xref: /freebsd/sys/vm/vm_pageout.c (revision 1b56bb4ca7afcf0f76feac28ddb97f5b78923913)
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  * $Id: vm_pageout.c,v 1.87 1996/11/28 23:15:06 dyson Exp $
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/proc.h>
79 #include <sys/resourcevar.h>
80 #include <sys/malloc.h>
81 #include <sys/kernel.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 <vm/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_kern.h>
96 #include <vm/vm_pager.h>
97 #include <vm/swap_pager.h>
98 #include <vm/vm_extern.h>
99 
100 /*
101  * System initialization
102  */
103 
104 /* the kernel process "vm_pageout"*/
105 static void vm_pageout __P((void));
106 static int vm_pageout_clean __P((vm_page_t, int));
107 static int vm_pageout_scan __P((void));
108 static int vm_pageout_free_page_calc __P((vm_size_t count));
109 struct proc *pageproc;
110 
111 static struct kproc_desc page_kp = {
112 	"pagedaemon",
113 	vm_pageout,
114 	&pageproc
115 };
116 SYSINIT_KT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
117 
118 #if !defined(NO_SWAPPING)
119 /* the kernel process "vm_daemon"*/
120 static void vm_daemon __P((void));
121 static struct	proc *vmproc;
122 
123 static struct kproc_desc vm_kp = {
124 	"vmdaemon",
125 	vm_daemon,
126 	&vmproc
127 };
128 SYSINIT_KT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
129 #endif
130 
131 
132 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
133 
134 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
135 
136 extern int npendingio;
137 #if !defined(NO_SWAPPING)
138 static int vm_pageout_req_swapout;	/* XXX */
139 static int vm_daemon_needed;
140 #endif
141 extern int nswiodone;
142 extern int vm_swap_size;
143 extern int vfs_update_wakeup;
144 int vm_pageout_algorithm_lru=0;
145 #if defined(NO_SWAPPING)
146 int vm_swapping_enabled=0;
147 #else
148 int vm_swapping_enabled=1;
149 #endif
150 
151 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
152 	CTLFLAG_RW, &vm_pageout_algorithm_lru, 0, "");
153 
154 #if defined(NO_SWAPPING)
155 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swapping_enabled,
156 	CTLFLAG_RD, &vm_swapping_enabled, 0, "");
157 #else
158 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swapping_enabled,
159 	CTLFLAG_RW, &vm_swapping_enabled, 0, "");
160 #endif
161 
162 #define MAXLAUNDER (cnt.v_page_count > 1800 ? 32 : 16)
163 
164 #define VM_PAGEOUT_PAGE_COUNT 16
165 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
166 
167 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
168 
169 #if !defined(NO_SWAPPING)
170 typedef void freeer_fcn_t __P((vm_map_t, vm_object_t, vm_pindex_t, int));
171 static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_pindex_t));
172 static freeer_fcn_t vm_pageout_object_deactivate_pages;
173 static void vm_req_vmdaemon __P((void));
174 #endif
175 
176 /*
177  * vm_pageout_clean:
178  *
179  * Clean the page and remove it from the laundry.
180  *
181  * We set the busy bit to cause potential page faults on this page to
182  * block.
183  *
184  * And we set pageout-in-progress to keep the object from disappearing
185  * during pageout.  This guarantees that the page won't move from the
186  * inactive queue.  (However, any other page on the inactive queue may
187  * move!)
188  */
189 static int
190 vm_pageout_clean(m, sync)
191 	vm_page_t m;
192 	int sync;
193 {
194 	register vm_object_t object;
195 	vm_page_t mc[2*vm_pageout_page_count];
196 	int pageout_count;
197 	int i, forward_okay, backward_okay, page_base;
198 	vm_pindex_t pindex = m->pindex;
199 
200 	object = m->object;
201 
202 	/*
203 	 * If not OBJT_SWAP, additional memory may be needed to do the pageout.
204 	 * Try to avoid the deadlock.
205 	 */
206 	if ((sync != VM_PAGEOUT_FORCE) &&
207 	    (object->type == OBJT_DEFAULT) &&
208 	    ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min))
209 		return 0;
210 
211 	/*
212 	 * Don't mess with the page if it's busy.
213 	 */
214 	if ((!sync && m->hold_count != 0) ||
215 	    ((m->busy != 0) || (m->flags & PG_BUSY)))
216 		return 0;
217 
218 #if defined(OLD_COLLAPSE_CODE)
219 	/*
220 	 * Try collapsing before it's too late.
221 	 */
222 	if (!sync && object->backing_object) {
223 		vm_object_collapse(object);
224 	}
225 #endif
226 
227 	mc[vm_pageout_page_count] = m;
228 	pageout_count = 1;
229 	page_base = vm_pageout_page_count;
230 	forward_okay = TRUE;
231 	if (pindex != 0)
232 		backward_okay = TRUE;
233 	else
234 		backward_okay = FALSE;
235 	/*
236 	 * Scan object for clusterable pages.
237 	 *
238 	 * We can cluster ONLY if: ->> the page is NOT
239 	 * clean, wired, busy, held, or mapped into a
240 	 * buffer, and one of the following:
241 	 * 1) The page is inactive, or a seldom used
242 	 *    active page.
243 	 * -or-
244 	 * 2) we force the issue.
245 	 */
246 	for (i = 1; (i < vm_pageout_page_count) && (forward_okay || backward_okay); i++) {
247 		vm_page_t p;
248 
249 		/*
250 		 * See if forward page is clusterable.
251 		 */
252 		if (forward_okay) {
253 			/*
254 			 * Stop forward scan at end of object.
255 			 */
256 			if ((pindex + i) > object->size) {
257 				forward_okay = FALSE;
258 				goto do_backward;
259 			}
260 			p = vm_page_lookup(object, pindex + i);
261 			if (p) {
262 				if (((p->queue - p->pc) == PQ_CACHE) ||
263 					(p->flags & PG_BUSY) || p->busy) {
264 					forward_okay = FALSE;
265 					goto do_backward;
266 				}
267 				vm_page_test_dirty(p);
268 				if ((p->dirty & p->valid) != 0 &&
269 				    ((p->queue == PQ_INACTIVE) ||
270 				     (sync == VM_PAGEOUT_FORCE)) &&
271 				    (p->wire_count == 0) &&
272 				    (p->hold_count == 0)) {
273 					mc[vm_pageout_page_count + i] = p;
274 					pageout_count++;
275 					if (pageout_count == vm_pageout_page_count)
276 						break;
277 				} else {
278 					forward_okay = FALSE;
279 				}
280 			} else {
281 				forward_okay = FALSE;
282 			}
283 		}
284 do_backward:
285 		/*
286 		 * See if backward page is clusterable.
287 		 */
288 		if (backward_okay) {
289 			/*
290 			 * Stop backward scan at beginning of object.
291 			 */
292 			if ((pindex - i) == 0) {
293 				backward_okay = FALSE;
294 			}
295 			p = vm_page_lookup(object, pindex - i);
296 			if (p) {
297 				if (((p->queue - p->pc) == PQ_CACHE) ||
298 					(p->flags & PG_BUSY) || p->busy) {
299 					backward_okay = FALSE;
300 					continue;
301 				}
302 				vm_page_test_dirty(p);
303 				if ((p->dirty & p->valid) != 0 &&
304 				    ((p->queue == PQ_INACTIVE) ||
305 				     (sync == VM_PAGEOUT_FORCE)) &&
306 				    (p->wire_count == 0) &&
307 				    (p->hold_count == 0)) {
308 					mc[vm_pageout_page_count - i] = p;
309 					pageout_count++;
310 					page_base--;
311 					if (pageout_count == vm_pageout_page_count)
312 						break;
313 				} else {
314 					backward_okay = FALSE;
315 				}
316 			} else {
317 				backward_okay = FALSE;
318 			}
319 		}
320 	}
321 
322 	/*
323 	 * we allow reads during pageouts...
324 	 */
325 	for (i = page_base; i < (page_base + pageout_count); i++) {
326 		mc[i]->flags |= PG_BUSY;
327 		vm_page_protect(mc[i], VM_PROT_READ);
328 	}
329 
330 	return vm_pageout_flush(&mc[page_base], pageout_count, sync);
331 }
332 
333 int
334 vm_pageout_flush(mc, count, sync)
335 	vm_page_t *mc;
336 	int count;
337 	int sync;
338 {
339 	register vm_object_t object;
340 	int pageout_status[count];
341 	int anyok = 0;
342 	int i;
343 
344 	object = mc[0]->object;
345 	object->paging_in_progress += count;
346 
347 	vm_pager_put_pages(object, mc, count,
348 	    ((sync || (object == kernel_object)) ? TRUE : FALSE),
349 	    pageout_status);
350 
351 	for (i = 0; i < count; i++) {
352 		vm_page_t mt = mc[i];
353 
354 		switch (pageout_status[i]) {
355 		case VM_PAGER_OK:
356 			++anyok;
357 			break;
358 		case VM_PAGER_PEND:
359 			++anyok;
360 			break;
361 		case VM_PAGER_BAD:
362 			/*
363 			 * Page outside of range of object. Right now we
364 			 * essentially lose the changes by pretending it
365 			 * worked.
366 			 */
367 			pmap_clear_modify(VM_PAGE_TO_PHYS(mt));
368 			mt->dirty = 0;
369 			break;
370 		case VM_PAGER_ERROR:
371 		case VM_PAGER_FAIL:
372 			/*
373 			 * If page couldn't be paged out, then reactivate the
374 			 * page so it doesn't clog the inactive list.  (We
375 			 * will try paging out it again later).
376 			 */
377 			if (mt->queue == PQ_INACTIVE)
378 				vm_page_activate(mt);
379 			break;
380 		case VM_PAGER_AGAIN:
381 			break;
382 		}
383 
384 
385 		/*
386 		 * If the operation is still going, leave the page busy to
387 		 * block all other accesses. Also, leave the paging in
388 		 * progress indicator set so that we don't attempt an object
389 		 * collapse.
390 		 */
391 		if (pageout_status[i] != VM_PAGER_PEND) {
392 			vm_object_pip_wakeup(object);
393 			PAGE_WAKEUP(mt);
394 		}
395 	}
396 	return anyok;
397 }
398 
399 #if !defined(NO_SWAPPING)
400 /*
401  *	vm_pageout_object_deactivate_pages
402  *
403  *	deactivate enough pages to satisfy the inactive target
404  *	requirements or if vm_page_proc_limit is set, then
405  *	deactivate all of the pages in the object and its
406  *	backing_objects.
407  *
408  *	The object and map must be locked.
409  */
410 static void
411 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only)
412 	vm_map_t map;
413 	vm_object_t object;
414 	vm_pindex_t desired;
415 	int map_remove_only;
416 {
417 	register vm_page_t p, next;
418 	int rcount;
419 	int remove_mode;
420 	int s;
421 
422 	if (object->type == OBJT_DEVICE)
423 		return;
424 
425 	while (object) {
426 		if (vm_map_pmap(map)->pm_stats.resident_count <= desired)
427 			return;
428 		if (object->paging_in_progress)
429 			return;
430 
431 		remove_mode = map_remove_only;
432 		if (object->shadow_count > 1)
433 			remove_mode = 1;
434 	/*
435 	 * scan the objects entire memory queue
436 	 */
437 		rcount = object->resident_page_count;
438 		p = TAILQ_FIRST(&object->memq);
439 		while (p && (rcount-- > 0)) {
440 			int refcount;
441 			if (vm_map_pmap(map)->pm_stats.resident_count <= desired)
442 				return;
443 			next = TAILQ_NEXT(p, listq);
444 			cnt.v_pdpages++;
445 			if (p->wire_count != 0 ||
446 			    p->hold_count != 0 ||
447 			    p->busy != 0 ||
448 			    (p->flags & PG_BUSY) ||
449 			    !pmap_page_exists(vm_map_pmap(map), VM_PAGE_TO_PHYS(p))) {
450 				p = next;
451 				continue;
452 			}
453 
454 			refcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(p));
455 			if (refcount) {
456 				p->flags |= PG_REFERENCED;
457 			} else if (p->flags & PG_REFERENCED) {
458 				refcount = 1;
459 			}
460 
461 			if ((p->queue != PQ_ACTIVE) &&
462 				(p->flags & PG_REFERENCED)) {
463 				vm_page_activate(p);
464 				p->act_count += refcount;
465 				p->flags &= ~PG_REFERENCED;
466 			} else if (p->queue == PQ_ACTIVE) {
467 				if ((p->flags & PG_REFERENCED) == 0) {
468 					p->act_count -= min(p->act_count, ACT_DECLINE);
469 					if (!remove_mode && (vm_pageout_algorithm_lru || (p->act_count == 0))) {
470 						vm_page_protect(p, VM_PROT_NONE);
471 						vm_page_deactivate(p);
472 					} else {
473 						s = splvm();
474 						TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
475 						TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
476 						splx(s);
477 					}
478 				} else {
479 					p->flags &= ~PG_REFERENCED;
480 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
481 						p->act_count += ACT_ADVANCE;
482 					s = splvm();
483 					TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
484 					TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
485 					splx(s);
486 				}
487 			} else if (p->queue == PQ_INACTIVE) {
488 				vm_page_protect(p, VM_PROT_NONE);
489 			}
490 			p = next;
491 		}
492 		object = object->backing_object;
493 	}
494 	return;
495 }
496 
497 /*
498  * deactivate some number of pages in a map, try to do it fairly, but
499  * that is really hard to do.
500  */
501 static void
502 vm_pageout_map_deactivate_pages(map, desired)
503 	vm_map_t map;
504 	vm_pindex_t desired;
505 {
506 	vm_map_entry_t tmpe;
507 	vm_object_t obj, bigobj;
508 
509 	vm_map_reference(map);
510 	if (!lock_try_write(&map->lock)) {
511 		vm_map_deallocate(map);
512 		return;
513 	}
514 
515 	bigobj = NULL;
516 
517 	/*
518 	 * first, search out the biggest object, and try to free pages from
519 	 * that.
520 	 */
521 	tmpe = map->header.next;
522 	while (tmpe != &map->header) {
523 		if ((tmpe->is_sub_map == 0) && (tmpe->is_a_map == 0)) {
524 			obj = tmpe->object.vm_object;
525 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
526 				((bigobj == NULL) ||
527 				 (bigobj->resident_page_count < obj->resident_page_count))) {
528 				bigobj = obj;
529 			}
530 		}
531 		tmpe = tmpe->next;
532 	}
533 
534 	if (bigobj)
535 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
536 
537 	/*
538 	 * Next, hunt around for other pages to deactivate.  We actually
539 	 * do this search sort of wrong -- .text first is not the best idea.
540 	 */
541 	tmpe = map->header.next;
542 	while (tmpe != &map->header) {
543 		if (vm_map_pmap(map)->pm_stats.resident_count <= desired)
544 			break;
545 		if ((tmpe->is_sub_map == 0) && (tmpe->is_a_map == 0)) {
546 			obj = tmpe->object.vm_object;
547 			if (obj)
548 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
549 		}
550 		tmpe = tmpe->next;
551 	};
552 
553 	/*
554 	 * Remove all mappings if a process is swapped out, this will free page
555 	 * table pages.
556 	 */
557 	if (desired == 0)
558 		pmap_remove(vm_map_pmap(map),
559 			VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
560 	vm_map_unlock(map);
561 	vm_map_deallocate(map);
562 	return;
563 }
564 #endif
565 
566 /*
567  *	vm_pageout_scan does the dirty work for the pageout daemon.
568  */
569 static int
570 vm_pageout_scan()
571 {
572 	vm_page_t m, next;
573 	int page_shortage, addl_page_shortage, maxscan, maxlaunder, pcount;
574 	int pages_freed;
575 	struct proc *p, *bigproc;
576 	vm_offset_t size, bigsize;
577 	vm_object_t object;
578 	int force_wakeup = 0;
579 	int vnodes_skipped = 0;
580 	int s;
581 
582 	/*
583 	 * Start scanning the inactive queue for pages we can free. We keep
584 	 * scanning until we have enough free pages or we have scanned through
585 	 * the entire queue.  If we encounter dirty pages, we start cleaning
586 	 * them.
587 	 */
588 
589 	pages_freed = 0;
590 	addl_page_shortage = 0;
591 
592 	maxlaunder = (cnt.v_inactive_target > MAXLAUNDER) ?
593 	    MAXLAUNDER : cnt.v_inactive_target;
594 rescan0:
595 	maxscan = cnt.v_inactive_count;
596 	for( m = TAILQ_FIRST(&vm_page_queue_inactive);
597 
598 		(m != NULL) && (maxscan-- > 0) &&
599 			((cnt.v_cache_count + cnt.v_free_count) <
600 			(cnt.v_cache_min + cnt.v_free_target));
601 
602 		m = next) {
603 
604 		cnt.v_pdpages++;
605 
606 		if (m->queue != PQ_INACTIVE) {
607 			goto rescan0;
608 		}
609 
610 		next = TAILQ_NEXT(m, pageq);
611 
612 		if (m->hold_count) {
613 			s = splvm();
614 			TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
615 			TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
616 			splx(s);
617 			addl_page_shortage++;
618 			continue;
619 		}
620 		/*
621 		 * Dont mess with busy pages, keep in the front of the
622 		 * queue, most likely are being paged out.
623 		 */
624 		if (m->busy || (m->flags & PG_BUSY)) {
625 			addl_page_shortage++;
626 			continue;
627 		}
628 
629 		if (m->object->ref_count == 0) {
630 			m->flags &= ~PG_REFERENCED;
631 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
632 		} else if (((m->flags & PG_REFERENCED) == 0) &&
633 			pmap_ts_referenced(VM_PAGE_TO_PHYS(m))) {
634 			vm_page_activate(m);
635 			continue;
636 		}
637 
638 		if ((m->flags & PG_REFERENCED) != 0) {
639 			m->flags &= ~PG_REFERENCED;
640 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
641 			vm_page_activate(m);
642 			continue;
643 		}
644 
645 		if (m->dirty == 0) {
646 			vm_page_test_dirty(m);
647 		} else if (m->dirty != 0) {
648 			m->dirty = VM_PAGE_BITS_ALL;
649 		}
650 
651 		if (m->valid == 0) {
652 			vm_page_protect(m, VM_PROT_NONE);
653 			vm_page_free(m);
654 			cnt.v_dfree++;
655 			++pages_freed;
656 		} else if (m->dirty == 0) {
657 			vm_page_cache(m);
658 			++pages_freed;
659 		} else if (maxlaunder > 0) {
660 			int written;
661 			struct vnode *vp = NULL;
662 
663 			object = m->object;
664 			if (object->flags & OBJ_DEAD) {
665 				s = splvm();
666 				TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
667 				TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
668 				splx(s);
669 				continue;
670 			}
671 
672 			if (object->type == OBJT_VNODE) {
673 				vp = object->handle;
674 				if (VOP_ISLOCKED(vp) || vget(vp, 1)) {
675 					if ((m->queue == PQ_INACTIVE) &&
676 						(m->hold_count == 0) &&
677 						(m->busy == 0) &&
678 						(m->flags & PG_BUSY) == 0) {
679 						s = splvm();
680 						TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
681 						TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
682 						splx(s);
683 					}
684 					if (object->flags & OBJ_MIGHTBEDIRTY)
685 						++vnodes_skipped;
686 					continue;
687 				}
688 
689 				/*
690 				 * The page might have been moved to another queue
691 				 * during potential blocking in vget() above.
692 				 */
693 				if (m->queue != PQ_INACTIVE) {
694 					if (object->flags & OBJ_MIGHTBEDIRTY)
695 						++vnodes_skipped;
696 					vput(vp);
697 					continue;
698 				}
699 
700 				/*
701 				 * The page may have been busied during the blocking in
702 				 * vput();  We don't move the page back onto the end of
703 				 * the queue so that statistics are more correct if we don't.
704 				 */
705 				if (m->busy || (m->flags & PG_BUSY)) {
706 					vput(vp);
707 					continue;
708 				}
709 
710 				/*
711 				 * If the page has become held, then skip it
712 				 */
713 				if (m->hold_count) {
714 					s = splvm();
715 					TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
716 					TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
717 					splx(s);
718 					if (object->flags & OBJ_MIGHTBEDIRTY)
719 						++vnodes_skipped;
720 					vput(vp);
721 					continue;
722 				}
723 			}
724 
725 			/*
726 			 * If a page is dirty, then it is either being washed
727 			 * (but not yet cleaned) or it is still in the
728 			 * laundry.  If it is still in the laundry, then we
729 			 * start the cleaning operation.
730 			 */
731 			written = vm_pageout_clean(m, 0);
732 
733 			if (vp)
734 				vput(vp);
735 
736 			maxlaunder -= written;
737 		}
738 	}
739 
740 	/*
741 	 * Compute the page shortage.  If we are still very low on memory be
742 	 * sure that we will move a minimal amount of pages from active to
743 	 * inactive.
744 	 */
745 
746 	page_shortage = (cnt.v_inactive_target + cnt.v_cache_min) -
747 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
748 	if (page_shortage <= 0) {
749 		if (pages_freed == 0) {
750 			page_shortage = cnt.v_free_min - cnt.v_free_count;
751 		} else {
752 			page_shortage = 1;
753 		}
754 	}
755 	if (addl_page_shortage) {
756 		if (page_shortage < 0)
757 			page_shortage = 0;
758 		page_shortage += addl_page_shortage;
759 	}
760 
761 	pcount = cnt.v_active_count;
762 	m = TAILQ_FIRST(&vm_page_queue_active);
763 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
764 		int refcount;
765 
766 		if (m->queue != PQ_ACTIVE) {
767 			break;
768 		}
769 
770 		next = TAILQ_NEXT(m, pageq);
771 		/*
772 		 * Don't deactivate pages that are busy.
773 		 */
774 		if ((m->busy != 0) ||
775 		    (m->flags & PG_BUSY) ||
776 		    (m->hold_count != 0)) {
777 			s = splvm();
778 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
779 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
780 			splx(s);
781 			m = next;
782 			continue;
783 		}
784 
785 		/*
786 		 * The count for pagedaemon pages is done after checking the
787 		 * page for eligbility...
788 		 */
789 		cnt.v_pdpages++;
790 
791 		refcount = 0;
792 		if (m->object->ref_count != 0) {
793 			if (m->flags & PG_REFERENCED) {
794 				refcount += 1;
795 			}
796 			refcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m));
797 			if (refcount) {
798 				m->act_count += ACT_ADVANCE + refcount;
799 				if (m->act_count > ACT_MAX)
800 					m->act_count = ACT_MAX;
801 			}
802 		}
803 
804 		m->flags &= ~PG_REFERENCED;
805 
806 		if (refcount && (m->object->ref_count != 0)) {
807 			s = splvm();
808 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
809 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
810 			splx(s);
811 		} else {
812 			m->act_count -= min(m->act_count, ACT_DECLINE);
813 			if (vm_pageout_algorithm_lru ||
814 				(m->object->ref_count == 0) || (m->act_count == 0)) {
815 				--page_shortage;
816 				vm_page_protect(m, VM_PROT_NONE);
817 				if ((m->dirty == 0) &&
818 					(m->object->ref_count == 0)) {
819 					vm_page_cache(m);
820 				} else {
821 					vm_page_deactivate(m);
822 				}
823 			} else {
824 				s = splvm();
825 				TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
826 				TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
827 				splx(s);
828 			}
829 		}
830 		m = next;
831 	}
832 
833 	s = splvm();
834 	/*
835 	 * We try to maintain some *really* free pages, this allows interrupt
836 	 * code to be guaranteed space.
837 	 */
838 	while (cnt.v_free_count < cnt.v_free_reserved) {
839 		static int cache_rover = 0;
840 		m = vm_page_list_find(PQ_CACHE, cache_rover);
841 		if (!m)
842 			break;
843 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
844 		vm_page_free(m);
845 		cnt.v_dfree++;
846 	}
847 	splx(s);
848 
849 	/*
850 	 * If we didn't get enough free pages, and we have skipped a vnode
851 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
852 	 * if we did not get enough free pages.
853 	 */
854 	if ((cnt.v_cache_count + cnt.v_free_count) <
855 		(cnt.v_free_target + cnt.v_cache_min) ) {
856 		if (vnodes_skipped &&
857 		    (cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_min) {
858 			if (!vfs_update_wakeup) {
859 				vfs_update_wakeup = 1;
860 				wakeup(&vfs_update_wakeup);
861 			}
862 		}
863 #if !defined(NO_SWAPPING)
864 		if (vm_swapping_enabled &&
865 			(cnt.v_free_count + cnt.v_cache_count < cnt.v_free_target)) {
866 			vm_req_vmdaemon();
867 			vm_pageout_req_swapout = 1;
868 		}
869 #endif
870 	}
871 
872 
873 	/*
874 	 * make sure that we have swap space -- if we are low on memory and
875 	 * swap -- then kill the biggest process.
876 	 */
877 	if ((vm_swap_size == 0 || swap_pager_full) &&
878 	    ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min)) {
879 		bigproc = NULL;
880 		bigsize = 0;
881 		for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
882 			/*
883 			 * if this is a system process, skip it
884 			 */
885 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
886 			    ((p->p_pid < 48) && (vm_swap_size != 0))) {
887 				continue;
888 			}
889 			/*
890 			 * if the process is in a non-running type state,
891 			 * don't touch it.
892 			 */
893 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
894 				continue;
895 			}
896 			/*
897 			 * get the process size
898 			 */
899 			size = p->p_vmspace->vm_pmap.pm_stats.resident_count;
900 			/*
901 			 * if the this process is bigger than the biggest one
902 			 * remember it.
903 			 */
904 			if (size > bigsize) {
905 				bigproc = p;
906 				bigsize = size;
907 			}
908 		}
909 		if (bigproc != NULL) {
910 			killproc(bigproc, "out of swap space");
911 			bigproc->p_estcpu = 0;
912 			bigproc->p_nice = PRIO_MIN;
913 			resetpriority(bigproc);
914 			wakeup(&cnt.v_free_count);
915 		}
916 	}
917 	return force_wakeup;
918 }
919 
920 static int
921 vm_pageout_free_page_calc(count)
922 vm_size_t count;
923 {
924 	if (count < cnt.v_page_count)
925 		 return 0;
926 	/*
927 	 * free_reserved needs to include enough for the largest swap pager
928 	 * structures plus enough for any pv_entry structs when paging.
929 	 */
930 	if (cnt.v_page_count > 1024)
931 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
932 	else
933 		cnt.v_free_min = 4;
934 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
935 		cnt.v_interrupt_free_min;
936 	cnt.v_free_reserved = vm_pageout_page_count +
937 		cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
938 	cnt.v_free_min += cnt.v_free_reserved;
939 	return 1;
940 }
941 
942 
943 #ifdef unused
944 int
945 vm_pageout_free_pages(object, add)
946 vm_object_t object;
947 int add;
948 {
949 	return vm_pageout_free_page_calc(object->size);
950 }
951 #endif
952 
953 /*
954  *	vm_pageout is the high level pageout daemon.
955  */
956 static void
957 vm_pageout()
958 {
959 	(void) spl0();
960 
961 	/*
962 	 * Initialize some paging parameters.
963 	 */
964 
965 	cnt.v_interrupt_free_min = 2;
966 	if (cnt.v_page_count < 2000)
967 		vm_pageout_page_count = 8;
968 
969 	vm_pageout_free_page_calc(cnt.v_page_count);
970 	/*
971 	 * free_reserved needs to include enough for the largest swap pager
972 	 * structures plus enough for any pv_entry structs when paging.
973 	 */
974 	cnt.v_free_target = 3 * cnt.v_free_min + cnt.v_free_reserved;
975 
976 	if (cnt.v_free_count > 1024) {
977 		cnt.v_cache_max = (cnt.v_free_count - 1024) / 2;
978 		cnt.v_cache_min = (cnt.v_free_count - 1024) / 8;
979 		cnt.v_inactive_target = 2*cnt.v_cache_min + 192;
980 	} else {
981 		cnt.v_cache_min = 0;
982 		cnt.v_cache_max = 0;
983 		cnt.v_inactive_target = cnt.v_free_count / 4;
984 	}
985 
986 	/* XXX does not really belong here */
987 	if (vm_page_max_wired == 0)
988 		vm_page_max_wired = cnt.v_free_count / 3;
989 
990 
991 	swap_pager_swap_init();
992 	/*
993 	 * The pageout daemon is never done, so loop forever.
994 	 */
995 	while (TRUE) {
996 		int inactive_target;
997 		int s = splvm();
998 		if (!vm_pages_needed ||
999 			((cnt.v_free_count + cnt.v_cache_count) > cnt.v_free_min)) {
1000 			vm_pages_needed = 0;
1001 			tsleep(&vm_pages_needed, PVM, "psleep", 0);
1002 		} else if (!vm_pages_needed) {
1003 			tsleep(&vm_pages_needed, PVM, "psleep", hz/10);
1004 		}
1005 		inactive_target =
1006 			(cnt.v_page_count - cnt.v_wire_count) / 4;
1007 		if (inactive_target < 2*cnt.v_free_min)
1008 			inactive_target = 2*cnt.v_free_min;
1009 		cnt.v_inactive_target = inactive_target;
1010 		if (vm_pages_needed)
1011 			cnt.v_pdwakeups++;
1012 		vm_pages_needed = 0;
1013 		splx(s);
1014 		vm_pager_sync();
1015 		vm_pageout_scan();
1016 		vm_pager_sync();
1017 		wakeup(&cnt.v_free_count);
1018 	}
1019 }
1020 
1021 void
1022 pagedaemon_wakeup()
1023 {
1024 	if (!vm_pages_needed && curproc != pageproc) {
1025 		vm_pages_needed++;
1026 		wakeup(&vm_pages_needed);
1027 	}
1028 }
1029 
1030 #if !defined(NO_SWAPPING)
1031 static void
1032 vm_req_vmdaemon()
1033 {
1034 	static int lastrun = 0;
1035 
1036 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1037 		wakeup(&vm_daemon_needed);
1038 		lastrun = ticks;
1039 	}
1040 }
1041 
1042 static void
1043 vm_daemon()
1044 {
1045 	vm_object_t object;
1046 	struct proc *p;
1047 
1048 	(void) spl0();
1049 
1050 	while (TRUE) {
1051 		tsleep(&vm_daemon_needed, PUSER, "psleep", 0);
1052 		if (vm_pageout_req_swapout) {
1053 			swapout_procs();
1054 			vm_pageout_req_swapout = 0;
1055 		}
1056 		/*
1057 		 * scan the processes for exceeding their rlimits or if
1058 		 * process is swapped out -- deactivate pages
1059 		 */
1060 
1061 		for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1062 			quad_t limit;
1063 			vm_offset_t size;
1064 
1065 			/*
1066 			 * if this is a system process or if we have already
1067 			 * looked at this process, skip it.
1068 			 */
1069 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1070 				continue;
1071 			}
1072 			/*
1073 			 * if the process is in a non-running type state,
1074 			 * don't touch it.
1075 			 */
1076 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1077 				continue;
1078 			}
1079 			/*
1080 			 * get a limit
1081 			 */
1082 			limit = qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1083 			    p->p_rlimit[RLIMIT_RSS].rlim_max);
1084 
1085 			/*
1086 			 * let processes that are swapped out really be
1087 			 * swapped out set the limit to nothing (will force a
1088 			 * swap-out.)
1089 			 */
1090 			if ((p->p_flag & P_INMEM) == 0)
1091 				limit = 0;	/* XXX */
1092 
1093 			size = p->p_vmspace->vm_pmap.pm_stats.resident_count * PAGE_SIZE;
1094 			if (limit >= 0 && size >= limit) {
1095 				vm_pageout_map_deactivate_pages(&p->p_vmspace->vm_map,
1096 				    (vm_pindex_t)(limit >> PAGE_SHIFT) );
1097 			}
1098 		}
1099 
1100 		/*
1101 		 * we remove cached objects that have no RSS...
1102 		 */
1103 restart:
1104 		object = TAILQ_FIRST(&vm_object_cached_list);
1105 		while (object) {
1106 			/*
1107 			 * if there are no resident pages -- get rid of the object
1108 			 */
1109 			if (object->resident_page_count == 0) {
1110 				vm_object_reference(object);
1111 				pager_cache(object, FALSE);
1112 				goto restart;
1113 			}
1114 			object = TAILQ_NEXT(object, cached_list);
1115 		}
1116 	}
1117 }
1118 #endif
1119