xref: /freebsd/sys/vm/vm_pageout.c (revision 30d239bc4c510432e65a84fa1c14ed67a3ab1c92)
1 /*-
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  * Copyright (c) 2005 Yahoo! Technologies Norway AS
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
77 
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
94 #include <sys/sx.h>
95 #include <sys/sysctl.h>
96 
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/uma.h>
107 
108 #include <machine/mutex.h>
109 
110 /*
111  * System initialization
112  */
113 
114 /* the kernel process "vm_pageout"*/
115 static void vm_pageout(void);
116 static int vm_pageout_clean(vm_page_t);
117 static void vm_pageout_scan(int pass);
118 
119 struct proc *pageproc;
120 
121 static struct kproc_desc page_kp = {
122 	"pagedaemon",
123 	vm_pageout,
124 	&pageproc
125 };
126 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
127 
128 #if !defined(NO_SWAPPING)
129 /* the kernel process "vm_daemon"*/
130 static void vm_daemon(void);
131 static struct	proc *vmproc;
132 
133 static struct kproc_desc vm_kp = {
134 	"vmdaemon",
135 	vm_daemon,
136 	&vmproc
137 };
138 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
139 #endif
140 
141 
142 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
143 int vm_pageout_deficit;		/* Estimated number of pages deficit */
144 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
145 
146 #if !defined(NO_SWAPPING)
147 static int vm_pageout_req_swapout;	/* XXX */
148 static int vm_daemon_needed;
149 static struct mtx vm_daemon_mtx;
150 /* Allow for use by vm_pageout before vm_daemon is initialized. */
151 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
152 #endif
153 static int vm_max_launder = 32;
154 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
155 static int vm_pageout_full_stats_interval = 0;
156 static int vm_pageout_algorithm=0;
157 static int defer_swap_pageouts=0;
158 static int disable_swap_pageouts=0;
159 
160 #if defined(NO_SWAPPING)
161 static int vm_swap_enabled=0;
162 static int vm_swap_idle_enabled=0;
163 #else
164 static int vm_swap_enabled=1;
165 static int vm_swap_idle_enabled=0;
166 #endif
167 
168 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
169 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
170 
171 SYSCTL_INT(_vm, OID_AUTO, max_launder,
172 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
173 
174 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
175 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
176 
177 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
178 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
179 
180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
181 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
182 
183 #if defined(NO_SWAPPING)
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
188 #else
189 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
190 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
191 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
192 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
193 #endif
194 
195 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
196 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
197 
198 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
199 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
200 
201 static int pageout_lock_miss;
202 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
203 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
204 
205 #define VM_PAGEOUT_PAGE_COUNT 16
206 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
207 
208 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
209 
210 #if !defined(NO_SWAPPING)
211 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
212 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
213 static void vm_req_vmdaemon(int req);
214 #endif
215 static void vm_pageout_page_stats(void);
216 
217 /*
218  * vm_pageout_fallback_object_lock:
219  *
220  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
221  * known to have failed and page queue must be either PQ_ACTIVE or
222  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
223  * while locking the vm object.  Use marker page to detect page queue
224  * changes and maintain notion of next page on page queue.  Return
225  * TRUE if no changes were detected, FALSE otherwise.  vm object is
226  * locked on return.
227  *
228  * This function depends on both the lock portion of struct vm_object
229  * and normal struct vm_page being type stable.
230  */
231 static boolean_t
232 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
233 {
234 	struct vm_page marker;
235 	boolean_t unchanged;
236 	u_short queue;
237 	vm_object_t object;
238 
239 	/*
240 	 * Initialize our marker
241 	 */
242 	bzero(&marker, sizeof(marker));
243 	marker.flags = PG_FICTITIOUS | PG_MARKER;
244 	marker.oflags = VPO_BUSY;
245 	marker.queue = m->queue;
246 	marker.wire_count = 1;
247 
248 	queue = m->queue;
249 	object = m->object;
250 
251 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
252 			   m, &marker, pageq);
253 	vm_page_unlock_queues();
254 	VM_OBJECT_LOCK(object);
255 	vm_page_lock_queues();
256 
257 	/* Page queue might have changed. */
258 	*next = TAILQ_NEXT(&marker, pageq);
259 	unchanged = (m->queue == queue &&
260 		     m->object == object &&
261 		     &marker == TAILQ_NEXT(m, pageq));
262 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
263 		     &marker, pageq);
264 	return (unchanged);
265 }
266 
267 /*
268  * vm_pageout_clean:
269  *
270  * Clean the page and remove it from the laundry.
271  *
272  * We set the busy bit to cause potential page faults on this page to
273  * block.  Note the careful timing, however, the busy bit isn't set till
274  * late and we cannot do anything that will mess with the page.
275  */
276 static int
277 vm_pageout_clean(m)
278 	vm_page_t m;
279 {
280 	vm_object_t object;
281 	vm_page_t mc[2*vm_pageout_page_count];
282 	int pageout_count;
283 	int ib, is, page_base;
284 	vm_pindex_t pindex = m->pindex;
285 
286 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
287 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
288 
289 	/*
290 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
291 	 * with the new swapper, but we could have serious problems paging
292 	 * out other object types if there is insufficient memory.
293 	 *
294 	 * Unfortunately, checking free memory here is far too late, so the
295 	 * check has been moved up a procedural level.
296 	 */
297 
298 	/*
299 	 * Can't clean the page if it's busy or held.
300 	 */
301 	if ((m->hold_count != 0) ||
302 	    ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
303 		return 0;
304 	}
305 
306 	mc[vm_pageout_page_count] = m;
307 	pageout_count = 1;
308 	page_base = vm_pageout_page_count;
309 	ib = 1;
310 	is = 1;
311 
312 	/*
313 	 * Scan object for clusterable pages.
314 	 *
315 	 * We can cluster ONLY if: ->> the page is NOT
316 	 * clean, wired, busy, held, or mapped into a
317 	 * buffer, and one of the following:
318 	 * 1) The page is inactive, or a seldom used
319 	 *    active page.
320 	 * -or-
321 	 * 2) we force the issue.
322 	 *
323 	 * During heavy mmap/modification loads the pageout
324 	 * daemon can really fragment the underlying file
325 	 * due to flushing pages out of order and not trying
326 	 * align the clusters (which leave sporatic out-of-order
327 	 * holes).  To solve this problem we do the reverse scan
328 	 * first and attempt to align our cluster, then do a
329 	 * forward scan if room remains.
330 	 */
331 	object = m->object;
332 more:
333 	while (ib && pageout_count < vm_pageout_page_count) {
334 		vm_page_t p;
335 
336 		if (ib > pindex) {
337 			ib = 0;
338 			break;
339 		}
340 
341 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
342 			ib = 0;
343 			break;
344 		}
345 		if ((p->oflags & VPO_BUSY) || p->busy) {
346 			ib = 0;
347 			break;
348 		}
349 		vm_page_test_dirty(p);
350 		if ((p->dirty & p->valid) == 0 ||
351 		    p->queue != PQ_INACTIVE ||
352 		    p->wire_count != 0 ||	/* may be held by buf cache */
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 & p->valid) == 0 ||
379 		    p->queue != PQ_INACTIVE ||
380 		    p->wire_count != 0 ||	/* may be held by buf cache */
381 		    p->hold_count != 0) {	/* may be undergoing I/O */
382 			break;
383 		}
384 		mc[page_base + pageout_count] = p;
385 		++pageout_count;
386 		++is;
387 	}
388 
389 	/*
390 	 * If we exhausted our forward scan, continue with the reverse scan
391 	 * when possible, even past a page boundry.  This catches boundry
392 	 * conditions.
393 	 */
394 	if (ib && pageout_count < vm_pageout_page_count)
395 		goto more;
396 
397 	/*
398 	 * we allow reads during pageouts...
399 	 */
400 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
401 }
402 
403 /*
404  * vm_pageout_flush() - launder the given pages
405  *
406  *	The given pages are laundered.  Note that we setup for the start of
407  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
408  *	reference count all in here rather then in the parent.  If we want
409  *	the parent to do more sophisticated things we may have to change
410  *	the ordering.
411  */
412 int
413 vm_pageout_flush(vm_page_t *mc, int count, int flags)
414 {
415 	vm_object_t object = mc[0]->object;
416 	int pageout_status[count];
417 	int numpagedout = 0;
418 	int i;
419 
420 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
421 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
422 	/*
423 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
424 	 * mark the pages read-only.
425 	 *
426 	 * We do not have to fixup the clean/dirty bits here... we can
427 	 * allow the pager to do it after the I/O completes.
428 	 *
429 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
430 	 * edge case with file fragments.
431 	 */
432 	for (i = 0; i < count; i++) {
433 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
434 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
435 			mc[i], i, count));
436 		vm_page_io_start(mc[i]);
437 		pmap_remove_write(mc[i]);
438 	}
439 	vm_page_unlock_queues();
440 	vm_object_pip_add(object, count);
441 
442 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
443 
444 	vm_page_lock_queues();
445 	for (i = 0; i < count; i++) {
446 		vm_page_t mt = mc[i];
447 
448 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
449 		    (mt->flags & PG_WRITEABLE) == 0,
450 		    ("vm_pageout_flush: page %p is not write protected", mt));
451 		switch (pageout_status[i]) {
452 		case VM_PAGER_OK:
453 		case VM_PAGER_PEND:
454 			numpagedout++;
455 			break;
456 		case VM_PAGER_BAD:
457 			/*
458 			 * Page outside of range of object. Right now we
459 			 * essentially lose the changes by pretending it
460 			 * worked.
461 			 */
462 			pmap_clear_modify(mt);
463 			vm_page_undirty(mt);
464 			break;
465 		case VM_PAGER_ERROR:
466 		case VM_PAGER_FAIL:
467 			/*
468 			 * If page couldn't be paged out, then reactivate the
469 			 * page so it doesn't clog the inactive list.  (We
470 			 * will try paging out it again later).
471 			 */
472 			vm_page_activate(mt);
473 			break;
474 		case VM_PAGER_AGAIN:
475 			break;
476 		}
477 
478 		/*
479 		 * If the operation is still going, leave the page busy to
480 		 * block all other accesses. Also, leave the paging in
481 		 * progress indicator set so that we don't attempt an object
482 		 * collapse.
483 		 */
484 		if (pageout_status[i] != VM_PAGER_PEND) {
485 			vm_object_pip_wakeup(object);
486 			vm_page_io_finish(mt);
487 			if (vm_page_count_severe())
488 				vm_page_try_to_cache(mt);
489 		}
490 	}
491 	return numpagedout;
492 }
493 
494 #if !defined(NO_SWAPPING)
495 /*
496  *	vm_pageout_object_deactivate_pages
497  *
498  *	deactivate enough pages to satisfy the inactive target
499  *	requirements or if vm_page_proc_limit is set, then
500  *	deactivate all of the pages in the object and its
501  *	backing_objects.
502  *
503  *	The object and map must be locked.
504  */
505 static void
506 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
507 	pmap_t pmap;
508 	vm_object_t first_object;
509 	long desired;
510 {
511 	vm_object_t backing_object, object;
512 	vm_page_t p, next;
513 	int actcount, rcount, remove_mode;
514 
515 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
516 	if (first_object->type == OBJT_DEVICE || 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_pageq_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_pageq_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 	struct proc *p, *bigproc;
682 	struct thread *td;
683 	vm_offset_t size, bigsize;
684 	vm_object_t object;
685 	int actcount;
686 	int vnodes_skipped = 0;
687 	int maxlaunder;
688 
689 	/*
690 	 * Decrease registered cache sizes.
691 	 */
692 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
693 	/*
694 	 * We do this explicitly after the caches have been drained above.
695 	 */
696 	uma_reclaim();
697 
698 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
699 
700 	/*
701 	 * Calculate the number of pages we want to either free or move
702 	 * to the cache.
703 	 */
704 	page_shortage = vm_paging_target() + addl_page_shortage_init;
705 
706 	/*
707 	 * Initialize our marker
708 	 */
709 	bzero(&marker, sizeof(marker));
710 	marker.flags = PG_FICTITIOUS | PG_MARKER;
711 	marker.oflags = VPO_BUSY;
712 	marker.queue = PQ_INACTIVE;
713 	marker.wire_count = 1;
714 
715 	/*
716 	 * Start scanning the inactive queue for pages we can move to the
717 	 * cache or free.  The scan will stop when the target is reached or
718 	 * we have scanned the entire inactive queue.  Note that m->act_count
719 	 * is not used to form decisions for the inactive queue, only for the
720 	 * active queue.
721 	 *
722 	 * maxlaunder limits the number of dirty pages we flush per scan.
723 	 * For most systems a smaller value (16 or 32) is more robust under
724 	 * extreme memory and disk pressure because any unnecessary writes
725 	 * to disk can result in extreme performance degredation.  However,
726 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
727 	 * used) will die horribly with limited laundering.  If the pageout
728 	 * daemon cannot clean enough pages in the first pass, we let it go
729 	 * all out in succeeding passes.
730 	 */
731 	if ((maxlaunder = vm_max_launder) <= 1)
732 		maxlaunder = 1;
733 	if (pass)
734 		maxlaunder = 10000;
735 	vm_page_lock_queues();
736 rescan0:
737 	addl_page_shortage = addl_page_shortage_init;
738 	maxscan = cnt.v_inactive_count;
739 
740 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
741 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
742 	     m = next) {
743 
744 		cnt.v_pdpages++;
745 
746 		if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
747 			goto rescan0;
748 		}
749 
750 		next = TAILQ_NEXT(m, pageq);
751 		object = m->object;
752 
753 		/*
754 		 * skip marker pages
755 		 */
756 		if (m->flags & PG_MARKER)
757 			continue;
758 
759 		/*
760 		 * A held page may be undergoing I/O, so skip it.
761 		 */
762 		if (m->hold_count) {
763 			vm_pageq_requeue(m);
764 			addl_page_shortage++;
765 			continue;
766 		}
767 		/*
768 		 * Don't mess with busy pages, keep in the front of the
769 		 * queue, most likely are being paged out.
770 		 */
771 		if (!VM_OBJECT_TRYLOCK(object) &&
772 		    (!vm_pageout_fallback_object_lock(m, &next) ||
773 		     m->hold_count != 0)) {
774 			VM_OBJECT_UNLOCK(object);
775 			addl_page_shortage++;
776 			continue;
777 		}
778 		if (m->busy || (m->oflags & VPO_BUSY)) {
779 			VM_OBJECT_UNLOCK(object);
780 			addl_page_shortage++;
781 			continue;
782 		}
783 
784 		/*
785 		 * If the object is not being used, we ignore previous
786 		 * references.
787 		 */
788 		if (object->ref_count == 0) {
789 			vm_page_flag_clear(m, PG_REFERENCED);
790 			pmap_clear_reference(m);
791 
792 		/*
793 		 * Otherwise, if the page has been referenced while in the
794 		 * inactive queue, we bump the "activation count" upwards,
795 		 * making it less likely that the page will be added back to
796 		 * the inactive queue prematurely again.  Here we check the
797 		 * page tables (or emulated bits, if any), given the upper
798 		 * level VM system not knowing anything about existing
799 		 * references.
800 		 */
801 		} else if (((m->flags & PG_REFERENCED) == 0) &&
802 			(actcount = pmap_ts_referenced(m))) {
803 			vm_page_activate(m);
804 			VM_OBJECT_UNLOCK(object);
805 			m->act_count += (actcount + ACT_ADVANCE);
806 			continue;
807 		}
808 
809 		/*
810 		 * If the upper level VM system knows about any page
811 		 * references, we activate the page.  We also set the
812 		 * "activation count" higher than normal so that we will less
813 		 * likely place pages back onto the inactive queue again.
814 		 */
815 		if ((m->flags & PG_REFERENCED) != 0) {
816 			vm_page_flag_clear(m, PG_REFERENCED);
817 			actcount = pmap_ts_referenced(m);
818 			vm_page_activate(m);
819 			VM_OBJECT_UNLOCK(object);
820 			m->act_count += (actcount + ACT_ADVANCE + 1);
821 			continue;
822 		}
823 
824 		/*
825 		 * If the upper level VM system doesn't know anything about
826 		 * the page being dirty, we have to check for it again.  As
827 		 * far as the VM code knows, any partially dirty pages are
828 		 * fully dirty.
829 		 */
830 		if (m->dirty == 0 && !pmap_is_modified(m)) {
831 			/*
832 			 * Avoid a race condition: Unless write access is
833 			 * removed from the page, another processor could
834 			 * modify it before all access is removed by the call
835 			 * to vm_page_cache() below.  If vm_page_cache() finds
836 			 * that the page has been modified when it removes all
837 			 * access, it panics because it cannot cache dirty
838 			 * pages.  In principle, we could eliminate just write
839 			 * access here rather than all access.  In the expected
840 			 * case, when there are no last instant modifications
841 			 * to the page, removing all access will be cheaper
842 			 * overall.
843 			 */
844 			if ((m->flags & PG_WRITEABLE) != 0)
845 				pmap_remove_all(m);
846 		} else {
847 			vm_page_dirty(m);
848 		}
849 
850 		if (m->valid == 0) {
851 			/*
852 			 * Invalid pages can be easily freed
853 			 */
854 			vm_page_free(m);
855 			cnt.v_dfree++;
856 			--page_shortage;
857 		} else if (m->dirty == 0) {
858 			/*
859 			 * Clean pages can be placed onto the cache queue.
860 			 * This effectively frees them.
861 			 */
862 			vm_page_cache(m);
863 			--page_shortage;
864 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
865 			/*
866 			 * Dirty pages need to be paged out, but flushing
867 			 * a page is extremely expensive verses freeing
868 			 * a clean page.  Rather then artificially limiting
869 			 * the number of pages we can flush, we instead give
870 			 * dirty pages extra priority on the inactive queue
871 			 * by forcing them to be cycled through the queue
872 			 * twice before being flushed, after which the
873 			 * (now clean) page will cycle through once more
874 			 * before being freed.  This significantly extends
875 			 * the thrash point for a heavily loaded machine.
876 			 */
877 			vm_page_flag_set(m, PG_WINATCFLS);
878 			vm_pageq_requeue(m);
879 		} else if (maxlaunder > 0) {
880 			/*
881 			 * We always want to try to flush some dirty pages if
882 			 * we encounter them, to keep the system stable.
883 			 * Normally this number is small, but under extreme
884 			 * pressure where there are insufficient clean pages
885 			 * on the inactive queue, we may have to go all out.
886 			 */
887 			int swap_pageouts_ok, vfslocked = 0;
888 			struct vnode *vp = NULL;
889 			struct mount *mp = NULL;
890 
891 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
892 				swap_pageouts_ok = 1;
893 			} else {
894 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
895 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
896 				vm_page_count_min());
897 
898 			}
899 
900 			/*
901 			 * We don't bother paging objects that are "dead".
902 			 * Those objects are in a "rundown" state.
903 			 */
904 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
905 				VM_OBJECT_UNLOCK(object);
906 				vm_pageq_requeue(m);
907 				continue;
908 			}
909 
910 			/*
911 			 * Following operations may unlock
912 			 * vm_page_queue_mtx, invalidating the 'next'
913 			 * pointer.  To prevent an inordinate number
914 			 * of restarts we use our marker to remember
915 			 * our place.
916 			 *
917 			 */
918 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
919 					   m, &marker, pageq);
920 			/*
921 			 * The object is already known NOT to be dead.   It
922 			 * is possible for the vget() to block the whole
923 			 * pageout daemon, but the new low-memory handling
924 			 * code should prevent it.
925 			 *
926 			 * The previous code skipped locked vnodes and, worse,
927 			 * reordered pages in the queue.  This results in
928 			 * completely non-deterministic operation and, on a
929 			 * busy system, can lead to extremely non-optimal
930 			 * pageouts.  For example, it can cause clean pages
931 			 * to be freed and dirty pages to be moved to the end
932 			 * of the queue.  Since dirty pages are also moved to
933 			 * the end of the queue once-cleaned, this gives
934 			 * way too large a weighting to defering the freeing
935 			 * of dirty pages.
936 			 *
937 			 * We can't wait forever for the vnode lock, we might
938 			 * deadlock due to a vn_read() getting stuck in
939 			 * vm_wait while holding this vnode.  We skip the
940 			 * vnode if we can't get it in a reasonable amount
941 			 * of time.
942 			 */
943 			if (object->type == OBJT_VNODE) {
944 				vp = object->handle;
945 				if (vp->v_type == VREG &&
946 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
947 					KASSERT(mp == NULL,
948 					    ("vm_pageout_scan: mp != NULL"));
949 					++pageout_lock_miss;
950 					if (object->flags & OBJ_MIGHTBEDIRTY)
951 						vnodes_skipped++;
952 					goto unlock_and_continue;
953 				}
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_pageq_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_pageq_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_pageq_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 					pmap_remove_all(m);
1126 					if (m->dirty == 0)
1127 						vm_page_cache(m);
1128 					else
1129 						vm_page_deactivate(m);
1130 				} else {
1131 					vm_page_deactivate(m);
1132 				}
1133 			} else {
1134 				vm_pageq_requeue(m);
1135 			}
1136 		}
1137 		VM_OBJECT_UNLOCK(object);
1138 		m = next;
1139 	}
1140 	vm_page_unlock_queues();
1141 #if !defined(NO_SWAPPING)
1142 	/*
1143 	 * Idle process swapout -- run once per second.
1144 	 */
1145 	if (vm_swap_idle_enabled) {
1146 		static long lsec;
1147 		if (time_second != lsec) {
1148 			vm_req_vmdaemon(VM_SWAP_IDLE);
1149 			lsec = time_second;
1150 		}
1151 	}
1152 #endif
1153 
1154 	/*
1155 	 * If we didn't get enough free pages, and we have skipped a vnode
1156 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1157 	 * if we did not get enough free pages.
1158 	 */
1159 	if (vm_paging_target() > 0) {
1160 		if (vnodes_skipped && vm_page_count_min())
1161 			(void) speedup_syncer();
1162 #if !defined(NO_SWAPPING)
1163 		if (vm_swap_enabled && vm_page_count_target())
1164 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1165 #endif
1166 	}
1167 
1168 	/*
1169 	 * If we are critically low on one of RAM or swap and low on
1170 	 * the other, kill the largest process.  However, we avoid
1171 	 * doing this on the first pass in order to give ourselves a
1172 	 * chance to flush out dirty vnode-backed pages and to allow
1173 	 * active pages to be moved to the inactive queue and reclaimed.
1174 	 *
1175 	 * We keep the process bigproc locked once we find it to keep anyone
1176 	 * from messing with it; however, there is a possibility of
1177 	 * deadlock if process B is bigproc and one of it's child processes
1178 	 * attempts to propagate a signal to B while we are waiting for A's
1179 	 * lock while walking this list.  To avoid this, we don't block on
1180 	 * the process lock but just skip a process if it is already locked.
1181 	 */
1182 	if (pass != 0 &&
1183 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1184 	     (swap_pager_full && vm_paging_target() > 0))) {
1185 		bigproc = NULL;
1186 		bigsize = 0;
1187 		sx_slock(&allproc_lock);
1188 		FOREACH_PROC_IN_SYSTEM(p) {
1189 			int breakout;
1190 
1191 			if (PROC_TRYLOCK(p) == 0)
1192 				continue;
1193 			/*
1194 			 * If this is a system or protected process, skip it.
1195 			 */
1196 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1197 			    (p->p_flag & P_PROTECTED) ||
1198 			    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1199 				PROC_UNLOCK(p);
1200 				continue;
1201 			}
1202 			/*
1203 			 * If the process is in a non-running type state,
1204 			 * don't touch it.  Check all the threads individually.
1205 			 */
1206 			PROC_SLOCK(p);
1207 			breakout = 0;
1208 			FOREACH_THREAD_IN_PROC(p, td) {
1209 				thread_lock(td);
1210 				if (!TD_ON_RUNQ(td) &&
1211 				    !TD_IS_RUNNING(td) &&
1212 				    !TD_IS_SLEEPING(td)) {
1213 					thread_unlock(td);
1214 					breakout = 1;
1215 					break;
1216 				}
1217 				thread_unlock(td);
1218 			}
1219 			PROC_SUNLOCK(p);
1220 			if (breakout) {
1221 				PROC_UNLOCK(p);
1222 				continue;
1223 			}
1224 			/*
1225 			 * get the process size
1226 			 */
1227 			if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1228 				PROC_UNLOCK(p);
1229 				continue;
1230 			}
1231 			size = vmspace_swap_count(p->p_vmspace);
1232 			vm_map_unlock_read(&p->p_vmspace->vm_map);
1233 			size += vmspace_resident_count(p->p_vmspace);
1234 			/*
1235 			 * if the this process is bigger than the biggest one
1236 			 * remember it.
1237 			 */
1238 			if (size > bigsize) {
1239 				if (bigproc != NULL)
1240 					PROC_UNLOCK(bigproc);
1241 				bigproc = p;
1242 				bigsize = size;
1243 			} else
1244 				PROC_UNLOCK(p);
1245 		}
1246 		sx_sunlock(&allproc_lock);
1247 		if (bigproc != NULL) {
1248 			killproc(bigproc, "out of swap space");
1249 			PROC_SLOCK(bigproc);
1250 			sched_nice(bigproc, PRIO_MIN);
1251 			PROC_SUNLOCK(bigproc);
1252 			PROC_UNLOCK(bigproc);
1253 			wakeup(&cnt.v_free_count);
1254 		}
1255 	}
1256 }
1257 
1258 /*
1259  * This routine tries to maintain the pseudo LRU active queue,
1260  * so that during long periods of time where there is no paging,
1261  * that some statistic accumulation still occurs.  This code
1262  * helps the situation where paging just starts to occur.
1263  */
1264 static void
1265 vm_pageout_page_stats()
1266 {
1267 	vm_object_t object;
1268 	vm_page_t m,next;
1269 	int pcount,tpcount;		/* Number of pages to check */
1270 	static int fullintervalcount = 0;
1271 	int page_shortage;
1272 
1273 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1274 	page_shortage =
1275 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1276 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1277 
1278 	if (page_shortage <= 0)
1279 		return;
1280 
1281 	pcount = cnt.v_active_count;
1282 	fullintervalcount += vm_pageout_stats_interval;
1283 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1284 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1285 		if (pcount > tpcount)
1286 			pcount = tpcount;
1287 	} else {
1288 		fullintervalcount = 0;
1289 	}
1290 
1291 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1292 	while ((m != NULL) && (pcount-- > 0)) {
1293 		int actcount;
1294 
1295 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1296 		    ("vm_pageout_page_stats: page %p isn't active", m));
1297 
1298 		next = TAILQ_NEXT(m, pageq);
1299 		object = m->object;
1300 
1301 		if ((m->flags & PG_MARKER) != 0) {
1302 			m = next;
1303 			continue;
1304 		}
1305 		if (!VM_OBJECT_TRYLOCK(object) &&
1306 		    !vm_pageout_fallback_object_lock(m, &next)) {
1307 			VM_OBJECT_UNLOCK(object);
1308 			m = next;
1309 			continue;
1310 		}
1311 
1312 		/*
1313 		 * Don't deactivate pages that are busy.
1314 		 */
1315 		if ((m->busy != 0) ||
1316 		    (m->oflags & VPO_BUSY) ||
1317 		    (m->hold_count != 0)) {
1318 			VM_OBJECT_UNLOCK(object);
1319 			vm_pageq_requeue(m);
1320 			m = next;
1321 			continue;
1322 		}
1323 
1324 		actcount = 0;
1325 		if (m->flags & PG_REFERENCED) {
1326 			vm_page_flag_clear(m, PG_REFERENCED);
1327 			actcount += 1;
1328 		}
1329 
1330 		actcount += pmap_ts_referenced(m);
1331 		if (actcount) {
1332 			m->act_count += ACT_ADVANCE + actcount;
1333 			if (m->act_count > ACT_MAX)
1334 				m->act_count = ACT_MAX;
1335 			vm_pageq_requeue(m);
1336 		} else {
1337 			if (m->act_count == 0) {
1338 				/*
1339 				 * We turn off page access, so that we have
1340 				 * more accurate RSS stats.  We don't do this
1341 				 * in the normal page deactivation when the
1342 				 * system is loaded VM wise, because the
1343 				 * cost of the large number of page protect
1344 				 * operations would be higher than the value
1345 				 * of doing the operation.
1346 				 */
1347 				pmap_remove_all(m);
1348 				vm_page_deactivate(m);
1349 			} else {
1350 				m->act_count -= min(m->act_count, ACT_DECLINE);
1351 				vm_pageq_requeue(m);
1352 			}
1353 		}
1354 		VM_OBJECT_UNLOCK(object);
1355 		m = next;
1356 	}
1357 }
1358 
1359 /*
1360  *	vm_pageout is the high level pageout daemon.
1361  */
1362 static void
1363 vm_pageout()
1364 {
1365 	int error, pass;
1366 
1367 	/*
1368 	 * Initialize some paging parameters.
1369 	 */
1370 	cnt.v_interrupt_free_min = 2;
1371 	if (cnt.v_page_count < 2000)
1372 		vm_pageout_page_count = 8;
1373 
1374 	/*
1375 	 * v_free_reserved needs to include enough for the largest
1376 	 * swap pager structures plus enough for any pv_entry structs
1377 	 * when paging.
1378 	 */
1379 	if (cnt.v_page_count > 1024)
1380 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1381 	else
1382 		cnt.v_free_min = 4;
1383 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1384 	    cnt.v_interrupt_free_min;
1385 	cnt.v_free_reserved = vm_pageout_page_count +
1386 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1387 	cnt.v_free_severe = cnt.v_free_min / 2;
1388 	cnt.v_free_min += cnt.v_free_reserved;
1389 	cnt.v_free_severe += cnt.v_free_reserved;
1390 
1391 	/*
1392 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1393 	 * that these are more a measure of the VM cache queue hysteresis
1394 	 * then the VM free queue.  Specifically, v_free_target is the
1395 	 * high water mark (free+cache pages).
1396 	 *
1397 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1398 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1399 	 * be big enough to handle memory needs while the pageout daemon
1400 	 * is signalled and run to free more pages.
1401 	 */
1402 	if (cnt.v_free_count > 6144)
1403 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1404 	else
1405 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1406 
1407 	if (cnt.v_free_count > 2048) {
1408 		cnt.v_cache_min = cnt.v_free_target;
1409 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1410 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1411 	} else {
1412 		cnt.v_cache_min = 0;
1413 		cnt.v_cache_max = 0;
1414 		cnt.v_inactive_target = cnt.v_free_count / 4;
1415 	}
1416 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1417 		cnt.v_inactive_target = cnt.v_free_count / 3;
1418 
1419 	/* XXX does not really belong here */
1420 	if (vm_page_max_wired == 0)
1421 		vm_page_max_wired = cnt.v_free_count / 3;
1422 
1423 	if (vm_pageout_stats_max == 0)
1424 		vm_pageout_stats_max = cnt.v_free_target;
1425 
1426 	/*
1427 	 * Set interval in seconds for stats scan.
1428 	 */
1429 	if (vm_pageout_stats_interval == 0)
1430 		vm_pageout_stats_interval = 5;
1431 	if (vm_pageout_full_stats_interval == 0)
1432 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1433 
1434 	swap_pager_swap_init();
1435 	pass = 0;
1436 	/*
1437 	 * The pageout daemon is never done, so loop forever.
1438 	 */
1439 	while (TRUE) {
1440 		/*
1441 		 * If we have enough free memory, wakeup waiters.  Do
1442 		 * not clear vm_pages_needed until we reach our target,
1443 		 * otherwise we may be woken up over and over again and
1444 		 * waste a lot of cpu.
1445 		 */
1446 		mtx_lock(&vm_page_queue_free_mtx);
1447 		if (vm_pages_needed && !vm_page_count_min()) {
1448 			if (!vm_paging_needed())
1449 				vm_pages_needed = 0;
1450 			wakeup(&cnt.v_free_count);
1451 		}
1452 		if (vm_pages_needed) {
1453 			/*
1454 			 * Still not done, take a second pass without waiting
1455 			 * (unlimited dirty cleaning), otherwise sleep a bit
1456 			 * and try again.
1457 			 */
1458 			++pass;
1459 			if (pass > 1)
1460 				msleep(&vm_pages_needed,
1461 				    &vm_page_queue_free_mtx, PVM, "psleep",
1462 				    hz / 2);
1463 		} else {
1464 			/*
1465 			 * Good enough, sleep & handle stats.  Prime the pass
1466 			 * for the next run.
1467 			 */
1468 			if (pass > 1)
1469 				pass = 1;
1470 			else
1471 				pass = 0;
1472 			error = msleep(&vm_pages_needed,
1473 			    &vm_page_queue_free_mtx, PVM, "psleep",
1474 			    vm_pageout_stats_interval * hz);
1475 			if (error && !vm_pages_needed) {
1476 				mtx_unlock(&vm_page_queue_free_mtx);
1477 				pass = 0;
1478 				vm_page_lock_queues();
1479 				vm_pageout_page_stats();
1480 				vm_page_unlock_queues();
1481 				continue;
1482 			}
1483 		}
1484 		if (vm_pages_needed)
1485 			cnt.v_pdwakeups++;
1486 		mtx_unlock(&vm_page_queue_free_mtx);
1487 		vm_pageout_scan(pass);
1488 	}
1489 }
1490 
1491 /*
1492  * Unless the free page queue lock is held by the caller, this function
1493  * should be regarded as advisory.  Specifically, the caller should
1494  * not msleep() on &cnt.v_free_count following this function unless
1495  * the free page queue lock is held until the msleep() is performed.
1496  */
1497 void
1498 pagedaemon_wakeup()
1499 {
1500 
1501 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1502 		vm_pages_needed = 1;
1503 		wakeup(&vm_pages_needed);
1504 	}
1505 }
1506 
1507 #if !defined(NO_SWAPPING)
1508 static void
1509 vm_req_vmdaemon(int req)
1510 {
1511 	static int lastrun = 0;
1512 
1513 	mtx_lock(&vm_daemon_mtx);
1514 	vm_pageout_req_swapout |= req;
1515 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1516 		wakeup(&vm_daemon_needed);
1517 		lastrun = ticks;
1518 	}
1519 	mtx_unlock(&vm_daemon_mtx);
1520 }
1521 
1522 static void
1523 vm_daemon()
1524 {
1525 	struct rlimit rsslim;
1526 	struct proc *p;
1527 	struct thread *td;
1528 	int breakout, swapout_flags;
1529 
1530 	while (TRUE) {
1531 		mtx_lock(&vm_daemon_mtx);
1532 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1533 		swapout_flags = vm_pageout_req_swapout;
1534 		vm_pageout_req_swapout = 0;
1535 		mtx_unlock(&vm_daemon_mtx);
1536 		if (swapout_flags)
1537 			swapout_procs(swapout_flags);
1538 
1539 		/*
1540 		 * scan the processes for exceeding their rlimits or if
1541 		 * process is swapped out -- deactivate pages
1542 		 */
1543 		sx_slock(&allproc_lock);
1544 		FOREACH_PROC_IN_SYSTEM(p) {
1545 			vm_pindex_t limit, size;
1546 
1547 			/*
1548 			 * if this is a system process or if we have already
1549 			 * looked at this process, skip it.
1550 			 */
1551 			PROC_LOCK(p);
1552 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1553 				PROC_UNLOCK(p);
1554 				continue;
1555 			}
1556 			/*
1557 			 * if the process is in a non-running type state,
1558 			 * don't touch it.
1559 			 */
1560 			PROC_SLOCK(p);
1561 			breakout = 0;
1562 			FOREACH_THREAD_IN_PROC(p, td) {
1563 				thread_lock(td);
1564 				if (!TD_ON_RUNQ(td) &&
1565 				    !TD_IS_RUNNING(td) &&
1566 				    !TD_IS_SLEEPING(td)) {
1567 					thread_unlock(td);
1568 					breakout = 1;
1569 					break;
1570 				}
1571 				thread_unlock(td);
1572 			}
1573 			PROC_SUNLOCK(p);
1574 			if (breakout) {
1575 				PROC_UNLOCK(p);
1576 				continue;
1577 			}
1578 			/*
1579 			 * get a limit
1580 			 */
1581 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1582 			limit = OFF_TO_IDX(
1583 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1584 
1585 			/*
1586 			 * let processes that are swapped out really be
1587 			 * swapped out set the limit to nothing (will force a
1588 			 * swap-out.)
1589 			 */
1590 			if ((p->p_flag & P_INMEM) == 0)
1591 				limit = 0;	/* XXX */
1592 			PROC_UNLOCK(p);
1593 
1594 			size = vmspace_resident_count(p->p_vmspace);
1595 			if (limit >= 0 && size >= limit) {
1596 				vm_pageout_map_deactivate_pages(
1597 				    &p->p_vmspace->vm_map, limit);
1598 			}
1599 		}
1600 		sx_sunlock(&allproc_lock);
1601 	}
1602 }
1603 #endif			/* !defined(NO_SWAPPING) */
1604