xref: /freebsd/sys/vm/vm_pageout.c (revision d086ded32300bc0f33fb1574d0bcfccfbc60881d)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
41  *
42  *
43  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44  * All rights reserved.
45  *
46  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 /*
70  *	The proverbial page-out daemon.
71  */
72 
73 #include <sys/cdefs.h>
74 __FBSDID("$FreeBSD$");
75 
76 #include "opt_vm.h"
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/eventhandler.h>
81 #include <sys/lock.h>
82 #include <sys/mutex.h>
83 #include <sys/proc.h>
84 #include <sys/kthread.h>
85 #include <sys/ktr.h>
86 #include <sys/resourcevar.h>
87 #include <sys/sched.h>
88 #include <sys/signalvar.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 #include <sys/sx.h>
92 #include <sys/sysctl.h>
93 
94 #include <vm/vm.h>
95 #include <vm/vm_param.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_pageout.h>
100 #include <vm/vm_pager.h>
101 #include <vm/swap_pager.h>
102 #include <vm/vm_extern.h>
103 #include <vm/uma.h>
104 
105 #include <machine/mutex.h>
106 
107 /*
108  * System initialization
109  */
110 
111 /* the kernel process "vm_pageout"*/
112 static void vm_pageout(void);
113 static int vm_pageout_clean(vm_page_t);
114 static void vm_pageout_page_free(vm_page_t);
115 static void vm_pageout_pmap_collect(void);
116 static void vm_pageout_scan(int pass);
117 static int vm_pageout_free_page_calc(vm_size_t count);
118 struct proc *pageproc;
119 
120 static struct kproc_desc page_kp = {
121 	"pagedaemon",
122 	vm_pageout,
123 	&pageproc
124 };
125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
126 
127 #if !defined(NO_SWAPPING)
128 /* the kernel process "vm_daemon"*/
129 static void vm_daemon(void);
130 static struct	proc *vmproc;
131 
132 static struct kproc_desc vm_kp = {
133 	"vmdaemon",
134 	vm_daemon,
135 	&vmproc
136 };
137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
138 #endif
139 
140 
141 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
142 int vm_pageout_deficit;		/* Estimated number of pages deficit */
143 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
144 
145 #if !defined(NO_SWAPPING)
146 static int vm_pageout_req_swapout;	/* XXX */
147 static int vm_daemon_needed;
148 #endif
149 static int vm_max_launder = 32;
150 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
151 static int vm_pageout_full_stats_interval = 0;
152 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
153 static int defer_swap_pageouts=0;
154 static int disable_swap_pageouts=0;
155 
156 #if defined(NO_SWAPPING)
157 static int vm_swap_enabled=0;
158 static int vm_swap_idle_enabled=0;
159 #else
160 static int vm_swap_enabled=1;
161 static int vm_swap_idle_enabled=0;
162 #endif
163 
164 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
165 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
166 
167 SYSCTL_INT(_vm, OID_AUTO, max_launder,
168 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
169 
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
171 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
172 
173 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
174 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
175 
176 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
177 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
178 
179 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
180 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
181 
182 #if defined(NO_SWAPPING)
183 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
184 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
185 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
186 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
187 #else
188 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
189 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
190 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
191 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
192 #endif
193 
194 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
195 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
196 
197 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
198 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
199 
200 static int pageout_lock_miss;
201 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
202 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
203 
204 #define VM_PAGEOUT_PAGE_COUNT 16
205 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
206 
207 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
208 
209 #if !defined(NO_SWAPPING)
210 typedef void freeer_fcn_t(vm_map_t, vm_object_t, vm_pindex_t);
211 static void vm_pageout_map_deactivate_pages(vm_map_t, vm_pindex_t);
212 static freeer_fcn_t vm_pageout_object_deactivate_pages;
213 static void vm_req_vmdaemon(void);
214 #endif
215 static void vm_pageout_page_stats(void);
216 
217 /*
218  * vm_pageout_clean:
219  *
220  * Clean the page and remove it from the laundry.
221  *
222  * We set the busy bit to cause potential page faults on this page to
223  * block.  Note the careful timing, however, the busy bit isn't set till
224  * late and we cannot do anything that will mess with the page.
225  */
226 static int
227 vm_pageout_clean(m)
228 	vm_page_t m;
229 {
230 	vm_object_t object;
231 	vm_page_t mc[2*vm_pageout_page_count];
232 	int pageout_count;
233 	int ib, is, page_base;
234 	vm_pindex_t pindex = m->pindex;
235 
236 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
237 
238 	object = m->object;
239 
240 	/*
241 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
242 	 * with the new swapper, but we could have serious problems paging
243 	 * out other object types if there is insufficient memory.
244 	 *
245 	 * Unfortunately, checking free memory here is far too late, so the
246 	 * check has been moved up a procedural level.
247 	 */
248 
249 	/*
250 	 * Don't mess with the page if it's busy, held, or special
251 	 */
252 	if ((m->hold_count != 0) ||
253 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
254 		return 0;
255 	}
256 
257 	mc[vm_pageout_page_count] = m;
258 	pageout_count = 1;
259 	page_base = vm_pageout_page_count;
260 	ib = 1;
261 	is = 1;
262 
263 	/*
264 	 * Scan object for clusterable pages.
265 	 *
266 	 * We can cluster ONLY if: ->> the page is NOT
267 	 * clean, wired, busy, held, or mapped into a
268 	 * buffer, and one of the following:
269 	 * 1) The page is inactive, or a seldom used
270 	 *    active page.
271 	 * -or-
272 	 * 2) we force the issue.
273 	 *
274 	 * During heavy mmap/modification loads the pageout
275 	 * daemon can really fragment the underlying file
276 	 * due to flushing pages out of order and not trying
277 	 * align the clusters (which leave sporatic out-of-order
278 	 * holes).  To solve this problem we do the reverse scan
279 	 * first and attempt to align our cluster, then do a
280 	 * forward scan if room remains.
281 	 */
282 more:
283 	while (ib && pageout_count < vm_pageout_page_count) {
284 		vm_page_t p;
285 
286 		if (ib > pindex) {
287 			ib = 0;
288 			break;
289 		}
290 
291 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
292 			ib = 0;
293 			break;
294 		}
295 		if (((p->queue - p->pc) == PQ_CACHE) ||
296 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
297 			ib = 0;
298 			break;
299 		}
300 		vm_page_test_dirty(p);
301 		if ((p->dirty & p->valid) == 0 ||
302 		    p->queue != PQ_INACTIVE ||
303 		    p->wire_count != 0 ||	/* may be held by buf cache */
304 		    p->hold_count != 0) {	/* may be undergoing I/O */
305 			ib = 0;
306 			break;
307 		}
308 		mc[--page_base] = p;
309 		++pageout_count;
310 		++ib;
311 		/*
312 		 * alignment boundry, stop here and switch directions.  Do
313 		 * not clear ib.
314 		 */
315 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
316 			break;
317 	}
318 
319 	while (pageout_count < vm_pageout_page_count &&
320 	    pindex + is < object->size) {
321 		vm_page_t p;
322 
323 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
324 			break;
325 		if (((p->queue - p->pc) == PQ_CACHE) ||
326 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
327 			break;
328 		}
329 		vm_page_test_dirty(p);
330 		if ((p->dirty & p->valid) == 0 ||
331 		    p->queue != PQ_INACTIVE ||
332 		    p->wire_count != 0 ||	/* may be held by buf cache */
333 		    p->hold_count != 0) {	/* may be undergoing I/O */
334 			break;
335 		}
336 		mc[page_base + pageout_count] = p;
337 		++pageout_count;
338 		++is;
339 	}
340 
341 	/*
342 	 * If we exhausted our forward scan, continue with the reverse scan
343 	 * when possible, even past a page boundry.  This catches boundry
344 	 * conditions.
345 	 */
346 	if (ib && pageout_count < vm_pageout_page_count)
347 		goto more;
348 
349 	/*
350 	 * we allow reads during pageouts...
351 	 */
352 	return vm_pageout_flush(&mc[page_base], pageout_count, 0, FALSE);
353 }
354 
355 /*
356  * vm_pageout_flush() - launder the given pages
357  *
358  *	The given pages are laundered.  Note that we setup for the start of
359  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
360  *	reference count all in here rather then in the parent.  If we want
361  *	the parent to do more sophisticated things we may have to change
362  *	the ordering.
363  */
364 int
365 vm_pageout_flush(mc, count, flags, is_object_locked)
366 	vm_page_t *mc;
367 	int count;
368 	int flags;
369 	int is_object_locked;
370 {
371 	vm_object_t object;
372 	int pageout_status[count];
373 	int numpagedout = 0;
374 	int i;
375 
376 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
377 	/*
378 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
379 	 * mark the pages read-only.
380 	 *
381 	 * We do not have to fixup the clean/dirty bits here... we can
382 	 * allow the pager to do it after the I/O completes.
383 	 *
384 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
385 	 * edge case with file fragments.
386 	 */
387 	for (i = 0; i < count; i++) {
388 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
389 		vm_page_io_start(mc[i]);
390 		pmap_page_protect(mc[i], VM_PROT_READ);
391 	}
392 	object = mc[0]->object;
393 	vm_page_unlock_queues();
394 	if (!is_object_locked)
395 		VM_OBJECT_LOCK(object);
396 	vm_object_pip_add(object, count);
397 	VM_OBJECT_UNLOCK(object);
398 
399 	vm_pager_put_pages(object, mc, count,
400 	    (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
401 	    pageout_status);
402 
403 	VM_OBJECT_LOCK(object);
404 	vm_page_lock_queues();
405 	for (i = 0; i < count; i++) {
406 		vm_page_t mt = mc[i];
407 
408 		switch (pageout_status[i]) {
409 		case VM_PAGER_OK:
410 		case VM_PAGER_PEND:
411 			numpagedout++;
412 			break;
413 		case VM_PAGER_BAD:
414 			/*
415 			 * Page outside of range of object. Right now we
416 			 * essentially lose the changes by pretending it
417 			 * worked.
418 			 */
419 			pmap_clear_modify(mt);
420 			vm_page_undirty(mt);
421 			break;
422 		case VM_PAGER_ERROR:
423 		case VM_PAGER_FAIL:
424 			/*
425 			 * If page couldn't be paged out, then reactivate the
426 			 * page so it doesn't clog the inactive list.  (We
427 			 * will try paging out it again later).
428 			 */
429 			vm_page_activate(mt);
430 			break;
431 		case VM_PAGER_AGAIN:
432 			break;
433 		}
434 
435 		/*
436 		 * If the operation is still going, leave the page busy to
437 		 * block all other accesses. Also, leave the paging in
438 		 * progress indicator set so that we don't attempt an object
439 		 * collapse.
440 		 */
441 		if (pageout_status[i] != VM_PAGER_PEND) {
442 			vm_object_pip_wakeup(object);
443 			vm_page_io_finish(mt);
444 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
445 				pmap_page_protect(mt, VM_PROT_READ);
446 		}
447 	}
448 	if (!is_object_locked)
449 		VM_OBJECT_UNLOCK(object);
450 	return numpagedout;
451 }
452 
453 #if !defined(NO_SWAPPING)
454 /*
455  *	vm_pageout_object_deactivate_pages
456  *
457  *	deactivate enough pages to satisfy the inactive target
458  *	requirements or if vm_page_proc_limit is set, then
459  *	deactivate all of the pages in the object and its
460  *	backing_objects.
461  *
462  *	The object and map must be locked.
463  */
464 static void
465 vm_pageout_object_deactivate_pages(map, object, desired)
466 	vm_map_t map;
467 	vm_object_t object;
468 	vm_pindex_t desired;
469 {
470 	vm_page_t p, next;
471 	int actcount, rcount, remove_mode;
472 
473 	GIANT_REQUIRED;
474 	if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
475 		return;
476 
477 	while (object) {
478 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
479 			return;
480 		if (object->paging_in_progress)
481 			return;
482 
483 		remove_mode = 0;
484 		if (object->shadow_count > 1)
485 			remove_mode = 1;
486 		/*
487 		 * scan the objects entire memory queue
488 		 */
489 		rcount = object->resident_page_count;
490 		p = TAILQ_FIRST(&object->memq);
491 		vm_page_lock_queues();
492 		while (p && (rcount-- > 0)) {
493 			if (pmap_resident_count(map->pmap) <= desired) {
494 				vm_page_unlock_queues();
495 				return;
496 			}
497 			next = TAILQ_NEXT(p, listq);
498 			cnt.v_pdpages++;
499 			if (p->wire_count != 0 ||
500 			    p->hold_count != 0 ||
501 			    p->busy != 0 ||
502 			    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
503 			    !pmap_page_exists_quick(vm_map_pmap(map), p)) {
504 				p = next;
505 				continue;
506 			}
507 			actcount = pmap_ts_referenced(p);
508 			if (actcount) {
509 				vm_page_flag_set(p, PG_REFERENCED);
510 			} else if (p->flags & PG_REFERENCED) {
511 				actcount = 1;
512 			}
513 			if ((p->queue != PQ_ACTIVE) &&
514 				(p->flags & PG_REFERENCED)) {
515 				vm_page_activate(p);
516 				p->act_count += actcount;
517 				vm_page_flag_clear(p, PG_REFERENCED);
518 			} else if (p->queue == PQ_ACTIVE) {
519 				if ((p->flags & PG_REFERENCED) == 0) {
520 					p->act_count -= min(p->act_count, ACT_DECLINE);
521 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
522 						pmap_remove_all(p);
523 						vm_page_deactivate(p);
524 					} else {
525 						vm_pageq_requeue(p);
526 					}
527 				} else {
528 					vm_page_activate(p);
529 					vm_page_flag_clear(p, PG_REFERENCED);
530 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
531 						p->act_count += ACT_ADVANCE;
532 					vm_pageq_requeue(p);
533 				}
534 			} else if (p->queue == PQ_INACTIVE) {
535 				pmap_remove_all(p);
536 			}
537 			p = next;
538 		}
539 		vm_page_unlock_queues();
540 		object = object->backing_object;
541 	}
542 }
543 
544 /*
545  * deactivate some number of pages in a map, try to do it fairly, but
546  * that is really hard to do.
547  */
548 static void
549 vm_pageout_map_deactivate_pages(map, desired)
550 	vm_map_t map;
551 	vm_pindex_t desired;
552 {
553 	vm_map_entry_t tmpe;
554 	vm_object_t obj, bigobj;
555 	int nothingwired;
556 
557 	GIANT_REQUIRED;
558 	if (!vm_map_trylock(map))
559 		return;
560 
561 	bigobj = NULL;
562 	nothingwired = TRUE;
563 
564 	/*
565 	 * first, search out the biggest object, and try to free pages from
566 	 * that.
567 	 */
568 	tmpe = map->header.next;
569 	while (tmpe != &map->header) {
570 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
571 			obj = tmpe->object.vm_object;
572 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
573 				((bigobj == NULL) ||
574 				 (bigobj->resident_page_count < obj->resident_page_count))) {
575 				bigobj = obj;
576 			}
577 		}
578 		if (tmpe->wired_count > 0)
579 			nothingwired = FALSE;
580 		tmpe = tmpe->next;
581 	}
582 
583 	if (bigobj)
584 		vm_pageout_object_deactivate_pages(map, bigobj, desired);
585 
586 	/*
587 	 * Next, hunt around for other pages to deactivate.  We actually
588 	 * do this search sort of wrong -- .text first is not the best idea.
589 	 */
590 	tmpe = map->header.next;
591 	while (tmpe != &map->header) {
592 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
593 			break;
594 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
595 			obj = tmpe->object.vm_object;
596 			if (obj)
597 				vm_pageout_object_deactivate_pages(map, obj, desired);
598 		}
599 		tmpe = tmpe->next;
600 	}
601 
602 	/*
603 	 * Remove all mappings if a process is swapped out, this will free page
604 	 * table pages.
605 	 */
606 	if (desired == 0 && nothingwired) {
607 		vm_page_lock_queues();
608 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
609 		    vm_map_max(map));
610 		vm_page_unlock_queues();
611 	}
612 	vm_map_unlock(map);
613 }
614 #endif		/* !defined(NO_SWAPPING) */
615 
616 /*
617  * Warning! The page queue lock is released and reacquired.
618  */
619 static void
620 vm_pageout_page_free(vm_page_t m)
621 {
622 	vm_object_t object = m->object;
623 
624 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
625 	vm_page_busy(m);
626 	vm_page_unlock_queues();
627 	/*
628 	 * Avoid a lock order reversal.  The page must be busy.
629 	 */
630 	VM_OBJECT_LOCK(object);
631 	vm_page_lock_queues();
632 	pmap_remove_all(m);
633 	vm_page_free(m);
634 	VM_OBJECT_UNLOCK(object);
635 	cnt.v_dfree++;
636 }
637 
638 /*
639  * This routine is very drastic, but can save the system
640  * in a pinch.
641  */
642 static void
643 vm_pageout_pmap_collect(void)
644 {
645 	int i;
646 	vm_page_t m;
647 	static int warningdone;
648 
649 	if (pmap_pagedaemon_waken == 0)
650 		return;
651 	if (warningdone < 5) {
652 		printf("collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
653 		warningdone++;
654 	}
655 	vm_page_lock_queues();
656 	for (i = 0; i < vm_page_array_size; i++) {
657 		m = &vm_page_array[i];
658 		if (m->wire_count || m->hold_count || m->busy ||
659 		    (m->flags & (PG_BUSY | PG_UNMANAGED)))
660 			continue;
661 		pmap_remove_all(m);
662 	}
663 	vm_page_unlock_queues();
664 	pmap_pagedaemon_waken = 0;
665 }
666 
667 /*
668  *	vm_pageout_scan does the dirty work for the pageout daemon.
669  */
670 static void
671 vm_pageout_scan(int pass)
672 {
673 	vm_page_t m, next;
674 	struct vm_page marker;
675 	int save_page_shortage;
676 	int save_inactive_count;
677 	int page_shortage, maxscan, pcount;
678 	int addl_page_shortage, addl_page_shortage_init;
679 	struct proc *p, *bigproc;
680 	vm_offset_t size, bigsize;
681 	vm_object_t object;
682 	int actcount;
683 	int vnodes_skipped = 0;
684 	int maxlaunder;
685 	int s;
686 	struct thread *td;
687 
688 	GIANT_REQUIRED;
689 	/*
690 	 * Decrease registered cache sizes.
691 	 */
692 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
693 	/*
694 	 * We do this explicitly after the caches have been drained above.
695 	 */
696 	uma_reclaim();
697 	/*
698 	 * Do whatever cleanup that the pmap code can.
699 	 */
700 	vm_pageout_pmap_collect();
701 
702 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
703 
704 	/*
705 	 * Calculate the number of pages we want to either free or move
706 	 * to the cache.
707 	 */
708 	page_shortage = vm_paging_target() + addl_page_shortage_init;
709 	save_page_shortage = page_shortage;
710 	save_inactive_count = cnt.v_inactive_count;
711 
712 	/*
713 	 * Initialize our marker
714 	 */
715 	bzero(&marker, sizeof(marker));
716 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
717 	marker.queue = PQ_INACTIVE;
718 	marker.wire_count = 1;
719 
720 	/*
721 	 * Start scanning the inactive queue for pages we can move to the
722 	 * cache or free.  The scan will stop when the target is reached or
723 	 * we have scanned the entire inactive queue.  Note that m->act_count
724 	 * is not used to form decisions for the inactive queue, only for the
725 	 * active queue.
726 	 *
727 	 * maxlaunder limits the number of dirty pages we flush per scan.
728 	 * For most systems a smaller value (16 or 32) is more robust under
729 	 * extreme memory and disk pressure because any unnecessary writes
730 	 * to disk can result in extreme performance degredation.  However,
731 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
732 	 * used) will die horribly with limited laundering.  If the pageout
733 	 * daemon cannot clean enough pages in the first pass, we let it go
734 	 * all out in succeeding passes.
735 	 */
736 	if ((maxlaunder = vm_max_launder) <= 1)
737 		maxlaunder = 1;
738 	if (pass)
739 		maxlaunder = 10000;
740 rescan0:
741 	addl_page_shortage = addl_page_shortage_init;
742 	maxscan = cnt.v_inactive_count;
743 
744 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
745 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
746 	     m = next) {
747 
748 		cnt.v_pdpages++;
749 
750 		if (m->queue != PQ_INACTIVE) {
751 			goto rescan0;
752 		}
753 
754 		next = TAILQ_NEXT(m, pageq);
755 
756 		/*
757 		 * skip marker pages
758 		 */
759 		if (m->flags & PG_MARKER)
760 			continue;
761 
762 		/*
763 		 * A held page may be undergoing I/O, so skip it.
764 		 */
765 		if (m->hold_count) {
766 			vm_pageq_requeue(m);
767 			addl_page_shortage++;
768 			continue;
769 		}
770 		/*
771 		 * Don't mess with busy pages, keep in the front of the
772 		 * queue, most likely are being paged out.
773 		 */
774 		if (m->busy || (m->flags & PG_BUSY)) {
775 			addl_page_shortage++;
776 			continue;
777 		}
778 
779 		vm_page_lock_queues();
780 		/*
781 		 * If the object is not being used, we ignore previous
782 		 * references.
783 		 */
784 		if (m->object->ref_count == 0) {
785 			vm_page_flag_clear(m, PG_REFERENCED);
786 			pmap_clear_reference(m);
787 
788 		/*
789 		 * Otherwise, if the page has been referenced while in the
790 		 * inactive queue, we bump the "activation count" upwards,
791 		 * making it less likely that the page will be added back to
792 		 * the inactive queue prematurely again.  Here we check the
793 		 * page tables (or emulated bits, if any), given the upper
794 		 * level VM system not knowing anything about existing
795 		 * references.
796 		 */
797 		} else if (((m->flags & PG_REFERENCED) == 0) &&
798 			(actcount = pmap_ts_referenced(m))) {
799 			vm_page_activate(m);
800 			vm_page_unlock_queues();
801 			m->act_count += (actcount + ACT_ADVANCE);
802 			continue;
803 		}
804 
805 		/*
806 		 * If the upper level VM system knows about any page
807 		 * references, we activate the page.  We also set the
808 		 * "activation count" higher than normal so that we will less
809 		 * likely place pages back onto the inactive queue again.
810 		 */
811 		if ((m->flags & PG_REFERENCED) != 0) {
812 			vm_page_flag_clear(m, PG_REFERENCED);
813 			actcount = pmap_ts_referenced(m);
814 			vm_page_activate(m);
815 			vm_page_unlock_queues();
816 			m->act_count += (actcount + ACT_ADVANCE + 1);
817 			continue;
818 		}
819 
820 		/*
821 		 * If the upper level VM system doesn't know anything about
822 		 * the page being dirty, we have to check for it again.  As
823 		 * far as the VM code knows, any partially dirty pages are
824 		 * fully dirty.
825 		 */
826 		if (m->dirty == 0) {
827 			vm_page_test_dirty(m);
828 		} else {
829 			vm_page_dirty(m);
830 		}
831 		vm_page_unlock_queues();
832 
833 		/*
834 		 * Invalid pages can be easily freed
835 		 */
836 		if (m->valid == 0) {
837 			vm_page_lock_queues();
838 			vm_pageout_page_free(m);
839 			vm_page_unlock_queues();
840 			--page_shortage;
841 
842 		/*
843 		 * Clean pages can be placed onto the cache queue.  This
844 		 * effectively frees them.
845 		 */
846 		} else if (m->dirty == 0) {
847 			vm_page_lock_queues();
848 			vm_page_cache(m);
849 			vm_page_unlock_queues();
850 			--page_shortage;
851 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
852 			/*
853 			 * Dirty pages need to be paged out, but flushing
854 			 * a page is extremely expensive verses freeing
855 			 * a clean page.  Rather then artificially limiting
856 			 * the number of pages we can flush, we instead give
857 			 * dirty pages extra priority on the inactive queue
858 			 * by forcing them to be cycled through the queue
859 			 * twice before being flushed, after which the
860 			 * (now clean) page will cycle through once more
861 			 * before being freed.  This significantly extends
862 			 * the thrash point for a heavily loaded machine.
863 			 */
864 			vm_page_lock_queues();
865 			vm_page_flag_set(m, PG_WINATCFLS);
866 			vm_pageq_requeue(m);
867 			vm_page_unlock_queues();
868 		} else if (maxlaunder > 0) {
869 			/*
870 			 * We always want to try to flush some dirty pages if
871 			 * we encounter them, to keep the system stable.
872 			 * Normally this number is small, but under extreme
873 			 * pressure where there are insufficient clean pages
874 			 * on the inactive queue, we may have to go all out.
875 			 */
876 			int swap_pageouts_ok;
877 			struct vnode *vp = NULL;
878 			struct mount *mp;
879 
880 			object = m->object;
881 
882 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
883 				swap_pageouts_ok = 1;
884 			} else {
885 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
886 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
887 				vm_page_count_min());
888 
889 			}
890 
891 			/*
892 			 * We don't bother paging objects that are "dead".
893 			 * Those objects are in a "rundown" state.
894 			 */
895 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
896 				vm_pageq_requeue(m);
897 				continue;
898 			}
899 
900 			/*
901 			 * The object is already known NOT to be dead.   It
902 			 * is possible for the vget() to block the whole
903 			 * pageout daemon, but the new low-memory handling
904 			 * code should prevent it.
905 			 *
906 			 * The previous code skipped locked vnodes and, worse,
907 			 * reordered pages in the queue.  This results in
908 			 * completely non-deterministic operation and, on a
909 			 * busy system, can lead to extremely non-optimal
910 			 * pageouts.  For example, it can cause clean pages
911 			 * to be freed and dirty pages to be moved to the end
912 			 * of the queue.  Since dirty pages are also moved to
913 			 * the end of the queue once-cleaned, this gives
914 			 * way too large a weighting to defering the freeing
915 			 * of dirty pages.
916 			 *
917 			 * We can't wait forever for the vnode lock, we might
918 			 * deadlock due to a vn_read() getting stuck in
919 			 * vm_wait while holding this vnode.  We skip the
920 			 * vnode if we can't get it in a reasonable amount
921 			 * of time.
922 			 */
923 			if (object->type == OBJT_VNODE) {
924 				vp = object->handle;
925 
926 				mp = NULL;
927 				if (vp->v_type == VREG)
928 					vn_start_write(vp, &mp, V_NOWAIT);
929 				if (vget(vp, LK_EXCLUSIVE|LK_TIMELOCK, curthread)) {
930 					++pageout_lock_miss;
931 					vn_finished_write(mp);
932 					if (object->flags & OBJ_MIGHTBEDIRTY)
933 						vnodes_skipped++;
934 					continue;
935 				}
936 
937 				/*
938 				 * The page might have been moved to another
939 				 * queue during potential blocking in vget()
940 				 * above.  The page might have been freed and
941 				 * reused for another vnode.  The object might
942 				 * have been reused for another vnode.
943 				 */
944 				if (m->queue != PQ_INACTIVE ||
945 				    m->object != object ||
946 				    object->handle != vp) {
947 					if (object->flags & OBJ_MIGHTBEDIRTY)
948 						vnodes_skipped++;
949 					vput(vp);
950 					vn_finished_write(mp);
951 					continue;
952 				}
953 
954 				/*
955 				 * The page may have been busied during the
956 				 * blocking in vput();  We don't move the
957 				 * page back onto the end of the queue so that
958 				 * statistics are more correct if we don't.
959 				 */
960 				if (m->busy || (m->flags & PG_BUSY)) {
961 					vput(vp);
962 					vn_finished_write(mp);
963 					continue;
964 				}
965 
966 				/*
967 				 * If the page has become held it might
968 				 * be undergoing I/O, so skip it
969 				 */
970 				if (m->hold_count) {
971 					vm_pageq_requeue(m);
972 					if (object->flags & OBJ_MIGHTBEDIRTY)
973 						vnodes_skipped++;
974 					vput(vp);
975 					vn_finished_write(mp);
976 					continue;
977 				}
978 			}
979 
980 			/*
981 			 * If a page is dirty, then it is either being washed
982 			 * (but not yet cleaned) or it is still in the
983 			 * laundry.  If it is still in the laundry, then we
984 			 * start the cleaning operation.
985 			 *
986 			 * This operation may cluster, invalidating the 'next'
987 			 * pointer.  To prevent an inordinate number of
988 			 * restarts we use our marker to remember our place.
989 			 *
990 			 * decrement page_shortage on success to account for
991 			 * the (future) cleaned page.  Otherwise we could wind
992 			 * up laundering or cleaning too many pages.
993 			 */
994 			vm_page_lock_queues();
995 			s = splvm();
996 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
997 			splx(s);
998 			if (vm_pageout_clean(m) != 0) {
999 				--page_shortage;
1000 				--maxlaunder;
1001 			}
1002 			s = splvm();
1003 			next = TAILQ_NEXT(&marker, pageq);
1004 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1005 			splx(s);
1006 			vm_page_unlock_queues();
1007 			if (vp) {
1008 				vput(vp);
1009 				vn_finished_write(mp);
1010 			}
1011 		}
1012 	}
1013 
1014 	/*
1015 	 * Compute the number of pages we want to try to move from the
1016 	 * active queue to the inactive queue.
1017 	 */
1018 	page_shortage = vm_paging_target() +
1019 		cnt.v_inactive_target - cnt.v_inactive_count;
1020 	page_shortage += addl_page_shortage;
1021 
1022 	vm_page_lock_queues();
1023 	/*
1024 	 * Scan the active queue for things we can deactivate. We nominally
1025 	 * track the per-page activity counter and use it to locate
1026 	 * deactivation candidates.
1027 	 */
1028 	pcount = cnt.v_active_count;
1029 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1030 
1031 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1032 
1033 		/*
1034 		 * This is a consistency check, and should likely be a panic
1035 		 * or warning.
1036 		 */
1037 		if (m->queue != PQ_ACTIVE) {
1038 			break;
1039 		}
1040 
1041 		next = TAILQ_NEXT(m, pageq);
1042 		/*
1043 		 * Don't deactivate pages that are busy.
1044 		 */
1045 		if ((m->busy != 0) ||
1046 		    (m->flags & PG_BUSY) ||
1047 		    (m->hold_count != 0)) {
1048 			vm_pageq_requeue(m);
1049 			m = next;
1050 			continue;
1051 		}
1052 
1053 		/*
1054 		 * The count for pagedaemon pages is done after checking the
1055 		 * page for eligibility...
1056 		 */
1057 		cnt.v_pdpages++;
1058 
1059 		/*
1060 		 * Check to see "how much" the page has been used.
1061 		 */
1062 		actcount = 0;
1063 		if (m->object->ref_count != 0) {
1064 			if (m->flags & PG_REFERENCED) {
1065 				actcount += 1;
1066 			}
1067 			actcount += pmap_ts_referenced(m);
1068 			if (actcount) {
1069 				m->act_count += ACT_ADVANCE + actcount;
1070 				if (m->act_count > ACT_MAX)
1071 					m->act_count = ACT_MAX;
1072 			}
1073 		}
1074 
1075 		/*
1076 		 * Since we have "tested" this bit, we need to clear it now.
1077 		 */
1078 		vm_page_flag_clear(m, PG_REFERENCED);
1079 
1080 		/*
1081 		 * Only if an object is currently being used, do we use the
1082 		 * page activation count stats.
1083 		 */
1084 		if (actcount && (m->object->ref_count != 0)) {
1085 			vm_pageq_requeue(m);
1086 		} else {
1087 			m->act_count -= min(m->act_count, ACT_DECLINE);
1088 			if (vm_pageout_algorithm ||
1089 			    m->object->ref_count == 0 ||
1090 			    m->act_count == 0) {
1091 				page_shortage--;
1092 				if (m->object->ref_count == 0) {
1093 					pmap_remove_all(m);
1094 					if (m->dirty == 0)
1095 						vm_page_cache(m);
1096 					else
1097 						vm_page_deactivate(m);
1098 				} else {
1099 					vm_page_deactivate(m);
1100 				}
1101 			} else {
1102 				vm_pageq_requeue(m);
1103 			}
1104 		}
1105 		m = next;
1106 	}
1107 	s = splvm();
1108 
1109 	/*
1110 	 * We try to maintain some *really* free pages, this allows interrupt
1111 	 * code to be guaranteed space.  Since both cache and free queues
1112 	 * are considered basically 'free', moving pages from cache to free
1113 	 * does not effect other calculations.
1114 	 */
1115 	while (cnt.v_free_count < cnt.v_free_reserved) {
1116 		static int cache_rover = 0;
1117 		m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE);
1118 		if (!m)
1119 			break;
1120 		if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1121 		    m->busy ||
1122 		    m->hold_count ||
1123 		    m->wire_count) {
1124 #ifdef INVARIANTS
1125 			printf("Warning: busy page %p found in cache\n", m);
1126 #endif
1127 			vm_page_deactivate(m);
1128 			continue;
1129 		}
1130 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1131 		vm_pageout_page_free(m);
1132 	}
1133 	splx(s);
1134 	vm_page_unlock_queues();
1135 #if !defined(NO_SWAPPING)
1136 	/*
1137 	 * Idle process swapout -- run once per second.
1138 	 */
1139 	if (vm_swap_idle_enabled) {
1140 		static long lsec;
1141 		if (time_second != lsec) {
1142 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1143 			vm_req_vmdaemon();
1144 			lsec = time_second;
1145 		}
1146 	}
1147 #endif
1148 
1149 	/*
1150 	 * If we didn't get enough free pages, and we have skipped a vnode
1151 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1152 	 * if we did not get enough free pages.
1153 	 */
1154 	if (vm_paging_target() > 0) {
1155 		if (vnodes_skipped && vm_page_count_min())
1156 			(void) speedup_syncer();
1157 #if !defined(NO_SWAPPING)
1158 		if (vm_swap_enabled && vm_page_count_target()) {
1159 			vm_req_vmdaemon();
1160 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1161 		}
1162 #endif
1163 	}
1164 
1165 	/*
1166 	 * If we are critically low on one of RAM or swap and low on
1167 	 * the other, kill the largest process.  However, we avoid
1168 	 * doing this on the first pass in order to give ourselves a
1169 	 * chance to flush out dirty vnode-backed pages and to allow
1170 	 * active pages to be moved to the inactive queue and reclaimed.
1171 	 *
1172 	 * We keep the process bigproc locked once we find it to keep anyone
1173 	 * from messing with it; however, there is a possibility of
1174 	 * deadlock if process B is bigproc and one of it's child processes
1175 	 * attempts to propagate a signal to B while we are waiting for A's
1176 	 * lock while walking this list.  To avoid this, we don't block on
1177 	 * the process lock but just skip a process if it is already locked.
1178 	 */
1179 	if (pass != 0 &&
1180 	    ((vm_swap_size < 64 && vm_page_count_min()) ||
1181 	     (swap_pager_full && vm_paging_target() > 0))) {
1182 		bigproc = NULL;
1183 		bigsize = 0;
1184 		sx_slock(&allproc_lock);
1185 		FOREACH_PROC_IN_SYSTEM(p) {
1186 			int breakout;
1187 			/*
1188 			 * If this process is already locked, skip it.
1189 			 */
1190 			if (PROC_TRYLOCK(p) == 0)
1191 				continue;
1192 			/*
1193 			 * If this is a system or protected process, skip it.
1194 			 */
1195 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1196 			    (p->p_flag & P_PROTECTED) ||
1197 			    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1198 				PROC_UNLOCK(p);
1199 				continue;
1200 			}
1201 			/*
1202 			 * if the process is in a non-running type state,
1203 			 * don't touch it. Check all the threads individually.
1204 			 */
1205 			mtx_lock_spin(&sched_lock);
1206 			breakout = 0;
1207 			FOREACH_THREAD_IN_PROC(p, td) {
1208 				if (!TD_ON_RUNQ(td) &&
1209 				    !TD_IS_RUNNING(td) &&
1210 				    !TD_IS_SLEEPING(td)) {
1211 					breakout = 1;
1212 					break;
1213 				}
1214 			}
1215 			if (breakout) {
1216 				mtx_unlock_spin(&sched_lock);
1217 				PROC_UNLOCK(p);
1218 				continue;
1219 			}
1220 			mtx_unlock_spin(&sched_lock);
1221 			/*
1222 			 * get the process size
1223 			 */
1224 			if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1225 				PROC_UNLOCK(p);
1226 				continue;
1227 			}
1228 			size = vmspace_swap_count(p->p_vmspace);
1229 			vm_map_unlock_read(&p->p_vmspace->vm_map);
1230 			size += vmspace_resident_count(p->p_vmspace);
1231 			/*
1232 			 * if the this process is bigger than the biggest one
1233 			 * remember it.
1234 			 */
1235 			if (size > bigsize) {
1236 				if (bigproc != NULL)
1237 					PROC_UNLOCK(bigproc);
1238 				bigproc = p;
1239 				bigsize = size;
1240 			} else
1241 				PROC_UNLOCK(p);
1242 		}
1243 		sx_sunlock(&allproc_lock);
1244 		if (bigproc != NULL) {
1245 			struct ksegrp *kg;
1246 			killproc(bigproc, "out of swap space");
1247 			mtx_lock_spin(&sched_lock);
1248 			FOREACH_KSEGRP_IN_PROC(bigproc, kg) {
1249 				sched_nice(kg, PRIO_MIN); /* XXXKSE ??? */
1250 			}
1251 			mtx_unlock_spin(&sched_lock);
1252 			PROC_UNLOCK(bigproc);
1253 			wakeup(&cnt.v_free_count);
1254 		}
1255 	}
1256 }
1257 
1258 /*
1259  * This routine tries to maintain the pseudo LRU active queue,
1260  * so that during long periods of time where there is no paging,
1261  * that some statistic accumulation still occurs.  This code
1262  * helps the situation where paging just starts to occur.
1263  */
1264 static void
1265 vm_pageout_page_stats()
1266 {
1267 	vm_page_t m,next;
1268 	int pcount,tpcount;		/* Number of pages to check */
1269 	static int fullintervalcount = 0;
1270 	int page_shortage;
1271 	int s0;
1272 
1273 	page_shortage =
1274 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1275 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1276 
1277 	if (page_shortage <= 0)
1278 		return;
1279 
1280 	s0 = splvm();
1281 	vm_page_lock_queues();
1282 	pcount = cnt.v_active_count;
1283 	fullintervalcount += vm_pageout_stats_interval;
1284 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1285 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1286 		if (pcount > tpcount)
1287 			pcount = tpcount;
1288 	} else {
1289 		fullintervalcount = 0;
1290 	}
1291 
1292 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1293 	while ((m != NULL) && (pcount-- > 0)) {
1294 		int actcount;
1295 
1296 		if (m->queue != PQ_ACTIVE) {
1297 			break;
1298 		}
1299 
1300 		next = TAILQ_NEXT(m, pageq);
1301 		/*
1302 		 * Don't deactivate pages that are busy.
1303 		 */
1304 		if ((m->busy != 0) ||
1305 		    (m->flags & PG_BUSY) ||
1306 		    (m->hold_count != 0)) {
1307 			vm_pageq_requeue(m);
1308 			m = next;
1309 			continue;
1310 		}
1311 
1312 		actcount = 0;
1313 		if (m->flags & PG_REFERENCED) {
1314 			vm_page_flag_clear(m, PG_REFERENCED);
1315 			actcount += 1;
1316 		}
1317 
1318 		actcount += pmap_ts_referenced(m);
1319 		if (actcount) {
1320 			m->act_count += ACT_ADVANCE + actcount;
1321 			if (m->act_count > ACT_MAX)
1322 				m->act_count = ACT_MAX;
1323 			vm_pageq_requeue(m);
1324 		} else {
1325 			if (m->act_count == 0) {
1326 				/*
1327 				 * We turn off page access, so that we have
1328 				 * more accurate RSS stats.  We don't do this
1329 				 * in the normal page deactivation when the
1330 				 * system is loaded VM wise, because the
1331 				 * cost of the large number of page protect
1332 				 * operations would be higher than the value
1333 				 * of doing the operation.
1334 				 */
1335 				pmap_remove_all(m);
1336 				vm_page_deactivate(m);
1337 			} else {
1338 				m->act_count -= min(m->act_count, ACT_DECLINE);
1339 				vm_pageq_requeue(m);
1340 			}
1341 		}
1342 
1343 		m = next;
1344 	}
1345 	vm_page_unlock_queues();
1346 	splx(s0);
1347 }
1348 
1349 static int
1350 vm_pageout_free_page_calc(count)
1351 vm_size_t count;
1352 {
1353 	if (count < cnt.v_page_count)
1354 		 return 0;
1355 	/*
1356 	 * free_reserved needs to include enough for the largest swap pager
1357 	 * structures plus enough for any pv_entry structs when paging.
1358 	 */
1359 	if (cnt.v_page_count > 1024)
1360 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1361 	else
1362 		cnt.v_free_min = 4;
1363 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1364 		cnt.v_interrupt_free_min;
1365 	cnt.v_free_reserved = vm_pageout_page_count +
1366 		cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1367 	cnt.v_free_severe = cnt.v_free_min / 2;
1368 	cnt.v_free_min += cnt.v_free_reserved;
1369 	cnt.v_free_severe += cnt.v_free_reserved;
1370 	return 1;
1371 }
1372 
1373 /*
1374  *	vm_pageout is the high level pageout daemon.
1375  */
1376 static void
1377 vm_pageout()
1378 {
1379 	int error, pass, s;
1380 
1381 	mtx_lock(&Giant);
1382 
1383 	/*
1384 	 * Initialize some paging parameters.
1385 	 */
1386 	cnt.v_interrupt_free_min = 2;
1387 	if (cnt.v_page_count < 2000)
1388 		vm_pageout_page_count = 8;
1389 
1390 	vm_pageout_free_page_calc(cnt.v_page_count);
1391 	/*
1392 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1393 	 * that these are more a measure of the VM cache queue hysteresis
1394 	 * then the VM free queue.  Specifically, v_free_target is the
1395 	 * high water mark (free+cache pages).
1396 	 *
1397 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1398 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1399 	 * be big enough to handle memory needs while the pageout daemon
1400 	 * is signalled and run to free more pages.
1401 	 */
1402 	if (cnt.v_free_count > 6144)
1403 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1404 	else
1405 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1406 
1407 	if (cnt.v_free_count > 2048) {
1408 		cnt.v_cache_min = cnt.v_free_target;
1409 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1410 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1411 	} else {
1412 		cnt.v_cache_min = 0;
1413 		cnt.v_cache_max = 0;
1414 		cnt.v_inactive_target = cnt.v_free_count / 4;
1415 	}
1416 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1417 		cnt.v_inactive_target = cnt.v_free_count / 3;
1418 
1419 	/* XXX does not really belong here */
1420 	if (vm_page_max_wired == 0)
1421 		vm_page_max_wired = cnt.v_free_count / 3;
1422 
1423 	if (vm_pageout_stats_max == 0)
1424 		vm_pageout_stats_max = cnt.v_free_target;
1425 
1426 	/*
1427 	 * Set interval in seconds for stats scan.
1428 	 */
1429 	if (vm_pageout_stats_interval == 0)
1430 		vm_pageout_stats_interval = 5;
1431 	if (vm_pageout_full_stats_interval == 0)
1432 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1433 
1434 	/*
1435 	 * Set maximum free per pass
1436 	 */
1437 	if (vm_pageout_stats_free_max == 0)
1438 		vm_pageout_stats_free_max = 5;
1439 
1440 	swap_pager_swap_init();
1441 	pass = 0;
1442 	/*
1443 	 * The pageout daemon is never done, so loop forever.
1444 	 */
1445 	while (TRUE) {
1446 		s = splvm();
1447 		vm_page_lock_queues();
1448 		/*
1449 		 * If we have enough free memory, wakeup waiters.  Do
1450 		 * not clear vm_pages_needed until we reach our target,
1451 		 * otherwise we may be woken up over and over again and
1452 		 * waste a lot of cpu.
1453 		 */
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, &vm_page_queue_mtx, PVM,
1468 				       "psleep", hz/2);
1469 		} else {
1470 			/*
1471 			 * Good enough, sleep & handle stats.  Prime the pass
1472 			 * for the next run.
1473 			 */
1474 			if (pass > 1)
1475 				pass = 1;
1476 			else
1477 				pass = 0;
1478 			error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM,
1479 				    "psleep", vm_pageout_stats_interval * hz);
1480 			if (error && !vm_pages_needed) {
1481 				vm_page_unlock_queues();
1482 				splx(s);
1483 				pass = 0;
1484 				vm_pageout_page_stats();
1485 				continue;
1486 			}
1487 		}
1488 		if (vm_pages_needed)
1489 			cnt.v_pdwakeups++;
1490 		vm_page_unlock_queues();
1491 		splx(s);
1492 		vm_pageout_scan(pass);
1493 	}
1494 }
1495 
1496 /*
1497  * Unless the page queue lock is held by the caller, this function
1498  * should be regarded as advisory.  Specifically, the caller should
1499  * not msleep() on &cnt.v_free_count following this function unless
1500  * the page queue lock is held until the msleep() is performed.
1501  */
1502 void
1503 pagedaemon_wakeup()
1504 {
1505 
1506 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1507 		vm_pages_needed = 1;
1508 		wakeup(&vm_pages_needed);
1509 	}
1510 }
1511 
1512 #if !defined(NO_SWAPPING)
1513 static void
1514 vm_req_vmdaemon()
1515 {
1516 	static int lastrun = 0;
1517 
1518 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1519 		wakeup(&vm_daemon_needed);
1520 		lastrun = ticks;
1521 	}
1522 }
1523 
1524 static void
1525 vm_daemon()
1526 {
1527 	struct proc *p;
1528 	int breakout;
1529 	struct thread *td;
1530 
1531 	mtx_lock(&Giant);
1532 	while (TRUE) {
1533 		tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1534 		if (vm_pageout_req_swapout) {
1535 			swapout_procs(vm_pageout_req_swapout);
1536 			vm_pageout_req_swapout = 0;
1537 		}
1538 		/*
1539 		 * scan the processes for exceeding their rlimits or if
1540 		 * process is swapped out -- deactivate pages
1541 		 */
1542 		sx_slock(&allproc_lock);
1543 		LIST_FOREACH(p, &allproc, p_list) {
1544 			vm_pindex_t limit, size;
1545 
1546 			/*
1547 			 * if this is a system process or if we have already
1548 			 * looked at this process, skip it.
1549 			 */
1550 			PROC_LOCK(p);
1551 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1552 				PROC_UNLOCK(p);
1553 				continue;
1554 			}
1555 			/*
1556 			 * if the process is in a non-running type state,
1557 			 * don't touch it.
1558 			 */
1559 			mtx_lock_spin(&sched_lock);
1560 			breakout = 0;
1561 			FOREACH_THREAD_IN_PROC(p, td) {
1562 				if (!TD_ON_RUNQ(td) &&
1563 				    !TD_IS_RUNNING(td) &&
1564 				    !TD_IS_SLEEPING(td)) {
1565 					breakout = 1;
1566 					break;
1567 				}
1568 			}
1569 			mtx_unlock_spin(&sched_lock);
1570 			if (breakout) {
1571 				PROC_UNLOCK(p);
1572 				continue;
1573 			}
1574 			/*
1575 			 * get a limit
1576 			 */
1577 			limit = OFF_TO_IDX(
1578 			    qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1579 				p->p_rlimit[RLIMIT_RSS].rlim_max));
1580 
1581 			/*
1582 			 * let processes that are swapped out really be
1583 			 * swapped out set the limit to nothing (will force a
1584 			 * swap-out.)
1585 			 */
1586 			if ((p->p_sflag & PS_INMEM) == 0)
1587 				limit = 0;	/* XXX */
1588 			PROC_UNLOCK(p);
1589 
1590 			size = vmspace_resident_count(p->p_vmspace);
1591 			if (limit >= 0 && size >= limit) {
1592 				vm_pageout_map_deactivate_pages(
1593 				    &p->p_vmspace->vm_map, limit);
1594 			}
1595 		}
1596 		sx_sunlock(&allproc_lock);
1597 	}
1598 }
1599 #endif			/* !defined(NO_SWAPPING) */
1600