xref: /freebsd/sys/vm/vm_pageout.c (revision c0020399a650364d0134f79f3fa319f84064372d)
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,
127     &page_kp);
128 
129 #if !defined(NO_SWAPPING)
130 /* the kernel process "vm_daemon"*/
131 static void vm_daemon(void);
132 static struct	proc *vmproc;
133 
134 static struct kproc_desc vm_kp = {
135 	"vmdaemon",
136 	vm_daemon,
137 	&vmproc
138 };
139 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
140 #endif
141 
142 
143 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
144 int vm_pageout_deficit;		/* Estimated number of pages deficit */
145 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
146 
147 #if !defined(NO_SWAPPING)
148 static int vm_pageout_req_swapout;	/* XXX */
149 static int vm_daemon_needed;
150 static struct mtx vm_daemon_mtx;
151 /* Allow for use by vm_pageout before vm_daemon is initialized. */
152 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
153 #endif
154 static int vm_max_launder = 32;
155 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
156 static int vm_pageout_full_stats_interval = 0;
157 static int vm_pageout_algorithm=0;
158 static int defer_swap_pageouts=0;
159 static int disable_swap_pageouts=0;
160 
161 #if defined(NO_SWAPPING)
162 static int vm_swap_enabled=0;
163 static int vm_swap_idle_enabled=0;
164 #else
165 static int vm_swap_enabled=1;
166 static int vm_swap_idle_enabled=0;
167 #endif
168 
169 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
170 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
171 
172 SYSCTL_INT(_vm, OID_AUTO, max_launder,
173 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
174 
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
176 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
177 
178 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
179 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
180 
181 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
182 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
183 
184 #if defined(NO_SWAPPING)
185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
186 	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
188 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
189 #else
190 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
191 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
192 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
193 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
194 #endif
195 
196 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
197 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
198 
199 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
200 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
201 
202 static int pageout_lock_miss;
203 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
204 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
205 
206 #define VM_PAGEOUT_PAGE_COUNT 16
207 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
208 
209 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
210 SYSCTL_INT(_vm, OID_AUTO, max_wired,
211 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
212 
213 #if !defined(NO_SWAPPING)
214 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
215 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
216 static void vm_req_vmdaemon(int req);
217 #endif
218 static void vm_pageout_page_stats(void);
219 
220 /*
221  * vm_pageout_fallback_object_lock:
222  *
223  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
224  * known to have failed and page queue must be either PQ_ACTIVE or
225  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
226  * while locking the vm object.  Use marker page to detect page queue
227  * changes and maintain notion of next page on page queue.  Return
228  * TRUE if no changes were detected, FALSE otherwise.  vm object is
229  * locked on return.
230  *
231  * This function depends on both the lock portion of struct vm_object
232  * and normal struct vm_page being type stable.
233  */
234 boolean_t
235 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
236 {
237 	struct vm_page marker;
238 	boolean_t unchanged;
239 	u_short queue;
240 	vm_object_t object;
241 
242 	/*
243 	 * Initialize our marker
244 	 */
245 	bzero(&marker, sizeof(marker));
246 	marker.flags = PG_FICTITIOUS | PG_MARKER;
247 	marker.oflags = VPO_BUSY;
248 	marker.queue = m->queue;
249 	marker.wire_count = 1;
250 
251 	queue = m->queue;
252 	object = m->object;
253 
254 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
255 			   m, &marker, pageq);
256 	vm_page_unlock_queues();
257 	VM_OBJECT_LOCK(object);
258 	vm_page_lock_queues();
259 
260 	/* Page queue might have changed. */
261 	*next = TAILQ_NEXT(&marker, pageq);
262 	unchanged = (m->queue == queue &&
263 		     m->object == object &&
264 		     &marker == TAILQ_NEXT(m, pageq));
265 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
266 		     &marker, pageq);
267 	return (unchanged);
268 }
269 
270 /*
271  * vm_pageout_clean:
272  *
273  * Clean the page and remove it from the laundry.
274  *
275  * We set the busy bit to cause potential page faults on this page to
276  * block.  Note the careful timing, however, the busy bit isn't set till
277  * late and we cannot do anything that will mess with the page.
278  */
279 static int
280 vm_pageout_clean(m)
281 	vm_page_t m;
282 {
283 	vm_object_t object;
284 	vm_page_t mc[2*vm_pageout_page_count];
285 	int pageout_count;
286 	int ib, is, page_base;
287 	vm_pindex_t pindex = m->pindex;
288 
289 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
290 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
291 
292 	/*
293 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
294 	 * with the new swapper, but we could have serious problems paging
295 	 * out other object types if there is insufficient memory.
296 	 *
297 	 * Unfortunately, checking free memory here is far too late, so the
298 	 * check has been moved up a procedural level.
299 	 */
300 
301 	/*
302 	 * Can't clean the page if it's busy or held.
303 	 */
304 	if ((m->hold_count != 0) ||
305 	    ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
306 		return 0;
307 	}
308 
309 	mc[vm_pageout_page_count] = m;
310 	pageout_count = 1;
311 	page_base = vm_pageout_page_count;
312 	ib = 1;
313 	is = 1;
314 
315 	/*
316 	 * Scan object for clusterable pages.
317 	 *
318 	 * We can cluster ONLY if: ->> the page is NOT
319 	 * clean, wired, busy, held, or mapped into a
320 	 * buffer, and one of the following:
321 	 * 1) The page is inactive, or a seldom used
322 	 *    active page.
323 	 * -or-
324 	 * 2) we force the issue.
325 	 *
326 	 * During heavy mmap/modification loads the pageout
327 	 * daemon can really fragment the underlying file
328 	 * due to flushing pages out of order and not trying
329 	 * align the clusters (which leave sporatic out-of-order
330 	 * holes).  To solve this problem we do the reverse scan
331 	 * first and attempt to align our cluster, then do a
332 	 * forward scan if room remains.
333 	 */
334 	object = m->object;
335 more:
336 	while (ib && pageout_count < vm_pageout_page_count) {
337 		vm_page_t p;
338 
339 		if (ib > pindex) {
340 			ib = 0;
341 			break;
342 		}
343 
344 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
345 			ib = 0;
346 			break;
347 		}
348 		if ((p->oflags & VPO_BUSY) || p->busy) {
349 			ib = 0;
350 			break;
351 		}
352 		vm_page_test_dirty(p);
353 		if ((p->dirty & p->valid) == 0 ||
354 		    p->queue != PQ_INACTIVE ||
355 		    p->wire_count != 0 ||	/* may be held by buf cache */
356 		    p->hold_count != 0) {	/* may be undergoing I/O */
357 			ib = 0;
358 			break;
359 		}
360 		mc[--page_base] = p;
361 		++pageout_count;
362 		++ib;
363 		/*
364 		 * alignment boundry, stop here and switch directions.  Do
365 		 * not clear ib.
366 		 */
367 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
368 			break;
369 	}
370 
371 	while (pageout_count < vm_pageout_page_count &&
372 	    pindex + is < object->size) {
373 		vm_page_t p;
374 
375 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
376 			break;
377 		if ((p->oflags & VPO_BUSY) || p->busy) {
378 			break;
379 		}
380 		vm_page_test_dirty(p);
381 		if ((p->dirty & p->valid) == 0 ||
382 		    p->queue != PQ_INACTIVE ||
383 		    p->wire_count != 0 ||	/* may be held by buf cache */
384 		    p->hold_count != 0) {	/* may be undergoing I/O */
385 			break;
386 		}
387 		mc[page_base + pageout_count] = p;
388 		++pageout_count;
389 		++is;
390 	}
391 
392 	/*
393 	 * If we exhausted our forward scan, continue with the reverse scan
394 	 * when possible, even past a page boundry.  This catches boundry
395 	 * conditions.
396 	 */
397 	if (ib && pageout_count < vm_pageout_page_count)
398 		goto more;
399 
400 	/*
401 	 * we allow reads during pageouts...
402 	 */
403 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
404 }
405 
406 /*
407  * vm_pageout_flush() - launder the given pages
408  *
409  *	The given pages are laundered.  Note that we setup for the start of
410  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
411  *	reference count all in here rather then in the parent.  If we want
412  *	the parent to do more sophisticated things we may have to change
413  *	the ordering.
414  */
415 int
416 vm_pageout_flush(vm_page_t *mc, int count, int flags)
417 {
418 	vm_object_t object = mc[0]->object;
419 	int pageout_status[count];
420 	int numpagedout = 0;
421 	int i;
422 
423 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
424 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
425 	/*
426 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
427 	 * mark the pages read-only.
428 	 *
429 	 * We do not have to fixup the clean/dirty bits here... we can
430 	 * allow the pager to do it after the I/O completes.
431 	 *
432 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
433 	 * edge case with file fragments.
434 	 */
435 	for (i = 0; i < count; i++) {
436 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
437 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
438 			mc[i], i, count));
439 		vm_page_io_start(mc[i]);
440 		pmap_remove_write(mc[i]);
441 	}
442 	vm_page_unlock_queues();
443 	vm_object_pip_add(object, count);
444 
445 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
446 
447 	vm_page_lock_queues();
448 	for (i = 0; i < count; i++) {
449 		vm_page_t mt = mc[i];
450 
451 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
452 		    (mt->flags & PG_WRITEABLE) == 0,
453 		    ("vm_pageout_flush: page %p is not write protected", mt));
454 		switch (pageout_status[i]) {
455 		case VM_PAGER_OK:
456 		case VM_PAGER_PEND:
457 			numpagedout++;
458 			break;
459 		case VM_PAGER_BAD:
460 			/*
461 			 * Page outside of range of object. Right now we
462 			 * essentially lose the changes by pretending it
463 			 * worked.
464 			 */
465 			vm_page_undirty(mt);
466 			break;
467 		case VM_PAGER_ERROR:
468 		case VM_PAGER_FAIL:
469 			/*
470 			 * If page couldn't be paged out, then reactivate the
471 			 * page so it doesn't clog the inactive list.  (We
472 			 * will try paging out it again later).
473 			 */
474 			vm_page_activate(mt);
475 			break;
476 		case VM_PAGER_AGAIN:
477 			break;
478 		}
479 
480 		/*
481 		 * If the operation is still going, leave the page busy to
482 		 * block all other accesses. Also, leave the paging in
483 		 * progress indicator set so that we don't attempt an object
484 		 * collapse.
485 		 */
486 		if (pageout_status[i] != VM_PAGER_PEND) {
487 			vm_object_pip_wakeup(object);
488 			vm_page_io_finish(mt);
489 			if (vm_page_count_severe())
490 				vm_page_try_to_cache(mt);
491 		}
492 	}
493 	return numpagedout;
494 }
495 
496 #if !defined(NO_SWAPPING)
497 /*
498  *	vm_pageout_object_deactivate_pages
499  *
500  *	deactivate enough pages to satisfy the inactive target
501  *	requirements or if vm_page_proc_limit is set, then
502  *	deactivate all of the pages in the object and its
503  *	backing_objects.
504  *
505  *	The object and map must be locked.
506  */
507 static void
508 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
509 	pmap_t pmap;
510 	vm_object_t first_object;
511 	long desired;
512 {
513 	vm_object_t backing_object, object;
514 	vm_page_t p, next;
515 	int actcount, rcount, remove_mode;
516 
517 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
518 	if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
519 		return;
520 	for (object = first_object;; object = backing_object) {
521 		if (pmap_resident_count(pmap) <= desired)
522 			goto unlock_return;
523 		if (object->paging_in_progress)
524 			goto unlock_return;
525 
526 		remove_mode = 0;
527 		if (object->shadow_count > 1)
528 			remove_mode = 1;
529 		/*
530 		 * scan the objects entire memory queue
531 		 */
532 		rcount = object->resident_page_count;
533 		p = TAILQ_FIRST(&object->memq);
534 		vm_page_lock_queues();
535 		while (p && (rcount-- > 0)) {
536 			if (pmap_resident_count(pmap) <= desired) {
537 				vm_page_unlock_queues();
538 				goto unlock_return;
539 			}
540 			next = TAILQ_NEXT(p, listq);
541 			cnt.v_pdpages++;
542 			if (p->wire_count != 0 ||
543 			    p->hold_count != 0 ||
544 			    p->busy != 0 ||
545 			    (p->oflags & VPO_BUSY) ||
546 			    (p->flags & PG_UNMANAGED) ||
547 			    !pmap_page_exists_quick(pmap, p)) {
548 				p = next;
549 				continue;
550 			}
551 			actcount = pmap_ts_referenced(p);
552 			if (actcount) {
553 				vm_page_flag_set(p, PG_REFERENCED);
554 			} else if (p->flags & PG_REFERENCED) {
555 				actcount = 1;
556 			}
557 			if ((p->queue != PQ_ACTIVE) &&
558 				(p->flags & PG_REFERENCED)) {
559 				vm_page_activate(p);
560 				p->act_count += actcount;
561 				vm_page_flag_clear(p, PG_REFERENCED);
562 			} else if (p->queue == PQ_ACTIVE) {
563 				if ((p->flags & PG_REFERENCED) == 0) {
564 					p->act_count -= min(p->act_count, ACT_DECLINE);
565 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
566 						pmap_remove_all(p);
567 						vm_page_deactivate(p);
568 					} else {
569 						vm_page_requeue(p);
570 					}
571 				} else {
572 					vm_page_activate(p);
573 					vm_page_flag_clear(p, PG_REFERENCED);
574 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
575 						p->act_count += ACT_ADVANCE;
576 					vm_page_requeue(p);
577 				}
578 			} else if (p->queue == PQ_INACTIVE) {
579 				pmap_remove_all(p);
580 			}
581 			p = next;
582 		}
583 		vm_page_unlock_queues();
584 		if ((backing_object = object->backing_object) == NULL)
585 			goto unlock_return;
586 		VM_OBJECT_LOCK(backing_object);
587 		if (object != first_object)
588 			VM_OBJECT_UNLOCK(object);
589 	}
590 unlock_return:
591 	if (object != first_object)
592 		VM_OBJECT_UNLOCK(object);
593 }
594 
595 /*
596  * deactivate some number of pages in a map, try to do it fairly, but
597  * that is really hard to do.
598  */
599 static void
600 vm_pageout_map_deactivate_pages(map, desired)
601 	vm_map_t map;
602 	long desired;
603 {
604 	vm_map_entry_t tmpe;
605 	vm_object_t obj, bigobj;
606 	int nothingwired;
607 
608 	if (!vm_map_trylock(map))
609 		return;
610 
611 	bigobj = NULL;
612 	nothingwired = TRUE;
613 
614 	/*
615 	 * first, search out the biggest object, and try to free pages from
616 	 * that.
617 	 */
618 	tmpe = map->header.next;
619 	while (tmpe != &map->header) {
620 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
621 			obj = tmpe->object.vm_object;
622 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
623 				if (obj->shadow_count <= 1 &&
624 				    (bigobj == NULL ||
625 				     bigobj->resident_page_count < obj->resident_page_count)) {
626 					if (bigobj != NULL)
627 						VM_OBJECT_UNLOCK(bigobj);
628 					bigobj = obj;
629 				} else
630 					VM_OBJECT_UNLOCK(obj);
631 			}
632 		}
633 		if (tmpe->wired_count > 0)
634 			nothingwired = FALSE;
635 		tmpe = tmpe->next;
636 	}
637 
638 	if (bigobj != NULL) {
639 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
640 		VM_OBJECT_UNLOCK(bigobj);
641 	}
642 	/*
643 	 * Next, hunt around for other pages to deactivate.  We actually
644 	 * do this search sort of wrong -- .text first is not the best idea.
645 	 */
646 	tmpe = map->header.next;
647 	while (tmpe != &map->header) {
648 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
649 			break;
650 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
651 			obj = tmpe->object.vm_object;
652 			if (obj != NULL) {
653 				VM_OBJECT_LOCK(obj);
654 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
655 				VM_OBJECT_UNLOCK(obj);
656 			}
657 		}
658 		tmpe = tmpe->next;
659 	}
660 
661 	/*
662 	 * Remove all mappings if a process is swapped out, this will free page
663 	 * table pages.
664 	 */
665 	if (desired == 0 && nothingwired) {
666 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
667 		    vm_map_max(map));
668 	}
669 	vm_map_unlock(map);
670 }
671 #endif		/* !defined(NO_SWAPPING) */
672 
673 /*
674  *	vm_pageout_scan does the dirty work for the pageout daemon.
675  */
676 static void
677 vm_pageout_scan(int pass)
678 {
679 	vm_page_t m, next;
680 	struct vm_page marker;
681 	int page_shortage, maxscan, pcount;
682 	int addl_page_shortage, addl_page_shortage_init;
683 	vm_object_t object;
684 	int actcount;
685 	int vnodes_skipped = 0;
686 	int maxlaunder;
687 
688 	/*
689 	 * Decrease registered cache sizes.
690 	 */
691 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
692 	/*
693 	 * We do this explicitly after the caches have been drained above.
694 	 */
695 	uma_reclaim();
696 
697 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
698 
699 	/*
700 	 * Calculate the number of pages we want to either free or move
701 	 * to the cache.
702 	 */
703 	page_shortage = vm_paging_target() + addl_page_shortage_init;
704 
705 	/*
706 	 * Initialize our marker
707 	 */
708 	bzero(&marker, sizeof(marker));
709 	marker.flags = PG_FICTITIOUS | PG_MARKER;
710 	marker.oflags = VPO_BUSY;
711 	marker.queue = PQ_INACTIVE;
712 	marker.wire_count = 1;
713 
714 	/*
715 	 * Start scanning the inactive queue for pages we can move to the
716 	 * cache or free.  The scan will stop when the target is reached or
717 	 * we have scanned the entire inactive queue.  Note that m->act_count
718 	 * is not used to form decisions for the inactive queue, only for the
719 	 * active queue.
720 	 *
721 	 * maxlaunder limits the number of dirty pages we flush per scan.
722 	 * For most systems a smaller value (16 or 32) is more robust under
723 	 * extreme memory and disk pressure because any unnecessary writes
724 	 * to disk can result in extreme performance degredation.  However,
725 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
726 	 * used) will die horribly with limited laundering.  If the pageout
727 	 * daemon cannot clean enough pages in the first pass, we let it go
728 	 * all out in succeeding passes.
729 	 */
730 	if ((maxlaunder = vm_max_launder) <= 1)
731 		maxlaunder = 1;
732 	if (pass)
733 		maxlaunder = 10000;
734 	vm_page_lock_queues();
735 rescan0:
736 	addl_page_shortage = addl_page_shortage_init;
737 	maxscan = cnt.v_inactive_count;
738 
739 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
740 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
741 	     m = next) {
742 
743 		cnt.v_pdpages++;
744 
745 		if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
746 			goto rescan0;
747 		}
748 
749 		next = TAILQ_NEXT(m, pageq);
750 		object = m->object;
751 
752 		/*
753 		 * skip marker pages
754 		 */
755 		if (m->flags & PG_MARKER)
756 			continue;
757 
758 		/*
759 		 * A held page may be undergoing I/O, so skip it.
760 		 */
761 		if (m->hold_count) {
762 			vm_page_requeue(m);
763 			addl_page_shortage++;
764 			continue;
765 		}
766 		/*
767 		 * Don't mess with busy pages, keep in the front of the
768 		 * queue, most likely are being paged out.
769 		 */
770 		if (!VM_OBJECT_TRYLOCK(object) &&
771 		    (!vm_pageout_fallback_object_lock(m, &next) ||
772 		     m->hold_count != 0)) {
773 			VM_OBJECT_UNLOCK(object);
774 			addl_page_shortage++;
775 			continue;
776 		}
777 		if (m->busy || (m->oflags & VPO_BUSY)) {
778 			VM_OBJECT_UNLOCK(object);
779 			addl_page_shortage++;
780 			continue;
781 		}
782 
783 		/*
784 		 * If the object is not being used, we ignore previous
785 		 * references.
786 		 */
787 		if (object->ref_count == 0) {
788 			vm_page_flag_clear(m, PG_REFERENCED);
789 			pmap_clear_reference(m);
790 
791 		/*
792 		 * Otherwise, if the page has been referenced while in the
793 		 * inactive queue, we bump the "activation count" upwards,
794 		 * making it less likely that the page will be added back to
795 		 * the inactive queue prematurely again.  Here we check the
796 		 * page tables (or emulated bits, if any), given the upper
797 		 * level VM system not knowing anything about existing
798 		 * references.
799 		 */
800 		} else if (((m->flags & PG_REFERENCED) == 0) &&
801 			(actcount = pmap_ts_referenced(m))) {
802 			vm_page_activate(m);
803 			VM_OBJECT_UNLOCK(object);
804 			m->act_count += (actcount + ACT_ADVANCE);
805 			continue;
806 		}
807 
808 		/*
809 		 * If the upper level VM system knows about any page
810 		 * references, we activate the page.  We also set the
811 		 * "activation count" higher than normal so that we will less
812 		 * likely place pages back onto the inactive queue again.
813 		 */
814 		if ((m->flags & PG_REFERENCED) != 0) {
815 			vm_page_flag_clear(m, PG_REFERENCED);
816 			actcount = pmap_ts_referenced(m);
817 			vm_page_activate(m);
818 			VM_OBJECT_UNLOCK(object);
819 			m->act_count += (actcount + ACT_ADVANCE + 1);
820 			continue;
821 		}
822 
823 		/*
824 		 * If the upper level VM system doesn't know anything about
825 		 * the page being dirty, we have to check for it again.  As
826 		 * far as the VM code knows, any partially dirty pages are
827 		 * fully dirty.
828 		 */
829 		if (m->dirty == 0 && !pmap_is_modified(m)) {
830 			/*
831 			 * Avoid a race condition: Unless write access is
832 			 * removed from the page, another processor could
833 			 * modify it before all access is removed by the call
834 			 * to vm_page_cache() below.  If vm_page_cache() finds
835 			 * that the page has been modified when it removes all
836 			 * access, it panics because it cannot cache dirty
837 			 * pages.  In principle, we could eliminate just write
838 			 * access here rather than all access.  In the expected
839 			 * case, when there are no last instant modifications
840 			 * to the page, removing all access will be cheaper
841 			 * overall.
842 			 */
843 			if ((m->flags & PG_WRITEABLE) != 0)
844 				pmap_remove_all(m);
845 		} else {
846 			vm_page_dirty(m);
847 		}
848 
849 		if (m->valid == 0) {
850 			/*
851 			 * Invalid pages can be easily freed
852 			 */
853 			vm_page_free(m);
854 			cnt.v_dfree++;
855 			--page_shortage;
856 		} else if (m->dirty == 0) {
857 			/*
858 			 * Clean pages can be placed onto the cache queue.
859 			 * This effectively frees them.
860 			 */
861 			vm_page_cache(m);
862 			--page_shortage;
863 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
864 			/*
865 			 * Dirty pages need to be paged out, but flushing
866 			 * a page is extremely expensive verses freeing
867 			 * a clean page.  Rather then artificially limiting
868 			 * the number of pages we can flush, we instead give
869 			 * dirty pages extra priority on the inactive queue
870 			 * by forcing them to be cycled through the queue
871 			 * twice before being flushed, after which the
872 			 * (now clean) page will cycle through once more
873 			 * before being freed.  This significantly extends
874 			 * the thrash point for a heavily loaded machine.
875 			 */
876 			vm_page_flag_set(m, PG_WINATCFLS);
877 			vm_page_requeue(m);
878 		} else if (maxlaunder > 0) {
879 			/*
880 			 * We always want to try to flush some dirty pages if
881 			 * we encounter them, to keep the system stable.
882 			 * Normally this number is small, but under extreme
883 			 * pressure where there are insufficient clean pages
884 			 * on the inactive queue, we may have to go all out.
885 			 */
886 			int swap_pageouts_ok, vfslocked = 0;
887 			struct vnode *vp = NULL;
888 			struct mount *mp = NULL;
889 
890 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
891 				swap_pageouts_ok = 1;
892 			} else {
893 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
894 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
895 				vm_page_count_min());
896 
897 			}
898 
899 			/*
900 			 * We don't bother paging objects that are "dead".
901 			 * Those objects are in a "rundown" state.
902 			 */
903 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
904 				VM_OBJECT_UNLOCK(object);
905 				vm_page_requeue(m);
906 				continue;
907 			}
908 
909 			/*
910 			 * Following operations may unlock
911 			 * vm_page_queue_mtx, invalidating the 'next'
912 			 * pointer.  To prevent an inordinate number
913 			 * of restarts we use our marker to remember
914 			 * our place.
915 			 *
916 			 */
917 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
918 					   m, &marker, pageq);
919 			/*
920 			 * The object is already known NOT to be dead.   It
921 			 * is possible for the vget() to block the whole
922 			 * pageout daemon, but the new low-memory handling
923 			 * code should prevent it.
924 			 *
925 			 * The previous code skipped locked vnodes and, worse,
926 			 * reordered pages in the queue.  This results in
927 			 * completely non-deterministic operation and, on a
928 			 * busy system, can lead to extremely non-optimal
929 			 * pageouts.  For example, it can cause clean pages
930 			 * to be freed and dirty pages to be moved to the end
931 			 * of the queue.  Since dirty pages are also moved to
932 			 * the end of the queue once-cleaned, this gives
933 			 * way too large a weighting to defering the freeing
934 			 * of dirty pages.
935 			 *
936 			 * We can't wait forever for the vnode lock, we might
937 			 * deadlock due to a vn_read() getting stuck in
938 			 * vm_wait while holding this vnode.  We skip the
939 			 * vnode if we can't get it in a reasonable amount
940 			 * of time.
941 			 */
942 			if (object->type == OBJT_VNODE) {
943 				vp = object->handle;
944 				if (vp->v_type == VREG &&
945 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
946 					mp = NULL;
947 					++pageout_lock_miss;
948 					if (object->flags & OBJ_MIGHTBEDIRTY)
949 						vnodes_skipped++;
950 					goto unlock_and_continue;
951 				}
952 				vm_page_unlock_queues();
953 				vm_object_reference_locked(object);
954 				VM_OBJECT_UNLOCK(object);
955 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
956 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
957 				    curthread)) {
958 					VM_OBJECT_LOCK(object);
959 					vm_page_lock_queues();
960 					++pageout_lock_miss;
961 					if (object->flags & OBJ_MIGHTBEDIRTY)
962 						vnodes_skipped++;
963 					vp = NULL;
964 					goto unlock_and_continue;
965 				}
966 				VM_OBJECT_LOCK(object);
967 				vm_page_lock_queues();
968 				/*
969 				 * The page might have been moved to another
970 				 * queue during potential blocking in vget()
971 				 * above.  The page might have been freed and
972 				 * reused for another vnode.
973 				 */
974 				if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
975 				    m->object != object ||
976 				    TAILQ_NEXT(m, pageq) != &marker) {
977 					if (object->flags & OBJ_MIGHTBEDIRTY)
978 						vnodes_skipped++;
979 					goto unlock_and_continue;
980 				}
981 
982 				/*
983 				 * The page may have been busied during the
984 				 * blocking in vget().  We don't move the
985 				 * page back onto the end of the queue so that
986 				 * statistics are more correct if we don't.
987 				 */
988 				if (m->busy || (m->oflags & VPO_BUSY)) {
989 					goto unlock_and_continue;
990 				}
991 
992 				/*
993 				 * If the page has become held it might
994 				 * be undergoing I/O, so skip it
995 				 */
996 				if (m->hold_count) {
997 					vm_page_requeue(m);
998 					if (object->flags & OBJ_MIGHTBEDIRTY)
999 						vnodes_skipped++;
1000 					goto unlock_and_continue;
1001 				}
1002 			}
1003 
1004 			/*
1005 			 * If a page is dirty, then it is either being washed
1006 			 * (but not yet cleaned) or it is still in the
1007 			 * laundry.  If it is still in the laundry, then we
1008 			 * start the cleaning operation.
1009 			 *
1010 			 * decrement page_shortage on success to account for
1011 			 * the (future) cleaned page.  Otherwise we could wind
1012 			 * up laundering or cleaning too many pages.
1013 			 */
1014 			if (vm_pageout_clean(m) != 0) {
1015 				--page_shortage;
1016 				--maxlaunder;
1017 			}
1018 unlock_and_continue:
1019 			VM_OBJECT_UNLOCK(object);
1020 			if (mp != NULL) {
1021 				vm_page_unlock_queues();
1022 				if (vp != NULL)
1023 					vput(vp);
1024 				VFS_UNLOCK_GIANT(vfslocked);
1025 				vm_object_deallocate(object);
1026 				vn_finished_write(mp);
1027 				vm_page_lock_queues();
1028 			}
1029 			next = TAILQ_NEXT(&marker, pageq);
1030 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1031 				     &marker, pageq);
1032 			continue;
1033 		}
1034 		VM_OBJECT_UNLOCK(object);
1035 	}
1036 
1037 	/*
1038 	 * Compute the number of pages we want to try to move from the
1039 	 * active queue to the inactive queue.
1040 	 */
1041 	page_shortage = vm_paging_target() +
1042 		cnt.v_inactive_target - cnt.v_inactive_count;
1043 	page_shortage += addl_page_shortage;
1044 
1045 	/*
1046 	 * Scan the active queue for things we can deactivate. We nominally
1047 	 * track the per-page activity counter and use it to locate
1048 	 * deactivation candidates.
1049 	 */
1050 	pcount = cnt.v_active_count;
1051 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1052 
1053 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1054 
1055 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1056 		    ("vm_pageout_scan: page %p isn't active", m));
1057 
1058 		next = TAILQ_NEXT(m, pageq);
1059 		object = m->object;
1060 		if ((m->flags & PG_MARKER) != 0) {
1061 			m = next;
1062 			continue;
1063 		}
1064 		if (!VM_OBJECT_TRYLOCK(object) &&
1065 		    !vm_pageout_fallback_object_lock(m, &next)) {
1066 			VM_OBJECT_UNLOCK(object);
1067 			m = next;
1068 			continue;
1069 		}
1070 
1071 		/*
1072 		 * Don't deactivate pages that are busy.
1073 		 */
1074 		if ((m->busy != 0) ||
1075 		    (m->oflags & VPO_BUSY) ||
1076 		    (m->hold_count != 0)) {
1077 			VM_OBJECT_UNLOCK(object);
1078 			vm_page_requeue(m);
1079 			m = next;
1080 			continue;
1081 		}
1082 
1083 		/*
1084 		 * The count for pagedaemon pages is done after checking the
1085 		 * page for eligibility...
1086 		 */
1087 		cnt.v_pdpages++;
1088 
1089 		/*
1090 		 * Check to see "how much" the page has been used.
1091 		 */
1092 		actcount = 0;
1093 		if (object->ref_count != 0) {
1094 			if (m->flags & PG_REFERENCED) {
1095 				actcount += 1;
1096 			}
1097 			actcount += pmap_ts_referenced(m);
1098 			if (actcount) {
1099 				m->act_count += ACT_ADVANCE + actcount;
1100 				if (m->act_count > ACT_MAX)
1101 					m->act_count = ACT_MAX;
1102 			}
1103 		}
1104 
1105 		/*
1106 		 * Since we have "tested" this bit, we need to clear it now.
1107 		 */
1108 		vm_page_flag_clear(m, PG_REFERENCED);
1109 
1110 		/*
1111 		 * Only if an object is currently being used, do we use the
1112 		 * page activation count stats.
1113 		 */
1114 		if (actcount && (object->ref_count != 0)) {
1115 			vm_page_requeue(m);
1116 		} else {
1117 			m->act_count -= min(m->act_count, ACT_DECLINE);
1118 			if (vm_pageout_algorithm ||
1119 			    object->ref_count == 0 ||
1120 			    m->act_count == 0) {
1121 				page_shortage--;
1122 				if (object->ref_count == 0) {
1123 					pmap_remove_all(m);
1124 					if (m->dirty == 0)
1125 						vm_page_cache(m);
1126 					else
1127 						vm_page_deactivate(m);
1128 				} else {
1129 					vm_page_deactivate(m);
1130 				}
1131 			} else {
1132 				vm_page_requeue(m);
1133 			}
1134 		}
1135 		VM_OBJECT_UNLOCK(object);
1136 		m = next;
1137 	}
1138 	vm_page_unlock_queues();
1139 #if !defined(NO_SWAPPING)
1140 	/*
1141 	 * Idle process swapout -- run once per second.
1142 	 */
1143 	if (vm_swap_idle_enabled) {
1144 		static long lsec;
1145 		if (time_second != lsec) {
1146 			vm_req_vmdaemon(VM_SWAP_IDLE);
1147 			lsec = time_second;
1148 		}
1149 	}
1150 #endif
1151 
1152 	/*
1153 	 * If we didn't get enough free pages, and we have skipped a vnode
1154 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1155 	 * if we did not get enough free pages.
1156 	 */
1157 	if (vm_paging_target() > 0) {
1158 		if (vnodes_skipped && vm_page_count_min())
1159 			(void) speedup_syncer();
1160 #if !defined(NO_SWAPPING)
1161 		if (vm_swap_enabled && vm_page_count_target())
1162 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1163 #endif
1164 	}
1165 
1166 	/*
1167 	 * If we are critically low on one of RAM or swap and low on
1168 	 * the other, kill the largest process.  However, we avoid
1169 	 * doing this on the first pass in order to give ourselves a
1170 	 * chance to flush out dirty vnode-backed pages and to allow
1171 	 * active pages to be moved to the inactive queue and reclaimed.
1172 	 */
1173 	if (pass != 0 &&
1174 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1175 	     (swap_pager_full && vm_paging_target() > 0)))
1176 		vm_pageout_oom(VM_OOM_MEM);
1177 }
1178 
1179 
1180 void
1181 vm_pageout_oom(int shortage)
1182 {
1183 	struct proc *p, *bigproc;
1184 	vm_offset_t size, bigsize;
1185 	struct thread *td;
1186 
1187 	/*
1188 	 * We keep the process bigproc locked once we find it to keep anyone
1189 	 * from messing with it; however, there is a possibility of
1190 	 * deadlock if process B is bigproc and one of it's child processes
1191 	 * attempts to propagate a signal to B while we are waiting for A's
1192 	 * lock while walking this list.  To avoid this, we don't block on
1193 	 * the process lock but just skip a process if it is already locked.
1194 	 */
1195 	bigproc = NULL;
1196 	bigsize = 0;
1197 	sx_slock(&allproc_lock);
1198 	FOREACH_PROC_IN_SYSTEM(p) {
1199 		int breakout;
1200 
1201 		if (PROC_TRYLOCK(p) == 0)
1202 			continue;
1203 		/*
1204 		 * If this is a system or protected process, skip it.
1205 		 */
1206 		if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1207 		    (p->p_flag & P_PROTECTED) ||
1208 		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1209 			PROC_UNLOCK(p);
1210 			continue;
1211 		}
1212 		/*
1213 		 * If the process is in a non-running type state,
1214 		 * don't touch it.  Check all the threads individually.
1215 		 */
1216 		breakout = 0;
1217 		FOREACH_THREAD_IN_PROC(p, td) {
1218 			thread_lock(td);
1219 			if (!TD_ON_RUNQ(td) &&
1220 			    !TD_IS_RUNNING(td) &&
1221 			    !TD_IS_SLEEPING(td)) {
1222 				thread_unlock(td);
1223 				breakout = 1;
1224 				break;
1225 			}
1226 			thread_unlock(td);
1227 		}
1228 		if (breakout) {
1229 			PROC_UNLOCK(p);
1230 			continue;
1231 		}
1232 		/*
1233 		 * get the process size
1234 		 */
1235 		if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1236 			PROC_UNLOCK(p);
1237 			continue;
1238 		}
1239 		size = vmspace_swap_count(p->p_vmspace);
1240 		vm_map_unlock_read(&p->p_vmspace->vm_map);
1241 		if (shortage == VM_OOM_MEM)
1242 			size += vmspace_resident_count(p->p_vmspace);
1243 		/*
1244 		 * if the this process is bigger than the biggest one
1245 		 * remember it.
1246 		 */
1247 		if (size > bigsize) {
1248 			if (bigproc != NULL)
1249 				PROC_UNLOCK(bigproc);
1250 			bigproc = p;
1251 			bigsize = size;
1252 		} else
1253 			PROC_UNLOCK(p);
1254 	}
1255 	sx_sunlock(&allproc_lock);
1256 	if (bigproc != NULL) {
1257 		killproc(bigproc, "out of swap space");
1258 		sched_nice(bigproc, PRIO_MIN);
1259 		PROC_UNLOCK(bigproc);
1260 		wakeup(&cnt.v_free_count);
1261 	}
1262 }
1263 
1264 /*
1265  * This routine tries to maintain the pseudo LRU active queue,
1266  * so that during long periods of time where there is no paging,
1267  * that some statistic accumulation still occurs.  This code
1268  * helps the situation where paging just starts to occur.
1269  */
1270 static void
1271 vm_pageout_page_stats()
1272 {
1273 	vm_object_t object;
1274 	vm_page_t m,next;
1275 	int pcount,tpcount;		/* Number of pages to check */
1276 	static int fullintervalcount = 0;
1277 	int page_shortage;
1278 
1279 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1280 	page_shortage =
1281 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1282 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1283 
1284 	if (page_shortage <= 0)
1285 		return;
1286 
1287 	pcount = cnt.v_active_count;
1288 	fullintervalcount += vm_pageout_stats_interval;
1289 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1290 		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1291 		    cnt.v_page_count;
1292 		if (pcount > tpcount)
1293 			pcount = tpcount;
1294 	} else {
1295 		fullintervalcount = 0;
1296 	}
1297 
1298 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1299 	while ((m != NULL) && (pcount-- > 0)) {
1300 		int actcount;
1301 
1302 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1303 		    ("vm_pageout_page_stats: page %p isn't active", m));
1304 
1305 		next = TAILQ_NEXT(m, pageq);
1306 		object = m->object;
1307 
1308 		if ((m->flags & PG_MARKER) != 0) {
1309 			m = next;
1310 			continue;
1311 		}
1312 		if (!VM_OBJECT_TRYLOCK(object) &&
1313 		    !vm_pageout_fallback_object_lock(m, &next)) {
1314 			VM_OBJECT_UNLOCK(object);
1315 			m = next;
1316 			continue;
1317 		}
1318 
1319 		/*
1320 		 * Don't deactivate pages that are busy.
1321 		 */
1322 		if ((m->busy != 0) ||
1323 		    (m->oflags & VPO_BUSY) ||
1324 		    (m->hold_count != 0)) {
1325 			VM_OBJECT_UNLOCK(object);
1326 			vm_page_requeue(m);
1327 			m = next;
1328 			continue;
1329 		}
1330 
1331 		actcount = 0;
1332 		if (m->flags & PG_REFERENCED) {
1333 			vm_page_flag_clear(m, PG_REFERENCED);
1334 			actcount += 1;
1335 		}
1336 
1337 		actcount += pmap_ts_referenced(m);
1338 		if (actcount) {
1339 			m->act_count += ACT_ADVANCE + actcount;
1340 			if (m->act_count > ACT_MAX)
1341 				m->act_count = ACT_MAX;
1342 			vm_page_requeue(m);
1343 		} else {
1344 			if (m->act_count == 0) {
1345 				/*
1346 				 * We turn off page access, so that we have
1347 				 * more accurate RSS stats.  We don't do this
1348 				 * in the normal page deactivation when the
1349 				 * system is loaded VM wise, because the
1350 				 * cost of the large number of page protect
1351 				 * operations would be higher than the value
1352 				 * of doing the operation.
1353 				 */
1354 				pmap_remove_all(m);
1355 				vm_page_deactivate(m);
1356 			} else {
1357 				m->act_count -= min(m->act_count, ACT_DECLINE);
1358 				vm_page_requeue(m);
1359 			}
1360 		}
1361 		VM_OBJECT_UNLOCK(object);
1362 		m = next;
1363 	}
1364 }
1365 
1366 /*
1367  *	vm_pageout is the high level pageout daemon.
1368  */
1369 static void
1370 vm_pageout()
1371 {
1372 	int error, pass;
1373 
1374 	/*
1375 	 * Initialize some paging parameters.
1376 	 */
1377 	cnt.v_interrupt_free_min = 2;
1378 	if (cnt.v_page_count < 2000)
1379 		vm_pageout_page_count = 8;
1380 
1381 	/*
1382 	 * v_free_reserved needs to include enough for the largest
1383 	 * swap pager structures plus enough for any pv_entry structs
1384 	 * when paging.
1385 	 */
1386 	if (cnt.v_page_count > 1024)
1387 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1388 	else
1389 		cnt.v_free_min = 4;
1390 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1391 	    cnt.v_interrupt_free_min;
1392 	cnt.v_free_reserved = vm_pageout_page_count +
1393 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1394 	cnt.v_free_severe = cnt.v_free_min / 2;
1395 	cnt.v_free_min += cnt.v_free_reserved;
1396 	cnt.v_free_severe += cnt.v_free_reserved;
1397 
1398 	/*
1399 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1400 	 * that these are more a measure of the VM cache queue hysteresis
1401 	 * then the VM free queue.  Specifically, v_free_target is the
1402 	 * high water mark (free+cache pages).
1403 	 *
1404 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1405 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1406 	 * be big enough to handle memory needs while the pageout daemon
1407 	 * is signalled and run to free more pages.
1408 	 */
1409 	if (cnt.v_free_count > 6144)
1410 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1411 	else
1412 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1413 
1414 	if (cnt.v_free_count > 2048) {
1415 		cnt.v_cache_min = cnt.v_free_target;
1416 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1417 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1418 	} else {
1419 		cnt.v_cache_min = 0;
1420 		cnt.v_cache_max = 0;
1421 		cnt.v_inactive_target = cnt.v_free_count / 4;
1422 	}
1423 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1424 		cnt.v_inactive_target = cnt.v_free_count / 3;
1425 
1426 	/* XXX does not really belong here */
1427 	if (vm_page_max_wired == 0)
1428 		vm_page_max_wired = cnt.v_free_count / 3;
1429 
1430 	if (vm_pageout_stats_max == 0)
1431 		vm_pageout_stats_max = cnt.v_free_target;
1432 
1433 	/*
1434 	 * Set interval in seconds for stats scan.
1435 	 */
1436 	if (vm_pageout_stats_interval == 0)
1437 		vm_pageout_stats_interval = 5;
1438 	if (vm_pageout_full_stats_interval == 0)
1439 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1440 
1441 	swap_pager_swap_init();
1442 	pass = 0;
1443 	/*
1444 	 * The pageout daemon is never done, so loop forever.
1445 	 */
1446 	while (TRUE) {
1447 		/*
1448 		 * If we have enough free memory, wakeup waiters.  Do
1449 		 * not clear vm_pages_needed until we reach our target,
1450 		 * otherwise we may be woken up over and over again and
1451 		 * waste a lot of cpu.
1452 		 */
1453 		mtx_lock(&vm_page_queue_free_mtx);
1454 		if (vm_pages_needed && !vm_page_count_min()) {
1455 			if (!vm_paging_needed())
1456 				vm_pages_needed = 0;
1457 			wakeup(&cnt.v_free_count);
1458 		}
1459 		if (vm_pages_needed) {
1460 			/*
1461 			 * Still not done, take a second pass without waiting
1462 			 * (unlimited dirty cleaning), otherwise sleep a bit
1463 			 * and try again.
1464 			 */
1465 			++pass;
1466 			if (pass > 1)
1467 				msleep(&vm_pages_needed,
1468 				    &vm_page_queue_free_mtx, PVM, "psleep",
1469 				    hz / 2);
1470 		} else {
1471 			/*
1472 			 * Good enough, sleep & handle stats.  Prime the pass
1473 			 * for the next run.
1474 			 */
1475 			if (pass > 1)
1476 				pass = 1;
1477 			else
1478 				pass = 0;
1479 			error = msleep(&vm_pages_needed,
1480 			    &vm_page_queue_free_mtx, PVM, "psleep",
1481 			    vm_pageout_stats_interval * hz);
1482 			if (error && !vm_pages_needed) {
1483 				mtx_unlock(&vm_page_queue_free_mtx);
1484 				pass = 0;
1485 				vm_page_lock_queues();
1486 				vm_pageout_page_stats();
1487 				vm_page_unlock_queues();
1488 				continue;
1489 			}
1490 		}
1491 		if (vm_pages_needed)
1492 			cnt.v_pdwakeups++;
1493 		mtx_unlock(&vm_page_queue_free_mtx);
1494 		vm_pageout_scan(pass);
1495 	}
1496 }
1497 
1498 /*
1499  * Unless the free page queue lock is held by the caller, this function
1500  * should be regarded as advisory.  Specifically, the caller should
1501  * not msleep() on &cnt.v_free_count following this function unless
1502  * the free page queue lock is held until the msleep() is performed.
1503  */
1504 void
1505 pagedaemon_wakeup()
1506 {
1507 
1508 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1509 		vm_pages_needed = 1;
1510 		wakeup(&vm_pages_needed);
1511 	}
1512 }
1513 
1514 #if !defined(NO_SWAPPING)
1515 static void
1516 vm_req_vmdaemon(int req)
1517 {
1518 	static int lastrun = 0;
1519 
1520 	mtx_lock(&vm_daemon_mtx);
1521 	vm_pageout_req_swapout |= req;
1522 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1523 		wakeup(&vm_daemon_needed);
1524 		lastrun = ticks;
1525 	}
1526 	mtx_unlock(&vm_daemon_mtx);
1527 }
1528 
1529 static void
1530 vm_daemon()
1531 {
1532 	struct rlimit rsslim;
1533 	struct proc *p;
1534 	struct thread *td;
1535 	int breakout, swapout_flags;
1536 
1537 	while (TRUE) {
1538 		mtx_lock(&vm_daemon_mtx);
1539 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1540 		swapout_flags = vm_pageout_req_swapout;
1541 		vm_pageout_req_swapout = 0;
1542 		mtx_unlock(&vm_daemon_mtx);
1543 		if (swapout_flags)
1544 			swapout_procs(swapout_flags);
1545 
1546 		/*
1547 		 * scan the processes for exceeding their rlimits or if
1548 		 * process is swapped out -- deactivate pages
1549 		 */
1550 		sx_slock(&allproc_lock);
1551 		FOREACH_PROC_IN_SYSTEM(p) {
1552 			vm_pindex_t limit, size;
1553 
1554 			/*
1555 			 * if this is a system process or if we have already
1556 			 * looked at this process, skip it.
1557 			 */
1558 			PROC_LOCK(p);
1559 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1560 				PROC_UNLOCK(p);
1561 				continue;
1562 			}
1563 			/*
1564 			 * if the process is in a non-running type state,
1565 			 * don't touch it.
1566 			 */
1567 			breakout = 0;
1568 			FOREACH_THREAD_IN_PROC(p, td) {
1569 				thread_lock(td);
1570 				if (!TD_ON_RUNQ(td) &&
1571 				    !TD_IS_RUNNING(td) &&
1572 				    !TD_IS_SLEEPING(td)) {
1573 					thread_unlock(td);
1574 					breakout = 1;
1575 					break;
1576 				}
1577 				thread_unlock(td);
1578 			}
1579 			if (breakout) {
1580 				PROC_UNLOCK(p);
1581 				continue;
1582 			}
1583 			/*
1584 			 * get a limit
1585 			 */
1586 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1587 			limit = OFF_TO_IDX(
1588 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1589 
1590 			/*
1591 			 * let processes that are swapped out really be
1592 			 * swapped out set the limit to nothing (will force a
1593 			 * swap-out.)
1594 			 */
1595 			if ((p->p_flag & P_INMEM) == 0)
1596 				limit = 0;	/* XXX */
1597 			PROC_UNLOCK(p);
1598 
1599 			size = vmspace_resident_count(p->p_vmspace);
1600 			if (limit >= 0 && size >= limit) {
1601 				vm_pageout_map_deactivate_pages(
1602 				    &p->p_vmspace->vm_map, limit);
1603 			}
1604 		}
1605 		sx_sunlock(&allproc_lock);
1606 	}
1607 }
1608 #endif			/* !defined(NO_SWAPPING) */
1609