xref: /freebsd/sys/vm/vm_pageout.c (revision 61afd5bb22d787b0641523e7b9b95c964d669bd5)
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 1.88 1997/01/01 04:45:05 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 	/*
219 	 * Try collapsing before it's too late.
220 	 */
221 	if (!sync && object->backing_object) {
222 		vm_object_collapse(object);
223 	}
224 
225 	mc[vm_pageout_page_count] = m;
226 	pageout_count = 1;
227 	page_base = vm_pageout_page_count;
228 	forward_okay = TRUE;
229 	if (pindex != 0)
230 		backward_okay = TRUE;
231 	else
232 		backward_okay = FALSE;
233 	/*
234 	 * Scan object for clusterable pages.
235 	 *
236 	 * We can cluster ONLY if: ->> the page is NOT
237 	 * clean, wired, busy, held, or mapped into a
238 	 * buffer, and one of the following:
239 	 * 1) The page is inactive, or a seldom used
240 	 *    active page.
241 	 * -or-
242 	 * 2) we force the issue.
243 	 */
244 	for (i = 1; (i < vm_pageout_page_count) && (forward_okay || backward_okay); i++) {
245 		vm_page_t p;
246 
247 		/*
248 		 * See if forward page is clusterable.
249 		 */
250 		if (forward_okay) {
251 			/*
252 			 * Stop forward scan at end of object.
253 			 */
254 			if ((pindex + i) > object->size) {
255 				forward_okay = FALSE;
256 				goto do_backward;
257 			}
258 			p = vm_page_lookup(object, pindex + i);
259 			if (p) {
260 				if (((p->queue - p->pc) == PQ_CACHE) ||
261 					(p->flags & PG_BUSY) || p->busy) {
262 					forward_okay = FALSE;
263 					goto do_backward;
264 				}
265 				vm_page_test_dirty(p);
266 				if ((p->dirty & p->valid) != 0 &&
267 				    ((p->queue == PQ_INACTIVE) ||
268 				     (sync == VM_PAGEOUT_FORCE)) &&
269 				    (p->wire_count == 0) &&
270 				    (p->hold_count == 0)) {
271 					mc[vm_pageout_page_count + i] = p;
272 					pageout_count++;
273 					if (pageout_count == vm_pageout_page_count)
274 						break;
275 				} else {
276 					forward_okay = FALSE;
277 				}
278 			} else {
279 				forward_okay = FALSE;
280 			}
281 		}
282 do_backward:
283 		/*
284 		 * See if backward page is clusterable.
285 		 */
286 		if (backward_okay) {
287 			/*
288 			 * Stop backward scan at beginning of object.
289 			 */
290 			if ((pindex - i) == 0) {
291 				backward_okay = FALSE;
292 			}
293 			p = vm_page_lookup(object, pindex - i);
294 			if (p) {
295 				if (((p->queue - p->pc) == PQ_CACHE) ||
296 					(p->flags & PG_BUSY) || p->busy) {
297 					backward_okay = FALSE;
298 					continue;
299 				}
300 				vm_page_test_dirty(p);
301 				if ((p->dirty & p->valid) != 0 &&
302 				    ((p->queue == PQ_INACTIVE) ||
303 				     (sync == VM_PAGEOUT_FORCE)) &&
304 				    (p->wire_count == 0) &&
305 				    (p->hold_count == 0)) {
306 					mc[vm_pageout_page_count - i] = p;
307 					pageout_count++;
308 					page_base--;
309 					if (pageout_count == vm_pageout_page_count)
310 						break;
311 				} else {
312 					backward_okay = FALSE;
313 				}
314 			} else {
315 				backward_okay = FALSE;
316 			}
317 		}
318 	}
319 
320 	/*
321 	 * we allow reads during pageouts...
322 	 */
323 	for (i = page_base; i < (page_base + pageout_count); i++) {
324 		mc[i]->flags |= PG_BUSY;
325 		vm_page_protect(mc[i], VM_PROT_READ);
326 	}
327 
328 	return vm_pageout_flush(&mc[page_base], pageout_count, sync);
329 }
330 
331 int
332 vm_pageout_flush(mc, count, sync)
333 	vm_page_t *mc;
334 	int count;
335 	int sync;
336 {
337 	register vm_object_t object;
338 	int pageout_status[count];
339 	int anyok = 0;
340 	int i;
341 
342 	object = mc[0]->object;
343 	object->paging_in_progress += count;
344 
345 	vm_pager_put_pages(object, mc, count,
346 	    ((sync || (object == kernel_object)) ? TRUE : FALSE),
347 	    pageout_status);
348 
349 	for (i = 0; i < count; i++) {
350 		vm_page_t mt = mc[i];
351 
352 		switch (pageout_status[i]) {
353 		case VM_PAGER_OK:
354 			++anyok;
355 			break;
356 		case VM_PAGER_PEND:
357 			++anyok;
358 			break;
359 		case VM_PAGER_BAD:
360 			/*
361 			 * Page outside of range of object. Right now we
362 			 * essentially lose the changes by pretending it
363 			 * worked.
364 			 */
365 			pmap_clear_modify(VM_PAGE_TO_PHYS(mt));
366 			mt->dirty = 0;
367 			break;
368 		case VM_PAGER_ERROR:
369 		case VM_PAGER_FAIL:
370 			/*
371 			 * If page couldn't be paged out, then reactivate the
372 			 * page so it doesn't clog the inactive list.  (We
373 			 * will try paging out it again later).
374 			 */
375 			if (mt->queue == PQ_INACTIVE)
376 				vm_page_activate(mt);
377 			break;
378 		case VM_PAGER_AGAIN:
379 			break;
380 		}
381 
382 
383 		/*
384 		 * If the operation is still going, leave the page busy to
385 		 * block all other accesses. Also, leave the paging in
386 		 * progress indicator set so that we don't attempt an object
387 		 * collapse.
388 		 */
389 		if (pageout_status[i] != VM_PAGER_PEND) {
390 			vm_object_pip_wakeup(object);
391 			PAGE_WAKEUP(mt);
392 		}
393 	}
394 	return anyok;
395 }
396 
397 #if !defined(NO_SWAPPING)
398 /*
399  *	vm_pageout_object_deactivate_pages
400  *
401  *	deactivate enough pages to satisfy the inactive target
402  *	requirements or if vm_page_proc_limit is set, then
403  *	deactivate all of the pages in the object and its
404  *	backing_objects.
405  *
406  *	The object and map must be locked.
407  */
408 static void
409 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only)
410 	vm_map_t map;
411 	vm_object_t object;
412 	vm_pindex_t desired;
413 	int map_remove_only;
414 {
415 	register vm_page_t p, next;
416 	int rcount;
417 	int remove_mode;
418 	int s;
419 
420 	if (object->type == OBJT_DEVICE)
421 		return;
422 
423 	while (object) {
424 		if (vm_map_pmap(map)->pm_stats.resident_count <= desired)
425 			return;
426 		if (object->paging_in_progress)
427 			return;
428 
429 		remove_mode = map_remove_only;
430 		if (object->shadow_count > 1)
431 			remove_mode = 1;
432 	/*
433 	 * scan the objects entire memory queue
434 	 */
435 		rcount = object->resident_page_count;
436 		p = TAILQ_FIRST(&object->memq);
437 		while (p && (rcount-- > 0)) {
438 			int refcount;
439 			if (vm_map_pmap(map)->pm_stats.resident_count <= desired)
440 				return;
441 			next = TAILQ_NEXT(p, listq);
442 			cnt.v_pdpages++;
443 			if (p->wire_count != 0 ||
444 			    p->hold_count != 0 ||
445 			    p->busy != 0 ||
446 			    (p->flags & PG_BUSY) ||
447 			    !pmap_page_exists(vm_map_pmap(map), VM_PAGE_TO_PHYS(p))) {
448 				p = next;
449 				continue;
450 			}
451 
452 			refcount = pmap_ts_referenced(VM_PAGE_TO_PHYS(p));
453 			if (refcount) {
454 				p->flags |= PG_REFERENCED;
455 			} else if (p->flags & PG_REFERENCED) {
456 				refcount = 1;
457 			}
458 
459 			if ((p->queue != PQ_ACTIVE) &&
460 				(p->flags & PG_REFERENCED)) {
461 				vm_page_activate(p);
462 				p->act_count += refcount;
463 				p->flags &= ~PG_REFERENCED;
464 			} else if (p->queue == PQ_ACTIVE) {
465 				if ((p->flags & PG_REFERENCED) == 0) {
466 					p->act_count -= min(p->act_count, ACT_DECLINE);
467 					if (!remove_mode && (vm_pageout_algorithm_lru || (p->act_count == 0))) {
468 						vm_page_protect(p, VM_PROT_NONE);
469 						vm_page_deactivate(p);
470 					} else {
471 						s = splvm();
472 						TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
473 						TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
474 						splx(s);
475 					}
476 				} else {
477 					p->flags &= ~PG_REFERENCED;
478 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
479 						p->act_count += ACT_ADVANCE;
480 					s = splvm();
481 					TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
482 					TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
483 					splx(s);
484 				}
485 			} else if (p->queue == PQ_INACTIVE) {
486 				vm_page_protect(p, VM_PROT_NONE);
487 			}
488 			p = next;
489 		}
490 		object = object->backing_object;
491 	}
492 	return;
493 }
494 
495 /*
496  * deactivate some number of pages in a map, try to do it fairly, but
497  * that is really hard to do.
498  */
499 static void
500 vm_pageout_map_deactivate_pages(map, desired)
501 	vm_map_t map;
502 	vm_pindex_t desired;
503 {
504 	vm_map_entry_t tmpe;
505 	vm_object_t obj, bigobj;
506 
507 	vm_map_reference(map);
508 	if (!lock_try_write(&map->lock)) {
509 		vm_map_deallocate(map);
510 		return;
511 	}
512 
513 	bigobj = NULL;
514 
515 	/*
516 	 * first, search out the biggest object, and try to free pages from
517 	 * that.
518 	 */
519 	tmpe = map->header.next;
520 	while (tmpe != &map->header) {
521 		if ((tmpe->is_sub_map == 0) && (tmpe->is_a_map == 0)) {
522 			obj = tmpe->object.vm_object;
523 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
524 				((bigobj == NULL) ||
525 				 (bigobj->resident_page_count < obj->resident_page_count))) {
526 				bigobj = obj;
527 			}
528 		}
529 		tmpe = tmpe->next;
530 	}
531 
532 	if (bigobj)
533 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
534 
535 	/*
536 	 * Next, hunt around for other pages to deactivate.  We actually
537 	 * do this search sort of wrong -- .text first is not the best idea.
538 	 */
539 	tmpe = map->header.next;
540 	while (tmpe != &map->header) {
541 		if (vm_map_pmap(map)->pm_stats.resident_count <= desired)
542 			break;
543 		if ((tmpe->is_sub_map == 0) && (tmpe->is_a_map == 0)) {
544 			obj = tmpe->object.vm_object;
545 			if (obj)
546 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
547 		}
548 		tmpe = tmpe->next;
549 	};
550 
551 	/*
552 	 * Remove all mappings if a process is swapped out, this will free page
553 	 * table pages.
554 	 */
555 	if (desired == 0)
556 		pmap_remove(vm_map_pmap(map),
557 			VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
558 	vm_map_unlock(map);
559 	vm_map_deallocate(map);
560 	return;
561 }
562 #endif
563 
564 /*
565  *	vm_pageout_scan does the dirty work for the pageout daemon.
566  */
567 static int
568 vm_pageout_scan()
569 {
570 	vm_page_t m, next;
571 	int page_shortage, addl_page_shortage, maxscan, maxlaunder, pcount;
572 	int pages_freed;
573 	struct proc *p, *bigproc;
574 	vm_offset_t size, bigsize;
575 	vm_object_t object;
576 	int force_wakeup = 0;
577 	int vnodes_skipped = 0;
578 	int s;
579 
580 	/*
581 	 * Start scanning the inactive queue for pages we can free. We keep
582 	 * scanning until we have enough free pages or we have scanned through
583 	 * the entire queue.  If we encounter dirty pages, we start cleaning
584 	 * them.
585 	 */
586 
587 	pages_freed = 0;
588 	addl_page_shortage = 0;
589 
590 	maxlaunder = (cnt.v_inactive_target > MAXLAUNDER) ?
591 	    MAXLAUNDER : cnt.v_inactive_target;
592 rescan0:
593 	maxscan = cnt.v_inactive_count;
594 	for( m = TAILQ_FIRST(&vm_page_queue_inactive);
595 
596 		(m != NULL) && (maxscan-- > 0) &&
597 			((cnt.v_cache_count + cnt.v_free_count) <
598 			(cnt.v_cache_min + cnt.v_free_target));
599 
600 		m = next) {
601 
602 		cnt.v_pdpages++;
603 
604 		if (m->queue != PQ_INACTIVE) {
605 			goto rescan0;
606 		}
607 
608 		next = TAILQ_NEXT(m, pageq);
609 
610 		if (m->hold_count) {
611 			s = splvm();
612 			TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
613 			TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
614 			splx(s);
615 			addl_page_shortage++;
616 			continue;
617 		}
618 		/*
619 		 * Dont mess with busy pages, keep in the front of the
620 		 * queue, most likely are being paged out.
621 		 */
622 		if (m->busy || (m->flags & PG_BUSY)) {
623 			addl_page_shortage++;
624 			continue;
625 		}
626 
627 		if (m->object->ref_count == 0) {
628 			m->flags &= ~PG_REFERENCED;
629 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
630 		} else if (((m->flags & PG_REFERENCED) == 0) &&
631 			pmap_ts_referenced(VM_PAGE_TO_PHYS(m))) {
632 			vm_page_activate(m);
633 			continue;
634 		}
635 
636 		if ((m->flags & PG_REFERENCED) != 0) {
637 			m->flags &= ~PG_REFERENCED;
638 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
639 			vm_page_activate(m);
640 			continue;
641 		}
642 
643 		if (m->dirty == 0) {
644 			vm_page_test_dirty(m);
645 		} else if (m->dirty != 0) {
646 			m->dirty = VM_PAGE_BITS_ALL;
647 		}
648 
649 		if (m->valid == 0) {
650 			vm_page_protect(m, VM_PROT_NONE);
651 			vm_page_free(m);
652 			cnt.v_dfree++;
653 			++pages_freed;
654 		} else if (m->dirty == 0) {
655 			vm_page_cache(m);
656 			++pages_freed;
657 		} else if (maxlaunder > 0) {
658 			int written;
659 			struct vnode *vp = NULL;
660 
661 			object = m->object;
662 			if (object->flags & OBJ_DEAD) {
663 				s = splvm();
664 				TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
665 				TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
666 				splx(s);
667 				continue;
668 			}
669 
670 			if (object->type == OBJT_VNODE) {
671 				vp = object->handle;
672 				if (VOP_ISLOCKED(vp) || vget(vp, 1)) {
673 					if ((m->queue == PQ_INACTIVE) &&
674 						(m->hold_count == 0) &&
675 						(m->busy == 0) &&
676 						(m->flags & PG_BUSY) == 0) {
677 						s = splvm();
678 						TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
679 						TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
680 						splx(s);
681 					}
682 					if (object->flags & OBJ_MIGHTBEDIRTY)
683 						++vnodes_skipped;
684 					continue;
685 				}
686 
687 				/*
688 				 * The page might have been moved to another queue
689 				 * during potential blocking in vget() above.
690 				 */
691 				if (m->queue != PQ_INACTIVE) {
692 					if (object->flags & OBJ_MIGHTBEDIRTY)
693 						++vnodes_skipped;
694 					vput(vp);
695 					continue;
696 				}
697 
698 				/*
699 				 * The page may have been busied during the blocking in
700 				 * vput();  We don't move the page back onto the end of
701 				 * the queue so that statistics are more correct if we don't.
702 				 */
703 				if (m->busy || (m->flags & PG_BUSY)) {
704 					vput(vp);
705 					continue;
706 				}
707 
708 				/*
709 				 * If the page has become held, then skip it
710 				 */
711 				if (m->hold_count) {
712 					s = splvm();
713 					TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
714 					TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
715 					splx(s);
716 					if (object->flags & OBJ_MIGHTBEDIRTY)
717 						++vnodes_skipped;
718 					vput(vp);
719 					continue;
720 				}
721 			}
722 
723 			/*
724 			 * If a page is dirty, then it is either being washed
725 			 * (but not yet cleaned) or it is still in the
726 			 * laundry.  If it is still in the laundry, then we
727 			 * start the cleaning operation.
728 			 */
729 			written = vm_pageout_clean(m, 0);
730 
731 			if (vp)
732 				vput(vp);
733 
734 			maxlaunder -= written;
735 		}
736 	}
737 
738 	/*
739 	 * Compute the page shortage.  If we are still very low on memory be
740 	 * sure that we will move a minimal amount of pages from active to
741 	 * inactive.
742 	 */
743 
744 	page_shortage = (cnt.v_inactive_target + cnt.v_cache_min) -
745 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
746 	if (page_shortage <= 0) {
747 		if (pages_freed == 0) {
748 			page_shortage = cnt.v_free_min - cnt.v_free_count;
749 		} else {
750 			page_shortage = 1;
751 		}
752 	}
753 	if (addl_page_shortage) {
754 		if (page_shortage < 0)
755 			page_shortage = 0;
756 		page_shortage += addl_page_shortage;
757 	}
758 
759 	pcount = cnt.v_active_count;
760 	m = TAILQ_FIRST(&vm_page_queue_active);
761 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
762 		int refcount;
763 
764 		if (m->queue != PQ_ACTIVE) {
765 			break;
766 		}
767 
768 		next = TAILQ_NEXT(m, pageq);
769 		/*
770 		 * Don't deactivate pages that are busy.
771 		 */
772 		if ((m->busy != 0) ||
773 		    (m->flags & PG_BUSY) ||
774 		    (m->hold_count != 0)) {
775 			s = splvm();
776 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
777 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
778 			splx(s);
779 			m = next;
780 			continue;
781 		}
782 
783 		/*
784 		 * The count for pagedaemon pages is done after checking the
785 		 * page for eligbility...
786 		 */
787 		cnt.v_pdpages++;
788 
789 		refcount = 0;
790 		if (m->object->ref_count != 0) {
791 			if (m->flags & PG_REFERENCED) {
792 				refcount += 1;
793 			}
794 			refcount += pmap_ts_referenced(VM_PAGE_TO_PHYS(m));
795 			if (refcount) {
796 				m->act_count += ACT_ADVANCE + refcount;
797 				if (m->act_count > ACT_MAX)
798 					m->act_count = ACT_MAX;
799 			}
800 		}
801 
802 		m->flags &= ~PG_REFERENCED;
803 
804 		if (refcount && (m->object->ref_count != 0)) {
805 			s = splvm();
806 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
807 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
808 			splx(s);
809 		} else {
810 			m->act_count -= min(m->act_count, ACT_DECLINE);
811 			if (vm_pageout_algorithm_lru ||
812 				(m->object->ref_count == 0) || (m->act_count == 0)) {
813 				--page_shortage;
814 				vm_page_protect(m, VM_PROT_NONE);
815 				if ((m->dirty == 0) &&
816 					(m->object->ref_count == 0)) {
817 					vm_page_cache(m);
818 				} else {
819 					vm_page_deactivate(m);
820 				}
821 			} else {
822 				s = splvm();
823 				TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
824 				TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
825 				splx(s);
826 			}
827 		}
828 		m = next;
829 	}
830 
831 	s = splvm();
832 	/*
833 	 * We try to maintain some *really* free pages, this allows interrupt
834 	 * code to be guaranteed space.
835 	 */
836 	while (cnt.v_free_count < cnt.v_free_reserved) {
837 		static int cache_rover = 0;
838 		m = vm_page_list_find(PQ_CACHE, cache_rover);
839 		if (!m)
840 			break;
841 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
842 		vm_page_free(m);
843 		cnt.v_dfree++;
844 	}
845 	splx(s);
846 
847 	/*
848 	 * If we didn't get enough free pages, and we have skipped a vnode
849 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
850 	 * if we did not get enough free pages.
851 	 */
852 	if ((cnt.v_cache_count + cnt.v_free_count) <
853 		(cnt.v_free_target + cnt.v_cache_min) ) {
854 		if (vnodes_skipped &&
855 		    (cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_min) {
856 			if (!vfs_update_wakeup) {
857 				vfs_update_wakeup = 1;
858 				wakeup(&vfs_update_wakeup);
859 			}
860 		}
861 #if !defined(NO_SWAPPING)
862 		if (vm_swapping_enabled &&
863 			(cnt.v_free_count + cnt.v_cache_count < cnt.v_free_target)) {
864 			vm_req_vmdaemon();
865 			vm_pageout_req_swapout = 1;
866 		}
867 #endif
868 	}
869 
870 
871 	/*
872 	 * make sure that we have swap space -- if we are low on memory and
873 	 * swap -- then kill the biggest process.
874 	 */
875 	if ((vm_swap_size == 0 || swap_pager_full) &&
876 	    ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min)) {
877 		bigproc = NULL;
878 		bigsize = 0;
879 		for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
880 			/*
881 			 * if this is a system process, skip it
882 			 */
883 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
884 			    ((p->p_pid < 48) && (vm_swap_size != 0))) {
885 				continue;
886 			}
887 			/*
888 			 * if the process is in a non-running type state,
889 			 * don't touch it.
890 			 */
891 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
892 				continue;
893 			}
894 			/*
895 			 * get the process size
896 			 */
897 			size = p->p_vmspace->vm_pmap.pm_stats.resident_count;
898 			/*
899 			 * if the this process is bigger than the biggest one
900 			 * remember it.
901 			 */
902 			if (size > bigsize) {
903 				bigproc = p;
904 				bigsize = size;
905 			}
906 		}
907 		if (bigproc != NULL) {
908 			killproc(bigproc, "out of swap space");
909 			bigproc->p_estcpu = 0;
910 			bigproc->p_nice = PRIO_MIN;
911 			resetpriority(bigproc);
912 			wakeup(&cnt.v_free_count);
913 		}
914 	}
915 	return force_wakeup;
916 }
917 
918 static int
919 vm_pageout_free_page_calc(count)
920 vm_size_t count;
921 {
922 	if (count < cnt.v_page_count)
923 		 return 0;
924 	/*
925 	 * free_reserved needs to include enough for the largest swap pager
926 	 * structures plus enough for any pv_entry structs when paging.
927 	 */
928 	if (cnt.v_page_count > 1024)
929 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
930 	else
931 		cnt.v_free_min = 4;
932 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
933 		cnt.v_interrupt_free_min;
934 	cnt.v_free_reserved = vm_pageout_page_count +
935 		cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
936 	cnt.v_free_min += cnt.v_free_reserved;
937 	return 1;
938 }
939 
940 
941 #ifdef unused
942 int
943 vm_pageout_free_pages(object, add)
944 vm_object_t object;
945 int add;
946 {
947 	return vm_pageout_free_page_calc(object->size);
948 }
949 #endif
950 
951 /*
952  *	vm_pageout is the high level pageout daemon.
953  */
954 static void
955 vm_pageout()
956 {
957 	(void) spl0();
958 
959 	/*
960 	 * Initialize some paging parameters.
961 	 */
962 
963 	cnt.v_interrupt_free_min = 2;
964 	if (cnt.v_page_count < 2000)
965 		vm_pageout_page_count = 8;
966 
967 	vm_pageout_free_page_calc(cnt.v_page_count);
968 	/*
969 	 * free_reserved needs to include enough for the largest swap pager
970 	 * structures plus enough for any pv_entry structs when paging.
971 	 */
972 	cnt.v_free_target = 3 * cnt.v_free_min + cnt.v_free_reserved;
973 
974 	if (cnt.v_free_count > 1024) {
975 		cnt.v_cache_max = (cnt.v_free_count - 1024) / 2;
976 		cnt.v_cache_min = (cnt.v_free_count - 1024) / 8;
977 		cnt.v_inactive_target = 2*cnt.v_cache_min + 192;
978 	} else {
979 		cnt.v_cache_min = 0;
980 		cnt.v_cache_max = 0;
981 		cnt.v_inactive_target = cnt.v_free_count / 4;
982 	}
983 
984 	/* XXX does not really belong here */
985 	if (vm_page_max_wired == 0)
986 		vm_page_max_wired = cnt.v_free_count / 3;
987 
988 
989 	swap_pager_swap_init();
990 	/*
991 	 * The pageout daemon is never done, so loop forever.
992 	 */
993 	while (TRUE) {
994 		int inactive_target;
995 		int s = splvm();
996 		if (!vm_pages_needed ||
997 			((cnt.v_free_count + cnt.v_cache_count) > cnt.v_free_min)) {
998 			vm_pages_needed = 0;
999 			tsleep(&vm_pages_needed, PVM, "psleep", 0);
1000 		} else if (!vm_pages_needed) {
1001 			tsleep(&vm_pages_needed, PVM, "psleep", hz/10);
1002 		}
1003 		inactive_target =
1004 			(cnt.v_page_count - cnt.v_wire_count) / 4;
1005 		if (inactive_target < 2*cnt.v_free_min)
1006 			inactive_target = 2*cnt.v_free_min;
1007 		cnt.v_inactive_target = inactive_target;
1008 		if (vm_pages_needed)
1009 			cnt.v_pdwakeups++;
1010 		vm_pages_needed = 0;
1011 		splx(s);
1012 		vm_pager_sync();
1013 		vm_pageout_scan();
1014 		vm_pager_sync();
1015 		wakeup(&cnt.v_free_count);
1016 	}
1017 }
1018 
1019 void
1020 pagedaemon_wakeup()
1021 {
1022 	if (!vm_pages_needed && curproc != pageproc) {
1023 		vm_pages_needed++;
1024 		wakeup(&vm_pages_needed);
1025 	}
1026 }
1027 
1028 #if !defined(NO_SWAPPING)
1029 static void
1030 vm_req_vmdaemon()
1031 {
1032 	static int lastrun = 0;
1033 
1034 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1035 		wakeup(&vm_daemon_needed);
1036 		lastrun = ticks;
1037 	}
1038 }
1039 
1040 static void
1041 vm_daemon()
1042 {
1043 	vm_object_t object;
1044 	struct proc *p;
1045 
1046 	(void) spl0();
1047 
1048 	while (TRUE) {
1049 		tsleep(&vm_daemon_needed, PUSER, "psleep", 0);
1050 		if (vm_pageout_req_swapout) {
1051 			swapout_procs();
1052 			vm_pageout_req_swapout = 0;
1053 		}
1054 		/*
1055 		 * scan the processes for exceeding their rlimits or if
1056 		 * process is swapped out -- deactivate pages
1057 		 */
1058 
1059 		for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1060 			quad_t limit;
1061 			vm_offset_t size;
1062 
1063 			/*
1064 			 * if this is a system process or if we have already
1065 			 * looked at this process, skip it.
1066 			 */
1067 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1068 				continue;
1069 			}
1070 			/*
1071 			 * if the process is in a non-running type state,
1072 			 * don't touch it.
1073 			 */
1074 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1075 				continue;
1076 			}
1077 			/*
1078 			 * get a limit
1079 			 */
1080 			limit = qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1081 			    p->p_rlimit[RLIMIT_RSS].rlim_max);
1082 
1083 			/*
1084 			 * let processes that are swapped out really be
1085 			 * swapped out set the limit to nothing (will force a
1086 			 * swap-out.)
1087 			 */
1088 			if ((p->p_flag & P_INMEM) == 0)
1089 				limit = 0;	/* XXX */
1090 
1091 			size = p->p_vmspace->vm_pmap.pm_stats.resident_count * PAGE_SIZE;
1092 			if (limit >= 0 && size >= limit) {
1093 				vm_pageout_map_deactivate_pages(&p->p_vmspace->vm_map,
1094 				    (vm_pindex_t)(limit >> PAGE_SHIFT) );
1095 			}
1096 		}
1097 
1098 		/*
1099 		 * we remove cached objects that have no RSS...
1100 		 */
1101 restart:
1102 		object = TAILQ_FIRST(&vm_object_cached_list);
1103 		while (object) {
1104 			/*
1105 			 * if there are no resident pages -- get rid of the object
1106 			 */
1107 			if (object->resident_page_count == 0) {
1108 				vm_object_reference(object);
1109 				pager_cache(object, FALSE);
1110 				goto restart;
1111 			}
1112 			object = TAILQ_NEXT(object, cached_list);
1113 		}
1114 	}
1115 }
1116 #endif
1117