xref: /dflybsd-src/sys/vm/vm_pageout.c (revision efda3bd00c039d6845508b47bb18d1687c72154e)
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  * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
69  * $DragonFly: src/sys/vm/vm_pageout.c,v 1.24 2006/08/12 00:26:22 dillon Exp $
70  */
71 
72 /*
73  *	The proverbial page-out daemon.
74  */
75 
76 #include "opt_vm.h"
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/proc.h>
81 #include <sys/kthread.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/vnode.h>
85 #include <sys/vmmeter.h>
86 #include <sys/sysctl.h>
87 
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <sys/lock.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pager.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_extern.h>
98 
99 #include <sys/thread2.h>
100 #include <vm/vm_page2.h>
101 
102 /*
103  * System initialization
104  */
105 
106 /* the kernel process "vm_pageout"*/
107 static void vm_pageout (void);
108 static int vm_pageout_clean (vm_page_t);
109 static void vm_pageout_scan (int pass);
110 static int vm_pageout_free_page_calc (vm_size_t count);
111 struct thread *pagethread;
112 
113 static struct kproc_desc page_kp = {
114 	"pagedaemon",
115 	vm_pageout,
116 	&pagethread
117 };
118 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
119 
120 #if !defined(NO_SWAPPING)
121 /* the kernel process "vm_daemon"*/
122 static void vm_daemon (void);
123 static struct	thread *vmthread;
124 
125 static struct kproc_desc vm_kp = {
126 	"vmdaemon",
127 	vm_daemon,
128 	&vmthread
129 };
130 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
131 #endif
132 
133 
134 int vm_pages_needed=0;		/* Event on which pageout daemon sleeps */
135 int vm_pageout_deficit=0;	/* Estimated number of pages deficit */
136 int vm_pageout_pages_needed=0;	/* flag saying that the pageout daemon needs pages */
137 
138 #if !defined(NO_SWAPPING)
139 static int vm_pageout_req_swapout;	/* XXX */
140 static int vm_daemon_needed;
141 #endif
142 extern int vm_swap_size;
143 static int vm_max_launder = 32;
144 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
145 static int vm_pageout_full_stats_interval = 0;
146 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
147 static int defer_swap_pageouts=0;
148 static int disable_swap_pageouts=0;
149 
150 #if defined(NO_SWAPPING)
151 static int vm_swap_enabled=0;
152 static int vm_swap_idle_enabled=0;
153 #else
154 static int vm_swap_enabled=1;
155 static int vm_swap_idle_enabled=0;
156 #endif
157 
158 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
159 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
160 
161 SYSCTL_INT(_vm, OID_AUTO, max_launder,
162 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
163 
164 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
165 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
166 
167 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
168 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
169 
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
171 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
172 
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
174 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
175 
176 #if defined(NO_SWAPPING)
177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
178 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
180 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
181 #else
182 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
183 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
184 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
185 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
186 #endif
187 
188 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
189 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
190 
191 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
192 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
193 
194 static int pageout_lock_miss;
195 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
196 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
197 
198 int vm_load;
199 SYSCTL_INT(_vm, OID_AUTO, vm_load,
200 	CTLFLAG_RD, &vm_load, 0, "load on the VM system");
201 int vm_load_enable = 1;
202 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
203 	CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
204 #ifdef INVARIANTS
205 int vm_load_debug;
206 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
207 	CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
208 #endif
209 
210 #define VM_PAGEOUT_PAGE_COUNT 16
211 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
212 
213 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
214 
215 #if !defined(NO_SWAPPING)
216 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
217 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
218 static freeer_fcn_t vm_pageout_object_deactivate_pages;
219 static void vm_req_vmdaemon (void);
220 #endif
221 static void vm_pageout_page_stats(void);
222 
223 /*
224  * Update
225  */
226 void
227 vm_fault_ratecheck(void)
228 {
229 	if (vm_pages_needed) {
230 		if (vm_load < 1000)
231 			++vm_load;
232 	} else {
233 		if (vm_load > 0)
234 			--vm_load;
235 	}
236 }
237 
238 /*
239  * vm_pageout_clean:
240  *
241  * Clean the page and remove it from the laundry.  The page must not be
242  * busy on-call.
243  *
244  * We set the busy bit to cause potential page faults on this page to
245  * block.  Note the careful timing, however, the busy bit isn't set till
246  * late and we cannot do anything that will mess with the page.
247  */
248 
249 static int
250 vm_pageout_clean(vm_page_t m)
251 {
252 	vm_object_t object;
253 	vm_page_t mc[2*vm_pageout_page_count];
254 	int pageout_count;
255 	int ib, is, page_base;
256 	vm_pindex_t pindex = m->pindex;
257 
258 	object = m->object;
259 
260 	/*
261 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
262 	 * with the new swapper, but we could have serious problems paging
263 	 * out other object types if there is insufficient memory.
264 	 *
265 	 * Unfortunately, checking free memory here is far too late, so the
266 	 * check has been moved up a procedural level.
267 	 */
268 
269 	/*
270 	 * Don't mess with the page if it's busy, held, or special
271 	 */
272 	if ((m->hold_count != 0) ||
273 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
274 		return 0;
275 	}
276 
277 	mc[vm_pageout_page_count] = m;
278 	pageout_count = 1;
279 	page_base = vm_pageout_page_count;
280 	ib = 1;
281 	is = 1;
282 
283 	/*
284 	 * Scan object for clusterable pages.
285 	 *
286 	 * We can cluster ONLY if: ->> the page is NOT
287 	 * clean, wired, busy, held, or mapped into a
288 	 * buffer, and one of the following:
289 	 * 1) The page is inactive, or a seldom used
290 	 *    active page.
291 	 * -or-
292 	 * 2) we force the issue.
293 	 *
294 	 * During heavy mmap/modification loads the pageout
295 	 * daemon can really fragment the underlying file
296 	 * due to flushing pages out of order and not trying
297 	 * align the clusters (which leave sporatic out-of-order
298 	 * holes).  To solve this problem we do the reverse scan
299 	 * first and attempt to align our cluster, then do a
300 	 * forward scan if room remains.
301 	 */
302 
303 more:
304 	while (ib && pageout_count < vm_pageout_page_count) {
305 		vm_page_t p;
306 
307 		if (ib > pindex) {
308 			ib = 0;
309 			break;
310 		}
311 
312 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
313 			ib = 0;
314 			break;
315 		}
316 		if (((p->queue - p->pc) == PQ_CACHE) ||
317 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
318 			ib = 0;
319 			break;
320 		}
321 		vm_page_test_dirty(p);
322 		if ((p->dirty & p->valid) == 0 ||
323 		    p->queue != PQ_INACTIVE ||
324 		    p->wire_count != 0 ||	/* may be held by buf cache */
325 		    p->hold_count != 0) {	/* may be undergoing I/O */
326 			ib = 0;
327 			break;
328 		}
329 		mc[--page_base] = p;
330 		++pageout_count;
331 		++ib;
332 		/*
333 		 * alignment boundry, stop here and switch directions.  Do
334 		 * not clear ib.
335 		 */
336 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
337 			break;
338 	}
339 
340 	while (pageout_count < vm_pageout_page_count &&
341 	    pindex + is < object->size) {
342 		vm_page_t p;
343 
344 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
345 			break;
346 		if (((p->queue - p->pc) == PQ_CACHE) ||
347 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
348 			break;
349 		}
350 		vm_page_test_dirty(p);
351 		if ((p->dirty & p->valid) == 0 ||
352 		    p->queue != PQ_INACTIVE ||
353 		    p->wire_count != 0 ||	/* may be held by buf cache */
354 		    p->hold_count != 0) {	/* may be undergoing I/O */
355 			break;
356 		}
357 		mc[page_base + pageout_count] = p;
358 		++pageout_count;
359 		++is;
360 	}
361 
362 	/*
363 	 * If we exhausted our forward scan, continue with the reverse scan
364 	 * when possible, even past a page boundry.  This catches boundry
365 	 * conditions.
366 	 */
367 	if (ib && pageout_count < vm_pageout_page_count)
368 		goto more;
369 
370 	/*
371 	 * we allow reads during pageouts...
372 	 */
373 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
374 }
375 
376 /*
377  * vm_pageout_flush() - launder the given pages
378  *
379  *	The given pages are laundered.  Note that we setup for the start of
380  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
381  *	reference count all in here rather then in the parent.  If we want
382  *	the parent to do more sophisticated things we may have to change
383  *	the ordering.
384  */
385 
386 int
387 vm_pageout_flush(vm_page_t *mc, int count, int flags)
388 {
389 	vm_object_t object;
390 	int pageout_status[count];
391 	int numpagedout = 0;
392 	int i;
393 
394 	/*
395 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
396 	 * mark the pages read-only.
397 	 *
398 	 * We do not have to fixup the clean/dirty bits here... we can
399 	 * allow the pager to do it after the I/O completes.
400 	 */
401 
402 	for (i = 0; i < count; i++) {
403 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
404 		vm_page_io_start(mc[i]);
405 		vm_page_protect(mc[i], VM_PROT_READ);
406 	}
407 
408 	object = mc[0]->object;
409 	vm_object_pip_add(object, count);
410 
411 	vm_pager_put_pages(object, mc, count,
412 	    (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
413 	    pageout_status);
414 
415 	for (i = 0; i < count; i++) {
416 		vm_page_t mt = mc[i];
417 
418 		switch (pageout_status[i]) {
419 		case VM_PAGER_OK:
420 			numpagedout++;
421 			break;
422 		case VM_PAGER_PEND:
423 			numpagedout++;
424 			break;
425 		case VM_PAGER_BAD:
426 			/*
427 			 * Page outside of range of object. Right now we
428 			 * essentially lose the changes by pretending it
429 			 * worked.
430 			 */
431 			pmap_clear_modify(mt);
432 			vm_page_undirty(mt);
433 			break;
434 		case VM_PAGER_ERROR:
435 		case VM_PAGER_FAIL:
436 			/*
437 			 * If page couldn't be paged out, then reactivate the
438 			 * page so it doesn't clog the inactive list.  (We
439 			 * will try paging out it again later).
440 			 */
441 			vm_page_activate(mt);
442 			break;
443 		case VM_PAGER_AGAIN:
444 			break;
445 		}
446 
447 		/*
448 		 * If the operation is still going, leave the page busy to
449 		 * block all other accesses. Also, leave the paging in
450 		 * progress indicator set so that we don't attempt an object
451 		 * collapse.
452 		 */
453 		if (pageout_status[i] != VM_PAGER_PEND) {
454 			vm_object_pip_wakeup(object);
455 			vm_page_io_finish(mt);
456 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
457 				vm_page_protect(mt, VM_PROT_READ);
458 		}
459 	}
460 	return numpagedout;
461 }
462 
463 #if !defined(NO_SWAPPING)
464 /*
465  *	vm_pageout_object_deactivate_pages
466  *
467  *	deactivate enough pages to satisfy the inactive target
468  *	requirements or if vm_page_proc_limit is set, then
469  *	deactivate all of the pages in the object and its
470  *	backing_objects.
471  *
472  *	The object and map must be locked.
473  */
474 static void
475 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
476 	vm_pindex_t desired, int map_remove_only)
477 {
478 	vm_page_t p, next;
479 	int rcount;
480 	int remove_mode;
481 
482 	if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
483 		return;
484 
485 	while (object) {
486 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
487 			return;
488 		if (object->paging_in_progress)
489 			return;
490 
491 		remove_mode = map_remove_only;
492 		if (object->shadow_count > 1)
493 			remove_mode = 1;
494 
495 		/*
496 		 * scan the objects entire memory queue.  spl protection is
497 		 * required to avoid an interrupt unbusy/free race against
498 		 * our busy check.
499 		 */
500 		crit_enter();
501 		rcount = object->resident_page_count;
502 		p = TAILQ_FIRST(&object->memq);
503 
504 		while (p && (rcount-- > 0)) {
505 			int actcount;
506 			if (pmap_resident_count(vm_map_pmap(map)) <= desired) {
507 				crit_exit();
508 				return;
509 			}
510 			next = TAILQ_NEXT(p, listq);
511 			mycpu->gd_cnt.v_pdpages++;
512 			if (p->wire_count != 0 ||
513 			    p->hold_count != 0 ||
514 			    p->busy != 0 ||
515 			    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
516 			    !pmap_page_exists_quick(vm_map_pmap(map), p)) {
517 				p = next;
518 				continue;
519 			}
520 
521 			actcount = pmap_ts_referenced(p);
522 			if (actcount) {
523 				vm_page_flag_set(p, PG_REFERENCED);
524 			} else if (p->flags & PG_REFERENCED) {
525 				actcount = 1;
526 			}
527 
528 			if ((p->queue != PQ_ACTIVE) &&
529 				(p->flags & PG_REFERENCED)) {
530 				vm_page_activate(p);
531 				p->act_count += actcount;
532 				vm_page_flag_clear(p, PG_REFERENCED);
533 			} else if (p->queue == PQ_ACTIVE) {
534 				if ((p->flags & PG_REFERENCED) == 0) {
535 					p->act_count -= min(p->act_count, ACT_DECLINE);
536 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
537 						vm_page_protect(p, VM_PROT_NONE);
538 						vm_page_deactivate(p);
539 					} else {
540 						TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
541 						TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
542 					}
543 				} else {
544 					vm_page_activate(p);
545 					vm_page_flag_clear(p, PG_REFERENCED);
546 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
547 						p->act_count += ACT_ADVANCE;
548 					TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
549 					TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
550 				}
551 			} else if (p->queue == PQ_INACTIVE) {
552 				vm_page_protect(p, VM_PROT_NONE);
553 			}
554 			p = next;
555 		}
556 		crit_exit();
557 		object = object->backing_object;
558 	}
559 }
560 
561 /*
562  * deactivate some number of pages in a map, try to do it fairly, but
563  * that is really hard to do.
564  */
565 static void
566 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
567 {
568 	vm_map_entry_t tmpe;
569 	vm_object_t obj, bigobj;
570 	int nothingwired;
571 
572 	if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
573 		return;
574 	}
575 
576 	bigobj = NULL;
577 	nothingwired = TRUE;
578 
579 	/*
580 	 * first, search out the biggest object, and try to free pages from
581 	 * that.
582 	 */
583 	tmpe = map->header.next;
584 	while (tmpe != &map->header) {
585 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
586 			obj = tmpe->object.vm_object;
587 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
588 				((bigobj == NULL) ||
589 				 (bigobj->resident_page_count < obj->resident_page_count))) {
590 				bigobj = obj;
591 			}
592 		}
593 		if (tmpe->wired_count > 0)
594 			nothingwired = FALSE;
595 		tmpe = tmpe->next;
596 	}
597 
598 	if (bigobj)
599 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
600 
601 	/*
602 	 * Next, hunt around for other pages to deactivate.  We actually
603 	 * do this search sort of wrong -- .text first is not the best idea.
604 	 */
605 	tmpe = map->header.next;
606 	while (tmpe != &map->header) {
607 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
608 			break;
609 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
610 			obj = tmpe->object.vm_object;
611 			if (obj)
612 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
613 		}
614 		tmpe = tmpe->next;
615 	};
616 
617 	/*
618 	 * Remove all mappings if a process is swapped out, this will free page
619 	 * table pages.
620 	 */
621 	if (desired == 0 && nothingwired)
622 		pmap_remove(vm_map_pmap(map),
623 			VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
624 	vm_map_unlock(map);
625 }
626 #endif
627 
628 /*
629  * Don't try to be fancy - being fancy can lead to vnode deadlocks.   We
630  * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
631  * be trivially freed.
632  */
633 void
634 vm_pageout_page_free(vm_page_t m) {
635 	vm_object_t object = m->object;
636 	int type = object->type;
637 
638 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
639 		vm_object_reference(object);
640 	vm_page_busy(m);
641 	vm_page_protect(m, VM_PROT_NONE);
642 	vm_page_free(m);
643 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
644 		vm_object_deallocate(object);
645 }
646 
647 /*
648  *	vm_pageout_scan does the dirty work for the pageout daemon.
649  */
650 
651 struct vm_pageout_scan_info {
652 	struct proc *bigproc;
653 	vm_offset_t bigsize;
654 };
655 
656 static int vm_pageout_scan_callback(struct proc *p, void *data);
657 
658 static void
659 vm_pageout_scan(int pass)
660 {
661 	struct vm_pageout_scan_info info;
662 	vm_page_t m, next;
663 	struct vm_page marker;
664 	int page_shortage, maxscan, pcount;
665 	int addl_page_shortage, addl_page_shortage_init;
666 	vm_object_t object;
667 	int actcount;
668 	int vnodes_skipped = 0;
669 	int maxlaunder;
670 
671 	/*
672 	 * Do whatever cleanup that the pmap code can.
673 	 */
674 	pmap_collect();
675 
676 	addl_page_shortage_init = vm_pageout_deficit;
677 	vm_pageout_deficit = 0;
678 
679 	/*
680 	 * Calculate the number of pages we want to either free or move
681 	 * to the cache.
682 	 */
683 	page_shortage = vm_paging_target() + addl_page_shortage_init;
684 
685 	/*
686 	 * Initialize our marker
687 	 */
688 	bzero(&marker, sizeof(marker));
689 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
690 	marker.queue = PQ_INACTIVE;
691 	marker.wire_count = 1;
692 
693 	/*
694 	 * Start scanning the inactive queue for pages we can move to the
695 	 * cache or free.  The scan will stop when the target is reached or
696 	 * we have scanned the entire inactive queue.  Note that m->act_count
697 	 * is not used to form decisions for the inactive queue, only for the
698 	 * active queue.
699 	 *
700 	 * maxlaunder limits the number of dirty pages we flush per scan.
701 	 * For most systems a smaller value (16 or 32) is more robust under
702 	 * extreme memory and disk pressure because any unnecessary writes
703 	 * to disk can result in extreme performance degredation.  However,
704 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
705 	 * used) will die horribly with limited laundering.  If the pageout
706 	 * daemon cannot clean enough pages in the first pass, we let it go
707 	 * all out in succeeding passes.
708 	 */
709 	if ((maxlaunder = vm_max_launder) <= 1)
710 		maxlaunder = 1;
711 	if (pass)
712 		maxlaunder = 10000;
713 
714 	/*
715 	 * We will generally be in a critical section throughout the
716 	 * scan, but we can release it temporarily when we are sitting on a
717 	 * non-busy page without fear.  this is required to prevent an
718 	 * interrupt from unbusying or freeing a page prior to our busy
719 	 * check, leaving us on the wrong queue or checking the wrong
720 	 * page.
721 	 */
722 	crit_enter();
723 rescan0:
724 	addl_page_shortage = addl_page_shortage_init;
725 	maxscan = vmstats.v_inactive_count;
726 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
727 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
728 	     m = next
729 	 ) {
730 		mycpu->gd_cnt.v_pdpages++;
731 
732 		/*
733 		 * Give interrupts a chance
734 		 */
735 		crit_exit();
736 		crit_enter();
737 
738 		/*
739 		 * It's easier for some of the conditions below to just loop
740 		 * and catch queue changes here rather then check everywhere
741 		 * else.
742 		 */
743 		if (m->queue != PQ_INACTIVE)
744 			goto rescan0;
745 		next = TAILQ_NEXT(m, pageq);
746 
747 		/*
748 		 * skip marker pages
749 		 */
750 		if (m->flags & PG_MARKER)
751 			continue;
752 
753 		/*
754 		 * A held page may be undergoing I/O, so skip it.
755 		 */
756 		if (m->hold_count) {
757 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
758 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
759 			addl_page_shortage++;
760 			continue;
761 		}
762 
763 		/*
764 		 * Dont mess with busy pages, keep in the front of the
765 		 * queue, most likely are being paged out.
766 		 */
767 		if (m->busy || (m->flags & PG_BUSY)) {
768 			addl_page_shortage++;
769 			continue;
770 		}
771 
772 		if (m->object->ref_count == 0) {
773 			/*
774 			 * If the object is not being used, we ignore previous
775 			 * references.
776 			 */
777 			vm_page_flag_clear(m, PG_REFERENCED);
778 			pmap_clear_reference(m);
779 
780 		} else if (((m->flags & PG_REFERENCED) == 0) &&
781 			    (actcount = pmap_ts_referenced(m))) {
782 			/*
783 			 * Otherwise, if the page has been referenced while
784 			 * in the inactive queue, we bump the "activation
785 			 * count" upwards, making it less likely that the
786 			 * page will be added back to the inactive queue
787 			 * prematurely again.  Here we check the page tables
788 			 * (or emulated bits, if any), given the upper level
789 			 * VM system not knowing anything about existing
790 			 * references.
791 			 */
792 			vm_page_activate(m);
793 			m->act_count += (actcount + ACT_ADVANCE);
794 			continue;
795 		}
796 
797 		/*
798 		 * If the upper level VM system knows about any page
799 		 * references, we activate the page.  We also set the
800 		 * "activation count" higher than normal so that we will less
801 		 * likely place pages back onto the inactive queue again.
802 		 */
803 		if ((m->flags & PG_REFERENCED) != 0) {
804 			vm_page_flag_clear(m, PG_REFERENCED);
805 			actcount = pmap_ts_referenced(m);
806 			vm_page_activate(m);
807 			m->act_count += (actcount + ACT_ADVANCE + 1);
808 			continue;
809 		}
810 
811 		/*
812 		 * If the upper level VM system doesn't know anything about
813 		 * the page being dirty, we have to check for it again.  As
814 		 * far as the VM code knows, any partially dirty pages are
815 		 * fully dirty.
816 		 *
817 		 * Pages marked PG_WRITEABLE may be mapped into the user
818 		 * address space of a process running on another cpu.  A
819 		 * user process (without holding the MP lock) running on
820 		 * another cpu may be able to touch the page while we are
821 		 * trying to remove it.  To prevent this from occuring we
822 		 * must call pmap_remove_all() or otherwise make the page
823 		 * read-only.  If the race occured pmap_remove_all() is
824 		 * responsible for setting m->dirty.
825 		 */
826 		if (m->dirty == 0) {
827 			vm_page_test_dirty(m);
828 #if 0
829 			if (m->dirty == 0 && (m->flags & PG_WRITEABLE) != 0)
830 				pmap_remove_all(m);
831 #endif
832 		} else {
833 			vm_page_dirty(m);
834 		}
835 
836 		if (m->valid == 0) {
837 			/*
838 			 * Invalid pages can be easily freed
839 			 */
840 			vm_pageout_page_free(m);
841 			mycpu->gd_cnt.v_dfree++;
842 			--page_shortage;
843 		} else if (m->dirty == 0) {
844 			/*
845 			 * Clean pages can be placed onto the cache queue.
846 			 * This effectively frees them.
847 			 */
848 			vm_page_cache(m);
849 			--page_shortage;
850 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
851 			/*
852 			 * Dirty pages need to be paged out, but flushing
853 			 * a page is extremely expensive verses freeing
854 			 * a clean page.  Rather then artificially limiting
855 			 * the number of pages we can flush, we instead give
856 			 * dirty pages extra priority on the inactive queue
857 			 * by forcing them to be cycled through the queue
858 			 * twice before being flushed, after which the
859 			 * (now clean) page will cycle through once more
860 			 * before being freed.  This significantly extends
861 			 * the thrash point for a heavily loaded machine.
862 			 */
863 			vm_page_flag_set(m, PG_WINATCFLS);
864 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
865 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
866 		} else if (maxlaunder > 0) {
867 			/*
868 			 * We always want to try to flush some dirty pages if
869 			 * we encounter them, to keep the system stable.
870 			 * Normally this number is small, but under extreme
871 			 * pressure where there are insufficient clean pages
872 			 * on the inactive queue, we may have to go all out.
873 			 */
874 			int swap_pageouts_ok;
875 			struct vnode *vp = NULL;
876 
877 			object = m->object;
878 
879 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
880 				swap_pageouts_ok = 1;
881 			} else {
882 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
883 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
884 				vm_page_count_min());
885 
886 			}
887 
888 			/*
889 			 * We don't bother paging objects that are "dead".
890 			 * Those objects are in a "rundown" state.
891 			 */
892 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
893 				TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
894 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
895 				continue;
896 			}
897 
898 			/*
899 			 * The object is already known NOT to be dead.   It
900 			 * is possible for the vget() to block the whole
901 			 * pageout daemon, but the new low-memory handling
902 			 * code should prevent it.
903 			 *
904 			 * The previous code skipped locked vnodes and, worse,
905 			 * reordered pages in the queue.  This results in
906 			 * completely non-deterministic operation because,
907 			 * quite often, a vm_fault has initiated an I/O and
908 			 * is holding a locked vnode at just the point where
909 			 * the pageout daemon is woken up.
910 			 *
911 			 * We can't wait forever for the vnode lock, we might
912 			 * deadlock due to a vn_read() getting stuck in
913 			 * vm_wait while holding this vnode.  We skip the
914 			 * vnode if we can't get it in a reasonable amount
915 			 * of time.
916 			 */
917 
918 			if (object->type == OBJT_VNODE) {
919 				vp = object->handle;
920 
921 				if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) {
922 					++pageout_lock_miss;
923 					if (object->flags & OBJ_MIGHTBEDIRTY)
924 						    vnodes_skipped++;
925 					continue;
926 				}
927 
928 				/*
929 				 * The page might have been moved to another
930 				 * queue during potential blocking in vget()
931 				 * above.  The page might have been freed and
932 				 * reused for another vnode.  The object might
933 				 * have been reused for another vnode.
934 				 */
935 				if (m->queue != PQ_INACTIVE ||
936 				    m->object != object ||
937 				    object->handle != vp) {
938 					if (object->flags & OBJ_MIGHTBEDIRTY)
939 						vnodes_skipped++;
940 					vput(vp);
941 					continue;
942 				}
943 
944 				/*
945 				 * The page may have been busied during the
946 				 * blocking in vput();  We don't move the
947 				 * page back onto the end of the queue so that
948 				 * statistics are more correct if we don't.
949 				 */
950 				if (m->busy || (m->flags & PG_BUSY)) {
951 					vput(vp);
952 					continue;
953 				}
954 
955 				/*
956 				 * If the page has become held it might
957 				 * be undergoing I/O, so skip it
958 				 */
959 				if (m->hold_count) {
960 					TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
961 					TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
962 					if (object->flags & OBJ_MIGHTBEDIRTY)
963 						vnodes_skipped++;
964 					vput(vp);
965 					continue;
966 				}
967 			}
968 
969 			/*
970 			 * If a page is dirty, then it is either being washed
971 			 * (but not yet cleaned) or it is still in the
972 			 * laundry.  If it is still in the laundry, then we
973 			 * start the cleaning operation.
974 			 *
975 			 * This operation may cluster, invalidating the 'next'
976 			 * pointer.  To prevent an inordinate number of
977 			 * restarts we use our marker to remember our place.
978 			 *
979 			 * decrement page_shortage on success to account for
980 			 * the (future) cleaned page.  Otherwise we could wind
981 			 * up laundering or cleaning too many pages.
982 			 */
983 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
984 			if (vm_pageout_clean(m) != 0) {
985 				--page_shortage;
986 				--maxlaunder;
987 			}
988 			next = TAILQ_NEXT(&marker, pageq);
989 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
990 			if (vp != NULL)
991 				vput(vp);
992 		}
993 	}
994 
995 	/*
996 	 * Compute the number of pages we want to try to move from the
997 	 * active queue to the inactive queue.
998 	 */
999 	page_shortage = vm_paging_target() +
1000 	    vmstats.v_inactive_target - vmstats.v_inactive_count;
1001 	page_shortage += addl_page_shortage;
1002 
1003 	/*
1004 	 * Scan the active queue for things we can deactivate. We nominally
1005 	 * track the per-page activity counter and use it to locate
1006 	 * deactivation candidates.
1007 	 *
1008 	 * NOTE: we are still in a critical section.
1009 	 */
1010 	pcount = vmstats.v_active_count;
1011 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1012 
1013 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1014 		/*
1015 		 * Give interrupts a chance.
1016 		 */
1017 		crit_exit();
1018 		crit_enter();
1019 
1020 		/*
1021 		 * If the page was ripped out from under us, just stop.
1022 		 */
1023 		if (m->queue != PQ_ACTIVE)
1024 			break;
1025 		next = TAILQ_NEXT(m, pageq);
1026 
1027 		/*
1028 		 * Don't deactivate pages that are busy.
1029 		 */
1030 		if ((m->busy != 0) ||
1031 		    (m->flags & PG_BUSY) ||
1032 		    (m->hold_count != 0)) {
1033 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1034 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1035 			m = next;
1036 			continue;
1037 		}
1038 
1039 		/*
1040 		 * The count for pagedaemon pages is done after checking the
1041 		 * page for eligibility...
1042 		 */
1043 		mycpu->gd_cnt.v_pdpages++;
1044 
1045 		/*
1046 		 * Check to see "how much" the page has been used.
1047 		 */
1048 		actcount = 0;
1049 		if (m->object->ref_count != 0) {
1050 			if (m->flags & PG_REFERENCED) {
1051 				actcount += 1;
1052 			}
1053 			actcount += pmap_ts_referenced(m);
1054 			if (actcount) {
1055 				m->act_count += ACT_ADVANCE + actcount;
1056 				if (m->act_count > ACT_MAX)
1057 					m->act_count = ACT_MAX;
1058 			}
1059 		}
1060 
1061 		/*
1062 		 * Since we have "tested" this bit, we need to clear it now.
1063 		 */
1064 		vm_page_flag_clear(m, PG_REFERENCED);
1065 
1066 		/*
1067 		 * Only if an object is currently being used, do we use the
1068 		 * page activation count stats.
1069 		 */
1070 		if (actcount && (m->object->ref_count != 0)) {
1071 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1072 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1073 		} else {
1074 			m->act_count -= min(m->act_count, ACT_DECLINE);
1075 			if (vm_pageout_algorithm ||
1076 			    m->object->ref_count == 0 ||
1077 			    m->act_count < pass) {
1078 				page_shortage--;
1079 				if (m->object->ref_count == 0) {
1080 					vm_page_protect(m, VM_PROT_NONE);
1081 					if (m->dirty == 0)
1082 						vm_page_cache(m);
1083 					else
1084 						vm_page_deactivate(m);
1085 				} else {
1086 					vm_page_deactivate(m);
1087 				}
1088 			} else {
1089 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1090 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1091 			}
1092 		}
1093 		m = next;
1094 	}
1095 
1096 	/*
1097 	 * We try to maintain some *really* free pages, this allows interrupt
1098 	 * code to be guaranteed space.  Since both cache and free queues
1099 	 * are considered basically 'free', moving pages from cache to free
1100 	 * does not effect other calculations.
1101 	 *
1102 	 * NOTE: we are still in a critical section.
1103 	 */
1104 
1105 	while (vmstats.v_free_count < vmstats.v_free_reserved) {
1106 		static int cache_rover = 0;
1107 		m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1108 		if (!m)
1109 			break;
1110 		if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1111 		    m->busy ||
1112 		    m->hold_count ||
1113 		    m->wire_count) {
1114 #ifdef INVARIANTS
1115 			printf("Warning: busy page %p found in cache\n", m);
1116 #endif
1117 			vm_page_deactivate(m);
1118 			continue;
1119 		}
1120 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1121 		vm_pageout_page_free(m);
1122 		mycpu->gd_cnt.v_dfree++;
1123 	}
1124 
1125 	crit_exit();
1126 
1127 #if !defined(NO_SWAPPING)
1128 	/*
1129 	 * Idle process swapout -- run once per second.
1130 	 */
1131 	if (vm_swap_idle_enabled) {
1132 		static long lsec;
1133 		if (time_second != lsec) {
1134 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1135 			vm_req_vmdaemon();
1136 			lsec = time_second;
1137 		}
1138 	}
1139 #endif
1140 
1141 	/*
1142 	 * If we didn't get enough free pages, and we have skipped a vnode
1143 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1144 	 * if we did not get enough free pages.
1145 	 */
1146 	if (vm_paging_target() > 0) {
1147 		if (vnodes_skipped && vm_page_count_min())
1148 			speedup_syncer();
1149 #if !defined(NO_SWAPPING)
1150 		if (vm_swap_enabled && vm_page_count_target()) {
1151 			vm_req_vmdaemon();
1152 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1153 		}
1154 #endif
1155 	}
1156 
1157 	/*
1158 	 * If we are out of swap and were not able to reach our paging
1159 	 * target, kill the largest process.
1160 	 */
1161 	if ((vm_swap_size < 64 && vm_page_count_min()) ||
1162 	    (swap_pager_full && vm_paging_target() > 0)) {
1163 #if 0
1164 	if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) {
1165 #endif
1166 		info.bigproc = NULL;
1167 		info.bigsize = 0;
1168 		allproc_scan(vm_pageout_scan_callback, &info);
1169 		if (info.bigproc != NULL) {
1170 			killproc(info.bigproc, "out of swap space");
1171 			info.bigproc->p_nice = PRIO_MIN;
1172 			info.bigproc->p_usched->resetpriority(&info.bigproc->p_lwp);
1173 			wakeup(&vmstats.v_free_count);
1174 			PRELE(info.bigproc);
1175 		}
1176 	}
1177 }
1178 
1179 static int
1180 vm_pageout_scan_callback(struct proc *p, void *data)
1181 {
1182 	struct vm_pageout_scan_info *info = data;
1183 	vm_offset_t size;
1184 
1185 	/*
1186 	 * if this is a system process, skip it
1187 	 */
1188 	if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1189 	    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1190 		return (0);
1191 	}
1192 
1193 	/*
1194 	 * if the process is in a non-running type state,
1195 	 * don't touch it.
1196 	 */
1197 	if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1198 		return (0);
1199 	}
1200 
1201 	/*
1202 	 * get the process size
1203 	 */
1204 	size = vmspace_resident_count(p->p_vmspace) +
1205 		vmspace_swap_count(p->p_vmspace);
1206 
1207 	/*
1208 	 * If the this process is bigger than the biggest one
1209 	 * remember it.
1210 	 */
1211 	if (size > info->bigsize) {
1212 		if (info->bigproc)
1213 			PRELE(info->bigproc);
1214 		PHOLD(p);
1215 		info->bigproc = p;
1216 		info->bigsize = size;
1217 	}
1218 	return(0);
1219 }
1220 
1221 /*
1222  * This routine tries to maintain the pseudo LRU active queue,
1223  * so that during long periods of time where there is no paging,
1224  * that some statistic accumulation still occurs.  This code
1225  * helps the situation where paging just starts to occur.
1226  */
1227 static void
1228 vm_pageout_page_stats(void)
1229 {
1230 	vm_page_t m,next;
1231 	int pcount,tpcount;		/* Number of pages to check */
1232 	static int fullintervalcount = 0;
1233 	int page_shortage;
1234 
1235 	page_shortage =
1236 	    (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1237 	    (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1238 
1239 	if (page_shortage <= 0)
1240 		return;
1241 
1242 	crit_enter();
1243 
1244 	pcount = vmstats.v_active_count;
1245 	fullintervalcount += vm_pageout_stats_interval;
1246 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1247 		tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1248 		if (pcount > tpcount)
1249 			pcount = tpcount;
1250 	} else {
1251 		fullintervalcount = 0;
1252 	}
1253 
1254 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1255 	while ((m != NULL) && (pcount-- > 0)) {
1256 		int actcount;
1257 
1258 		if (m->queue != PQ_ACTIVE) {
1259 			break;
1260 		}
1261 
1262 		next = TAILQ_NEXT(m, pageq);
1263 		/*
1264 		 * Don't deactivate pages that are busy.
1265 		 */
1266 		if ((m->busy != 0) ||
1267 		    (m->flags & PG_BUSY) ||
1268 		    (m->hold_count != 0)) {
1269 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1270 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1271 			m = next;
1272 			continue;
1273 		}
1274 
1275 		actcount = 0;
1276 		if (m->flags & PG_REFERENCED) {
1277 			vm_page_flag_clear(m, PG_REFERENCED);
1278 			actcount += 1;
1279 		}
1280 
1281 		actcount += pmap_ts_referenced(m);
1282 		if (actcount) {
1283 			m->act_count += ACT_ADVANCE + actcount;
1284 			if (m->act_count > ACT_MAX)
1285 				m->act_count = ACT_MAX;
1286 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1287 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1288 		} else {
1289 			if (m->act_count == 0) {
1290 				/*
1291 				 * We turn off page access, so that we have
1292 				 * more accurate RSS stats.  We don't do this
1293 				 * in the normal page deactivation when the
1294 				 * system is loaded VM wise, because the
1295 				 * cost of the large number of page protect
1296 				 * operations would be higher than the value
1297 				 * of doing the operation.
1298 				 */
1299 				vm_page_protect(m, VM_PROT_NONE);
1300 				vm_page_deactivate(m);
1301 			} else {
1302 				m->act_count -= min(m->act_count, ACT_DECLINE);
1303 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1304 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1305 			}
1306 		}
1307 
1308 		m = next;
1309 	}
1310 	crit_exit();
1311 }
1312 
1313 static int
1314 vm_pageout_free_page_calc(vm_size_t count)
1315 {
1316 	if (count < vmstats.v_page_count)
1317 		 return 0;
1318 	/*
1319 	 * free_reserved needs to include enough for the largest swap pager
1320 	 * structures plus enough for any pv_entry structs when paging.
1321 	 */
1322 	if (vmstats.v_page_count > 1024)
1323 		vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1324 	else
1325 		vmstats.v_free_min = 4;
1326 	vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1327 		vmstats.v_interrupt_free_min;
1328 	vmstats.v_free_reserved = vm_pageout_page_count +
1329 		vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1330 	vmstats.v_free_severe = vmstats.v_free_min / 2;
1331 	vmstats.v_free_min += vmstats.v_free_reserved;
1332 	vmstats.v_free_severe += vmstats.v_free_reserved;
1333 	return 1;
1334 }
1335 
1336 
1337 /*
1338  *	vm_pageout is the high level pageout daemon.
1339  */
1340 static void
1341 vm_pageout(void)
1342 {
1343 	int pass;
1344 
1345 	/*
1346 	 * Initialize some paging parameters.
1347 	 */
1348 
1349 	vmstats.v_interrupt_free_min = 2;
1350 	if (vmstats.v_page_count < 2000)
1351 		vm_pageout_page_count = 8;
1352 
1353 	vm_pageout_free_page_calc(vmstats.v_page_count);
1354 	/*
1355 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1356 	 * that these are more a measure of the VM cache queue hysteresis
1357 	 * then the VM free queue.  Specifically, v_free_target is the
1358 	 * high water mark (free+cache pages).
1359 	 *
1360 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1361 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1362 	 * be big enough to handle memory needs while the pageout daemon
1363 	 * is signalled and run to free more pages.
1364 	 */
1365 	if (vmstats.v_free_count > 6144)
1366 		vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1367 	else
1368 		vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1369 
1370 	if (vmstats.v_free_count > 2048) {
1371 		vmstats.v_cache_min = vmstats.v_free_target;
1372 		vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1373 		vmstats.v_inactive_target = (3 * vmstats.v_free_target) / 2;
1374 	} else {
1375 		vmstats.v_cache_min = 0;
1376 		vmstats.v_cache_max = 0;
1377 		vmstats.v_inactive_target = vmstats.v_free_count / 4;
1378 	}
1379 	if (vmstats.v_inactive_target > vmstats.v_free_count / 3)
1380 		vmstats.v_inactive_target = vmstats.v_free_count / 3;
1381 
1382 	/* XXX does not really belong here */
1383 	if (vm_page_max_wired == 0)
1384 		vm_page_max_wired = vmstats.v_free_count / 3;
1385 
1386 	if (vm_pageout_stats_max == 0)
1387 		vm_pageout_stats_max = vmstats.v_free_target;
1388 
1389 	/*
1390 	 * Set interval in seconds for stats scan.
1391 	 */
1392 	if (vm_pageout_stats_interval == 0)
1393 		vm_pageout_stats_interval = 5;
1394 	if (vm_pageout_full_stats_interval == 0)
1395 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1396 
1397 
1398 	/*
1399 	 * Set maximum free per pass
1400 	 */
1401 	if (vm_pageout_stats_free_max == 0)
1402 		vm_pageout_stats_free_max = 5;
1403 
1404 	swap_pager_swap_init();
1405 	pass = 0;
1406 	/*
1407 	 * The pageout daemon is never done, so loop forever.
1408 	 */
1409 	while (TRUE) {
1410 		int error;
1411 
1412 		/*
1413 		 * If we have enough free memory, wakeup waiters.  Do
1414 		 * not clear vm_pages_needed until we reach our target,
1415 		 * otherwise we may be woken up over and over again and
1416 		 * waste a lot of cpu.
1417 		 */
1418 		crit_enter();
1419 		if (vm_pages_needed && !vm_page_count_min()) {
1420 			if (vm_paging_needed() <= 0)
1421 				vm_pages_needed = 0;
1422 			wakeup(&vmstats.v_free_count);
1423 		}
1424 		if (vm_pages_needed) {
1425 			/*
1426 			 * Still not done, take a second pass without waiting
1427 			 * (unlimited dirty cleaning), otherwise sleep a bit
1428 			 * and try again.
1429 			 */
1430 			++pass;
1431 			if (pass > 1)
1432 				tsleep(&vm_pages_needed, 0, "psleep", hz/2);
1433 		} else {
1434 			/*
1435 			 * Good enough, sleep & handle stats.  Prime the pass
1436 			 * for the next run.
1437 			 */
1438 			if (pass > 1)
1439 				pass = 1;
1440 			else
1441 				pass = 0;
1442 			error = tsleep(&vm_pages_needed,
1443 				0, "psleep", vm_pageout_stats_interval * hz);
1444 			if (error && !vm_pages_needed) {
1445 				crit_exit();
1446 				pass = 0;
1447 				vm_pageout_page_stats();
1448 				continue;
1449 			}
1450 		}
1451 
1452 		if (vm_pages_needed)
1453 			mycpu->gd_cnt.v_pdwakeups++;
1454 		crit_exit();
1455 		vm_pageout_scan(pass);
1456 		vm_pageout_deficit = 0;
1457 	}
1458 }
1459 
1460 void
1461 pagedaemon_wakeup(void)
1462 {
1463 	if (!vm_pages_needed && curthread != pagethread) {
1464 		vm_pages_needed++;
1465 		wakeup(&vm_pages_needed);
1466 	}
1467 }
1468 
1469 #if !defined(NO_SWAPPING)
1470 static void
1471 vm_req_vmdaemon(void)
1472 {
1473 	static int lastrun = 0;
1474 
1475 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1476 		wakeup(&vm_daemon_needed);
1477 		lastrun = ticks;
1478 	}
1479 }
1480 
1481 static int vm_daemon_callback(struct proc *p, void *data __unused);
1482 
1483 static void
1484 vm_daemon(void)
1485 {
1486 	while (TRUE) {
1487 		tsleep(&vm_daemon_needed, 0, "psleep", 0);
1488 		if (vm_pageout_req_swapout) {
1489 			swapout_procs(vm_pageout_req_swapout);
1490 			vm_pageout_req_swapout = 0;
1491 		}
1492 		/*
1493 		 * scan the processes for exceeding their rlimits or if
1494 		 * process is swapped out -- deactivate pages
1495 		 */
1496 		allproc_scan(vm_daemon_callback, NULL);
1497 	}
1498 }
1499 
1500 static int
1501 vm_daemon_callback(struct proc *p, void *data __unused)
1502 {
1503 	vm_pindex_t limit, size;
1504 
1505 	/*
1506 	 * if this is a system process or if we have already
1507 	 * looked at this process, skip it.
1508 	 */
1509 	if (p->p_flag & (P_SYSTEM | P_WEXIT))
1510 		return (0);
1511 
1512 	/*
1513 	 * if the process is in a non-running type state,
1514 	 * don't touch it.
1515 	 */
1516 	if (p->p_stat != SRUN && p->p_stat != SSLEEP)
1517 		return (0);
1518 
1519 	/*
1520 	 * get a limit
1521 	 */
1522 	limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1523 			        p->p_rlimit[RLIMIT_RSS].rlim_max));
1524 
1525 	/*
1526 	 * let processes that are swapped out really be
1527 	 * swapped out.  Set the limit to nothing to get as
1528 	 * many pages out to swap as possible.
1529 	 */
1530 	if (p->p_flag & P_SWAPPEDOUT)
1531 		limit = 0;
1532 
1533 	size = vmspace_resident_count(p->p_vmspace);
1534 	if (limit >= 0 && size >= limit) {
1535 		vm_pageout_map_deactivate_pages(
1536 		    &p->p_vmspace->vm_map, limit);
1537 	}
1538 	return (0);
1539 }
1540 
1541 #endif
1542