xref: /freebsd/sys/vm/vm_pageout.c (revision a1a4f1a0d87b594d3f17a97dc0127eec1417e6f6)
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  * $FreeBSD$
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include "opt_vm.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/proc.h>
80 #include <sys/kthread.h>
81 #include <sys/resourcevar.h>
82 #include <sys/signalvar.h>
83 #include <sys/vnode.h>
84 #include <sys/vmmeter.h>
85 #include <sys/sysctl.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_param.h>
89 #include <vm/vm_prot.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 /*
100  * System initialization
101  */
102 
103 /* the kernel process "vm_pageout"*/
104 static void vm_pageout __P((void));
105 static int vm_pageout_clean __P((vm_page_t));
106 static int vm_pageout_scan __P((void));
107 static int vm_pageout_free_page_calc __P((vm_size_t count));
108 struct proc *pageproc;
109 
110 static struct kproc_desc page_kp = {
111 	"pagedaemon",
112 	vm_pageout,
113 	&pageproc
114 };
115 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
116 
117 #if !defined(NO_SWAPPING)
118 /* the kernel process "vm_daemon"*/
119 static void vm_daemon __P((void));
120 static struct	proc *vmproc;
121 
122 static struct kproc_desc vm_kp = {
123 	"vmdaemon",
124 	vm_daemon,
125 	&vmproc
126 };
127 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
128 #endif
129 
130 
131 int vm_pages_needed=0;		/* Event on which pageout daemon sleeps */
132 int vm_pageout_deficit=0;	/* Estimated number of pages deficit */
133 int vm_pageout_pages_needed=0;	/* flag saying that the pageout daemon needs pages */
134 
135 #if !defined(NO_SWAPPING)
136 static int vm_pageout_req_swapout;	/* XXX */
137 static int vm_daemon_needed;
138 #endif
139 extern int vm_swap_size;
140 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
141 static int vm_pageout_full_stats_interval = 0;
142 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm_lru=0;
143 static int defer_swap_pageouts=0;
144 static int disable_swap_pageouts=0;
145 
146 static int max_page_launder=100;
147 #if defined(NO_SWAPPING)
148 static int vm_swap_enabled=0;
149 static int vm_swap_idle_enabled=0;
150 #else
151 static int vm_swap_enabled=1;
152 static int vm_swap_idle_enabled=0;
153 #endif
154 
155 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
156 	CTLFLAG_RW, &vm_pageout_algorithm_lru, 0, "LRU page mgmt");
157 
158 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
159 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
160 
161 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
162 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
163 
164 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
165 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
166 
167 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
168 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
169 
170 #if defined(NO_SWAPPING)
171 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
172 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
173 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
174 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
175 #else
176 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
177 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
178 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
179 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
180 #endif
181 
182 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
183 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
184 
185 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
186 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
187 
188 SYSCTL_INT(_vm, OID_AUTO, max_page_launder,
189 	CTLFLAG_RW, &max_page_launder, 0, "Maximum number of pages to clean per pass");
190 
191 
192 #define VM_PAGEOUT_PAGE_COUNT 16
193 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
194 
195 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
196 
197 #if !defined(NO_SWAPPING)
198 typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int));
199 static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t));
200 static freeer_fcn_t vm_pageout_object_deactivate_pages;
201 static void vm_req_vmdaemon __P((void));
202 #endif
203 static void vm_pageout_page_stats(void);
204 
205 /*
206  * vm_pageout_clean:
207  *
208  * Clean the page and remove it from the laundry.
209  *
210  * We set the busy bit to cause potential page faults on this page to
211  * block.  Note the careful timing, however, the busy bit isn't set till
212  * late and we cannot do anything that will mess with the page.
213  */
214 
215 static int
216 vm_pageout_clean(m)
217 	vm_page_t m;
218 {
219 	register vm_object_t object;
220 	vm_page_t mc[2*vm_pageout_page_count];
221 	int pageout_count;
222 	int i, forward_okay, backward_okay, page_base;
223 	vm_pindex_t pindex = m->pindex;
224 
225 	object = m->object;
226 
227 	/*
228 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
229 	 * with the new swapper, but we could have serious problems paging
230 	 * out other object types if there is insufficient memory.
231 	 *
232 	 * Unfortunately, checking free memory here is far too late, so the
233 	 * check has been moved up a procedural level.
234 	 */
235 
236 	/*
237 	 * Don't mess with the page if it's busy.
238 	 */
239 	if ((m->hold_count != 0) ||
240 	    ((m->busy != 0) || (m->flags & PG_BUSY)))
241 		return 0;
242 
243 	mc[vm_pageout_page_count] = m;
244 	pageout_count = 1;
245 	page_base = vm_pageout_page_count;
246 	forward_okay = TRUE;
247 	if (pindex != 0)
248 		backward_okay = TRUE;
249 	else
250 		backward_okay = FALSE;
251 	/*
252 	 * Scan object for clusterable pages.
253 	 *
254 	 * We can cluster ONLY if: ->> the page is NOT
255 	 * clean, wired, busy, held, or mapped into a
256 	 * buffer, and one of the following:
257 	 * 1) The page is inactive, or a seldom used
258 	 *    active page.
259 	 * -or-
260 	 * 2) we force the issue.
261 	 */
262 	for (i = 1; (i < vm_pageout_page_count) && (forward_okay || backward_okay); i++) {
263 		vm_page_t p;
264 
265 		/*
266 		 * See if forward page is clusterable.
267 		 */
268 		if (forward_okay) {
269 			/*
270 			 * Stop forward scan at end of object.
271 			 */
272 			if ((pindex + i) > object->size) {
273 				forward_okay = FALSE;
274 				goto do_backward;
275 			}
276 			p = vm_page_lookup(object, pindex + i);
277 			if (p) {
278 				if (((p->queue - p->pc) == PQ_CACHE) ||
279 					(p->flags & PG_BUSY) || p->busy) {
280 					forward_okay = FALSE;
281 					goto do_backward;
282 				}
283 				vm_page_test_dirty(p);
284 				if ((p->dirty & p->valid) != 0 &&
285 				    (p->queue == PQ_INACTIVE) &&
286 				    (p->wire_count == 0) &&
287 				    (p->hold_count == 0)) {
288 					mc[vm_pageout_page_count + i] = p;
289 					pageout_count++;
290 					if (pageout_count == vm_pageout_page_count)
291 						break;
292 				} else {
293 					forward_okay = FALSE;
294 				}
295 			} else {
296 				forward_okay = FALSE;
297 			}
298 		}
299 do_backward:
300 		/*
301 		 * See if backward page is clusterable.
302 		 */
303 		if (backward_okay) {
304 			/*
305 			 * Stop backward scan at beginning of object.
306 			 */
307 			if ((pindex - i) == 0) {
308 				backward_okay = FALSE;
309 			}
310 			p = vm_page_lookup(object, pindex - i);
311 			if (p) {
312 				if (((p->queue - p->pc) == PQ_CACHE) ||
313 					(p->flags & PG_BUSY) || p->busy) {
314 					backward_okay = FALSE;
315 					continue;
316 				}
317 				vm_page_test_dirty(p);
318 				if ((p->dirty & p->valid) != 0 &&
319 				    (p->queue == PQ_INACTIVE) &&
320 				    (p->wire_count == 0) &&
321 				    (p->hold_count == 0)) {
322 					mc[vm_pageout_page_count - i] = p;
323 					pageout_count++;
324 					page_base--;
325 					if (pageout_count == vm_pageout_page_count)
326 						break;
327 				} else {
328 					backward_okay = FALSE;
329 				}
330 			} else {
331 				backward_okay = FALSE;
332 			}
333 		}
334 	}
335 
336 	/*
337 	 * we allow reads during pageouts...
338 	 */
339 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
340 }
341 
342 /*
343  * vm_pageout_flush() - launder the given pages
344  *
345  *	The given pages are laundered.  Note that we setup for the start of
346  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
347  *	reference count all in here rather then in the parent.  If we want
348  *	the parent to do more sophisticated things we may have to change
349  *	the ordering.
350  */
351 
352 int
353 vm_pageout_flush(mc, count, flags)
354 	vm_page_t *mc;
355 	int count;
356 	int flags;
357 {
358 	register vm_object_t object;
359 	int pageout_status[count];
360 	int numpagedout = 0;
361 	int i;
362 
363 	/*
364 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
365 	 * mark the pages read-only.
366 	 *
367 	 * We do not have to fixup the clean/dirty bits here... we can
368 	 * allow the pager to do it after the I/O completes.
369 	 */
370 
371 	for (i = 0; i < count; i++) {
372 		vm_page_io_start(mc[i]);
373 		vm_page_protect(mc[i], VM_PROT_READ);
374 	}
375 
376 	object = mc[0]->object;
377 	vm_object_pip_add(object, count);
378 
379 	vm_pager_put_pages(object, mc, count,
380 	    (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)),
381 	    pageout_status);
382 
383 	for (i = 0; i < count; i++) {
384 		vm_page_t mt = mc[i];
385 
386 		switch (pageout_status[i]) {
387 		case VM_PAGER_OK:
388 			numpagedout++;
389 			break;
390 		case VM_PAGER_PEND:
391 			numpagedout++;
392 			break;
393 		case VM_PAGER_BAD:
394 			/*
395 			 * Page outside of range of object. Right now we
396 			 * essentially lose the changes by pretending it
397 			 * worked.
398 			 */
399 			pmap_clear_modify(VM_PAGE_TO_PHYS(mt));
400 			mt->dirty = 0;
401 			break;
402 		case VM_PAGER_ERROR:
403 		case VM_PAGER_FAIL:
404 			/*
405 			 * If page couldn't be paged out, then reactivate the
406 			 * page so it doesn't clog the inactive list.  (We
407 			 * will try paging out it again later).
408 			 */
409 			vm_page_activate(mt);
410 			break;
411 		case VM_PAGER_AGAIN:
412 			break;
413 		}
414 
415 		/*
416 		 * If the operation is still going, leave the page busy to
417 		 * block all other accesses. Also, leave the paging in
418 		 * progress indicator set so that we don't attempt an object
419 		 * collapse.
420 		 */
421 		if (pageout_status[i] != VM_PAGER_PEND) {
422 			vm_object_pip_wakeup(object);
423 			vm_page_io_finish(mt);
424 		}
425 	}
426 	return numpagedout;
427 }
428 
429 #if !defined(NO_SWAPPING)
430 /*
431  *	vm_pageout_object_deactivate_pages
432  *
433  *	deactivate enough pages to satisfy the inactive target
434  *	requirements or if vm_page_proc_limit is set, then
435  *	deactivate all of the pages in the object and its
436  *	backing_objects.
437  *
438  *	The object and map must be locked.
439  */
440 static void
441 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only)
442 	vm_map_t map;
443 	vm_object_t object;
444 	vm_pindex_t desired;
445 	int map_remove_only;
446 {
447 	register vm_page_t p, next;
448 	int rcount;
449 	int remove_mode;
450 	int s;
451 
452 	if (object->type == OBJT_DEVICE)
453 		return;
454 
455 	while (object) {
456 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
457 			return;
458 		if (object->paging_in_progress)
459 			return;
460 
461 		remove_mode = map_remove_only;
462 		if (object->shadow_count > 1)
463 			remove_mode = 1;
464 	/*
465 	 * scan the objects entire memory queue
466 	 */
467 		rcount = object->resident_page_count;
468 		p = TAILQ_FIRST(&object->memq);
469 		while (p && (rcount-- > 0)) {
470 			int actcount;
471 			if (pmap_resident_count(vm_map_pmap(map)) <= desired)
472 				return;
473 			next = TAILQ_NEXT(p, listq);
474 			cnt.v_pdpages++;
475 			if (p->wire_count != 0 ||
476 			    p->hold_count != 0 ||
477 			    p->busy != 0 ||
478 			    (p->flags & PG_BUSY) ||
479 			    !pmap_page_exists(vm_map_pmap(map), VM_PAGE_TO_PHYS(p))) {
480 				p = next;
481 				continue;
482 			}
483 
484 			actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(p));
485 			if (actcount) {
486 				vm_page_flag_set(p, PG_REFERENCED);
487 			} else if (p->flags & PG_REFERENCED) {
488 				actcount = 1;
489 			}
490 
491 			if ((p->queue != PQ_ACTIVE) &&
492 				(p->flags & PG_REFERENCED)) {
493 				vm_page_activate(p);
494 				p->act_count += actcount;
495 				vm_page_flag_clear(p, PG_REFERENCED);
496 			} else if (p->queue == PQ_ACTIVE) {
497 				if ((p->flags & PG_REFERENCED) == 0) {
498 					p->act_count -= min(p->act_count, ACT_DECLINE);
499 					if (!remove_mode && (vm_pageout_algorithm_lru || (p->act_count == 0))) {
500 						vm_page_protect(p, VM_PROT_NONE);
501 						vm_page_deactivate(p);
502 					} else {
503 						s = splvm();
504 						TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
505 						TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
506 						splx(s);
507 					}
508 				} else {
509 					vm_page_activate(p);
510 					vm_page_flag_clear(p, PG_REFERENCED);
511 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
512 						p->act_count += ACT_ADVANCE;
513 					s = splvm();
514 					TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
515 					TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
516 					splx(s);
517 				}
518 			} else if (p->queue == PQ_INACTIVE) {
519 				vm_page_protect(p, VM_PROT_NONE);
520 			}
521 			p = next;
522 		}
523 		object = object->backing_object;
524 	}
525 	return;
526 }
527 
528 /*
529  * deactivate some number of pages in a map, try to do it fairly, but
530  * that is really hard to do.
531  */
532 static void
533 vm_pageout_map_deactivate_pages(map, desired)
534 	vm_map_t map;
535 	vm_pindex_t desired;
536 {
537 	vm_map_entry_t tmpe;
538 	vm_object_t obj, bigobj;
539 
540 	if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT, (void *)0, curproc)) {
541 		return;
542 	}
543 
544 	bigobj = NULL;
545 
546 	/*
547 	 * first, search out the biggest object, and try to free pages from
548 	 * that.
549 	 */
550 	tmpe = map->header.next;
551 	while (tmpe != &map->header) {
552 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
553 			obj = tmpe->object.vm_object;
554 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
555 				((bigobj == NULL) ||
556 				 (bigobj->resident_page_count < obj->resident_page_count))) {
557 				bigobj = obj;
558 			}
559 		}
560 		tmpe = tmpe->next;
561 	}
562 
563 	if (bigobj)
564 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
565 
566 	/*
567 	 * Next, hunt around for other pages to deactivate.  We actually
568 	 * do this search sort of wrong -- .text first is not the best idea.
569 	 */
570 	tmpe = map->header.next;
571 	while (tmpe != &map->header) {
572 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
573 			break;
574 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
575 			obj = tmpe->object.vm_object;
576 			if (obj)
577 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
578 		}
579 		tmpe = tmpe->next;
580 	};
581 
582 	/*
583 	 * Remove all mappings if a process is swapped out, this will free page
584 	 * table pages.
585 	 */
586 	if (desired == 0)
587 		pmap_remove(vm_map_pmap(map),
588 			VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
589 	vm_map_unlock(map);
590 	return;
591 }
592 #endif
593 
594 /*
595  * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore
596  * to vnode deadlocks.  We only do it for OBJT_DEFAULT and OBJT_SWAP objects
597  * which we know can be trivially freed.
598  */
599 
600 void
601 vm_pageout_page_free(vm_page_t m) {
602 	vm_object_t object = m->object;
603 	int type = object->type;
604 
605 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
606 		vm_object_reference(object);
607 	vm_page_busy(m);
608 	vm_page_protect(m, VM_PROT_NONE);
609 	vm_page_free(m);
610 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
611 		vm_object_deallocate(object);
612 }
613 
614 /*
615  *	vm_pageout_scan does the dirty work for the pageout daemon.
616  */
617 static int
618 vm_pageout_scan()
619 {
620 	vm_page_t m, next;
621 	int page_shortage, maxscan, pcount;
622 	int addl_page_shortage, addl_page_shortage_init;
623 	int maxlaunder;
624 	int launder_loop = 0;
625 	struct proc *p, *bigproc;
626 	vm_offset_t size, bigsize;
627 	vm_object_t object;
628 	int force_wakeup = 0;
629 	int actcount;
630 	int vnodes_skipped = 0;
631 	int s;
632 
633 	/*
634 	 * Do whatever cleanup that the pmap code can.
635 	 */
636 	pmap_collect();
637 
638 	addl_page_shortage_init = vm_pageout_deficit;
639 	vm_pageout_deficit = 0;
640 
641 	if (max_page_launder == 0)
642 		max_page_launder = 1;
643 
644 	/*
645 	 * Calculate the number of pages we want to either free or move
646 	 * to the cache.
647 	 */
648 
649 	page_shortage = (cnt.v_free_target + cnt.v_cache_min) -
650 	    (cnt.v_free_count + cnt.v_cache_count);
651 	page_shortage += addl_page_shortage_init;
652 
653 	/*
654 	 * Figure out what to do with dirty pages when they are encountered.
655 	 * Assume that 1/3 of the pages on the inactive list are clean.  If
656 	 * we think we can reach our target, disable laundering (do not
657 	 * clean any dirty pages).  If we miss the target we will loop back
658 	 * up and do a laundering run.
659 	 */
660 
661 	if (cnt.v_inactive_count / 3 > page_shortage) {
662 		maxlaunder = 0;
663 		launder_loop = 0;
664 	} else {
665 		maxlaunder =
666 		    (cnt.v_inactive_target > max_page_launder) ?
667 		    max_page_launder : cnt.v_inactive_target;
668 		launder_loop = 1;
669 	}
670 
671 	/*
672 	 * Start scanning the inactive queue for pages we can move to the
673 	 * cache or free.  The scan will stop when the target is reached or
674 	 * we have scanned the entire inactive queue.
675 	 */
676 
677 rescan0:
678 	addl_page_shortage = addl_page_shortage_init;
679 	maxscan = cnt.v_inactive_count;
680 	for (m = TAILQ_FIRST(&vm_page_queue_inactive);
681 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
682 	     m = next) {
683 
684 		cnt.v_pdpages++;
685 
686 		if (m->queue != PQ_INACTIVE) {
687 			goto rescan0;
688 		}
689 
690 		next = TAILQ_NEXT(m, pageq);
691 
692 		if (m->hold_count) {
693 			s = splvm();
694 			TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
695 			TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
696 			splx(s);
697 			addl_page_shortage++;
698 			continue;
699 		}
700 		/*
701 		 * Dont mess with busy pages, keep in the front of the
702 		 * queue, most likely are being paged out.
703 		 */
704 		if (m->busy || (m->flags & PG_BUSY)) {
705 			addl_page_shortage++;
706 			continue;
707 		}
708 
709 		/*
710 		 * If the object is not being used, we ignore previous
711 		 * references.
712 		 */
713 		if (m->object->ref_count == 0) {
714 			vm_page_flag_clear(m, PG_REFERENCED);
715 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
716 
717 		/*
718 		 * Otherwise, if the page has been referenced while in the
719 		 * inactive queue, we bump the "activation count" upwards,
720 		 * making it less likely that the page will be added back to
721 		 * the inactive queue prematurely again.  Here we check the
722 		 * page tables (or emulated bits, if any), given the upper
723 		 * level VM system not knowing anything about existing
724 		 * references.
725 		 */
726 		} else if (((m->flags & PG_REFERENCED) == 0) &&
727 			(actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m)))) {
728 			vm_page_activate(m);
729 			m->act_count += (actcount + ACT_ADVANCE);
730 			continue;
731 		}
732 
733 		/*
734 		 * If the upper level VM system knows about any page
735 		 * references, we activate the page.  We also set the
736 		 * "activation count" higher than normal so that we will less
737 		 * likely place pages back onto the inactive queue again.
738 		 */
739 		if ((m->flags & PG_REFERENCED) != 0) {
740 			vm_page_flag_clear(m, PG_REFERENCED);
741 			actcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(m));
742 			vm_page_activate(m);
743 			m->act_count += (actcount + ACT_ADVANCE + 1);
744 			continue;
745 		}
746 
747 		/*
748 		 * If the upper level VM system doesn't know anything about
749 		 * the page being dirty, we have to check for it again.  As
750 		 * far as the VM code knows, any partially dirty pages are
751 		 * fully dirty.
752 		 */
753 		if (m->dirty == 0) {
754 			vm_page_test_dirty(m);
755 		} else {
756 			vm_page_dirty(m);
757 		}
758 
759 		/*
760 		 * Invalid pages can be easily freed
761 		 */
762 		if (m->valid == 0) {
763 			vm_pageout_page_free(m);
764 			cnt.v_dfree++;
765 			--page_shortage;
766 
767 		/*
768 		 * Clean pages can be placed onto the cache queue.
769 		 */
770 		} else if (m->dirty == 0) {
771 			vm_page_cache(m);
772 			--page_shortage;
773 
774 		/*
775 		 * Dirty pages need to be paged out.  Note that we clean
776 		 * only a limited number of pages per pagedaemon pass.
777 		 */
778 		} else if (maxlaunder > 0) {
779 			int written;
780 			int swap_pageouts_ok;
781 			struct vnode *vp = NULL;
782 
783 			object = m->object;
784 
785 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
786 				swap_pageouts_ok = 1;
787 			} else {
788 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
789 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
790 					(cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min);
791 
792 			}
793 
794 			/*
795 			 * We don't bother paging objects that are "dead".
796 			 * Those objects are in a "rundown" state.
797 			 */
798 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
799 				s = splvm();
800 				TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
801 				TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
802 				splx(s);
803 				continue;
804 			}
805 
806 			/*
807 			 * For now we protect against potential memory
808 			 * deadlocks by requiring significant memory to be
809 			 * free if the object is not OBJT_DEFAULT or OBJT_SWAP.
810 			 * We do not 'trust' any other object type to operate
811 			 * with low memory, not even OBJT_DEVICE.  The VM
812 			 * allocator will special case allocations done by
813 			 * the pageout daemon so the check below actually
814 			 * does have some hysteresis in it.  It isn't the best
815 			 * solution, though.
816 			 */
817 
818 			if (object->type != OBJT_DEFAULT &&
819 			    object->type != OBJT_SWAP &&
820 			    cnt.v_free_count < cnt.v_free_reserved) {
821 				s = splvm();
822 				TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
823 				TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m,
824 				    pageq);
825 				splx(s);
826 				continue;
827 			}
828 
829 			/*
830 			 * Presumably we have sufficient free memory to do
831 			 * the more sophisticated checks and locking required
832 			 * for vnodes.
833 			 *
834 			 * The object is already known NOT to be dead.  The
835 			 * vget() may still block, though, because
836 			 * VOP_ISLOCKED() doesn't check to see if an inode
837 			 * (v_data) is associated with the vnode.  If it isn't,
838 			 * vget() will load in it from disk.  Worse, vget()
839 			 * may actually get stuck waiting on "inode" if another
840 			 * process is in the process of bringing the inode in.
841 			 * This is bad news for us either way.
842 			 *
843 			 * So for the moment we check v_data == NULL as a
844 			 * workaround.  This means that vnodes which do not
845 			 * use v_data in the way we expect probably will not
846 			 * wind up being paged out by the pager and it will be
847 			 * up to the syncer to get them.  That's better then
848 			 * us blocking here.
849 			 *
850 			 * This whole code section is bogus - we need to fix
851 			 * the vnode pager to handle vm_page_t's without us
852 			 * having to do any sophisticated VOP tests.
853 			 */
854 
855 			if (object->type == OBJT_VNODE) {
856 				vp = object->handle;
857 
858 				if (VOP_ISLOCKED(vp) ||
859 				    vp->v_data == NULL ||
860 				    vget(vp, LK_EXCLUSIVE|LK_NOOBJ, curproc)) {
861 					if ((m->queue == PQ_INACTIVE) &&
862 						(m->hold_count == 0) &&
863 						(m->busy == 0) &&
864 						(m->flags & PG_BUSY) == 0) {
865 						s = splvm();
866 						TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
867 						TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
868 						splx(s);
869 					}
870 					if (object->flags & OBJ_MIGHTBEDIRTY)
871 						vnodes_skipped++;
872 					continue;
873 				}
874 
875 				/*
876 				 * The page might have been moved to another queue
877 				 * during potential blocking in vget() above.
878 				 */
879 				if (m->queue != PQ_INACTIVE) {
880 					if (object->flags & OBJ_MIGHTBEDIRTY)
881 						vnodes_skipped++;
882 					vput(vp);
883 					continue;
884 				}
885 
886 				/*
887 				 * The page may have been busied during the blocking in
888 				 * vput();  We don't move the page back onto the end of
889 				 * the queue so that statistics are more correct if we don't.
890 				 */
891 				if (m->busy || (m->flags & PG_BUSY)) {
892 					vput(vp);
893 					continue;
894 				}
895 
896 				/*
897 				 * If the page has become held, then skip it
898 				 */
899 				if (m->hold_count) {
900 					s = splvm();
901 					TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
902 					TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
903 					splx(s);
904 					if (object->flags & OBJ_MIGHTBEDIRTY)
905 						vnodes_skipped++;
906 					vput(vp);
907 					continue;
908 				}
909 			}
910 
911 			/*
912 			 * If a page is dirty, then it is either being washed
913 			 * (but not yet cleaned) or it is still in the
914 			 * laundry.  If it is still in the laundry, then we
915 			 * start the cleaning operation.
916 			 */
917 			written = vm_pageout_clean(m);
918 			if (vp)
919 				vput(vp);
920 
921 			maxlaunder -= written;
922 		}
923 	}
924 
925 	/*
926 	 * If we still have a page shortage and we didn't launder anything,
927 	 * run the inactive scan again and launder something this time.
928 	 */
929 
930 	if (launder_loop == 0 && page_shortage > 0) {
931 		launder_loop = 1;
932 		maxlaunder =
933 		    (cnt.v_inactive_target > max_page_launder) ?
934 		    max_page_launder : cnt.v_inactive_target;
935 		goto rescan0;
936 	}
937 
938 	/*
939 	 * Compute the page shortage from the point of view of having to
940 	 * move pages from the active queue to the inactive queue.
941 	 */
942 
943 	page_shortage = (cnt.v_inactive_target + cnt.v_cache_min) -
944 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
945 	page_shortage += addl_page_shortage;
946 
947 	/*
948 	 * Scan the active queue for things we can deactivate
949 	 */
950 
951 	pcount = cnt.v_active_count;
952 	m = TAILQ_FIRST(&vm_page_queue_active);
953 
954 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
955 
956 		/*
957 		 * This is a consistancy check, and should likely be a panic
958 		 * or warning.
959 		 */
960 		if (m->queue != PQ_ACTIVE) {
961 			break;
962 		}
963 
964 		next = TAILQ_NEXT(m, pageq);
965 		/*
966 		 * Don't deactivate pages that are busy.
967 		 */
968 		if ((m->busy != 0) ||
969 		    (m->flags & PG_BUSY) ||
970 		    (m->hold_count != 0)) {
971 			s = splvm();
972 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
973 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
974 			splx(s);
975 			m = next;
976 			continue;
977 		}
978 
979 		/*
980 		 * The count for pagedaemon pages is done after checking the
981 		 * page for eligbility...
982 		 */
983 		cnt.v_pdpages++;
984 
985 		/*
986 		 * Check to see "how much" the page has been used.
987 		 */
988 		actcount = 0;
989 		if (m->object->ref_count != 0) {
990 			if (m->flags & PG_REFERENCED) {
991 				actcount += 1;
992 			}
993 			actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m));
994 			if (actcount) {
995 				m->act_count += ACT_ADVANCE + actcount;
996 				if (m->act_count > ACT_MAX)
997 					m->act_count = ACT_MAX;
998 			}
999 		}
1000 
1001 		/*
1002 		 * Since we have "tested" this bit, we need to clear it now.
1003 		 */
1004 		vm_page_flag_clear(m, PG_REFERENCED);
1005 
1006 		/*
1007 		 * Only if an object is currently being used, do we use the
1008 		 * page activation count stats.
1009 		 */
1010 		if (actcount && (m->object->ref_count != 0)) {
1011 			s = splvm();
1012 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
1013 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1014 			splx(s);
1015 		} else {
1016 			m->act_count -= min(m->act_count, ACT_DECLINE);
1017 			if (vm_pageout_algorithm_lru ||
1018 				(m->object->ref_count == 0) || (m->act_count == 0)) {
1019 				page_shortage--;
1020 				if (m->object->ref_count == 0) {
1021 					vm_page_protect(m, VM_PROT_NONE);
1022 					if (m->dirty == 0)
1023 						vm_page_cache(m);
1024 					else
1025 						vm_page_deactivate(m);
1026 				} else {
1027 					vm_page_deactivate(m);
1028 				}
1029 			} else {
1030 				s = splvm();
1031 				TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
1032 				TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1033 				splx(s);
1034 			}
1035 		}
1036 		m = next;
1037 	}
1038 
1039 	s = splvm();
1040 
1041 	/*
1042 	 * We try to maintain some *really* free pages, this allows interrupt
1043 	 * code to be guaranteed space.  Since both cache and free queues
1044 	 * are considered basically 'free', moving pages from cache to free
1045 	 * does not effect other calculations.
1046 	 */
1047 
1048 	while (cnt.v_free_count < cnt.v_free_reserved) {
1049 		static int cache_rover = 0;
1050 		m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1051 		if (!m)
1052 			break;
1053 		if ((m->flags & PG_BUSY) || m->busy || m->hold_count || m->wire_count) {
1054 #ifdef INVARIANTS
1055 			printf("Warning: busy page %p found in cache\n", m);
1056 #endif
1057 			vm_page_deactivate(m);
1058 			continue;
1059 		}
1060 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1061 		vm_pageout_page_free(m);
1062 		cnt.v_dfree++;
1063 	}
1064 	splx(s);
1065 
1066 #if !defined(NO_SWAPPING)
1067 	/*
1068 	 * Idle process swapout -- run once per second.
1069 	 */
1070 	if (vm_swap_idle_enabled) {
1071 		static long lsec;
1072 		if (time_second != lsec) {
1073 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1074 			vm_req_vmdaemon();
1075 			lsec = time_second;
1076 		}
1077 	}
1078 #endif
1079 
1080 	/*
1081 	 * If we didn't get enough free pages, and we have skipped a vnode
1082 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1083 	 * if we did not get enough free pages.
1084 	 */
1085 	if ((cnt.v_cache_count + cnt.v_free_count) <
1086 		(cnt.v_free_target + cnt.v_cache_min) ) {
1087 		if (vnodes_skipped &&
1088 		    (cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_min) {
1089 			(void) speedup_syncer();
1090 		}
1091 #if !defined(NO_SWAPPING)
1092 		if (vm_swap_enabled &&
1093 			(cnt.v_free_count + cnt.v_cache_count < cnt.v_free_target)) {
1094 			vm_req_vmdaemon();
1095 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1096 		}
1097 #endif
1098 	}
1099 
1100 	/*
1101 	 * make sure that we have swap space -- if we are low on memory and
1102 	 * swap -- then kill the biggest process.
1103 	 */
1104 	if ((vm_swap_size == 0 || swap_pager_full) &&
1105 	    ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min)) {
1106 		bigproc = NULL;
1107 		bigsize = 0;
1108 		for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1109 			/*
1110 			 * if this is a system process, skip it
1111 			 */
1112 			if ((p->p_flag & P_SYSTEM) || (p->p_lock > 0) ||
1113 			    (p->p_pid == 1) ||
1114 			    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1115 				continue;
1116 			}
1117 			/*
1118 			 * if the process is in a non-running type state,
1119 			 * don't touch it.
1120 			 */
1121 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1122 				continue;
1123 			}
1124 			/*
1125 			 * get the process size
1126 			 */
1127 			size = vmspace_resident_count(p->p_vmspace);
1128 			/*
1129 			 * if the this process is bigger than the biggest one
1130 			 * remember it.
1131 			 */
1132 			if (size > bigsize) {
1133 				bigproc = p;
1134 				bigsize = size;
1135 			}
1136 		}
1137 		if (bigproc != NULL) {
1138 			killproc(bigproc, "out of swap space");
1139 			bigproc->p_estcpu = 0;
1140 			bigproc->p_nice = PRIO_MIN;
1141 			resetpriority(bigproc);
1142 			wakeup(&cnt.v_free_count);
1143 		}
1144 	}
1145 	return force_wakeup;
1146 }
1147 
1148 /*
1149  * This routine tries to maintain the pseudo LRU active queue,
1150  * so that during long periods of time where there is no paging,
1151  * that some statistic accumlation still occurs.  This code
1152  * helps the situation where paging just starts to occur.
1153  */
1154 static void
1155 vm_pageout_page_stats()
1156 {
1157 	int s;
1158 	vm_page_t m,next;
1159 	int pcount,tpcount;		/* Number of pages to check */
1160 	static int fullintervalcount = 0;
1161 	int page_shortage;
1162 
1163 	page_shortage = (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1164 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1165 	if (page_shortage <= 0)
1166 		return;
1167 
1168 	pcount = cnt.v_active_count;
1169 	fullintervalcount += vm_pageout_stats_interval;
1170 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1171 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1172 		if (pcount > tpcount)
1173 			pcount = tpcount;
1174 	}
1175 
1176 	m = TAILQ_FIRST(&vm_page_queue_active);
1177 	while ((m != NULL) && (pcount-- > 0)) {
1178 		int actcount;
1179 
1180 		if (m->queue != PQ_ACTIVE) {
1181 			break;
1182 		}
1183 
1184 		next = TAILQ_NEXT(m, pageq);
1185 		/*
1186 		 * Don't deactivate pages that are busy.
1187 		 */
1188 		if ((m->busy != 0) ||
1189 		    (m->flags & PG_BUSY) ||
1190 		    (m->hold_count != 0)) {
1191 			s = splvm();
1192 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
1193 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1194 			splx(s);
1195 			m = next;
1196 			continue;
1197 		}
1198 
1199 		actcount = 0;
1200 		if (m->flags & PG_REFERENCED) {
1201 			vm_page_flag_clear(m, PG_REFERENCED);
1202 			actcount += 1;
1203 		}
1204 
1205 		actcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m));
1206 		if (actcount) {
1207 			m->act_count += ACT_ADVANCE + actcount;
1208 			if (m->act_count > ACT_MAX)
1209 				m->act_count = ACT_MAX;
1210 			s = splvm();
1211 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
1212 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1213 			splx(s);
1214 		} else {
1215 			if (m->act_count == 0) {
1216 				/*
1217 				 * We turn off page access, so that we have more accurate
1218 				 * RSS stats.  We don't do this in the normal page deactivation
1219 				 * when the system is loaded VM wise, because the cost of
1220 				 * the large number of page protect operations would be higher
1221 				 * than the value of doing the operation.
1222 				 */
1223 				vm_page_protect(m, VM_PROT_NONE);
1224 				vm_page_deactivate(m);
1225 			} else {
1226 				m->act_count -= min(m->act_count, ACT_DECLINE);
1227 				s = splvm();
1228 				TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
1229 				TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
1230 				splx(s);
1231 			}
1232 		}
1233 
1234 		m = next;
1235 	}
1236 }
1237 
1238 static int
1239 vm_pageout_free_page_calc(count)
1240 vm_size_t count;
1241 {
1242 	if (count < cnt.v_page_count)
1243 		 return 0;
1244 	/*
1245 	 * free_reserved needs to include enough for the largest swap pager
1246 	 * structures plus enough for any pv_entry structs when paging.
1247 	 */
1248 	if (cnt.v_page_count > 1024)
1249 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1250 	else
1251 		cnt.v_free_min = 4;
1252 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1253 		cnt.v_interrupt_free_min;
1254 	cnt.v_free_reserved = vm_pageout_page_count +
1255 		cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1256 	cnt.v_free_min += cnt.v_free_reserved;
1257 	return 1;
1258 }
1259 
1260 
1261 /*
1262  *	vm_pageout is the high level pageout daemon.
1263  */
1264 static void
1265 vm_pageout()
1266 {
1267 	/*
1268 	 * Initialize some paging parameters.
1269 	 */
1270 
1271 	cnt.v_interrupt_free_min = 2;
1272 	if (cnt.v_page_count < 2000)
1273 		vm_pageout_page_count = 8;
1274 
1275 	vm_pageout_free_page_calc(cnt.v_page_count);
1276 	/*
1277 	 * free_reserved needs to include enough for the largest swap pager
1278 	 * structures plus enough for any pv_entry structs when paging.
1279 	 */
1280 	if (cnt.v_free_count > 6144)
1281 		cnt.v_free_target = 3 * cnt.v_free_min + cnt.v_free_reserved;
1282 	else
1283 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1284 
1285 	if (cnt.v_free_count > 2048) {
1286 		cnt.v_cache_min = cnt.v_free_target;
1287 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1288 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1289 	} else {
1290 		cnt.v_cache_min = 0;
1291 		cnt.v_cache_max = 0;
1292 		cnt.v_inactive_target = cnt.v_free_count / 4;
1293 	}
1294 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1295 		cnt.v_inactive_target = cnt.v_free_count / 3;
1296 
1297 	/* XXX does not really belong here */
1298 	if (vm_page_max_wired == 0)
1299 		vm_page_max_wired = cnt.v_free_count / 3;
1300 
1301 	if (vm_pageout_stats_max == 0)
1302 		vm_pageout_stats_max = cnt.v_free_target;
1303 
1304 	/*
1305 	 * Set interval in seconds for stats scan.
1306 	 */
1307 	if (vm_pageout_stats_interval == 0)
1308 		vm_pageout_stats_interval = 5;
1309 	if (vm_pageout_full_stats_interval == 0)
1310 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1311 
1312 
1313 	/*
1314 	 * Set maximum free per pass
1315 	 */
1316 	if (vm_pageout_stats_free_max == 0)
1317 		vm_pageout_stats_free_max = 5;
1318 
1319 	max_page_launder = (cnt.v_page_count > 1800 ? 32 : 16);
1320 
1321 	curproc->p_flag |= P_BUFEXHAUST;
1322 	swap_pager_swap_init();
1323 	/*
1324 	 * The pageout daemon is never done, so loop forever.
1325 	 */
1326 	while (TRUE) {
1327 		int error;
1328 		int s = splvm();
1329 		if (!vm_pages_needed ||
1330 			((cnt.v_free_count + cnt.v_cache_count) > cnt.v_free_min)) {
1331 			vm_pages_needed = 0;
1332 			error = tsleep(&vm_pages_needed,
1333 				PVM, "psleep", vm_pageout_stats_interval * hz);
1334 			if (error && !vm_pages_needed) {
1335 				splx(s);
1336 				vm_pageout_page_stats();
1337 				continue;
1338 			}
1339 		} else if (vm_pages_needed) {
1340 			vm_pages_needed = 0;
1341 			tsleep(&vm_pages_needed, PVM, "psleep", hz/2);
1342 		}
1343 
1344 		if (vm_pages_needed)
1345 			cnt.v_pdwakeups++;
1346 		vm_pages_needed = 0;
1347 		splx(s);
1348 		vm_pageout_scan();
1349 		vm_pageout_deficit = 0;
1350 		wakeup(&cnt.v_free_count);
1351 	}
1352 }
1353 
1354 void
1355 pagedaemon_wakeup()
1356 {
1357 	if (!vm_pages_needed && curproc != pageproc) {
1358 		vm_pages_needed++;
1359 		wakeup(&vm_pages_needed);
1360 	}
1361 }
1362 
1363 #if !defined(NO_SWAPPING)
1364 static void
1365 vm_req_vmdaemon()
1366 {
1367 	static int lastrun = 0;
1368 
1369 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1370 		wakeup(&vm_daemon_needed);
1371 		lastrun = ticks;
1372 	}
1373 }
1374 
1375 static void
1376 vm_daemon()
1377 {
1378 	struct proc *p;
1379 
1380 	while (TRUE) {
1381 		tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1382 		if (vm_pageout_req_swapout) {
1383 			swapout_procs(vm_pageout_req_swapout);
1384 			vm_pageout_req_swapout = 0;
1385 		}
1386 		/*
1387 		 * scan the processes for exceeding their rlimits or if
1388 		 * process is swapped out -- deactivate pages
1389 		 */
1390 
1391 		for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1392 			vm_pindex_t limit, size;
1393 
1394 			/*
1395 			 * if this is a system process or if we have already
1396 			 * looked at this process, skip it.
1397 			 */
1398 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1399 				continue;
1400 			}
1401 			/*
1402 			 * if the process is in a non-running type state,
1403 			 * don't touch it.
1404 			 */
1405 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1406 				continue;
1407 			}
1408 			/*
1409 			 * get a limit
1410 			 */
1411 			limit = OFF_TO_IDX(
1412 			    qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1413 				p->p_rlimit[RLIMIT_RSS].rlim_max));
1414 
1415 			/*
1416 			 * let processes that are swapped out really be
1417 			 * swapped out set the limit to nothing (will force a
1418 			 * swap-out.)
1419 			 */
1420 			if ((p->p_flag & P_INMEM) == 0)
1421 				limit = 0;	/* XXX */
1422 
1423 			size = vmspace_resident_count(p->p_vmspace);
1424 			if (limit >= 0 && size >= limit) {
1425 				vm_pageout_map_deactivate_pages(
1426 				    &p->p_vmspace->vm_map, limit);
1427 			}
1428 		}
1429 	}
1430 }
1431 #endif
1432