xref: /freebsd/sys/vm/vm_pageout.c (revision 2546665afcaf0d53dc2c7058fee96354b3680f5a)
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  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
41  *
42  *
43  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44  * All rights reserved.
45  *
46  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 /*
70  *	The proverbial page-out daemon.
71  */
72 
73 #include <sys/cdefs.h>
74 __FBSDID("$FreeBSD$");
75 
76 #include "opt_vm.h"
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/eventhandler.h>
81 #include <sys/lock.h>
82 #include <sys/mutex.h>
83 #include <sys/proc.h>
84 #include <sys/kthread.h>
85 #include <sys/ktr.h>
86 #include <sys/resourcevar.h>
87 #include <sys/sched.h>
88 #include <sys/signalvar.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 #include <sys/sx.h>
92 #include <sys/sysctl.h>
93 
94 #include <vm/vm.h>
95 #include <vm/vm_param.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_pageout.h>
100 #include <vm/vm_pager.h>
101 #include <vm/swap_pager.h>
102 #include <vm/vm_extern.h>
103 #include <vm/uma.h>
104 
105 #include <machine/mutex.h>
106 
107 /*
108  * System initialization
109  */
110 
111 /* the kernel process "vm_pageout"*/
112 static void vm_pageout(void);
113 static int vm_pageout_clean(vm_page_t);
114 static void vm_pageout_pmap_collect(void);
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, &page_kp)
125 
126 #if !defined(NO_SWAPPING)
127 /* the kernel process "vm_daemon"*/
128 static void vm_daemon(void);
129 static struct	proc *vmproc;
130 
131 static struct kproc_desc vm_kp = {
132 	"vmdaemon",
133 	vm_daemon,
134 	&vmproc
135 };
136 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
137 #endif
138 
139 
140 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
141 int vm_pageout_deficit;		/* Estimated number of pages deficit */
142 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
143 
144 #if !defined(NO_SWAPPING)
145 static int vm_pageout_req_swapout;	/* XXX */
146 static int vm_daemon_needed;
147 #endif
148 static int vm_max_launder = 32;
149 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
150 static int vm_pageout_full_stats_interval = 0;
151 static int vm_pageout_algorithm=0;
152 static int defer_swap_pageouts=0;
153 static int disable_swap_pageouts=0;
154 
155 #if defined(NO_SWAPPING)
156 static int vm_swap_enabled=0;
157 static int vm_swap_idle_enabled=0;
158 #else
159 static int vm_swap_enabled=1;
160 static int vm_swap_idle_enabled=0;
161 #endif
162 
163 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
164 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
165 
166 SYSCTL_INT(_vm, OID_AUTO, max_launder,
167 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
168 
169 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
170 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
171 
172 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
173 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
174 
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
176 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
177 
178 #if defined(NO_SWAPPING)
179 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
180 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
181 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
182 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
183 #else
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 #endif
189 
190 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
191 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
192 
193 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
194 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
195 
196 static int pageout_lock_miss;
197 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
198 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
199 
200 #define VM_PAGEOUT_PAGE_COUNT 16
201 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
202 
203 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
204 
205 #if !defined(NO_SWAPPING)
206 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
207 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
208 static void vm_req_vmdaemon(void);
209 #endif
210 static void vm_pageout_page_stats(void);
211 
212 /*
213  * vm_pageout_clean:
214  *
215  * Clean the page and remove it from the laundry.
216  *
217  * We set the busy bit to cause potential page faults on this page to
218  * block.  Note the careful timing, however, the busy bit isn't set till
219  * late and we cannot do anything that will mess with the page.
220  */
221 static int
222 vm_pageout_clean(m)
223 	vm_page_t m;
224 {
225 	vm_object_t object;
226 	vm_page_t mc[2*vm_pageout_page_count];
227 	int pageout_count;
228 	int ib, is, page_base;
229 	vm_pindex_t pindex = m->pindex;
230 
231 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
232 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
233 
234 	/*
235 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
236 	 * with the new swapper, but we could have serious problems paging
237 	 * out other object types if there is insufficient memory.
238 	 *
239 	 * Unfortunately, checking free memory here is far too late, so the
240 	 * check has been moved up a procedural level.
241 	 */
242 
243 	/*
244 	 * Don't mess with the page if it's busy, held, or special
245 	 */
246 	if ((m->hold_count != 0) ||
247 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
248 		return 0;
249 	}
250 
251 	mc[vm_pageout_page_count] = m;
252 	pageout_count = 1;
253 	page_base = vm_pageout_page_count;
254 	ib = 1;
255 	is = 1;
256 
257 	/*
258 	 * Scan object for clusterable pages.
259 	 *
260 	 * We can cluster ONLY if: ->> the page is NOT
261 	 * clean, wired, busy, held, or mapped into a
262 	 * buffer, and one of the following:
263 	 * 1) The page is inactive, or a seldom used
264 	 *    active page.
265 	 * -or-
266 	 * 2) we force the issue.
267 	 *
268 	 * During heavy mmap/modification loads the pageout
269 	 * daemon can really fragment the underlying file
270 	 * due to flushing pages out of order and not trying
271 	 * align the clusters (which leave sporatic out-of-order
272 	 * holes).  To solve this problem we do the reverse scan
273 	 * first and attempt to align our cluster, then do a
274 	 * forward scan if room remains.
275 	 */
276 	object = m->object;
277 more:
278 	while (ib && pageout_count < vm_pageout_page_count) {
279 		vm_page_t p;
280 
281 		if (ib > pindex) {
282 			ib = 0;
283 			break;
284 		}
285 
286 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
287 			ib = 0;
288 			break;
289 		}
290 		if (((p->queue - p->pc) == PQ_CACHE) ||
291 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
292 			ib = 0;
293 			break;
294 		}
295 		vm_page_test_dirty(p);
296 		if ((p->dirty & p->valid) == 0 ||
297 		    p->queue != PQ_INACTIVE ||
298 		    p->wire_count != 0 ||	/* may be held by buf cache */
299 		    p->hold_count != 0) {	/* may be undergoing I/O */
300 			ib = 0;
301 			break;
302 		}
303 		mc[--page_base] = p;
304 		++pageout_count;
305 		++ib;
306 		/*
307 		 * alignment boundry, stop here and switch directions.  Do
308 		 * not clear ib.
309 		 */
310 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
311 			break;
312 	}
313 
314 	while (pageout_count < vm_pageout_page_count &&
315 	    pindex + is < object->size) {
316 		vm_page_t p;
317 
318 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
319 			break;
320 		if (((p->queue - p->pc) == PQ_CACHE) ||
321 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
322 			break;
323 		}
324 		vm_page_test_dirty(p);
325 		if ((p->dirty & p->valid) == 0 ||
326 		    p->queue != PQ_INACTIVE ||
327 		    p->wire_count != 0 ||	/* may be held by buf cache */
328 		    p->hold_count != 0) {	/* may be undergoing I/O */
329 			break;
330 		}
331 		mc[page_base + pageout_count] = p;
332 		++pageout_count;
333 		++is;
334 	}
335 
336 	/*
337 	 * If we exhausted our forward scan, continue with the reverse scan
338 	 * when possible, even past a page boundry.  This catches boundry
339 	 * conditions.
340 	 */
341 	if (ib && pageout_count < vm_pageout_page_count)
342 		goto more;
343 
344 	/*
345 	 * we allow reads during pageouts...
346 	 */
347 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
348 }
349 
350 /*
351  * vm_pageout_flush() - launder the given pages
352  *
353  *	The given pages are laundered.  Note that we setup for the start of
354  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
355  *	reference count all in here rather then in the parent.  If we want
356  *	the parent to do more sophisticated things we may have to change
357  *	the ordering.
358  */
359 int
360 vm_pageout_flush(vm_page_t *mc, int count, int flags)
361 {
362 	vm_object_t object = mc[0]->object;
363 	int pageout_status[count];
364 	int numpagedout = 0;
365 	int i;
366 
367 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
368 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
369 	/*
370 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
371 	 * mark the pages read-only.
372 	 *
373 	 * We do not have to fixup the clean/dirty bits here... we can
374 	 * allow the pager to do it after the I/O completes.
375 	 *
376 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
377 	 * edge case with file fragments.
378 	 */
379 	for (i = 0; i < count; i++) {
380 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
381 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
382 			mc[i], i, count));
383 		vm_page_io_start(mc[i]);
384 		pmap_page_protect(mc[i], VM_PROT_READ);
385 	}
386 	vm_page_unlock_queues();
387 	vm_object_pip_add(object, count);
388 
389 	vm_pager_put_pages(object, mc, count,
390 	    (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
391 	    pageout_status);
392 
393 	vm_page_lock_queues();
394 	for (i = 0; i < count; i++) {
395 		vm_page_t mt = mc[i];
396 
397 		KASSERT((mt->flags & PG_WRITEABLE) == 0,
398 		    ("vm_pageout_flush: page %p is not write protected", mt));
399 		switch (pageout_status[i]) {
400 		case VM_PAGER_OK:
401 		case VM_PAGER_PEND:
402 			numpagedout++;
403 			break;
404 		case VM_PAGER_BAD:
405 			/*
406 			 * Page outside of range of object. Right now we
407 			 * essentially lose the changes by pretending it
408 			 * worked.
409 			 */
410 			pmap_clear_modify(mt);
411 			vm_page_undirty(mt);
412 			break;
413 		case VM_PAGER_ERROR:
414 		case VM_PAGER_FAIL:
415 			/*
416 			 * If page couldn't be paged out, then reactivate the
417 			 * page so it doesn't clog the inactive list.  (We
418 			 * will try paging out it again later).
419 			 */
420 			vm_page_activate(mt);
421 			break;
422 		case VM_PAGER_AGAIN:
423 			break;
424 		}
425 
426 		/*
427 		 * If the operation is still going, leave the page busy to
428 		 * block all other accesses. Also, leave the paging in
429 		 * progress indicator set so that we don't attempt an object
430 		 * collapse.
431 		 */
432 		if (pageout_status[i] != VM_PAGER_PEND) {
433 			vm_object_pip_wakeup(object);
434 			vm_page_io_finish(mt);
435 			if (vm_page_count_severe())
436 				vm_page_try_to_cache(mt);
437 		}
438 	}
439 	return numpagedout;
440 }
441 
442 #if !defined(NO_SWAPPING)
443 /*
444  *	vm_pageout_object_deactivate_pages
445  *
446  *	deactivate enough pages to satisfy the inactive target
447  *	requirements or if vm_page_proc_limit is set, then
448  *	deactivate all of the pages in the object and its
449  *	backing_objects.
450  *
451  *	The object and map must be locked.
452  */
453 static void
454 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
455 	pmap_t pmap;
456 	vm_object_t first_object;
457 	long desired;
458 {
459 	vm_object_t backing_object, object;
460 	vm_page_t p, next;
461 	int actcount, rcount, remove_mode;
462 
463 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
464 	if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
465 		return;
466 	for (object = first_object;; object = backing_object) {
467 		if (pmap_resident_count(pmap) <= desired)
468 			goto unlock_return;
469 		if (object->paging_in_progress)
470 			goto unlock_return;
471 
472 		remove_mode = 0;
473 		if (object->shadow_count > 1)
474 			remove_mode = 1;
475 		/*
476 		 * scan the objects entire memory queue
477 		 */
478 		rcount = object->resident_page_count;
479 		p = TAILQ_FIRST(&object->memq);
480 		vm_page_lock_queues();
481 		while (p && (rcount-- > 0)) {
482 			if (pmap_resident_count(pmap) <= desired) {
483 				vm_page_unlock_queues();
484 				goto unlock_return;
485 			}
486 			next = TAILQ_NEXT(p, listq);
487 			cnt.v_pdpages++;
488 			if (p->wire_count != 0 ||
489 			    p->hold_count != 0 ||
490 			    p->busy != 0 ||
491 			    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
492 			    !pmap_page_exists_quick(pmap, p)) {
493 				p = next;
494 				continue;
495 			}
496 			actcount = pmap_ts_referenced(p);
497 			if (actcount) {
498 				vm_page_flag_set(p, PG_REFERENCED);
499 			} else if (p->flags & PG_REFERENCED) {
500 				actcount = 1;
501 			}
502 			if ((p->queue != PQ_ACTIVE) &&
503 				(p->flags & PG_REFERENCED)) {
504 				vm_page_activate(p);
505 				p->act_count += actcount;
506 				vm_page_flag_clear(p, PG_REFERENCED);
507 			} else if (p->queue == PQ_ACTIVE) {
508 				if ((p->flags & PG_REFERENCED) == 0) {
509 					p->act_count -= min(p->act_count, ACT_DECLINE);
510 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
511 						pmap_remove_all(p);
512 						vm_page_deactivate(p);
513 					} else {
514 						vm_pageq_requeue(p);
515 					}
516 				} else {
517 					vm_page_activate(p);
518 					vm_page_flag_clear(p, PG_REFERENCED);
519 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
520 						p->act_count += ACT_ADVANCE;
521 					vm_pageq_requeue(p);
522 				}
523 			} else if (p->queue == PQ_INACTIVE) {
524 				pmap_remove_all(p);
525 			}
526 			p = next;
527 		}
528 		vm_page_unlock_queues();
529 		if ((backing_object = object->backing_object) == NULL)
530 			goto unlock_return;
531 		VM_OBJECT_LOCK(backing_object);
532 		if (object != first_object)
533 			VM_OBJECT_UNLOCK(object);
534 	}
535 unlock_return:
536 	if (object != first_object)
537 		VM_OBJECT_UNLOCK(object);
538 }
539 
540 /*
541  * deactivate some number of pages in a map, try to do it fairly, but
542  * that is really hard to do.
543  */
544 static void
545 vm_pageout_map_deactivate_pages(map, desired)
546 	vm_map_t map;
547 	long desired;
548 {
549 	vm_map_entry_t tmpe;
550 	vm_object_t obj, bigobj;
551 	int nothingwired;
552 
553 	if (!vm_map_trylock(map))
554 		return;
555 
556 	bigobj = NULL;
557 	nothingwired = TRUE;
558 
559 	/*
560 	 * first, search out the biggest object, and try to free pages from
561 	 * that.
562 	 */
563 	tmpe = map->header.next;
564 	while (tmpe != &map->header) {
565 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
566 			obj = tmpe->object.vm_object;
567 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
568 				if (obj->shadow_count <= 1 &&
569 				    (bigobj == NULL ||
570 				     bigobj->resident_page_count < obj->resident_page_count)) {
571 					if (bigobj != NULL)
572 						VM_OBJECT_UNLOCK(bigobj);
573 					bigobj = obj;
574 				} else
575 					VM_OBJECT_UNLOCK(obj);
576 			}
577 		}
578 		if (tmpe->wired_count > 0)
579 			nothingwired = FALSE;
580 		tmpe = tmpe->next;
581 	}
582 
583 	if (bigobj != NULL) {
584 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
585 		VM_OBJECT_UNLOCK(bigobj);
586 	}
587 	/*
588 	 * Next, hunt around for other pages to deactivate.  We actually
589 	 * do this search sort of wrong -- .text first is not the best idea.
590 	 */
591 	tmpe = map->header.next;
592 	while (tmpe != &map->header) {
593 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
594 			break;
595 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
596 			obj = tmpe->object.vm_object;
597 			if (obj != NULL) {
598 				VM_OBJECT_LOCK(obj);
599 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
600 				VM_OBJECT_UNLOCK(obj);
601 			}
602 		}
603 		tmpe = tmpe->next;
604 	}
605 
606 	/*
607 	 * Remove all mappings if a process is swapped out, this will free page
608 	 * table pages.
609 	 */
610 	if (desired == 0 && nothingwired) {
611 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
612 		    vm_map_max(map));
613 	}
614 	vm_map_unlock(map);
615 }
616 #endif		/* !defined(NO_SWAPPING) */
617 
618 /*
619  * This routine is very drastic, but can save the system
620  * in a pinch.
621  */
622 static void
623 vm_pageout_pmap_collect(void)
624 {
625 	int i;
626 	vm_page_t m;
627 	static int warningdone;
628 
629 	if (pmap_pagedaemon_waken == 0)
630 		return;
631 	if (warningdone < 5) {
632 		printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
633 		warningdone++;
634 	}
635 	vm_page_lock_queues();
636 	for (i = 0; i < vm_page_array_size; i++) {
637 		m = &vm_page_array[i];
638 		if (m->wire_count || m->hold_count || m->busy ||
639 		    (m->flags & (PG_BUSY | PG_UNMANAGED)))
640 			continue;
641 		pmap_remove_all(m);
642 	}
643 	vm_page_unlock_queues();
644 	pmap_pagedaemon_waken = 0;
645 }
646 
647 /*
648  *	vm_pageout_scan does the dirty work for the pageout daemon.
649  */
650 static void
651 vm_pageout_scan(int pass)
652 {
653 	vm_page_t m, next;
654 	struct vm_page marker;
655 	int page_shortage, maxscan, pcount;
656 	int addl_page_shortage, addl_page_shortage_init;
657 	struct proc *p, *bigproc;
658 	struct thread *td;
659 	vm_offset_t size, bigsize;
660 	vm_object_t object;
661 	int actcount;
662 	int vnodes_skipped = 0;
663 	int maxlaunder;
664 
665 	mtx_lock(&Giant);
666 	/*
667 	 * Decrease registered cache sizes.
668 	 */
669 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
670 	/*
671 	 * We do this explicitly after the caches have been drained above.
672 	 */
673 	uma_reclaim();
674 	/*
675 	 * Do whatever cleanup that the pmap code can.
676 	 */
677 	vm_pageout_pmap_collect();
678 
679 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
680 
681 	/*
682 	 * Calculate the number of pages we want to either free or move
683 	 * to the cache.
684 	 */
685 	page_shortage = vm_paging_target() + addl_page_shortage_init;
686 
687 	/*
688 	 * Initialize our marker
689 	 */
690 	bzero(&marker, sizeof(marker));
691 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
692 	marker.queue = PQ_INACTIVE;
693 	marker.wire_count = 1;
694 
695 	/*
696 	 * Start scanning the inactive queue for pages we can move to the
697 	 * cache or free.  The scan will stop when the target is reached or
698 	 * we have scanned the entire inactive queue.  Note that m->act_count
699 	 * is not used to form decisions for the inactive queue, only for the
700 	 * active queue.
701 	 *
702 	 * maxlaunder limits the number of dirty pages we flush per scan.
703 	 * For most systems a smaller value (16 or 32) is more robust under
704 	 * extreme memory and disk pressure because any unnecessary writes
705 	 * to disk can result in extreme performance degredation.  However,
706 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
707 	 * used) will die horribly with limited laundering.  If the pageout
708 	 * daemon cannot clean enough pages in the first pass, we let it go
709 	 * all out in succeeding passes.
710 	 */
711 	if ((maxlaunder = vm_max_launder) <= 1)
712 		maxlaunder = 1;
713 	if (pass)
714 		maxlaunder = 10000;
715 	vm_page_lock_queues();
716 rescan0:
717 	addl_page_shortage = addl_page_shortage_init;
718 	maxscan = cnt.v_inactive_count;
719 
720 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
721 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
722 	     m = next) {
723 
724 		cnt.v_pdpages++;
725 
726 		if (m->queue != PQ_INACTIVE) {
727 			goto rescan0;
728 		}
729 
730 		next = TAILQ_NEXT(m, pageq);
731 
732 		/*
733 		 * skip marker pages
734 		 */
735 		if (m->flags & PG_MARKER)
736 			continue;
737 
738 		/*
739 		 * A held page may be undergoing I/O, so skip it.
740 		 */
741 		if (m->hold_count) {
742 			vm_pageq_requeue(m);
743 			addl_page_shortage++;
744 			continue;
745 		}
746 		/*
747 		 * Don't mess with busy pages, keep in the front of the
748 		 * queue, most likely are being paged out.
749 		 */
750 		if (m->busy || (m->flags & PG_BUSY)) {
751 			addl_page_shortage++;
752 			continue;
753 		}
754 
755 		/*
756 		 * If the object is not being used, we ignore previous
757 		 * references.
758 		 */
759 		if (m->object->ref_count == 0) {
760 			vm_page_flag_clear(m, PG_REFERENCED);
761 			pmap_clear_reference(m);
762 
763 		/*
764 		 * Otherwise, if the page has been referenced while in the
765 		 * inactive queue, we bump the "activation count" upwards,
766 		 * making it less likely that the page will be added back to
767 		 * the inactive queue prematurely again.  Here we check the
768 		 * page tables (or emulated bits, if any), given the upper
769 		 * level VM system not knowing anything about existing
770 		 * references.
771 		 */
772 		} else if (((m->flags & PG_REFERENCED) == 0) &&
773 			(actcount = pmap_ts_referenced(m))) {
774 			vm_page_activate(m);
775 			m->act_count += (actcount + ACT_ADVANCE);
776 			continue;
777 		}
778 
779 		/*
780 		 * If the upper level VM system knows about any page
781 		 * references, we activate the page.  We also set the
782 		 * "activation count" higher than normal so that we will less
783 		 * likely place pages back onto the inactive queue again.
784 		 */
785 		if ((m->flags & PG_REFERENCED) != 0) {
786 			vm_page_flag_clear(m, PG_REFERENCED);
787 			actcount = pmap_ts_referenced(m);
788 			vm_page_activate(m);
789 			m->act_count += (actcount + ACT_ADVANCE + 1);
790 			continue;
791 		}
792 
793 		/*
794 		 * If the upper level VM system doesn't know anything about
795 		 * the page being dirty, we have to check for it again.  As
796 		 * far as the VM code knows, any partially dirty pages are
797 		 * fully dirty.
798 		 */
799 		if (m->dirty == 0 && !pmap_is_modified(m)) {
800 			/*
801 			 * Avoid a race condition: Unless write access is
802 			 * removed from the page, another processor could
803 			 * modify it before all access is removed by the call
804 			 * to vm_page_cache() below.  If vm_page_cache() finds
805 			 * that the page has been modified when it removes all
806 			 * access, it panics because it cannot cache dirty
807 			 * pages.  In principle, we could eliminate just write
808 			 * access here rather than all access.  In the expected
809 			 * case, when there are no last instant modifications
810 			 * to the page, removing all access will be cheaper
811 			 * overall.
812 			 */
813 			if ((m->flags & PG_WRITEABLE) != 0)
814 				pmap_remove_all(m);
815 		} else {
816 			vm_page_dirty(m);
817 		}
818 
819 		object = m->object;
820 		if (!VM_OBJECT_TRYLOCK(object))
821 			continue;
822 		if (m->valid == 0) {
823 			/*
824 			 * Invalid pages can be easily freed
825 			 */
826 			vm_page_busy(m);
827 			pmap_remove_all(m);
828 			vm_page_free(m);
829 			cnt.v_dfree++;
830 			--page_shortage;
831 		} else if (m->dirty == 0) {
832 			/*
833 			 * Clean pages can be placed onto the cache queue.
834 			 * This effectively frees them.
835 			 */
836 			vm_page_cache(m);
837 			--page_shortage;
838 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
839 			/*
840 			 * Dirty pages need to be paged out, but flushing
841 			 * a page is extremely expensive verses freeing
842 			 * a clean page.  Rather then artificially limiting
843 			 * the number of pages we can flush, we instead give
844 			 * dirty pages extra priority on the inactive queue
845 			 * by forcing them to be cycled through the queue
846 			 * twice before being flushed, after which the
847 			 * (now clean) page will cycle through once more
848 			 * before being freed.  This significantly extends
849 			 * the thrash point for a heavily loaded machine.
850 			 */
851 			vm_page_flag_set(m, PG_WINATCFLS);
852 			vm_pageq_requeue(m);
853 		} else if (maxlaunder > 0) {
854 			/*
855 			 * We always want to try to flush some dirty pages if
856 			 * we encounter them, to keep the system stable.
857 			 * Normally this number is small, but under extreme
858 			 * pressure where there are insufficient clean pages
859 			 * on the inactive queue, we may have to go all out.
860 			 */
861 			int swap_pageouts_ok;
862 			struct vnode *vp = NULL;
863 			struct mount *mp;
864 
865 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
866 				swap_pageouts_ok = 1;
867 			} else {
868 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
869 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
870 				vm_page_count_min());
871 
872 			}
873 
874 			/*
875 			 * We don't bother paging objects that are "dead".
876 			 * Those objects are in a "rundown" state.
877 			 */
878 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
879 				VM_OBJECT_UNLOCK(object);
880 				vm_pageq_requeue(m);
881 				continue;
882 			}
883 
884 			/*
885 			 * The object is already known NOT to be dead.   It
886 			 * is possible for the vget() to block the whole
887 			 * pageout daemon, but the new low-memory handling
888 			 * code should prevent it.
889 			 *
890 			 * The previous code skipped locked vnodes and, worse,
891 			 * reordered pages in the queue.  This results in
892 			 * completely non-deterministic operation and, on a
893 			 * busy system, can lead to extremely non-optimal
894 			 * pageouts.  For example, it can cause clean pages
895 			 * to be freed and dirty pages to be moved to the end
896 			 * of the queue.  Since dirty pages are also moved to
897 			 * the end of the queue once-cleaned, this gives
898 			 * way too large a weighting to defering the freeing
899 			 * of dirty pages.
900 			 *
901 			 * We can't wait forever for the vnode lock, we might
902 			 * deadlock due to a vn_read() getting stuck in
903 			 * vm_wait while holding this vnode.  We skip the
904 			 * vnode if we can't get it in a reasonable amount
905 			 * of time.
906 			 */
907 			if (object->type == OBJT_VNODE) {
908 				vp = object->handle;
909 				mp = NULL;
910 				if (vp->v_type == VREG)
911 					vn_start_write(vp, &mp, V_NOWAIT);
912 				vm_page_unlock_queues();
913 				VI_LOCK(vp);
914 				VM_OBJECT_UNLOCK(object);
915 				if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK |
916 				    LK_TIMELOCK, curthread)) {
917 					VM_OBJECT_LOCK(object);
918 					vm_page_lock_queues();
919 					++pageout_lock_miss;
920 					vn_finished_write(mp);
921 					if (object->flags & OBJ_MIGHTBEDIRTY)
922 						vnodes_skipped++;
923 					VM_OBJECT_UNLOCK(object);
924 					continue;
925 				}
926 				VM_OBJECT_LOCK(object);
927 				vm_page_lock_queues();
928 				/*
929 				 * The page might have been moved to another
930 				 * queue during potential blocking in vget()
931 				 * above.  The page might have been freed and
932 				 * reused for another vnode.  The object might
933 				 * have been reused for another vnode.
934 				 */
935 				if (m->queue != PQ_INACTIVE ||
936 				    m->object != object ||
937 				    object->handle != vp) {
938 					if (object->flags & OBJ_MIGHTBEDIRTY)
939 						vnodes_skipped++;
940 					goto unlock_and_continue;
941 				}
942 
943 				/*
944 				 * The page may have been busied during the
945 				 * blocking in vput();  We don't move the
946 				 * page back onto the end of the queue so that
947 				 * statistics are more correct if we don't.
948 				 */
949 				if (m->busy || (m->flags & PG_BUSY)) {
950 					goto unlock_and_continue;
951 				}
952 
953 				/*
954 				 * If the page has become held it might
955 				 * be undergoing I/O, so skip it
956 				 */
957 				if (m->hold_count) {
958 					vm_pageq_requeue(m);
959 					if (object->flags & OBJ_MIGHTBEDIRTY)
960 						vnodes_skipped++;
961 					goto unlock_and_continue;
962 				}
963 			}
964 
965 			/*
966 			 * If a page is dirty, then it is either being washed
967 			 * (but not yet cleaned) or it is still in the
968 			 * laundry.  If it is still in the laundry, then we
969 			 * start the cleaning operation.
970 			 *
971 			 * This operation may cluster, invalidating the 'next'
972 			 * pointer.  To prevent an inordinate number of
973 			 * restarts we use our marker to remember our place.
974 			 *
975 			 * decrement page_shortage on success to account for
976 			 * the (future) cleaned page.  Otherwise we could wind
977 			 * up laundering or cleaning too many pages.
978 			 */
979 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
980 			if (vm_pageout_clean(m) != 0) {
981 				--page_shortage;
982 				--maxlaunder;
983 			}
984 			next = TAILQ_NEXT(&marker, pageq);
985 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
986 unlock_and_continue:
987 			VM_OBJECT_UNLOCK(object);
988 			if (vp) {
989 				vm_page_unlock_queues();
990 				vput(vp);
991 				vn_finished_write(mp);
992 				vm_page_lock_queues();
993 			}
994 			continue;
995 		}
996 		VM_OBJECT_UNLOCK(object);
997 	}
998 
999 	/*
1000 	 * Compute the number of pages we want to try to move from the
1001 	 * active queue to the inactive queue.
1002 	 */
1003 	page_shortage = vm_paging_target() +
1004 		cnt.v_inactive_target - cnt.v_inactive_count;
1005 	page_shortage += addl_page_shortage;
1006 
1007 	/*
1008 	 * Scan the active queue for things we can deactivate. We nominally
1009 	 * track the per-page activity counter and use it to locate
1010 	 * deactivation candidates.
1011 	 */
1012 	pcount = cnt.v_active_count;
1013 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1014 
1015 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1016 
1017 		KASSERT(m->queue == PQ_ACTIVE,
1018 		    ("vm_pageout_scan: page %p isn't active", m));
1019 
1020 		next = TAILQ_NEXT(m, pageq);
1021 		/*
1022 		 * Don't deactivate pages that are busy.
1023 		 */
1024 		if ((m->busy != 0) ||
1025 		    (m->flags & PG_BUSY) ||
1026 		    (m->hold_count != 0)) {
1027 			vm_pageq_requeue(m);
1028 			m = next;
1029 			continue;
1030 		}
1031 
1032 		/*
1033 		 * The count for pagedaemon pages is done after checking the
1034 		 * page for eligibility...
1035 		 */
1036 		cnt.v_pdpages++;
1037 
1038 		/*
1039 		 * Check to see "how much" the page has been used.
1040 		 */
1041 		actcount = 0;
1042 		if (m->object->ref_count != 0) {
1043 			if (m->flags & PG_REFERENCED) {
1044 				actcount += 1;
1045 			}
1046 			actcount += pmap_ts_referenced(m);
1047 			if (actcount) {
1048 				m->act_count += ACT_ADVANCE + actcount;
1049 				if (m->act_count > ACT_MAX)
1050 					m->act_count = ACT_MAX;
1051 			}
1052 		}
1053 
1054 		/*
1055 		 * Since we have "tested" this bit, we need to clear it now.
1056 		 */
1057 		vm_page_flag_clear(m, PG_REFERENCED);
1058 
1059 		/*
1060 		 * Only if an object is currently being used, do we use the
1061 		 * page activation count stats.
1062 		 */
1063 		if (actcount && (m->object->ref_count != 0)) {
1064 			vm_pageq_requeue(m);
1065 		} else {
1066 			m->act_count -= min(m->act_count, ACT_DECLINE);
1067 			if (vm_pageout_algorithm ||
1068 			    m->object->ref_count == 0 ||
1069 			    m->act_count == 0) {
1070 				page_shortage--;
1071 				if (m->object->ref_count == 0) {
1072 					pmap_remove_all(m);
1073 					if (m->dirty == 0)
1074 						vm_page_cache(m);
1075 					else
1076 						vm_page_deactivate(m);
1077 				} else {
1078 					vm_page_deactivate(m);
1079 				}
1080 			} else {
1081 				vm_pageq_requeue(m);
1082 			}
1083 		}
1084 		m = next;
1085 	}
1086 
1087 	/*
1088 	 * We try to maintain some *really* free pages, this allows interrupt
1089 	 * code to be guaranteed space.  Since both cache and free queues
1090 	 * are considered basically 'free', moving pages from cache to free
1091 	 * does not effect other calculations.
1092 	 */
1093 	while (cnt.v_free_count < cnt.v_free_reserved) {
1094 		static int cache_rover = 0;
1095 
1096 		if ((m = vm_page_select_cache(cache_rover)) == NULL)
1097 			break;
1098 		cache_rover = (m->pc + PQ_PRIME2) & PQ_L2_MASK;
1099 		object = m->object;
1100 		VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1101 		vm_page_busy(m);
1102 		vm_page_free(m);
1103 		VM_OBJECT_UNLOCK(object);
1104 		cnt.v_dfree++;
1105 	}
1106 	vm_page_unlock_queues();
1107 #if !defined(NO_SWAPPING)
1108 	/*
1109 	 * Idle process swapout -- run once per second.
1110 	 */
1111 	if (vm_swap_idle_enabled) {
1112 		static long lsec;
1113 		if (time_second != lsec) {
1114 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1115 			vm_req_vmdaemon();
1116 			lsec = time_second;
1117 		}
1118 	}
1119 #endif
1120 
1121 	/*
1122 	 * If we didn't get enough free pages, and we have skipped a vnode
1123 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1124 	 * if we did not get enough free pages.
1125 	 */
1126 	if (vm_paging_target() > 0) {
1127 		if (vnodes_skipped && vm_page_count_min())
1128 			(void) speedup_syncer();
1129 #if !defined(NO_SWAPPING)
1130 		if (vm_swap_enabled && vm_page_count_target()) {
1131 			vm_req_vmdaemon();
1132 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1133 		}
1134 #endif
1135 	}
1136 
1137 	/*
1138 	 * If we are critically low on one of RAM or swap and low on
1139 	 * the other, kill the largest process.  However, we avoid
1140 	 * doing this on the first pass in order to give ourselves a
1141 	 * chance to flush out dirty vnode-backed pages and to allow
1142 	 * active pages to be moved to the inactive queue and reclaimed.
1143 	 *
1144 	 * We keep the process bigproc locked once we find it to keep anyone
1145 	 * from messing with it; however, there is a possibility of
1146 	 * deadlock if process B is bigproc and one of it's child processes
1147 	 * attempts to propagate a signal to B while we are waiting for A's
1148 	 * lock while walking this list.  To avoid this, we don't block on
1149 	 * the process lock but just skip a process if it is already locked.
1150 	 */
1151 	if (pass != 0 &&
1152 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1153 	     (swap_pager_full && vm_paging_target() > 0))) {
1154 		bigproc = NULL;
1155 		bigsize = 0;
1156 		sx_slock(&allproc_lock);
1157 		FOREACH_PROC_IN_SYSTEM(p) {
1158 			int breakout;
1159 
1160 			if (PROC_TRYLOCK(p) == 0)
1161 				continue;
1162 			/*
1163 			 * If this is a system or protected process, skip it.
1164 			 */
1165 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1166 			    (p->p_flag & P_PROTECTED) ||
1167 			    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1168 				PROC_UNLOCK(p);
1169 				continue;
1170 			}
1171 			/*
1172 			 * If the process is in a non-running type state,
1173 			 * don't touch it.  Check all the threads individually.
1174 			 */
1175 			mtx_lock_spin(&sched_lock);
1176 			breakout = 0;
1177 			FOREACH_THREAD_IN_PROC(p, td) {
1178 				if (!TD_ON_RUNQ(td) &&
1179 				    !TD_IS_RUNNING(td) &&
1180 				    !TD_IS_SLEEPING(td)) {
1181 					breakout = 1;
1182 					break;
1183 				}
1184 			}
1185 			if (breakout) {
1186 				mtx_unlock_spin(&sched_lock);
1187 				PROC_UNLOCK(p);
1188 				continue;
1189 			}
1190 			mtx_unlock_spin(&sched_lock);
1191 			/*
1192 			 * get the process size
1193 			 */
1194 			if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1195 				PROC_UNLOCK(p);
1196 				continue;
1197 			}
1198 			size = vmspace_swap_count(p->p_vmspace);
1199 			vm_map_unlock_read(&p->p_vmspace->vm_map);
1200 			size += vmspace_resident_count(p->p_vmspace);
1201 			/*
1202 			 * if the this process is bigger than the biggest one
1203 			 * remember it.
1204 			 */
1205 			if (size > bigsize) {
1206 				if (bigproc != NULL)
1207 					PROC_UNLOCK(bigproc);
1208 				bigproc = p;
1209 				bigsize = size;
1210 			} else
1211 				PROC_UNLOCK(p);
1212 		}
1213 		sx_sunlock(&allproc_lock);
1214 		if (bigproc != NULL) {
1215 			killproc(bigproc, "out of swap space");
1216 			mtx_lock_spin(&sched_lock);
1217 			sched_nice(bigproc, PRIO_MIN);
1218 			mtx_unlock_spin(&sched_lock);
1219 			PROC_UNLOCK(bigproc);
1220 			wakeup(&cnt.v_free_count);
1221 		}
1222 	}
1223 	mtx_unlock(&Giant);
1224 }
1225 
1226 /*
1227  * This routine tries to maintain the pseudo LRU active queue,
1228  * so that during long periods of time where there is no paging,
1229  * that some statistic accumulation still occurs.  This code
1230  * helps the situation where paging just starts to occur.
1231  */
1232 static void
1233 vm_pageout_page_stats()
1234 {
1235 	vm_page_t m,next;
1236 	int pcount,tpcount;		/* Number of pages to check */
1237 	static int fullintervalcount = 0;
1238 	int page_shortage;
1239 
1240 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1241 	page_shortage =
1242 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1243 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1244 
1245 	if (page_shortage <= 0)
1246 		return;
1247 
1248 	pcount = cnt.v_active_count;
1249 	fullintervalcount += vm_pageout_stats_interval;
1250 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1251 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1252 		if (pcount > tpcount)
1253 			pcount = tpcount;
1254 	} else {
1255 		fullintervalcount = 0;
1256 	}
1257 
1258 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1259 	while ((m != NULL) && (pcount-- > 0)) {
1260 		int actcount;
1261 
1262 		KASSERT(m->queue == PQ_ACTIVE,
1263 		    ("vm_pageout_page_stats: page %p isn't active", m));
1264 
1265 		next = TAILQ_NEXT(m, pageq);
1266 		/*
1267 		 * Don't deactivate pages that are busy.
1268 		 */
1269 		if ((m->busy != 0) ||
1270 		    (m->flags & PG_BUSY) ||
1271 		    (m->hold_count != 0)) {
1272 			vm_pageq_requeue(m);
1273 			m = next;
1274 			continue;
1275 		}
1276 
1277 		actcount = 0;
1278 		if (m->flags & PG_REFERENCED) {
1279 			vm_page_flag_clear(m, PG_REFERENCED);
1280 			actcount += 1;
1281 		}
1282 
1283 		actcount += pmap_ts_referenced(m);
1284 		if (actcount) {
1285 			m->act_count += ACT_ADVANCE + actcount;
1286 			if (m->act_count > ACT_MAX)
1287 				m->act_count = ACT_MAX;
1288 			vm_pageq_requeue(m);
1289 		} else {
1290 			if (m->act_count == 0) {
1291 				/*
1292 				 * We turn off page access, so that we have
1293 				 * more accurate RSS stats.  We don't do this
1294 				 * in the normal page deactivation when the
1295 				 * system is loaded VM wise, because the
1296 				 * cost of the large number of page protect
1297 				 * operations would be higher than the value
1298 				 * of doing the operation.
1299 				 */
1300 				pmap_remove_all(m);
1301 				vm_page_deactivate(m);
1302 			} else {
1303 				m->act_count -= min(m->act_count, ACT_DECLINE);
1304 				vm_pageq_requeue(m);
1305 			}
1306 		}
1307 
1308 		m = next;
1309 	}
1310 }
1311 
1312 /*
1313  *	vm_pageout is the high level pageout daemon.
1314  */
1315 static void
1316 vm_pageout()
1317 {
1318 	int error, pass;
1319 
1320 	/*
1321 	 * Initialize some paging parameters.
1322 	 */
1323 	cnt.v_interrupt_free_min = 2;
1324 	if (cnt.v_page_count < 2000)
1325 		vm_pageout_page_count = 8;
1326 
1327 	/*
1328 	 * v_free_reserved needs to include enough for the largest
1329 	 * swap pager structures plus enough for any pv_entry structs
1330 	 * when paging.
1331 	 */
1332 	if (cnt.v_page_count > 1024)
1333 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1334 	else
1335 		cnt.v_free_min = 4;
1336 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1337 	    cnt.v_interrupt_free_min;
1338 	cnt.v_free_reserved = vm_pageout_page_count +
1339 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_L2_SIZE;
1340 	cnt.v_free_severe = cnt.v_free_min / 2;
1341 	cnt.v_free_min += cnt.v_free_reserved;
1342 	cnt.v_free_severe += cnt.v_free_reserved;
1343 
1344 	/*
1345 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1346 	 * that these are more a measure of the VM cache queue hysteresis
1347 	 * then the VM free queue.  Specifically, v_free_target is the
1348 	 * high water mark (free+cache pages).
1349 	 *
1350 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1351 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1352 	 * be big enough to handle memory needs while the pageout daemon
1353 	 * is signalled and run to free more pages.
1354 	 */
1355 	if (cnt.v_free_count > 6144)
1356 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1357 	else
1358 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1359 
1360 	if (cnt.v_free_count > 2048) {
1361 		cnt.v_cache_min = cnt.v_free_target;
1362 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1363 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1364 	} else {
1365 		cnt.v_cache_min = 0;
1366 		cnt.v_cache_max = 0;
1367 		cnt.v_inactive_target = cnt.v_free_count / 4;
1368 	}
1369 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1370 		cnt.v_inactive_target = cnt.v_free_count / 3;
1371 
1372 	/* XXX does not really belong here */
1373 	if (vm_page_max_wired == 0)
1374 		vm_page_max_wired = cnt.v_free_count / 3;
1375 
1376 	if (vm_pageout_stats_max == 0)
1377 		vm_pageout_stats_max = cnt.v_free_target;
1378 
1379 	/*
1380 	 * Set interval in seconds for stats scan.
1381 	 */
1382 	if (vm_pageout_stats_interval == 0)
1383 		vm_pageout_stats_interval = 5;
1384 	if (vm_pageout_full_stats_interval == 0)
1385 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1386 
1387 	swap_pager_swap_init();
1388 	pass = 0;
1389 	/*
1390 	 * The pageout daemon is never done, so loop forever.
1391 	 */
1392 	while (TRUE) {
1393 		vm_page_lock_queues();
1394 		/*
1395 		 * If we have enough free memory, wakeup waiters.  Do
1396 		 * not clear vm_pages_needed until we reach our target,
1397 		 * otherwise we may be woken up over and over again and
1398 		 * waste a lot of cpu.
1399 		 */
1400 		if (vm_pages_needed && !vm_page_count_min()) {
1401 			if (!vm_paging_needed())
1402 				vm_pages_needed = 0;
1403 			wakeup(&cnt.v_free_count);
1404 		}
1405 		if (vm_pages_needed) {
1406 			/*
1407 			 * Still not done, take a second pass without waiting
1408 			 * (unlimited dirty cleaning), otherwise sleep a bit
1409 			 * and try again.
1410 			 */
1411 			++pass;
1412 			if (pass > 1)
1413 				msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1414 				       "psleep", hz/2);
1415 		} else {
1416 			/*
1417 			 * Good enough, sleep & handle stats.  Prime the pass
1418 			 * for the next run.
1419 			 */
1420 			if (pass > 1)
1421 				pass = 1;
1422 			else
1423 				pass = 0;
1424 			error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1425 				    "psleep", vm_pageout_stats_interval * hz);
1426 			if (error && !vm_pages_needed) {
1427 				pass = 0;
1428 				vm_pageout_page_stats();
1429 				vm_page_unlock_queues();
1430 				continue;
1431 			}
1432 		}
1433 		if (vm_pages_needed)
1434 			cnt.v_pdwakeups++;
1435 		vm_page_unlock_queues();
1436 		vm_pageout_scan(pass);
1437 	}
1438 }
1439 
1440 /*
1441  * Unless the page queue lock is held by the caller, this function
1442  * should be regarded as advisory.  Specifically, the caller should
1443  * not msleep() on &cnt.v_free_count following this function unless
1444  * the page queue lock is held until the msleep() is performed.
1445  */
1446 void
1447 pagedaemon_wakeup()
1448 {
1449 
1450 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1451 		vm_pages_needed = 1;
1452 		wakeup(&vm_pages_needed);
1453 	}
1454 }
1455 
1456 #if !defined(NO_SWAPPING)
1457 static void
1458 vm_req_vmdaemon()
1459 {
1460 	static int lastrun = 0;
1461 
1462 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1463 		wakeup(&vm_daemon_needed);
1464 		lastrun = ticks;
1465 	}
1466 }
1467 
1468 static void
1469 vm_daemon()
1470 {
1471 	struct rlimit rsslim;
1472 	struct proc *p;
1473 	struct thread *td;
1474 	int breakout;
1475 
1476 	mtx_lock(&Giant);
1477 	while (TRUE) {
1478 		tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1479 		if (vm_pageout_req_swapout) {
1480 			swapout_procs(vm_pageout_req_swapout);
1481 			vm_pageout_req_swapout = 0;
1482 		}
1483 		/*
1484 		 * scan the processes for exceeding their rlimits or if
1485 		 * process is swapped out -- deactivate pages
1486 		 */
1487 		sx_slock(&allproc_lock);
1488 		LIST_FOREACH(p, &allproc, p_list) {
1489 			vm_pindex_t limit, size;
1490 
1491 			/*
1492 			 * if this is a system process or if we have already
1493 			 * looked at this process, skip it.
1494 			 */
1495 			PROC_LOCK(p);
1496 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1497 				PROC_UNLOCK(p);
1498 				continue;
1499 			}
1500 			/*
1501 			 * if the process is in a non-running type state,
1502 			 * don't touch it.
1503 			 */
1504 			mtx_lock_spin(&sched_lock);
1505 			breakout = 0;
1506 			FOREACH_THREAD_IN_PROC(p, td) {
1507 				if (!TD_ON_RUNQ(td) &&
1508 				    !TD_IS_RUNNING(td) &&
1509 				    !TD_IS_SLEEPING(td)) {
1510 					breakout = 1;
1511 					break;
1512 				}
1513 			}
1514 			mtx_unlock_spin(&sched_lock);
1515 			if (breakout) {
1516 				PROC_UNLOCK(p);
1517 				continue;
1518 			}
1519 			/*
1520 			 * get a limit
1521 			 */
1522 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1523 			limit = OFF_TO_IDX(
1524 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1525 
1526 			/*
1527 			 * let processes that are swapped out really be
1528 			 * swapped out set the limit to nothing (will force a
1529 			 * swap-out.)
1530 			 */
1531 			if ((p->p_sflag & PS_INMEM) == 0)
1532 				limit = 0;	/* XXX */
1533 			PROC_UNLOCK(p);
1534 
1535 			size = vmspace_resident_count(p->p_vmspace);
1536 			if (limit >= 0 && size >= limit) {
1537 				vm_pageout_map_deactivate_pages(
1538 				    &p->p_vmspace->vm_map, limit);
1539 			}
1540 		}
1541 		sx_sunlock(&allproc_lock);
1542 	}
1543 }
1544 #endif			/* !defined(NO_SWAPPING) */
1545