xref: /freebsd/sys/vm/vm_pageout.c (revision 282a3889ebf826db9839be296ff1dd903f6d6d6e)
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 (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
346 		    (p->oflags & VPO_BUSY) || p->busy) {
347 			ib = 0;
348 			break;
349 		}
350 		vm_page_test_dirty(p);
351 		if ((p->dirty & p->valid) == 0 ||
352 		    p->queue != PQ_INACTIVE ||
353 		    p->wire_count != 0 ||	/* may be held by buf cache */
354 		    p->hold_count != 0) {	/* may be undergoing I/O */
355 			ib = 0;
356 			break;
357 		}
358 		mc[--page_base] = p;
359 		++pageout_count;
360 		++ib;
361 		/*
362 		 * alignment boundry, stop here and switch directions.  Do
363 		 * not clear ib.
364 		 */
365 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
366 			break;
367 	}
368 
369 	while (pageout_count < vm_pageout_page_count &&
370 	    pindex + is < object->size) {
371 		vm_page_t p;
372 
373 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
374 			break;
375 		if (VM_PAGE_INQUEUE1(p, PQ_CACHE) ||
376 		    (p->oflags & VPO_BUSY) || p->busy) {
377 			break;
378 		}
379 		vm_page_test_dirty(p);
380 		if ((p->dirty & p->valid) == 0 ||
381 		    p->queue != PQ_INACTIVE ||
382 		    p->wire_count != 0 ||	/* may be held by buf cache */
383 		    p->hold_count != 0) {	/* may be undergoing I/O */
384 			break;
385 		}
386 		mc[page_base + pageout_count] = p;
387 		++pageout_count;
388 		++is;
389 	}
390 
391 	/*
392 	 * If we exhausted our forward scan, continue with the reverse scan
393 	 * when possible, even past a page boundry.  This catches boundry
394 	 * conditions.
395 	 */
396 	if (ib && pageout_count < vm_pageout_page_count)
397 		goto more;
398 
399 	/*
400 	 * we allow reads during pageouts...
401 	 */
402 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
403 }
404 
405 /*
406  * vm_pageout_flush() - launder the given pages
407  *
408  *	The given pages are laundered.  Note that we setup for the start of
409  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
410  *	reference count all in here rather then in the parent.  If we want
411  *	the parent to do more sophisticated things we may have to change
412  *	the ordering.
413  */
414 int
415 vm_pageout_flush(vm_page_t *mc, int count, int flags)
416 {
417 	vm_object_t object = mc[0]->object;
418 	int pageout_status[count];
419 	int numpagedout = 0;
420 	int i;
421 
422 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
423 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
424 	/*
425 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
426 	 * mark the pages read-only.
427 	 *
428 	 * We do not have to fixup the clean/dirty bits here... we can
429 	 * allow the pager to do it after the I/O completes.
430 	 *
431 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
432 	 * edge case with file fragments.
433 	 */
434 	for (i = 0; i < count; i++) {
435 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
436 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
437 			mc[i], i, count));
438 		vm_page_io_start(mc[i]);
439 		pmap_remove_write(mc[i]);
440 	}
441 	vm_page_unlock_queues();
442 	vm_object_pip_add(object, count);
443 
444 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
445 
446 	vm_page_lock_queues();
447 	for (i = 0; i < count; i++) {
448 		vm_page_t mt = mc[i];
449 
450 		KASSERT((mt->flags & PG_WRITEABLE) == 0,
451 		    ("vm_pageout_flush: page %p is not write protected", mt));
452 		switch (pageout_status[i]) {
453 		case VM_PAGER_OK:
454 		case VM_PAGER_PEND:
455 			numpagedout++;
456 			break;
457 		case VM_PAGER_BAD:
458 			/*
459 			 * Page outside of range of object. Right now we
460 			 * essentially lose the changes by pretending it
461 			 * worked.
462 			 */
463 			pmap_clear_modify(mt);
464 			vm_page_undirty(mt);
465 			break;
466 		case VM_PAGER_ERROR:
467 		case VM_PAGER_FAIL:
468 			/*
469 			 * If page couldn't be paged out, then reactivate the
470 			 * page so it doesn't clog the inactive list.  (We
471 			 * will try paging out it again later).
472 			 */
473 			vm_page_activate(mt);
474 			break;
475 		case VM_PAGER_AGAIN:
476 			break;
477 		}
478 
479 		/*
480 		 * If the operation is still going, leave the page busy to
481 		 * block all other accesses. Also, leave the paging in
482 		 * progress indicator set so that we don't attempt an object
483 		 * collapse.
484 		 */
485 		if (pageout_status[i] != VM_PAGER_PEND) {
486 			vm_object_pip_wakeup(object);
487 			vm_page_io_finish(mt);
488 			if (vm_page_count_severe())
489 				vm_page_try_to_cache(mt);
490 		}
491 	}
492 	return numpagedout;
493 }
494 
495 #if !defined(NO_SWAPPING)
496 /*
497  *	vm_pageout_object_deactivate_pages
498  *
499  *	deactivate enough pages to satisfy the inactive target
500  *	requirements or if vm_page_proc_limit is set, then
501  *	deactivate all of the pages in the object and its
502  *	backing_objects.
503  *
504  *	The object and map must be locked.
505  */
506 static void
507 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
508 	pmap_t pmap;
509 	vm_object_t first_object;
510 	long desired;
511 {
512 	vm_object_t backing_object, object;
513 	vm_page_t p, next;
514 	int actcount, rcount, remove_mode;
515 
516 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
517 	if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
518 		return;
519 	for (object = first_object;; object = backing_object) {
520 		if (pmap_resident_count(pmap) <= desired)
521 			goto unlock_return;
522 		if (object->paging_in_progress)
523 			goto unlock_return;
524 
525 		remove_mode = 0;
526 		if (object->shadow_count > 1)
527 			remove_mode = 1;
528 		/*
529 		 * scan the objects entire memory queue
530 		 */
531 		rcount = object->resident_page_count;
532 		p = TAILQ_FIRST(&object->memq);
533 		vm_page_lock_queues();
534 		while (p && (rcount-- > 0)) {
535 			if (pmap_resident_count(pmap) <= desired) {
536 				vm_page_unlock_queues();
537 				goto unlock_return;
538 			}
539 			next = TAILQ_NEXT(p, listq);
540 			cnt.v_pdpages++;
541 			if (p->wire_count != 0 ||
542 			    p->hold_count != 0 ||
543 			    p->busy != 0 ||
544 			    (p->oflags & VPO_BUSY) ||
545 			    (p->flags & PG_UNMANAGED) ||
546 			    !pmap_page_exists_quick(pmap, p)) {
547 				p = next;
548 				continue;
549 			}
550 			actcount = pmap_ts_referenced(p);
551 			if (actcount) {
552 				vm_page_flag_set(p, PG_REFERENCED);
553 			} else if (p->flags & PG_REFERENCED) {
554 				actcount = 1;
555 			}
556 			if ((p->queue != PQ_ACTIVE) &&
557 				(p->flags & PG_REFERENCED)) {
558 				vm_page_activate(p);
559 				p->act_count += actcount;
560 				vm_page_flag_clear(p, PG_REFERENCED);
561 			} else if (p->queue == PQ_ACTIVE) {
562 				if ((p->flags & PG_REFERENCED) == 0) {
563 					p->act_count -= min(p->act_count, ACT_DECLINE);
564 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
565 						pmap_remove_all(p);
566 						vm_page_deactivate(p);
567 					} else {
568 						vm_pageq_requeue(p);
569 					}
570 				} else {
571 					vm_page_activate(p);
572 					vm_page_flag_clear(p, PG_REFERENCED);
573 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
574 						p->act_count += ACT_ADVANCE;
575 					vm_pageq_requeue(p);
576 				}
577 			} else if (p->queue == PQ_INACTIVE) {
578 				pmap_remove_all(p);
579 			}
580 			p = next;
581 		}
582 		vm_page_unlock_queues();
583 		if ((backing_object = object->backing_object) == NULL)
584 			goto unlock_return;
585 		VM_OBJECT_LOCK(backing_object);
586 		if (object != first_object)
587 			VM_OBJECT_UNLOCK(object);
588 	}
589 unlock_return:
590 	if (object != first_object)
591 		VM_OBJECT_UNLOCK(object);
592 }
593 
594 /*
595  * deactivate some number of pages in a map, try to do it fairly, but
596  * that is really hard to do.
597  */
598 static void
599 vm_pageout_map_deactivate_pages(map, desired)
600 	vm_map_t map;
601 	long desired;
602 {
603 	vm_map_entry_t tmpe;
604 	vm_object_t obj, bigobj;
605 	int nothingwired;
606 
607 	if (!vm_map_trylock(map))
608 		return;
609 
610 	bigobj = NULL;
611 	nothingwired = TRUE;
612 
613 	/*
614 	 * first, search out the biggest object, and try to free pages from
615 	 * that.
616 	 */
617 	tmpe = map->header.next;
618 	while (tmpe != &map->header) {
619 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
620 			obj = tmpe->object.vm_object;
621 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
622 				if (obj->shadow_count <= 1 &&
623 				    (bigobj == NULL ||
624 				     bigobj->resident_page_count < obj->resident_page_count)) {
625 					if (bigobj != NULL)
626 						VM_OBJECT_UNLOCK(bigobj);
627 					bigobj = obj;
628 				} else
629 					VM_OBJECT_UNLOCK(obj);
630 			}
631 		}
632 		if (tmpe->wired_count > 0)
633 			nothingwired = FALSE;
634 		tmpe = tmpe->next;
635 	}
636 
637 	if (bigobj != NULL) {
638 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
639 		VM_OBJECT_UNLOCK(bigobj);
640 	}
641 	/*
642 	 * Next, hunt around for other pages to deactivate.  We actually
643 	 * do this search sort of wrong -- .text first is not the best idea.
644 	 */
645 	tmpe = map->header.next;
646 	while (tmpe != &map->header) {
647 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
648 			break;
649 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
650 			obj = tmpe->object.vm_object;
651 			if (obj != NULL) {
652 				VM_OBJECT_LOCK(obj);
653 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
654 				VM_OBJECT_UNLOCK(obj);
655 			}
656 		}
657 		tmpe = tmpe->next;
658 	}
659 
660 	/*
661 	 * Remove all mappings if a process is swapped out, this will free page
662 	 * table pages.
663 	 */
664 	if (desired == 0 && nothingwired) {
665 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
666 		    vm_map_max(map));
667 	}
668 	vm_map_unlock(map);
669 }
670 #endif		/* !defined(NO_SWAPPING) */
671 
672 /*
673  *	vm_pageout_scan does the dirty work for the pageout daemon.
674  */
675 static void
676 vm_pageout_scan(int pass)
677 {
678 	vm_page_t m, next;
679 	struct vm_page marker;
680 	int page_shortage, maxscan, pcount;
681 	int addl_page_shortage, addl_page_shortage_init;
682 	struct proc *p, *bigproc;
683 	struct thread *td;
684 	vm_offset_t size, bigsize;
685 	vm_object_t object;
686 	int actcount;
687 	int vnodes_skipped = 0;
688 	int maxlaunder;
689 
690 	/*
691 	 * Decrease registered cache sizes.
692 	 */
693 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
694 	/*
695 	 * We do this explicitly after the caches have been drained above.
696 	 */
697 	uma_reclaim();
698 
699 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
700 
701 	/*
702 	 * Calculate the number of pages we want to either free or move
703 	 * to the cache.
704 	 */
705 	page_shortage = vm_paging_target() + addl_page_shortage_init;
706 
707 	/*
708 	 * Initialize our marker
709 	 */
710 	bzero(&marker, sizeof(marker));
711 	marker.flags = PG_FICTITIOUS | PG_MARKER;
712 	marker.oflags = VPO_BUSY;
713 	marker.queue = PQ_INACTIVE;
714 	marker.wire_count = 1;
715 
716 	/*
717 	 * Start scanning the inactive queue for pages we can move to the
718 	 * cache or free.  The scan will stop when the target is reached or
719 	 * we have scanned the entire inactive queue.  Note that m->act_count
720 	 * is not used to form decisions for the inactive queue, only for the
721 	 * active queue.
722 	 *
723 	 * maxlaunder limits the number of dirty pages we flush per scan.
724 	 * For most systems a smaller value (16 or 32) is more robust under
725 	 * extreme memory and disk pressure because any unnecessary writes
726 	 * to disk can result in extreme performance degredation.  However,
727 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
728 	 * used) will die horribly with limited laundering.  If the pageout
729 	 * daemon cannot clean enough pages in the first pass, we let it go
730 	 * all out in succeeding passes.
731 	 */
732 	if ((maxlaunder = vm_max_launder) <= 1)
733 		maxlaunder = 1;
734 	if (pass)
735 		maxlaunder = 10000;
736 	vm_page_lock_queues();
737 rescan0:
738 	addl_page_shortage = addl_page_shortage_init;
739 	maxscan = cnt.v_inactive_count;
740 
741 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
742 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
743 	     m = next) {
744 
745 		cnt.v_pdpages++;
746 
747 		if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
748 			goto rescan0;
749 		}
750 
751 		next = TAILQ_NEXT(m, pageq);
752 		object = m->object;
753 
754 		/*
755 		 * skip marker pages
756 		 */
757 		if (m->flags & PG_MARKER)
758 			continue;
759 
760 		/*
761 		 * A held page may be undergoing I/O, so skip it.
762 		 */
763 		if (m->hold_count) {
764 			vm_pageq_requeue(m);
765 			addl_page_shortage++;
766 			continue;
767 		}
768 		/*
769 		 * Don't mess with busy pages, keep in the front of the
770 		 * queue, most likely are being paged out.
771 		 */
772 		if (!VM_OBJECT_TRYLOCK(object) &&
773 		    (!vm_pageout_fallback_object_lock(m, &next) ||
774 		     m->hold_count != 0)) {
775 			VM_OBJECT_UNLOCK(object);
776 			addl_page_shortage++;
777 			continue;
778 		}
779 		if (m->busy || (m->oflags & VPO_BUSY)) {
780 			VM_OBJECT_UNLOCK(object);
781 			addl_page_shortage++;
782 			continue;
783 		}
784 
785 		/*
786 		 * If the object is not being used, we ignore previous
787 		 * references.
788 		 */
789 		if (object->ref_count == 0) {
790 			vm_page_flag_clear(m, PG_REFERENCED);
791 			pmap_clear_reference(m);
792 
793 		/*
794 		 * Otherwise, if the page has been referenced while in the
795 		 * inactive queue, we bump the "activation count" upwards,
796 		 * making it less likely that the page will be added back to
797 		 * the inactive queue prematurely again.  Here we check the
798 		 * page tables (or emulated bits, if any), given the upper
799 		 * level VM system not knowing anything about existing
800 		 * references.
801 		 */
802 		} else if (((m->flags & PG_REFERENCED) == 0) &&
803 			(actcount = pmap_ts_referenced(m))) {
804 			vm_page_activate(m);
805 			VM_OBJECT_UNLOCK(object);
806 			m->act_count += (actcount + ACT_ADVANCE);
807 			continue;
808 		}
809 
810 		/*
811 		 * If the upper level VM system knows about any page
812 		 * references, we activate the page.  We also set the
813 		 * "activation count" higher than normal so that we will less
814 		 * likely place pages back onto the inactive queue again.
815 		 */
816 		if ((m->flags & PG_REFERENCED) != 0) {
817 			vm_page_flag_clear(m, PG_REFERENCED);
818 			actcount = pmap_ts_referenced(m);
819 			vm_page_activate(m);
820 			VM_OBJECT_UNLOCK(object);
821 			m->act_count += (actcount + ACT_ADVANCE + 1);
822 			continue;
823 		}
824 
825 		/*
826 		 * If the upper level VM system doesn't know anything about
827 		 * the page being dirty, we have to check for it again.  As
828 		 * far as the VM code knows, any partially dirty pages are
829 		 * fully dirty.
830 		 */
831 		if (m->dirty == 0 && !pmap_is_modified(m)) {
832 			/*
833 			 * Avoid a race condition: Unless write access is
834 			 * removed from the page, another processor could
835 			 * modify it before all access is removed by the call
836 			 * to vm_page_cache() below.  If vm_page_cache() finds
837 			 * that the page has been modified when it removes all
838 			 * access, it panics because it cannot cache dirty
839 			 * pages.  In principle, we could eliminate just write
840 			 * access here rather than all access.  In the expected
841 			 * case, when there are no last instant modifications
842 			 * to the page, removing all access will be cheaper
843 			 * overall.
844 			 */
845 			if ((m->flags & PG_WRITEABLE) != 0)
846 				pmap_remove_all(m);
847 		} else {
848 			vm_page_dirty(m);
849 		}
850 
851 		if (m->valid == 0) {
852 			/*
853 			 * Invalid pages can be easily freed
854 			 */
855 			vm_page_free(m);
856 			cnt.v_dfree++;
857 			--page_shortage;
858 		} else if (m->dirty == 0) {
859 			/*
860 			 * Clean pages can be placed onto the cache queue.
861 			 * This effectively frees them.
862 			 */
863 			vm_page_cache(m);
864 			--page_shortage;
865 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
866 			/*
867 			 * Dirty pages need to be paged out, but flushing
868 			 * a page is extremely expensive verses freeing
869 			 * a clean page.  Rather then artificially limiting
870 			 * the number of pages we can flush, we instead give
871 			 * dirty pages extra priority on the inactive queue
872 			 * by forcing them to be cycled through the queue
873 			 * twice before being flushed, after which the
874 			 * (now clean) page will cycle through once more
875 			 * before being freed.  This significantly extends
876 			 * the thrash point for a heavily loaded machine.
877 			 */
878 			vm_page_flag_set(m, PG_WINATCFLS);
879 			vm_pageq_requeue(m);
880 		} else if (maxlaunder > 0) {
881 			/*
882 			 * We always want to try to flush some dirty pages if
883 			 * we encounter them, to keep the system stable.
884 			 * Normally this number is small, but under extreme
885 			 * pressure where there are insufficient clean pages
886 			 * on the inactive queue, we may have to go all out.
887 			 */
888 			int swap_pageouts_ok, vfslocked = 0;
889 			struct vnode *vp = NULL;
890 			struct mount *mp = NULL;
891 
892 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
893 				swap_pageouts_ok = 1;
894 			} else {
895 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
896 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
897 				vm_page_count_min());
898 
899 			}
900 
901 			/*
902 			 * We don't bother paging objects that are "dead".
903 			 * Those objects are in a "rundown" state.
904 			 */
905 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
906 				VM_OBJECT_UNLOCK(object);
907 				vm_pageq_requeue(m);
908 				continue;
909 			}
910 
911 			/*
912 			 * Following operations may unlock
913 			 * vm_page_queue_mtx, invalidating the 'next'
914 			 * pointer.  To prevent an inordinate number
915 			 * of restarts we use our marker to remember
916 			 * our place.
917 			 *
918 			 */
919 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
920 					   m, &marker, pageq);
921 			/*
922 			 * The object is already known NOT to be dead.   It
923 			 * is possible for the vget() to block the whole
924 			 * pageout daemon, but the new low-memory handling
925 			 * code should prevent it.
926 			 *
927 			 * The previous code skipped locked vnodes and, worse,
928 			 * reordered pages in the queue.  This results in
929 			 * completely non-deterministic operation and, on a
930 			 * busy system, can lead to extremely non-optimal
931 			 * pageouts.  For example, it can cause clean pages
932 			 * to be freed and dirty pages to be moved to the end
933 			 * of the queue.  Since dirty pages are also moved to
934 			 * the end of the queue once-cleaned, this gives
935 			 * way too large a weighting to defering the freeing
936 			 * of dirty pages.
937 			 *
938 			 * We can't wait forever for the vnode lock, we might
939 			 * deadlock due to a vn_read() getting stuck in
940 			 * vm_wait while holding this vnode.  We skip the
941 			 * vnode if we can't get it in a reasonable amount
942 			 * of time.
943 			 */
944 			if (object->type == OBJT_VNODE) {
945 				vp = object->handle;
946 				if (vp->v_type == VREG &&
947 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
948 					KASSERT(mp == NULL,
949 					    ("vm_pageout_scan: mp != NULL"));
950 					++pageout_lock_miss;
951 					if (object->flags & OBJ_MIGHTBEDIRTY)
952 						vnodes_skipped++;
953 					goto unlock_and_continue;
954 				}
955 				vm_page_unlock_queues();
956 				vm_object_reference_locked(object);
957 				VM_OBJECT_UNLOCK(object);
958 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
959 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
960 				    curthread)) {
961 					VM_OBJECT_LOCK(object);
962 					vm_page_lock_queues();
963 					++pageout_lock_miss;
964 					if (object->flags & OBJ_MIGHTBEDIRTY)
965 						vnodes_skipped++;
966 					vp = NULL;
967 					goto unlock_and_continue;
968 				}
969 				VM_OBJECT_LOCK(object);
970 				vm_page_lock_queues();
971 				/*
972 				 * The page might have been moved to another
973 				 * queue during potential blocking in vget()
974 				 * above.  The page might have been freed and
975 				 * reused for another vnode.
976 				 */
977 				if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
978 				    m->object != object ||
979 				    TAILQ_NEXT(m, pageq) != &marker) {
980 					if (object->flags & OBJ_MIGHTBEDIRTY)
981 						vnodes_skipped++;
982 					goto unlock_and_continue;
983 				}
984 
985 				/*
986 				 * The page may have been busied during the
987 				 * blocking in vget().  We don't move the
988 				 * page back onto the end of the queue so that
989 				 * statistics are more correct if we don't.
990 				 */
991 				if (m->busy || (m->oflags & VPO_BUSY)) {
992 					goto unlock_and_continue;
993 				}
994 
995 				/*
996 				 * If the page has become held it might
997 				 * be undergoing I/O, so skip it
998 				 */
999 				if (m->hold_count) {
1000 					vm_pageq_requeue(m);
1001 					if (object->flags & OBJ_MIGHTBEDIRTY)
1002 						vnodes_skipped++;
1003 					goto unlock_and_continue;
1004 				}
1005 			}
1006 
1007 			/*
1008 			 * If a page is dirty, then it is either being washed
1009 			 * (but not yet cleaned) or it is still in the
1010 			 * laundry.  If it is still in the laundry, then we
1011 			 * start the cleaning operation.
1012 			 *
1013 			 * decrement page_shortage on success to account for
1014 			 * the (future) cleaned page.  Otherwise we could wind
1015 			 * up laundering or cleaning too many pages.
1016 			 */
1017 			if (vm_pageout_clean(m) != 0) {
1018 				--page_shortage;
1019 				--maxlaunder;
1020 			}
1021 unlock_and_continue:
1022 			VM_OBJECT_UNLOCK(object);
1023 			if (mp != NULL) {
1024 				vm_page_unlock_queues();
1025 				if (vp != NULL)
1026 					vput(vp);
1027 				VFS_UNLOCK_GIANT(vfslocked);
1028 				vm_object_deallocate(object);
1029 				vn_finished_write(mp);
1030 				vm_page_lock_queues();
1031 			}
1032 			next = TAILQ_NEXT(&marker, pageq);
1033 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1034 				     &marker, pageq);
1035 			continue;
1036 		}
1037 		VM_OBJECT_UNLOCK(object);
1038 	}
1039 
1040 	/*
1041 	 * Compute the number of pages we want to try to move from the
1042 	 * active queue to the inactive queue.
1043 	 */
1044 	page_shortage = vm_paging_target() +
1045 		cnt.v_inactive_target - cnt.v_inactive_count;
1046 	page_shortage += addl_page_shortage;
1047 
1048 	/*
1049 	 * Scan the active queue for things we can deactivate. We nominally
1050 	 * track the per-page activity counter and use it to locate
1051 	 * deactivation candidates.
1052 	 */
1053 	pcount = cnt.v_active_count;
1054 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1055 
1056 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1057 
1058 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1059 		    ("vm_pageout_scan: page %p isn't active", m));
1060 
1061 		next = TAILQ_NEXT(m, pageq);
1062 		object = m->object;
1063 		if ((m->flags & PG_MARKER) != 0) {
1064 			m = next;
1065 			continue;
1066 		}
1067 		if (!VM_OBJECT_TRYLOCK(object) &&
1068 		    !vm_pageout_fallback_object_lock(m, &next)) {
1069 			VM_OBJECT_UNLOCK(object);
1070 			m = next;
1071 			continue;
1072 		}
1073 
1074 		/*
1075 		 * Don't deactivate pages that are busy.
1076 		 */
1077 		if ((m->busy != 0) ||
1078 		    (m->oflags & VPO_BUSY) ||
1079 		    (m->hold_count != 0)) {
1080 			VM_OBJECT_UNLOCK(object);
1081 			vm_pageq_requeue(m);
1082 			m = next;
1083 			continue;
1084 		}
1085 
1086 		/*
1087 		 * The count for pagedaemon pages is done after checking the
1088 		 * page for eligibility...
1089 		 */
1090 		cnt.v_pdpages++;
1091 
1092 		/*
1093 		 * Check to see "how much" the page has been used.
1094 		 */
1095 		actcount = 0;
1096 		if (object->ref_count != 0) {
1097 			if (m->flags & PG_REFERENCED) {
1098 				actcount += 1;
1099 			}
1100 			actcount += pmap_ts_referenced(m);
1101 			if (actcount) {
1102 				m->act_count += ACT_ADVANCE + actcount;
1103 				if (m->act_count > ACT_MAX)
1104 					m->act_count = ACT_MAX;
1105 			}
1106 		}
1107 
1108 		/*
1109 		 * Since we have "tested" this bit, we need to clear it now.
1110 		 */
1111 		vm_page_flag_clear(m, PG_REFERENCED);
1112 
1113 		/*
1114 		 * Only if an object is currently being used, do we use the
1115 		 * page activation count stats.
1116 		 */
1117 		if (actcount && (object->ref_count != 0)) {
1118 			vm_pageq_requeue(m);
1119 		} else {
1120 			m->act_count -= min(m->act_count, ACT_DECLINE);
1121 			if (vm_pageout_algorithm ||
1122 			    object->ref_count == 0 ||
1123 			    m->act_count == 0) {
1124 				page_shortage--;
1125 				if (object->ref_count == 0) {
1126 					pmap_remove_all(m);
1127 					if (m->dirty == 0)
1128 						vm_page_cache(m);
1129 					else
1130 						vm_page_deactivate(m);
1131 				} else {
1132 					vm_page_deactivate(m);
1133 				}
1134 			} else {
1135 				vm_pageq_requeue(m);
1136 			}
1137 		}
1138 		VM_OBJECT_UNLOCK(object);
1139 		m = next;
1140 	}
1141 
1142 	/*
1143 	 * We try to maintain some *really* free pages, this allows interrupt
1144 	 * code to be guaranteed space.  Since both cache and free queues
1145 	 * are considered basically 'free', moving pages from cache to free
1146 	 * does not effect other calculations.
1147 	 */
1148 	while (cnt.v_free_count < cnt.v_free_reserved) {
1149 		TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE].pl, pageq) {
1150 			KASSERT(m->dirty == 0,
1151 			    ("Found dirty cache page %p", m));
1152 			KASSERT(!pmap_page_is_mapped(m),
1153 			    ("Found mapped cache page %p", m));
1154 			KASSERT((m->flags & PG_UNMANAGED) == 0,
1155 			    ("Found unmanaged cache page %p", m));
1156 			KASSERT(m->wire_count == 0,
1157 			    ("Found wired cache page %p", m));
1158 			if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object =
1159 			    m->object)) {
1160 				KASSERT((m->oflags & VPO_BUSY) == 0 &&
1161 				    m->busy == 0, ("Found busy cache page %p",
1162 				    m));
1163 				vm_page_free(m);
1164 				VM_OBJECT_UNLOCK(object);
1165 				cnt.v_dfree++;
1166 				break;
1167 			}
1168 		}
1169 		if (m == NULL)
1170 			break;
1171 	}
1172 	vm_page_unlock_queues();
1173 #if !defined(NO_SWAPPING)
1174 	/*
1175 	 * Idle process swapout -- run once per second.
1176 	 */
1177 	if (vm_swap_idle_enabled) {
1178 		static long lsec;
1179 		if (time_second != lsec) {
1180 			vm_req_vmdaemon(VM_SWAP_IDLE);
1181 			lsec = time_second;
1182 		}
1183 	}
1184 #endif
1185 
1186 	/*
1187 	 * If we didn't get enough free pages, and we have skipped a vnode
1188 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1189 	 * if we did not get enough free pages.
1190 	 */
1191 	if (vm_paging_target() > 0) {
1192 		if (vnodes_skipped && vm_page_count_min())
1193 			(void) speedup_syncer();
1194 #if !defined(NO_SWAPPING)
1195 		if (vm_swap_enabled && vm_page_count_target())
1196 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1197 #endif
1198 	}
1199 
1200 	/*
1201 	 * If we are critically low on one of RAM or swap and low on
1202 	 * the other, kill the largest process.  However, we avoid
1203 	 * doing this on the first pass in order to give ourselves a
1204 	 * chance to flush out dirty vnode-backed pages and to allow
1205 	 * active pages to be moved to the inactive queue and reclaimed.
1206 	 *
1207 	 * We keep the process bigproc locked once we find it to keep anyone
1208 	 * from messing with it; however, there is a possibility of
1209 	 * deadlock if process B is bigproc and one of it's child processes
1210 	 * attempts to propagate a signal to B while we are waiting for A's
1211 	 * lock while walking this list.  To avoid this, we don't block on
1212 	 * the process lock but just skip a process if it is already locked.
1213 	 */
1214 	if (pass != 0 &&
1215 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1216 	     (swap_pager_full && vm_paging_target() > 0))) {
1217 		bigproc = NULL;
1218 		bigsize = 0;
1219 		sx_slock(&allproc_lock);
1220 		FOREACH_PROC_IN_SYSTEM(p) {
1221 			int breakout;
1222 
1223 			if (PROC_TRYLOCK(p) == 0)
1224 				continue;
1225 			/*
1226 			 * If this is a system or protected process, skip it.
1227 			 */
1228 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1229 			    (p->p_flag & P_PROTECTED) ||
1230 			    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1231 				PROC_UNLOCK(p);
1232 				continue;
1233 			}
1234 			/*
1235 			 * If the process is in a non-running type state,
1236 			 * don't touch it.  Check all the threads individually.
1237 			 */
1238 			PROC_SLOCK(p);
1239 			breakout = 0;
1240 			FOREACH_THREAD_IN_PROC(p, td) {
1241 				thread_lock(td);
1242 				if (!TD_ON_RUNQ(td) &&
1243 				    !TD_IS_RUNNING(td) &&
1244 				    !TD_IS_SLEEPING(td)) {
1245 					thread_unlock(td);
1246 					breakout = 1;
1247 					break;
1248 				}
1249 				thread_unlock(td);
1250 			}
1251 			PROC_SUNLOCK(p);
1252 			if (breakout) {
1253 				PROC_UNLOCK(p);
1254 				continue;
1255 			}
1256 			/*
1257 			 * get the process size
1258 			 */
1259 			if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1260 				PROC_UNLOCK(p);
1261 				continue;
1262 			}
1263 			size = vmspace_swap_count(p->p_vmspace);
1264 			vm_map_unlock_read(&p->p_vmspace->vm_map);
1265 			size += vmspace_resident_count(p->p_vmspace);
1266 			/*
1267 			 * if the this process is bigger than the biggest one
1268 			 * remember it.
1269 			 */
1270 			if (size > bigsize) {
1271 				if (bigproc != NULL)
1272 					PROC_UNLOCK(bigproc);
1273 				bigproc = p;
1274 				bigsize = size;
1275 			} else
1276 				PROC_UNLOCK(p);
1277 		}
1278 		sx_sunlock(&allproc_lock);
1279 		if (bigproc != NULL) {
1280 			killproc(bigproc, "out of swap space");
1281 			PROC_SLOCK(bigproc);
1282 			sched_nice(bigproc, PRIO_MIN);
1283 			PROC_SUNLOCK(bigproc);
1284 			PROC_UNLOCK(bigproc);
1285 			wakeup(&cnt.v_free_count);
1286 		}
1287 	}
1288 }
1289 
1290 /*
1291  * This routine tries to maintain the pseudo LRU active queue,
1292  * so that during long periods of time where there is no paging,
1293  * that some statistic accumulation still occurs.  This code
1294  * helps the situation where paging just starts to occur.
1295  */
1296 static void
1297 vm_pageout_page_stats()
1298 {
1299 	vm_object_t object;
1300 	vm_page_t m,next;
1301 	int pcount,tpcount;		/* Number of pages to check */
1302 	static int fullintervalcount = 0;
1303 	int page_shortage;
1304 
1305 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1306 	page_shortage =
1307 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1308 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1309 
1310 	if (page_shortage <= 0)
1311 		return;
1312 
1313 	pcount = cnt.v_active_count;
1314 	fullintervalcount += vm_pageout_stats_interval;
1315 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1316 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1317 		if (pcount > tpcount)
1318 			pcount = tpcount;
1319 	} else {
1320 		fullintervalcount = 0;
1321 	}
1322 
1323 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1324 	while ((m != NULL) && (pcount-- > 0)) {
1325 		int actcount;
1326 
1327 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1328 		    ("vm_pageout_page_stats: page %p isn't active", m));
1329 
1330 		next = TAILQ_NEXT(m, pageq);
1331 		object = m->object;
1332 
1333 		if ((m->flags & PG_MARKER) != 0) {
1334 			m = next;
1335 			continue;
1336 		}
1337 		if (!VM_OBJECT_TRYLOCK(object) &&
1338 		    !vm_pageout_fallback_object_lock(m, &next)) {
1339 			VM_OBJECT_UNLOCK(object);
1340 			m = next;
1341 			continue;
1342 		}
1343 
1344 		/*
1345 		 * Don't deactivate pages that are busy.
1346 		 */
1347 		if ((m->busy != 0) ||
1348 		    (m->oflags & VPO_BUSY) ||
1349 		    (m->hold_count != 0)) {
1350 			VM_OBJECT_UNLOCK(object);
1351 			vm_pageq_requeue(m);
1352 			m = next;
1353 			continue;
1354 		}
1355 
1356 		actcount = 0;
1357 		if (m->flags & PG_REFERENCED) {
1358 			vm_page_flag_clear(m, PG_REFERENCED);
1359 			actcount += 1;
1360 		}
1361 
1362 		actcount += pmap_ts_referenced(m);
1363 		if (actcount) {
1364 			m->act_count += ACT_ADVANCE + actcount;
1365 			if (m->act_count > ACT_MAX)
1366 				m->act_count = ACT_MAX;
1367 			vm_pageq_requeue(m);
1368 		} else {
1369 			if (m->act_count == 0) {
1370 				/*
1371 				 * We turn off page access, so that we have
1372 				 * more accurate RSS stats.  We don't do this
1373 				 * in the normal page deactivation when the
1374 				 * system is loaded VM wise, because the
1375 				 * cost of the large number of page protect
1376 				 * operations would be higher than the value
1377 				 * of doing the operation.
1378 				 */
1379 				pmap_remove_all(m);
1380 				vm_page_deactivate(m);
1381 			} else {
1382 				m->act_count -= min(m->act_count, ACT_DECLINE);
1383 				vm_pageq_requeue(m);
1384 			}
1385 		}
1386 		VM_OBJECT_UNLOCK(object);
1387 		m = next;
1388 	}
1389 }
1390 
1391 /*
1392  *	vm_pageout is the high level pageout daemon.
1393  */
1394 static void
1395 vm_pageout()
1396 {
1397 	int error, pass;
1398 
1399 	/*
1400 	 * Initialize some paging parameters.
1401 	 */
1402 	cnt.v_interrupt_free_min = 2;
1403 	if (cnt.v_page_count < 2000)
1404 		vm_pageout_page_count = 8;
1405 
1406 	/*
1407 	 * v_free_reserved needs to include enough for the largest
1408 	 * swap pager structures plus enough for any pv_entry structs
1409 	 * when paging.
1410 	 */
1411 	if (cnt.v_page_count > 1024)
1412 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1413 	else
1414 		cnt.v_free_min = 4;
1415 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1416 	    cnt.v_interrupt_free_min;
1417 	cnt.v_free_reserved = vm_pageout_page_count +
1418 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1419 	cnt.v_free_severe = cnt.v_free_min / 2;
1420 	cnt.v_free_min += cnt.v_free_reserved;
1421 	cnt.v_free_severe += cnt.v_free_reserved;
1422 
1423 	/*
1424 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1425 	 * that these are more a measure of the VM cache queue hysteresis
1426 	 * then the VM free queue.  Specifically, v_free_target is the
1427 	 * high water mark (free+cache pages).
1428 	 *
1429 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1430 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1431 	 * be big enough to handle memory needs while the pageout daemon
1432 	 * is signalled and run to free more pages.
1433 	 */
1434 	if (cnt.v_free_count > 6144)
1435 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1436 	else
1437 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1438 
1439 	if (cnt.v_free_count > 2048) {
1440 		cnt.v_cache_min = cnt.v_free_target;
1441 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1442 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1443 	} else {
1444 		cnt.v_cache_min = 0;
1445 		cnt.v_cache_max = 0;
1446 		cnt.v_inactive_target = cnt.v_free_count / 4;
1447 	}
1448 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1449 		cnt.v_inactive_target = cnt.v_free_count / 3;
1450 
1451 	/* XXX does not really belong here */
1452 	if (vm_page_max_wired == 0)
1453 		vm_page_max_wired = cnt.v_free_count / 3;
1454 
1455 	if (vm_pageout_stats_max == 0)
1456 		vm_pageout_stats_max = cnt.v_free_target;
1457 
1458 	/*
1459 	 * Set interval in seconds for stats scan.
1460 	 */
1461 	if (vm_pageout_stats_interval == 0)
1462 		vm_pageout_stats_interval = 5;
1463 	if (vm_pageout_full_stats_interval == 0)
1464 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1465 
1466 	swap_pager_swap_init();
1467 	pass = 0;
1468 	/*
1469 	 * The pageout daemon is never done, so loop forever.
1470 	 */
1471 	while (TRUE) {
1472 		/*
1473 		 * If we have enough free memory, wakeup waiters.  Do
1474 		 * not clear vm_pages_needed until we reach our target,
1475 		 * otherwise we may be woken up over and over again and
1476 		 * waste a lot of cpu.
1477 		 */
1478 		mtx_lock(&vm_page_queue_free_mtx);
1479 		if (vm_pages_needed && !vm_page_count_min()) {
1480 			if (!vm_paging_needed())
1481 				vm_pages_needed = 0;
1482 			wakeup(&cnt.v_free_count);
1483 		}
1484 		if (vm_pages_needed) {
1485 			/*
1486 			 * Still not done, take a second pass without waiting
1487 			 * (unlimited dirty cleaning), otherwise sleep a bit
1488 			 * and try again.
1489 			 */
1490 			++pass;
1491 			if (pass > 1)
1492 				msleep(&vm_pages_needed,
1493 				    &vm_page_queue_free_mtx, PVM, "psleep",
1494 				    hz / 2);
1495 		} else {
1496 			/*
1497 			 * Good enough, sleep & handle stats.  Prime the pass
1498 			 * for the next run.
1499 			 */
1500 			if (pass > 1)
1501 				pass = 1;
1502 			else
1503 				pass = 0;
1504 			error = msleep(&vm_pages_needed,
1505 			    &vm_page_queue_free_mtx, PVM, "psleep",
1506 			    vm_pageout_stats_interval * hz);
1507 			if (error && !vm_pages_needed) {
1508 				mtx_unlock(&vm_page_queue_free_mtx);
1509 				pass = 0;
1510 				vm_page_lock_queues();
1511 				vm_pageout_page_stats();
1512 				vm_page_unlock_queues();
1513 				continue;
1514 			}
1515 		}
1516 		if (vm_pages_needed)
1517 			cnt.v_pdwakeups++;
1518 		mtx_unlock(&vm_page_queue_free_mtx);
1519 		vm_pageout_scan(pass);
1520 	}
1521 }
1522 
1523 /*
1524  * Unless the free page queue lock is held by the caller, this function
1525  * should be regarded as advisory.  Specifically, the caller should
1526  * not msleep() on &cnt.v_free_count following this function unless
1527  * the free page queue lock is held until the msleep() is performed.
1528  */
1529 void
1530 pagedaemon_wakeup()
1531 {
1532 
1533 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1534 		vm_pages_needed = 1;
1535 		wakeup(&vm_pages_needed);
1536 	}
1537 }
1538 
1539 #if !defined(NO_SWAPPING)
1540 static void
1541 vm_req_vmdaemon(int req)
1542 {
1543 	static int lastrun = 0;
1544 
1545 	mtx_lock(&vm_daemon_mtx);
1546 	vm_pageout_req_swapout |= req;
1547 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1548 		wakeup(&vm_daemon_needed);
1549 		lastrun = ticks;
1550 	}
1551 	mtx_unlock(&vm_daemon_mtx);
1552 }
1553 
1554 static void
1555 vm_daemon()
1556 {
1557 	struct rlimit rsslim;
1558 	struct proc *p;
1559 	struct thread *td;
1560 	int breakout, swapout_flags;
1561 
1562 	while (TRUE) {
1563 		mtx_lock(&vm_daemon_mtx);
1564 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1565 		swapout_flags = vm_pageout_req_swapout;
1566 		vm_pageout_req_swapout = 0;
1567 		mtx_unlock(&vm_daemon_mtx);
1568 		if (swapout_flags)
1569 			swapout_procs(swapout_flags);
1570 
1571 		/*
1572 		 * scan the processes for exceeding their rlimits or if
1573 		 * process is swapped out -- deactivate pages
1574 		 */
1575 		sx_slock(&allproc_lock);
1576 		FOREACH_PROC_IN_SYSTEM(p) {
1577 			vm_pindex_t limit, size;
1578 
1579 			/*
1580 			 * if this is a system process or if we have already
1581 			 * looked at this process, skip it.
1582 			 */
1583 			PROC_LOCK(p);
1584 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1585 				PROC_UNLOCK(p);
1586 				continue;
1587 			}
1588 			/*
1589 			 * if the process is in a non-running type state,
1590 			 * don't touch it.
1591 			 */
1592 			PROC_SLOCK(p);
1593 			breakout = 0;
1594 			FOREACH_THREAD_IN_PROC(p, td) {
1595 				thread_lock(td);
1596 				if (!TD_ON_RUNQ(td) &&
1597 				    !TD_IS_RUNNING(td) &&
1598 				    !TD_IS_SLEEPING(td)) {
1599 					thread_unlock(td);
1600 					breakout = 1;
1601 					break;
1602 				}
1603 				thread_unlock(td);
1604 			}
1605 			PROC_SUNLOCK(p);
1606 			if (breakout) {
1607 				PROC_UNLOCK(p);
1608 				continue;
1609 			}
1610 			/*
1611 			 * get a limit
1612 			 */
1613 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1614 			limit = OFF_TO_IDX(
1615 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1616 
1617 			/*
1618 			 * let processes that are swapped out really be
1619 			 * swapped out set the limit to nothing (will force a
1620 			 * swap-out.)
1621 			 */
1622 			if ((p->p_sflag & PS_INMEM) == 0)
1623 				limit = 0;	/* XXX */
1624 			PROC_UNLOCK(p);
1625 
1626 			size = vmspace_resident_count(p->p_vmspace);
1627 			if (limit >= 0 && size >= limit) {
1628 				vm_pageout_map_deactivate_pages(
1629 				    &p->p_vmspace->vm_map, limit);
1630 			}
1631 		}
1632 		sx_sunlock(&allproc_lock);
1633 	}
1634 }
1635 #endif			/* !defined(NO_SWAPPING) */
1636