };
137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
138 #endif
139 
140 
141 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
142 int vm_pageout_deficit;		/* Estimated number of pages deficit */
143 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
144 
145 #if !defined(NO_SWAPPING)
146 static int vm_pageout_req_swapout;	/* XXX */
147 static int vm_daemon_needed;
148 static struct mtx vm_daemon_mtx;
149 /* Allow for use by vm_pageout before vm_daemon is initialized. */
150 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
151 #endif
152 static int vm_max_launder = 32;
153 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
154 static int vm_pageout_full_stats_interval = 0;
155 static int vm_pageout_algorithm=0;
156 static int defer_swap_pageouts=0;
157 static int disable_swap_pageouts=0;
158 
159 #if defined(NO_SWAPPING)
160 static int vm_swap_enabled=0;
161 static int vm_swap_idle_enabled=0;
162 #else
163 static int vm_swap_enabled=1;
164 static int vm_swap_idle_enabled=0;
165 #endif
166 
167 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
168 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
169 
170 SYSCTL_INT(_vm, OID_AUTO, max_launder,
171 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
172 
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
174 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
175 
176 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
177 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
178 
179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
180 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
181 
182 #if defined(NO_SWAPPING)
183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
184 	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
186 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
187 #else
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
192 #endif
193 
194 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
195 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
196 
197 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
198 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
199 
200 static int pageout_lock_miss;
201 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
202 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
203 
204 #define VM_PAGEOUT_PAGE_COUNT 16
205 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
206 
207 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
208 SYSCTL_INT(_vm, OID_AUTO, max_wired,
209 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
210 
211 #if !defined(NO_SWAPPING)
212 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
213 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
214 static void vm_req_vmdaemon(int req);
215 #endif
216 static void vm_pageout_page_stats(void);
217 
218 /*
219  * vm_pageout_fallback_object_lock:
220  *
221  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
222  * known to have failed and page queue must be either PQ_ACTIVE or
223  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
224  * while locking the vm object.  Use marker page to detect page queue
225  * changes and maintain notion of next page on page queue.  Return
226  * TRUE if no changes were detected, FALSE otherwise.  vm object is
227  * locked on return.
228  *
229  * This function depends on both the lock portion of struct vm_object
230  * and normal struct vm_page being type stable.
231  */
232 boolean_t
233 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
234 {
235 	struct vm_page marker;
236 	boolean_t unchanged;
237 	u_short queue;
238 	vm_object_t object;
239 
240 	/*
241 	 * Initialize our marker
242 	 */
243 	bzero(&marker, sizeof(marker));
244 	marker.flags = PG_FICTITIOUS | PG_MARKER;
245 	marker.oflags = VPO_BUSY;
246 	marker.queue = m->queue;
247 	marker.wire_count = 1;
248 
249 	queue = m->queue;
250 	object = m->object;
251 
252 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
253 			   m, &marker, pageq);
254 	vm_page_unlock_queues();
255 	VM_OBJECT_LOCK(object);
256 	vm_page_lock_queues();
257 
258 	/* Page queue might have changed. */
259 	*next = TAILQ_NEXT(&marker, pageq);
260 	unchanged = (m->queue == queue &&
261 		     m->object == object &&
262 		     &marker == TAILQ_NEXT(m, pageq));
263 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
264 		     &marker, pageq);
265 	return (unchanged);
266 }
267 
268 /*
269  * vm_pageout_clean:
270  *
271  * Clean the page and remove it from the laundry.
272  *
273  * We set the busy bit to cause potential page faults on this page to
274  * block.  Note the careful timing, however, the busy bit isn't set till
275  * late and we cannot do anything that will mess with the page.
276  */
277 static int
278 vm_pageout_clean(m)
279 	vm_page_t m;
280 {
281 	vm_object_t object;
282 	vm_page_t mc[2*vm_pageout_page_count];
283 	int pageout_count;
284 	int ib, is, page_base;
285 	vm_pindex_t pindex = m->pindex;
286 
287 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
288 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
289 
290 	/*
291 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
292 	 * with the new swapper, but we could have serious problems paging
293 	 * out other object types if there is insufficient memory.
294 	 *
295 	 * Unfortunately, checking free memory here is far too late, so the
296 	 * check has been moved up a procedural level.
297 	 */
298 
299 	/*
300 	 * Can't clean the page if it's busy or held.
301 	 */
302 	if ((m->hold_count != 0) ||
303 	    ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
304 		return 0;
305 	}
306 
307 	mc[vm_pageout_page_count] = m;
308 	pageout_count = 1;
309 	page_base = vm_pageout_page_count;
310 	ib = 1;
311 	is = 1;
312 
313 	/*
314 	 * Scan object for clusterable pages.
315 	 *
316 	 * We can cluster ONLY if: ->> the page is NOT
317 	 * clean, wired, busy, held, or mapped into a
318 	 * buffer, and one of the following:
319 	 * 1) The page is inactive, or a seldom used
320 	 *    active page.
321 	 * -or-
322 	 * 2) we force the issue.
323 	 *
324 	 * During heavy mmap/modification loads the pageout
325 	 * daemon can really fragment the underlying file
326 	 * due to flushing pages out of order and not trying
327 	 * align the clusters (which leave sporatic out-of-order
328 	 * holes).  To solve this problem we do the reverse scan
329 	 * first and attempt to align our cluster, then do a
330 	 * forward scan if room remains.
331 	 */
332 	object = m->object;
333 more:
334 	while (ib && pageout_count < vm_pageout_page_count) {
335 		vm_page_t p;
336 
337 		if (ib > pindex) {
338 			ib = 0;
339 			break;
340 		}
341 
342 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
343 			ib = 0;
344 			break;
345 		}
346 		if ((p->oflags & VPO_BUSY) || p->busy) {
347 			ib = 0;
348 			break;
349 		}
350 		vm_page_test_dirty(p);
351 		if (p->dirty == 0 ||
352 		    p->queue != PQ_INACTIVE ||
353 		    p->hold_count != 0) {	/* may be undergoing I/O */
354 			ib = 0;
355 			break;
356 		}
357 		mc[--page_base] = p;
358 		++pageout_count;
359 		++ib;
360 		/*
361 		 * alignment boundry, stop here and switch directions.  Do
362 		 * not clear ib.
363 		 */
364 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
365 			break;
366 	}
367 
368 	while (pageout_count < vm_pageout_page_count &&
369 	    pindex + is < object->size) {
370 		vm_page_t p;
371 
372 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
373 			break;
374 		if ((p->oflags & VPO_BUSY) || p->busy) {
375 			break;
376 		}
377 		vm_page_test_dirty(p);
378 		if (p->dirty == 0 ||
379 		    p->queue != PQ_INACTIVE ||
380 		    p->hold_count != 0) {	/* may be undergoing I/O */
381 			break;
382 		}
383 		mc[page_base + pageout_count] = p;
384 		++pageout_count;
385 		++is;
386 	}
387 
388 	/*
389 	 * If we exhausted our forward scan, continue with the reverse scan
390 	 * when possible, even past a page boundry.  This catches boundry
391 	 * conditions.
392 	 */
393 	if (ib && pageout_count < vm_pageout_page_count)
394 		goto more;
395 
396 	/*
397 	 * we allow reads during pageouts...
398 	 */
399 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
400 }
401 
402 /*
403  * vm_pageout_flush() - launder the given pages
404  *
405  *	The given pages are laundered.  Note that we setup for the start of
406  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
407  *	reference count all in here rather then in the parent.  If we want
408  *	the parent to do more sophisticated things we may have to change
409  *	the ordering.
410  */
411 int
412 vm_pageout_flush(vm_page_t *mc, int count, int flags)
413 {
414 	vm_object_t object = mc[0]->object;
415 	int pageout_status[count];
416 	int numpagedout = 0;
417 	int i;
418 
419 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
420 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
421 	/*
422 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
423 	 * mark the pages read-only.
424 	 *
425 	 * We do not have to fixup the clean/dirty bits here... we can
426 	 * allow the pager to do it after the I/O completes.
427 	 *
428 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
429 	 * edge case with file fragments.
430 	 */
431 	for (i = 0; i < count; i++) {
432 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
433 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
434 			mc[i], i, count));
435 		vm_page_io_start(mc[i]);
436 		pmap_remove_write(mc[i]);
437 	}
438 	vm_page_unlock_queues();
439 	vm_object_pip_add(object, count);
440 
441 	vm_pager_put_pages(object, mc, count, flags, 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(pageout_status[i] == VM_PAGER_PEND ||
448 		    (mt->flags & PG_WRITEABLE) == 0,
449 		    ("vm_pageout_flush: page %p is not write protected", mt));
450 		switch (pageout_status[i]) {
451 		case VM_PAGER_OK:
452 		case VM_PAGER_PEND:
453 			numpagedout++;
454 			break;
455 		case VM_PAGER_BAD:
456 			/*
457 			 * Page outside of range of object. Right now we
458 			 * essentially lose the changes by pretending it
459 			 * worked.
460 			 */
461 			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 ||
515 	    first_object->type == OBJT_SG ||
516 	    first_object->type == OBJT_PHYS)
517 		return;
518 	for (object = first_object;; object = backing_object) {
519 		if (pmap_resident_count(pmap) <= desired)
520 			goto unlock_return;
521 		if (object->paging_in_progress)
522 			goto unlock_return;
523 
524 		remove_mode = 0;
525 		if (object->shadow_count > 1)
526 			remove_mode = 1;
527 		/*
528 		 * scan the objects entire memory queue
529 		 */
530 		rcount = object->resident_page_count;
531 		p = TAILQ_FIRST(&object->memq);
532 		vm_page_lock_queues();
533 		while (p && (rcount-- > 0)) {
534 			if (pmap_resident_count(pmap) <= desired) {
535 				vm_page_unlock_queues();
536 				goto unlock_return;
537 			}
538 			next = TAILQ_NEXT(p, listq);
539 			cnt.v_pdpages++;
540 			if (p->wire_count != 0 ||
541 			    p->hold_count != 0 ||
542 			    p->busy != 0 ||
543 			    (p->oflags & VPO_BUSY) ||
544 			    (p->flags & PG_UNMANAGED) ||
545 			    !pmap_page_exists_quick(pmap, p)) {
546 				p = next;
547 				continue;
548 			}
549 			actcount = pmap_ts_referenced(p);
550 			if (actcount) {
551 				vm_page_flag_set(p, PG_REFERENCED);
552 			} else if (p->flags & PG_REFERENCED) {
553 				actcount = 1;
554 			}
555 			if ((p->queue != PQ_ACTIVE) &&
556 				(p->flags & PG_REFERENCED)) {
557 				vm_page_activate(p);
558 				p->act_count += actcount;
559 				vm_page_flag_clear(p, PG_REFERENCED);
560 			} else if (p->queue == PQ_ACTIVE) {
561 				if ((p->flags & PG_REFERENCED) == 0) {
562 					p->act_count -= min(p->act_count, ACT_DECLINE);
563 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
564 						pmap_remove_all(p);
565 						vm_page_deactivate(p);
566 					} else {
567 						vm_page_requeue(p);
568 					}
569 				} else {
570 					vm_page_activate(p);
571 					vm_page_flag_clear(p, PG_REFERENCED);
572 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
573 						p->act_count += ACT_ADVANCE;
574 					vm_page_requeue(p);
575 				}
576 			} else if (p->queue == PQ_INACTIVE) {
577 				pmap_remove_all(p);
578 			}
579 			p = next;
580 		}
581 		vm_page_unlock_queues();
582 		if ((backing_object = object->backing_object) == NULL)
583 			goto unlock_return;
584 		VM_OBJECT_LOCK(backing_object);
585 		if (object != first_object)
586 			VM_OBJECT_UNLOCK(object);
587 	}
588 unlock_return:
589 	if (object != first_object)
590 		VM_OBJECT_UNLOCK(object);
591 }
592 
593 /*
594  * deactivate some number of pages in a map, try to do it fairly, but
595  * that is really hard to do.
596  */
597 static void
598 vm_pageout_map_deactivate_pages(map, desired)
599 	vm_map_t map;
600 	long desired;
601 {
602 	vm_map_entry_t tmpe;
603 	vm_object_t obj, bigobj;
604 	int nothingwired;
605 
606 	if (!vm_map_trylock(map))
607 		return;
608 
609 	bigobj = NULL;
610 	nothingwired = TRUE;
611 
612 	/*
613 	 * first, search out the biggest object, and try to free pages from
614 	 * that.
615 	 */
616 	tmpe = map->header.next;
617 	while (tmpe != &map->header) {
618 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
619 			obj = tmpe->object.vm_object;
620 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
621 				if (obj->shadow_count <= 1 &&
622 				    (bigobj == NULL ||
623 				     bigobj->resident_page_count < obj->resident_page_count)) {
624 					if (bigobj != NULL)
625 						VM_OBJECT_UNLOCK(bigobj);
626 					bigobj = obj;
627 				} else
628 					VM_OBJECT_UNLOCK(obj);
629 			}
630 		}
631 		if (tmpe->wired_count > 0)
632 			nothingwired = FALSE;
633 		tmpe = tmpe->next;
634 	}
635 
636 	if (bigobj != NULL) {
637 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
638 		VM_OBJECT_UNLOCK(bigobj);
639 	}
640 	/*
641 	 * Next, hunt around for other pages to deactivate.  We actually
642 	 * do this search sort of wrong -- .text first is not the best idea.
643 	 */
644 	tmpe = map->header.next;
645 	while (tmpe != &map->header) {
646 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
647 			break;
648 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
649 			obj = tmpe->object.vm_object;
650 			if (obj != NULL) {
651 				VM_OBJECT_LOCK(obj);
652 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
653 				VM_OBJECT_UNLOCK(obj);
654 			}
655 		}
656 		tmpe = tmpe->next;
657 	}
658 
659 	/*
660 	 * Remove all mappings if a process is swapped out, this will free page
661 	 * table pages.
662 	 */
663 	if (desired == 0 && nothingwired) {
664 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
665 		    vm_map_max(map));
666 	}
667 	vm_map_unlock(map);
668 }
669 #endif		/* !defined(NO_SWAPPING) */
670 
671 /*
672  *	vm_pageout_scan does the dirty work for the pageout daemon.
673  */
674 static void
675 vm_pageout_scan(int pass)
676 {
677 	vm_page_t m, next;
678 	struct vm_page marker;
679 	int page_shortage, maxscan, pcount;
680 	int addl_page_shortage, addl_page_shortage_init;
681 	vm_object_t object;
682 	int actcount;
683 	int vnodes_skipped = 0;
684 	int maxlaunder;
685 
686 	/*
687 	 * Decrease registered cache sizes.
688 	 */
689 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
690 	/*
691 	 * We do this explicitly after the caches have been drained above.
692 	 */
693 	uma_reclaim();
694 
695 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
696 
697 	/*
698 	 * Calculate the number of pages we want to either free or move
699 	 * to the cache.
700 	 */
701 	page_shortage = vm_paging_target() + addl_page_shortage_init;
702 
703 	/*
704 	 * Initialize our marker
705 	 */
706 	bzero(&marker, sizeof(marker));
707 	marker.flags = PG_FICTITIOUS | PG_MARKER;
708 	marker.oflags = VPO_BUSY;
709 	marker.queue = PQ_INACTIVE;
710 	marker.wire_count = 1;
711 
712 	/*
713 	 * Start scanning the inactive queue for pages we can move to the
714 	 * cache or free.  The scan will stop when the target is reached or
715 	 * we have scanned the entire inactive queue.  Note that m->act_count
716 	 * is not used to form decisions for the inactive queue, only for the
717 	 * active queue.
718 	 *
719 	 * maxlaunder limits the number of dirty pages we flush per scan.
720 	 * For most systems a smaller value (16 or 32) is more robust under
721 	 * extreme memory and disk pressure because any unnecessary writes
722 	 * to disk can result in extreme performance degredation.  However,
723 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
724 	 * used) will die horribly with limited laundering.  If the pageout
725 	 * daemon cannot clean enough pages in the first pass, we let it go
726 	 * all out in succeeding passes.
727 	 */
728 	if ((maxlaunder = vm_max_launder) <= 1)
729 		maxlaunder = 1;
730 	if (pass)
731 		maxlaunder = 10000;
732 	vm_page_lock_queues();
733 rescan0:
734 	addl_page_shortage = addl_page_shortage_init;
735 	maxscan = cnt.v_inactive_count;
736 
737 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
738 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
739 	     m = next) {
740 
741 		cnt.v_pdpages++;
742 
743 		if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
744 			goto rescan0;
745 		}
746 
747 		next = TAILQ_NEXT(m, pageq);
748 		object = m->object;
749 
750 		/*
751 		 * skip marker pages
752 		 */
753 		if (m->flags & PG_MARKER)
754 			continue;
755 
756 		/*
757 		 * A held page may be undergoing I/O, so skip it.
758 		 */
759 		if (m->hold_count) {
760 			vm_page_requeue(m);
761 			addl_page_shortage++;
762 			continue;
763 		}
764 		/*
765 		 * Don't mess with busy pages, keep in the front of the
766 		 * queue, most likely are being paged out.
767 		 */
768 		if (!VM_OBJECT_TRYLOCK(object) &&
769 		    (!vm_pageout_fallback_object_lock(m, &next) ||
770 		     m->hold_count != 0)) {
771 			VM_OBJECT_UNLOCK(object);
772 			addl_page_shortage++;
773 			continue;
774 		}
775 		if (m->busy || (m->oflags & VPO_BUSY)) {
776 			VM_OBJECT_UNLOCK(object);
777 			addl_page_shortage++;
778 			continue;
779 		}
780 
781 		/*
782 		 * If the object is not being used, we ignore previous
783 		 * references.
784 		 */
785 		if (object->ref_count == 0) {
786 			vm_page_flag_clear(m, PG_REFERENCED);
787 			KASSERT(!pmap_page_is_mapped(m),
788 			    ("vm_pageout_scan: page %p is mapped", 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 does not believe that the page
824 		 * is fully dirty, but it is mapped for write access, then we
825 		 * consult the pmap to see if the page's dirty status should
826 		 * be updated.
827 		 */
828 		if (m->dirty != VM_PAGE_BITS_ALL &&
829 		    (m->flags & PG_WRITEABLE) != 0) {
830 			/*
831 			 * Avoid a race condition: Unless write access is
832 			 * removed from the page, another processor could
833 			 * modify it before all access is removed by the call
834 			 * to vm_page_cache() below.  If vm_page_cache() finds
835 			 * that the page has been modified when it removes all
836 			 * access, it panics because it cannot cache dirty
837 			 * pages.  In principle, we could eliminate just write
838 			 * access here rather than all access.  In the expected
839 			 * case, when there are no last instant modifications
840 			 * to the page, removing all access will be cheaper
841 			 * overall.
842 			 */
843 			if (pmap_is_modified(m))
844 				vm_page_dirty(m);
845 			else if (m->dirty == 0)
846 				pmap_remove_all(m);
847 		}
848 
849 		if (m->valid == 0) {
850 			/*
851 			 * Invalid pages can be easily freed
852 			 */
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_page_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, vfslocked = 0;
887 			struct vnode *vp = NULL;
888 			struct mount *mp = NULL;
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_page_requeue(m);
906 				continue;
907 			}
908 
909 			/*
910 			 * Following operations may unlock
911 			 * vm_page_queue_mtx, invalidating the 'next'
912 			 * pointer.  To prevent an inordinate number
913 			 * of restarts we use our marker to remember
914 			 * our place.
915 			 *
916 			 */
917 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
918 					   m, &marker, pageq);
919 			/*
920 			 * The object is already known NOT to be dead.   It
921 			 * is possible for the vget() to block the whole
922 			 * pageout daemon, but the new low-memory handling
923 			 * code should prevent it.
924 			 *
925 			 * The previous code skipped locked vnodes and, worse,
926 			 * reordered pages in the queue.  This results in
927 			 * completely non-deterministic operation and, on a
928 			 * busy system, can lead to extremely non-optimal
929 			 * pageouts.  For example, it can cause clean pages
930 			 * to be freed and dirty pages to be moved to the end
931 			 * of the queue.  Since dirty pages are also moved to
932 			 * the end of the queue once-cleaned, this gives
933 			 * way too large a weighting to defering the freeing
934 			 * of dirty pages.
935 			 *
936 			 * We can't wait forever for the vnode lock, we might
937 			 * deadlock due to a vn_read() getting stuck in
938 			 * vm_wait while holding this vnode.  We skip the
939 			 * vnode if we can't get it in a reasonable amount
940 			 * of time.
941 			 */
942 			if (object->type == OBJT_VNODE) {
943 				vp = object->handle;
944 				if (vp->v_type == VREG &&
945 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
946 					mp = NULL;
947 					++pageout_lock_miss;
948 					if (object->flags & OBJ_MIGHTBEDIRTY)
949 						vnodes_skipped++;
950 					goto unlock_and_continue;
951 				}
952 				KASSERT(mp != NULL,
953 				    ("vp %p with NULL v_mount", vp));
954 				vm_page_unlock_queues();
955 				vm_object_reference_locked(object);
956 				VM_OBJECT_UNLOCK(object);
957 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
958 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
959 				    curthread)) {
960 					VM_OBJECT_LOCK(object);
961 					vm_page_lock_queues();
962 					++pageout_lock_miss;
963 					if (object->flags & OBJ_MIGHTBEDIRTY)
964 						vnodes_skipped++;
965 					vp = NULL;
966 					goto unlock_and_continue;
967 				}
968 				VM_OBJECT_LOCK(object);
969 				vm_page_lock_queues();
970 				/*
971 				 * The page might have been moved to another
972 				 * queue during potential blocking in vget()
973 				 * above.  The page might have been freed and
974 				 * reused for another vnode.
975 				 */
976 				if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
977 				    m->object != object ||
978 				    TAILQ_NEXT(m, pageq) != &marker) {
979 					if (object->flags & OBJ_MIGHTBEDIRTY)
980 						vnodes_skipped++;
981 					goto unlock_and_continue;
982 				}
983 
984 				/*
985 				 * The page may have been busied during the
986 				 * blocking in vget().  We don't move the
987 				 * page back onto the end of the queue so that
988 				 * statistics are more correct if we don't.
989 				 */
990 				if (m->busy || (m->oflags & VPO_BUSY)) {
991 					goto unlock_and_continue;
992 				}
993 
994 				/*
995 				 * If the page has become held it might
996 				 * be undergoing I/O, so skip it
997 				 */
998 				if (m->hold_count) {
999 					vm_page_requeue(m);
1000 					if (object->flags & OBJ_MIGHTBEDIRTY)
1001 						vnodes_skipped++;
1002 					goto unlock_and_continue;
1003 				}
1004 			}
1005 
1006 			/*
1007 			 * If a page is dirty, then it is either being washed
1008 			 * (but not yet cleaned) or it is still in the
1009 			 * laundry.  If it is still in the laundry, then we
1010 			 * start the cleaning operation.
1011 			 *
1012 			 * decrement page_shortage on success to account for
1013 			 * the (future) cleaned page.  Otherwise we could wind
1014 			 * up laundering or cleaning too many pages.
1015 			 */
1016 			if (vm_pageout_clean(m) != 0) {
1017 				--page_shortage;
1018 				--maxlaunder;
1019 			}
1020 unlock_and_continue:
1021 			VM_OBJECT_UNLOCK(object);
1022 			if (mp != NULL) {
1023 				vm_page_unlock_queues();
1024 				if (vp != NULL)
1025 					vput(vp);
1026 				VFS_UNLOCK_GIANT(vfslocked);
1027 				vm_object_deallocate(object);
1028 				vn_finished_write(mp);
1029 				vm_page_lock_queues();
1030 			}
1031 			next = TAILQ_NEXT(&marker, pageq);
1032 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1033 				     &marker, pageq);
1034 			continue;
1035 		}
1036 		VM_OBJECT_UNLOCK(object);
1037 	}
1038 
1039 	/*
1040 	 * Compute the number of pages we want to try to move from the
1041 	 * active queue to the inactive queue.
1042 	 */
1043 	page_shortage = vm_paging_target() +
1044 		cnt.v_inactive_target - cnt.v_inactive_count;
1045 	page_shortage += addl_page_shortage;
1046 
1047 	/*
1048 	 * Scan the active queue for things we can deactivate. We nominally
1049 	 * track the per-page activity counter and use it to locate
1050 	 * deactivation candidates.
1051 	 */
1052 	pcount = cnt.v_active_count;
1053 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1054 
1055 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1056 
1057 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1058 		    ("vm_pageout_scan: page %p isn't active", m));
1059 
1060 		next = TAILQ_NEXT(m, pageq);
1061 		object = m->object;
1062 		if ((m->flags & PG_MARKER) != 0) {
1063 			m = next;
1064 			continue;
1065 		}
1066 		if (!VM_OBJECT_TRYLOCK(object) &&
1067 		    !vm_pageout_fallback_object_lock(m, &next)) {
1068 			VM_OBJECT_UNLOCK(object);
1069 			m = next;
1070 			continue;
1071 		}
1072 
1073 		/*
1074 		 * Don't deactivate pages that are busy.
1075 		 */
1076 		if ((m->busy != 0) ||
1077 		    (m->oflags & VPO_BUSY) ||
1078 		    (m->hold_count != 0)) {
1079 			VM_OBJECT_UNLOCK(object);
1080 			vm_page_requeue(m);
1081 			m = next;
1082 			continue;
1083 		}
1084 
1085 		/*
1086 		 * The count for pagedaemon pages is done after checking the
1087 		 * page for eligibility...
1088 		 */
1089 		cnt.v_pdpages++;
1090 
1091 		/*
1092 		 * Check to see "how much" the page has been used.
1093 		 */
1094 		actcount = 0;
1095 		if (object->ref_count != 0) {
1096 			if (m->flags & PG_REFERENCED) {
1097 				actcount += 1;
1098 			}
1099 			actcount += pmap_ts_referenced(m);
1100 			if (actcount) {
1101 				m->act_count += ACT_ADVANCE + actcount;
1102 				if (m->act_count > ACT_MAX)
1103 					m->act_count = ACT_MAX;
1104 			}
1105 		}
1106 
1107 		/*
1108 		 * Since we have "tested" this bit, we need to clear it now.
1109 		 */
1110 		vm_page_flag_clear(m, PG_REFERENCED);
1111 
1112 		/*
1113 		 * Only if an object is currently being used, do we use the
1114 		 * page activation count stats.
1115 		 */
1116 		if (actcount && (object->ref_count != 0)) {
1117 			vm_page_requeue(m);
1118 		} else {
1119 			m->act_count -= min(m->act_count, ACT_DECLINE);
1120 			if (vm_pageout_algorithm ||
1121 			    object->ref_count == 0 ||
1122 			    m->act_count == 0) {
1123 				page_shortage--;
1124 				if (object->ref_count == 0) {
1125 					KASSERT(!pmap_page_is_mapped(m),
1126 				    ("vm_pageout_scan: page %p is mapped", m));
1127 					if (m->dirty == 0)
1128 						vm_page_cache(m);
1129 					else
1130 						vm_page_deactivate(m);
1131 				} else {
1132 					vm_page_deactivate(m);
1133 				}
1134 			} else {
1135 				vm_page_requeue(m);
1136 			}
1137 		}
1138 		VM_OBJECT_UNLOCK(object);
1139 		m = next;
1140 	}
1141 	vm_page_unlock_queues();
1142 #if !defined(NO_SWAPPING)
1143 	/*
1144 	 * Idle process swapout -- run once per second.
1145 	 */
1146 	if (vm_swap_idle_enabled) {
1147 		static long lsec;
1148 		if (time_second != lsec) {
1149 			vm_req_vmdaemon(VM_SWAP_IDLE);
1150 			lsec = time_second;
1151 		}
1152 	}
1153 #endif
1154 
1155 	/*
1156 	 * If we didn't get enough free pages, and we have skipped a vnode
1157 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1158 	 * if we did not get enough free pages.
1159 	 */
1160 	if (vm_paging_target() > 0) {
1161 		if (vnodes_skipped && vm_page_count_min())
1162 			(void) speedup_syncer();
1163 #if !defined(NO_SWAPPING)
1164 		if (vm_swap_enabled && vm_page_count_target())
1165 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1166 #endif
1167 	}
1168 
1169 	/*
1170 	 * If we are critically low on one of RAM or swap and low on
1171 	 * the other, kill the largest process.  However, we avoid
1172 	 * doing this on the first pass in order to give ourselves a
1173 	 * chance to flush out dirty vnode-backed pages and to allow
1174 	 * active pages to be moved to the inactive queue and reclaimed.
1175 	 */
1176 	if (pass != 0 &&
1177 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1178 	     (swap_pager_full && vm_paging_target() > 0)))
1179 		vm_pageout_oom(VM_OOM_MEM);
1180 }
1181 
1182 
1183 void
1184 vm_pageout_oom(int shortage)
1185 {
1186 	struct proc *p, *bigproc;
1187 	vm_offset_t size, bigsize;
1188 	struct thread *td;
1189 	struct vmspace *vm;
1190 
1191 	/*
1192 	 * We keep the process bigproc locked once we find it to keep anyone
1193 	 * from messing with it; however, there is a possibility of
1194 	 * deadlock if process B is bigproc and one of it's child processes
1195 	 * attempts to propagate a signal to B while we are waiting for A's
1196 	 * lock while walking this list.  To avoid this, we don't block on
1197 	 * the process lock but just skip a process if it is already locked.
1198 	 */
1199 	bigproc = NULL;
1200 	bigsize = 0;
1201 	sx_slock(&allproc_lock);
1202 	FOREACH_PROC_IN_SYSTEM(p) {
1203 		int breakout;
1204 
1205 		if (PROC_TRYLOCK(p) == 0)
1206 			continue;
1207 		/*
1208 		 * If this is a system, protected or killed process, skip it.
1209 		 */
1210 		if ((p->p_flag & (P_INEXEC | P_PROTECTED | P_SYSTEM)) ||
1211 		    (p->p_pid == 1) || P_KILLED(p) ||
1212 		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1213 			PROC_UNLOCK(p);
1214 			continue;
1215 		}
1216 		/*
1217 		 * If the process is in a non-running type state,
1218 		 * don't touch it.  Check all the threads individually.
1219 		 */
1220 		breakout = 0;
1221 		FOREACH_THREAD_IN_PROC(p, td) {
1222 			thread_lock(td);
1223 			if (!TD_ON_RUNQ(td) &&
1224 			    !TD_IS_RUNNING(td) &&
1225 			    !TD_IS_SLEEPING(td)) {
1226 				thread_unlock(td);
1227 				breakout = 1;
1228 				break;
1229 			}
1230 			thread_unlock(td);
1231 		}
1232 		if (breakout) {
1233 			PROC_UNLOCK(p);
1234 			continue;
1235 		}
1236 		/*
1237 		 * get the process size
1238 		 */
1239 		vm = vmspace_acquire_ref(p);
1240 		if (vm == NULL) {
1241 			PROC_UNLOCK(p);
1242 			continue;
1243 		}
1244 		if (!vm_map_trylock_read(&vm->vm_map)) {
1245 			vmspace_free(vm);
1246 			PROC_UNLOCK(p);
1247 			continue;
1248 		}
1249 		size = vmspace_swap_count(vm);
1250 		vm_map_unlock_read(&vm->vm_map);
1251 		if (shortage == VM_OOM_MEM)
1252 			size += vmspace_resident_count(vm);
1253 		vmspace_free(vm);
1254 		/*
1255 		 * if the this process is bigger than the biggest one
1256 		 * remember it.
1257 		 */
1258 		if (size > bigsize) {
1259 			if (bigproc != NULL)
1260 				PROC_UNLOCK(bigproc);
1261 			bigproc = p;
1262 			bigsize = size;
1263 		} else
1264 			PROC_UNLOCK(p);
1265 	}
1266 	sx_sunlock(&allproc_lock);
1267 	if (bigproc != NULL) {
1268 		killproc(bigproc, "out of swap space");
1269 		sched_nice(bigproc, PRIO_MIN);
1270 		PROC_UNLOCK(bigproc);
1271 		wakeup(&cnt.v_free_count);
1272 	}
1273 }
1274 
1275 /*
1276  * This routine tries to maintain the pseudo LRU active queue,
1277  * so that during long periods of time where there is no paging,
1278  * that some statistic accumulation still occurs.  This code
1279  * helps the situation where paging just starts to occur.
1280  */
1281 static void
1282 vm_pageout_page_stats()
1283 {
1284 	vm_object_t object;
1285 	vm_page_t m,next;
1286 	int pcount,tpcount;		/* Number of pages to check */
1287 	static int fullintervalcount = 0;
1288 	int page_shortage;
1289 
1290 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1291 	page_shortage =
1292 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1293 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1294 
1295 	if (page_shortage <= 0)
1296 		return;
1297 
1298 	pcount = cnt.v_active_count;
1299 	fullintervalcount += vm_pageout_stats_interval;
1300 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1301 		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1302 		    cnt.v_page_count;
1303 		if (pcount > tpcount)
1304 			pcount = tpcount;
1305 	} else {
1306 		fullintervalcount = 0;
1307 	}
1308 
1309 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1310 	while ((m != NULL) && (pcount-- > 0)) {
1311 		int actcount;
1312 
1313 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1314 		    ("vm_pageout_page_stats: page %p isn't active", m));
1315 
1316 		next = TAILQ_NEXT(m, pageq);
1317 		object = m->object;
1318 
1319 		if ((m->flags & PG_MARKER) != 0) {
1320 			m = next;
1321 			continue;
1322 		}
1323 		if (!VM_OBJECT_TRYLOCK(object) &&
1324 		    !vm_pageout_fallback_object_lock(m, &next)) {
1325 			VM_OBJECT_UNLOCK(object);
1326 			m = next;
1327 			continue;
1328 		}
1329 
1330 		/*
1331 		 * Don't deactivate pages that are busy.
1332 		 */
1333 		if ((m->busy != 0) ||
1334 		    (m->oflags & VPO_BUSY) ||
1335 		    (m->hold_count != 0)) {
1336 			VM_OBJECT_UNLOCK(object);
1337 			vm_page_requeue(m);
1338 			m = next;
1339 			continue;
1340 		}
1341 
1342 		actcount = 0;
1343 		if (m->flags & PG_REFERENCED) {
1344 			vm_page_flag_clear(m, PG_REFERENCED);
1345 			actcount += 1;
1346 		}
1347 
1348 		actcount += pmap_ts_referenced(m);
1349 		if (actcount) {
1350 			m->act_count += ACT_ADVANCE + actcount;
1351 			if (m->act_count > ACT_MAX)
1352 				m->act_count = ACT_MAX;
1353 			vm_page_requeue(m);
1354 		} else {
1355 			if (m->act_count == 0) {
1356 				/*
1357 				 * We turn off page access, so that we have
1358 				 * more accurate RSS stats.  We don't do this
1359 				 * in the normal page deactivation when the
1360 				 * system is loaded VM wise, because the
1361 				 * cost of the large number of page protect
1362 				 * operations would be higher than the value
1363 				 * of doing the operation.
1364 				 */
1365 				pmap_remove_all(m);
1366 				vm_page_deactivate(m);
1367 			} else {
1368 				m->act_count -= min(m->act_count, ACT_DECLINE);
1369 				vm_page_requeue(m);
1370 			}
1371 		}
1372 		VM_OBJECT_UNLOCK(object);
1373 		m = next;
1374 	}
1375 }
1376 
1377 /*
1378  *	vm_pageout is the high level pageout daemon.
1379  */
1380 static void
1381 vm_pageout()
1382 {
1383 	int error, pass;
1384 
1385 	/*
1386 	 * Initialize some paging parameters.
1387 	 */
1388 	cnt.v_interrupt_free_min = 2;
1389 	if (cnt.v_page_count < 2000)
1390 		vm_pageout_page_count = 8;
1391 
1392 	/*
1393 	 * v_free_reserved needs to include enough for the largest
1394 	 * swap pager structures plus enough for any pv_entry structs
1395 	 * when paging.
1396 	 */
1397 	if (cnt.v_page_count > 1024)
1398 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1399 	else
1400 		cnt.v_free_min = 4;
1401 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1402 	    cnt.v_interrupt_free_min;
1403 	cnt.v_free_reserved = vm_pageout_page_count +
1404 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1405 	cnt.v_free_severe = cnt.v_free_min / 2;
1406 	cnt.v_free_min += cnt.v_free_reserved;
1407 	cnt.v_free_severe += cnt.v_free_reserved;
1408 
1409 	/*
1410 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1411 	 * that these are more a measure of the VM cache queue hysteresis
1412 	 * then the VM free queue.  Specifically, v_free_target is the
1413 	 * high water mark (free+cache pages).
1414 	 *
1415 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1416 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1417 	 * be big enough to handle memory needs while the pageout daemon
1418 	 * is signalled and run to free more pages.
1419 	 */
1420 	if (cnt.v_free_count > 6144)
1421 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1422 	else
1423 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1424 
1425 	if (cnt.v_free_count > 2048) {
1426 		cnt.v_cache_min = cnt.v_free_target;
1427 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1428 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1429 	} else {
1430 		cnt.v_cache_min = 0;
1431 		cnt.v_cache_max = 0;
1432 		cnt.v_inactive_target = cnt.v_free_count / 4;
1433 	}
1434 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1435 		cnt.v_inactive_target = cnt.v_free_count / 3;
1436 
1437 	/* XXX does not really belong here */
1438 	if (vm_page_max_wired == 0)
1439 		vm_page_max_wired = cnt.v_free_count / 3;
1440 
1441 	if (vm_pageout_stats_max == 0)
1442 		vm_pageout_stats_max = cnt.v_free_target;
1443 
1444 	/*
1445 	 * Set interval in seconds for stats scan.
1446 	 */
1447 	if (vm_pageout_stats_interval == 0)
1448 		vm_pageout_stats_interval = 5;
1449 	if (vm_pageout_full_stats_interval == 0)
1450 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1451 
1452 	swap_pager_swap_init();
1453 	pass = 0;
1454 	/*
1455 	 * The pageout daemon is never done, so loop forever.
1456 	 */
1457 	while (TRUE) {
1458 		/*
1459 		 * If we have enough free memory, wakeup waiters.  Do
1460 		 * not clear vm_pages_needed until we reach our target,
1461 		 * otherwise we may be woken up over and over again and
1462 		 * waste a lot of cpu.
1463 		 */
1464 		mtx_lock(&vm_page_queue_free_mtx);
1465 		if (vm_pages_needed && !vm_page_count_min()) {
1466 			if (!vm_paging_needed())
1467 				vm_pages_needed = 0;
1468 			wakeup(&cnt.v_free_count);
1469 		}
1470 		if (vm_pages_needed) {
1471 			/*
1472 			 * Still not done, take a second pass without waiting
1473 			 * (unlimited dirty cleaning), otherwise sleep a bit
1474 			 * and try again.
1475 			 */
1476 			++pass;
1477 			if (pass > 1)
1478 				msleep(&vm_pages_needed,
1479 				    &vm_page_queue_free_mtx, PVM, "psleep",
1480 				    hz / 2);
1481 		} else {
1482 			/*
1483 			 * Good enough, sleep & handle stats.  Prime the pass
1484 			 * for the next run.
1485 			 */
1486 			if (pass > 1)
1487 				pass = 1;
1488 			else
1489 				pass = 0;
1490 			error = msleep(&vm_pages_needed,
1491 			    &vm_page_queue_free_mtx, PVM, "psleep",
1492 			    vm_pageout_stats_interval * hz);
1493 			if (error && !vm_pages_needed) {
1494 				mtx_unlock(&vm_page_queue_free_mtx);
1495 				pass = 0;
1496 				vm_page_lock_queues();
1497 				vm_pageout_page_stats();
1498 				vm_page_unlock_queues();
1499 				continue;
1500 			}
1501 		}
1502 		if (vm_pages_needed)
1503 			cnt.v_pdwakeups++;
1504 		mtx_unlock(&vm_page_queue_free_mtx);
1505 		vm_pageout_scan(pass);
1506 	}
1507 }
1508 
1509 /*
1510  * Unless the free page queue lock is held by the caller, this function
1511  * should be regarded as advisory.  Specifically, the caller should
1512  * not msleep() on &cnt.v_free_count following this function unless
1513  * the free page queue lock is held until the msleep() is performed.
1514  */
1515 void
1516 pagedaemon_wakeup()
1517 {
1518 
1519 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1520 		vm_pages_needed = 1;
1521 		wakeup(&vm_pages_needed);
1522 	}
1523 }
1524 
1525 #if !defined(NO_SWAPPING)
1526 static void
1527 vm_req_vmdaemon(int req)
1528 {
1529 	static int lastrun = 0;
1530 
1531 	mtx_lock(&vm_daemon_mtx);
1532 	vm_pageout_req_swapout |= req;
1533 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1534 		wakeup(&vm_daemon_needed);
1535 		lastrun = ticks;
1536 	}
1537 	mtx_unlock(&vm_daemon_mtx);
1538 }
1539 
1540 static void
1541 vm_daemon()
1542 {
1543 	struct rlimit rsslim;
1544 	struct proc *p;
1545 	struct thread *td;
1546 	struct vmspace *vm;
1547 	int breakout, swapout_flags;
1548 
1549 	while (TRUE) {
1550 		mtx_lock(&vm_daemon_mtx);
1551 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1552 		swapout_flags = vm_pageout_req_swapout;
1553 		vm_pageout_req_swapout = 0;
1554 		mtx_unlock(&vm_daemon_mtx);
1555 		if (swapout_flags)
1556 			swapout_procs(swapout_flags);
1557 
1558 		/*
1559 		 * scan the processes for exceeding their rlimits or if
1560 		 * process is swapped out -- deactivate pages
1561 		 */
1562 		sx_slock(&allproc_lock);
1563 		FOREACH_PROC_IN_SYSTEM(p) {
1564 			vm_pindex_t limit, size;
1565 
1566 			/*
1567 			 * if this is a system process or if we have already
1568 			 * looked at this process, skip it.
1569 			 */
1570 			PROC_LOCK(p);
1571 			if (p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1572 				PROC_UNLOCK(p);
1573 				continue;
1574 			}
1575 			/*
1576 			 * if the process is in a non-running type state,
1577 			 * don't touch it.
1578 			 */
1579 			breakout = 0;
1580 			FOREACH_THREAD_IN_PROC(p, td) {
1581 				thread_lock(td);
1582 				if (!TD_ON_RUNQ(td) &&
1583 				    !TD_IS_RUNNING(td) &&
1584 				    !TD_IS_SLEEPING(td)) {
1585 					thread_unlock(td);
1586 					breakout = 1;
1587 					break;
1588 				}
1589 				thread_unlock(td);
1590 			}
1591 			if (breakout) {
1592 				PROC_UNLOCK(p);
1593 				continue;
1594 			}
1595 			/*
1596 			 * get a limit
1597 			 */
1598 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1599 			limit = OFF_TO_IDX(
1600 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1601 
1602 			/*
1603 			 * let processes that are swapped out really be
1604 			 * swapped out set the limit to nothing (will force a
1605 			 * swap-out.)
1606 			 */
1607 			if ((p->p_flag & P_INMEM) == 0)
1608 				limit = 0;	/* XXX */
1609 			vm = vmspace_acquire_ref(p);
1610 			PROC_UNLOCK(p);
1611 			if (vm == NULL)
1612 				continue;
1613 
1614 			size = vmspace_resident_count(vm);
1615 			if (limit >= 0 && size >= limit) {
1616 				vm_pageout_map_deactivate_pages(
1617 				    &vm->vm_map, limit);
1618 			}
1619 			vmspace_free(vm);
1620 		}
1621 		sx_sunlock(&allproc_lock);
1622 	}
1623 }
1624 #endif			/* !defined(NO_SWAPPING) */
1625