xref: /freebsd/sys/vm/vm_pageout.c (revision 0ea3482342b4d7d6e71f3007ce4dafe445c639fd)
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 1.58 1995/10/23 05:35:48 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 
85 #include <vm/vm.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_pager.h>
90 #include <vm/swap_pager.h>
91 
92 /*
93  * System initialization
94  */
95 
96 /* the kernel process "vm_pageout"*/
97 static void vm_pageout __P((void));
98 struct proc *pageproc;
99 
100 static struct kproc_desc page_kp = {
101 	"pagedaemon",
102 	vm_pageout,
103 	&pageproc
104 };
105 SYSINIT_KT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
106 
107 /* the kernel process "vm_daemon"*/
108 static void vm_daemon __P((void));
109 struct	proc *vmproc;
110 
111 static struct kproc_desc vm_kp = {
112 	"vmdaemon",
113 	vm_daemon,
114 	&vmproc
115 };
116 SYSINIT_KT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
117 
118 
119 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
120 
121 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
122 
123 extern int npendingio;
124 int vm_pageout_req_swapout;	/* XXX */
125 int vm_daemon_needed;
126 extern int nswiodone;
127 extern int vm_swap_size;
128 extern int vfs_update_wakeup;
129 
130 #define MAXSCAN 1024		/* maximum number of pages to scan in queues */
131 
132 #define MAXLAUNDER (cnt.v_page_count > 1800 ? 32 : 16)
133 
134 #define VM_PAGEOUT_PAGE_COUNT 8
135 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
136 
137 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
138 
139 typedef int freeer_fcn_t __P((vm_map_t, vm_object_t, int, int));
140 static void vm_pageout_map_deactivate_pages __P((vm_map_t, vm_map_entry_t,
141 						 int *, freeer_fcn_t *));
142 static freeer_fcn_t vm_pageout_object_deactivate_pages;
143 static void vm_req_vmdaemon __P((void));
144 
145 /*
146  * vm_pageout_clean:
147  *
148  * Clean the page and remove it from the laundry.
149  *
150  * We set the busy bit to cause potential page faults on this page to
151  * block.
152  *
153  * And we set pageout-in-progress to keep the object from disappearing
154  * during pageout.  This guarantees that the page won't move from the
155  * inactive queue.  (However, any other page on the inactive queue may
156  * move!)
157  */
158 int
159 vm_pageout_clean(m, sync)
160 	vm_page_t m;
161 	int sync;
162 {
163 	register vm_object_t object;
164 	int pageout_status[VM_PAGEOUT_PAGE_COUNT];
165 	vm_page_t mc[2*VM_PAGEOUT_PAGE_COUNT];
166 	int pageout_count;
167 	int i, forward_okay, backward_okay, page_base;
168 	vm_offset_t offset = m->offset;
169 
170 	object = m->object;
171 
172 	/*
173 	 * If not OBJT_SWAP, additional memory may be needed to do the pageout.
174 	 * Try to avoid the deadlock.
175 	 */
176 	if ((sync != VM_PAGEOUT_FORCE) &&
177 	    (object->type != OBJT_SWAP) &&
178 	    ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_pageout_free_min))
179 		return 0;
180 
181 	/*
182 	 * Don't mess with the page if it's busy.
183 	 */
184 	if ((!sync && m->hold_count != 0) ||
185 	    ((m->busy != 0) || (m->flags & PG_BUSY)))
186 		return 0;
187 
188 	/*
189 	 * Try collapsing before it's too late.
190 	 */
191 	if (!sync && object->backing_object) {
192 		vm_object_collapse(object);
193 	}
194 	mc[VM_PAGEOUT_PAGE_COUNT] = m;
195 	pageout_count = 1;
196 	page_base = VM_PAGEOUT_PAGE_COUNT;
197 	forward_okay = TRUE;
198 	if (offset != 0)
199 		backward_okay = TRUE;
200 	else
201 		backward_okay = FALSE;
202 	/*
203 	 * Scan object for clusterable pages.
204 	 *
205 	 * We can cluster ONLY if: ->> the page is NOT
206 	 * clean, wired, busy, held, or mapped into a
207 	 * buffer, and one of the following:
208 	 * 1) The page is inactive, or a seldom used
209 	 *    active page.
210 	 * -or-
211 	 * 2) we force the issue.
212 	 */
213 	for (i = 1; (i < vm_pageout_page_count) && (forward_okay || backward_okay); i++) {
214 		vm_page_t p;
215 
216 		/*
217 		 * See if forward page is clusterable.
218 		 */
219 		if (forward_okay) {
220 			/*
221 			 * Stop forward scan at end of object.
222 			 */
223 			if ((offset + i * PAGE_SIZE) > object->size) {
224 				forward_okay = FALSE;
225 				goto do_backward;
226 			}
227 			p = vm_page_lookup(object, offset + i * PAGE_SIZE);
228 			if (p) {
229 				if ((p->flags & (PG_BUSY|PG_CACHE)) || p->busy) {
230 					forward_okay = FALSE;
231 					goto do_backward;
232 				}
233 				vm_page_test_dirty(p);
234 				if ((p->dirty & p->valid) != 0 &&
235 				    ((p->flags & PG_INACTIVE) ||
236 				     (sync == VM_PAGEOUT_FORCE)) &&
237 				    (p->wire_count == 0) &&
238 				    (p->hold_count == 0)) {
239 					mc[VM_PAGEOUT_PAGE_COUNT + i] = p;
240 					pageout_count++;
241 					if (pageout_count == vm_pageout_page_count)
242 						break;
243 				} else {
244 					forward_okay = FALSE;
245 				}
246 			} else {
247 				forward_okay = FALSE;
248 			}
249 		}
250 do_backward:
251 		/*
252 		 * See if backward page is clusterable.
253 		 */
254 		if (backward_okay) {
255 			/*
256 			 * Stop backward scan at beginning of object.
257 			 */
258 			if ((offset - i * PAGE_SIZE) == 0) {
259 				backward_okay = FALSE;
260 			}
261 			p = vm_page_lookup(object, offset - i * PAGE_SIZE);
262 			if (p) {
263 				if ((p->flags & (PG_BUSY|PG_CACHE)) || p->busy) {
264 					backward_okay = FALSE;
265 					continue;
266 				}
267 				vm_page_test_dirty(p);
268 				if ((p->dirty & p->valid) != 0 &&
269 				    ((p->flags & PG_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 					page_base--;
276 					if (pageout_count == vm_pageout_page_count)
277 						break;
278 				} else {
279 					backward_okay = FALSE;
280 				}
281 			} else {
282 				backward_okay = FALSE;
283 			}
284 		}
285 	}
286 
287 	/*
288 	 * we allow reads during pageouts...
289 	 */
290 	for (i = page_base; i < (page_base + pageout_count); i++) {
291 		mc[i]->flags |= PG_BUSY;
292 		vm_page_protect(mc[i], VM_PROT_READ);
293 	}
294 
295 	return vm_pageout_flush(&mc[page_base], pageout_count, sync);
296 }
297 
298 int
299 vm_pageout_flush(mc, count, sync)
300 	vm_page_t *mc;
301 	int count;
302 	int sync;
303 {
304 	register vm_object_t object;
305 	int pageout_status[count];
306 	int anyok = 0;
307 	int i;
308 
309 	object = mc[0]->object;
310 	object->paging_in_progress += count;
311 
312 	vm_pager_put_pages(object, mc, count,
313 	    ((sync || (object == kernel_object)) ? TRUE : FALSE),
314 	    pageout_status);
315 
316 
317 	for (i = 0; i < count; i++) {
318 		vm_page_t mt = mc[i];
319 
320 		switch (pageout_status[i]) {
321 		case VM_PAGER_OK:
322 			++anyok;
323 			break;
324 		case VM_PAGER_PEND:
325 			++anyok;
326 			break;
327 		case VM_PAGER_BAD:
328 			/*
329 			 * Page outside of range of object. Right now we
330 			 * essentially lose the changes by pretending it
331 			 * worked.
332 			 */
333 			pmap_clear_modify(VM_PAGE_TO_PHYS(mt));
334 			mt->dirty = 0;
335 			break;
336 		case VM_PAGER_ERROR:
337 		case VM_PAGER_FAIL:
338 			/*
339 			 * If page couldn't be paged out, then reactivate the
340 			 * page so it doesn't clog the inactive list.  (We
341 			 * will try paging out it again later).
342 			 */
343 			if (mt->flags & PG_INACTIVE)
344 				vm_page_activate(mt);
345 			break;
346 		case VM_PAGER_AGAIN:
347 			break;
348 		}
349 
350 
351 		/*
352 		 * If the operation is still going, leave the page busy to
353 		 * block all other accesses. Also, leave the paging in
354 		 * progress indicator set so that we don't attempt an object
355 		 * collapse.
356 		 */
357 		if (pageout_status[i] != VM_PAGER_PEND) {
358 			vm_object_pip_wakeup(object);
359 			if ((mt->flags & (PG_REFERENCED|PG_WANTED)) ||
360 			    pmap_is_referenced(VM_PAGE_TO_PHYS(mt))) {
361 				pmap_clear_reference(VM_PAGE_TO_PHYS(mt));
362 				mt->flags &= ~PG_REFERENCED;
363 				if (mt->flags & PG_INACTIVE)
364 					vm_page_activate(mt);
365 			}
366 			PAGE_WAKEUP(mt);
367 		}
368 	}
369 	return anyok;
370 }
371 
372 /*
373  *	vm_pageout_object_deactivate_pages
374  *
375  *	deactivate enough pages to satisfy the inactive target
376  *	requirements or if vm_page_proc_limit is set, then
377  *	deactivate all of the pages in the object and its
378  *	backing_objects.
379  *
380  *	The object and map must be locked.
381  */
382 static int
383 vm_pageout_object_deactivate_pages(map, object, count, map_remove_only)
384 	vm_map_t map;
385 	vm_object_t object;
386 	int count;
387 	int map_remove_only;
388 {
389 	register vm_page_t p, next;
390 	int rcount;
391 	int dcount;
392 
393 	dcount = 0;
394 	if (count == 0)
395 		count = 1;
396 
397 	if (object->type == OBJT_DEVICE)
398 		return 0;
399 
400 	if (object->backing_object) {
401 		if (object->backing_object->ref_count == 1)
402 			dcount += vm_pageout_object_deactivate_pages(map,
403 			    object->backing_object, count / 2 + 1, map_remove_only);
404 		else
405 			vm_pageout_object_deactivate_pages(map,
406 			    object->backing_object, count, 1);
407 	}
408 	if (object->paging_in_progress)
409 		return dcount;
410 
411 	/*
412 	 * scan the objects entire memory queue
413 	 */
414 	rcount = object->resident_page_count;
415 	p = object->memq.tqh_first;
416 	while (p && (rcount-- > 0)) {
417 		next = p->listq.tqe_next;
418 		cnt.v_pdpages++;
419 		if (p->wire_count != 0 ||
420 		    p->hold_count != 0 ||
421 		    p->busy != 0 ||
422 		    !pmap_page_exists(vm_map_pmap(map), VM_PAGE_TO_PHYS(p))) {
423 			p = next;
424 			continue;
425 		}
426 		/*
427 		 * if a page is active, not wired and is in the processes
428 		 * pmap, then deactivate the page.
429 		 */
430 		if ((p->flags & (PG_ACTIVE | PG_BUSY)) == PG_ACTIVE) {
431 			if (!pmap_is_referenced(VM_PAGE_TO_PHYS(p)) &&
432 			    (p->flags & (PG_REFERENCED|PG_WANTED)) == 0) {
433 				p->act_count -= min(p->act_count, ACT_DECLINE);
434 				/*
435 				 * if the page act_count is zero -- then we
436 				 * deactivate
437 				 */
438 				if (!p->act_count) {
439 					if (!map_remove_only)
440 						vm_page_deactivate(p);
441 					vm_page_protect(p, VM_PROT_NONE);
442 					/*
443 					 * else if on the next go-around we
444 					 * will deactivate the page we need to
445 					 * place the page on the end of the
446 					 * queue to age the other pages in
447 					 * memory.
448 					 */
449 				} else {
450 					TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
451 					TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
452 				}
453 				/*
454 				 * see if we are done yet
455 				 */
456 				if (p->flags & PG_INACTIVE) {
457 					--count;
458 					++dcount;
459 					if (count <= 0 &&
460 					    cnt.v_inactive_count > cnt.v_inactive_target) {
461 						return dcount;
462 					}
463 				}
464 			} else {
465 				/*
466 				 * Move the page to the bottom of the queue.
467 				 */
468 				pmap_clear_reference(VM_PAGE_TO_PHYS(p));
469 				p->flags &= ~PG_REFERENCED;
470 				if (p->act_count < ACT_MAX)
471 					p->act_count += ACT_ADVANCE;
472 
473 				TAILQ_REMOVE(&vm_page_queue_active, p, pageq);
474 				TAILQ_INSERT_TAIL(&vm_page_queue_active, p, pageq);
475 			}
476 		} else if ((p->flags & (PG_INACTIVE | PG_BUSY)) == PG_INACTIVE) {
477 			vm_page_protect(p, VM_PROT_NONE);
478 		}
479 		p = next;
480 	}
481 	return dcount;
482 }
483 
484 
485 /*
486  * deactivate some number of pages in a map, try to do it fairly, but
487  * that is really hard to do.
488  */
489 
490 static void
491 vm_pageout_map_deactivate_pages(map, entry, count, freeer)
492 	vm_map_t map;
493 	vm_map_entry_t entry;
494 	int *count;
495 	freeer_fcn_t *freeer;
496 {
497 	vm_map_t tmpm;
498 	vm_map_entry_t tmpe;
499 	vm_object_t obj;
500 
501 	if (*count <= 0)
502 		return;
503 	vm_map_reference(map);
504 	if (!lock_try_read(&map->lock)) {
505 		vm_map_deallocate(map);
506 		return;
507 	}
508 	if (entry == 0) {
509 		tmpe = map->header.next;
510 		while (tmpe != &map->header && *count > 0) {
511 			vm_pageout_map_deactivate_pages(map, tmpe, count, freeer);
512 			tmpe = tmpe->next;
513 		};
514 	} else if (entry->is_sub_map || entry->is_a_map) {
515 		tmpm = entry->object.share_map;
516 		tmpe = tmpm->header.next;
517 		while (tmpe != &tmpm->header && *count > 0) {
518 			vm_pageout_map_deactivate_pages(tmpm, tmpe, count, freeer);
519 			tmpe = tmpe->next;
520 		};
521 	} else if ((obj = entry->object.vm_object) != 0) {
522 		*count -= (*freeer) (map, obj, *count, TRUE);
523 	}
524 	lock_read_done(&map->lock);
525 	vm_map_deallocate(map);
526 	return;
527 }
528 
529 static void
530 vm_req_vmdaemon()
531 {
532 	static int lastrun = 0;
533 
534 	if ((ticks > (lastrun + hz / 10)) || (ticks < lastrun)) {
535 		wakeup(&vm_daemon_needed);
536 		lastrun = ticks;
537 	}
538 }
539 
540 /*
541  *	vm_pageout_scan does the dirty work for the pageout daemon.
542  */
543 int
544 vm_pageout_scan()
545 {
546 	vm_page_t m;
547 	int page_shortage, maxscan, maxlaunder, pcount;
548 	int pages_freed;
549 	vm_page_t next;
550 	struct proc *p, *bigproc;
551 	vm_offset_t size, bigsize;
552 	vm_object_t object;
553 	int force_wakeup = 0;
554 	int vnodes_skipped = 0;
555 
556 	pages_freed = 0;
557 
558 	/*
559 	 * Start scanning the inactive queue for pages we can free. We keep
560 	 * scanning until we have enough free pages or we have scanned through
561 	 * the entire queue.  If we encounter dirty pages, we start cleaning
562 	 * them.
563 	 */
564 
565 	maxlaunder = (cnt.v_inactive_target > MAXLAUNDER) ?
566 	    MAXLAUNDER : cnt.v_inactive_target;
567 
568 rescan1:
569 	maxscan = cnt.v_inactive_count;
570 	m = vm_page_queue_inactive.tqh_first;
571 	while ((m != NULL) && (maxscan-- > 0) &&
572 	    ((cnt.v_cache_count + cnt.v_free_count) < (cnt.v_cache_min + cnt.v_free_target))) {
573 		vm_page_t next;
574 
575 		cnt.v_pdpages++;
576 		next = m->pageq.tqe_next;
577 
578 #if defined(VM_DIAGNOSE)
579 		if ((m->flags & PG_INACTIVE) == 0) {
580 			printf("vm_pageout_scan: page not inactive?\n");
581 			break;
582 		}
583 #endif
584 
585 		/*
586 		 * dont mess with busy pages
587 		 */
588 		if (m->hold_count || m->busy || (m->flags & PG_BUSY)) {
589 			TAILQ_REMOVE(&vm_page_queue_inactive, m, pageq);
590 			TAILQ_INSERT_TAIL(&vm_page_queue_inactive, m, pageq);
591 			m = next;
592 			continue;
593 		}
594 		if (((m->flags & PG_REFERENCED) == 0) &&
595 		    pmap_is_referenced(VM_PAGE_TO_PHYS(m))) {
596 			m->flags |= PG_REFERENCED;
597 		}
598 		if (m->object->ref_count == 0) {
599 			m->flags &= ~PG_REFERENCED;
600 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
601 		}
602 		if ((m->flags & (PG_REFERENCED|PG_WANTED)) != 0) {
603 			m->flags &= ~PG_REFERENCED;
604 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
605 			vm_page_activate(m);
606 			if (m->act_count < ACT_MAX)
607 				m->act_count += ACT_ADVANCE;
608 			m = next;
609 			continue;
610 		}
611 
612 		vm_page_test_dirty(m);
613 		if (m->dirty == 0) {
614 			if (m->bmapped == 0) {
615 				if (m->valid == 0) {
616 					pmap_page_protect(VM_PAGE_TO_PHYS(m), VM_PROT_NONE);
617 					vm_page_free(m);
618 					cnt.v_dfree++;
619 				} else {
620 					vm_page_cache(m);
621 				}
622 				++pages_freed;
623 			} else {
624 				m = next;
625 				continue;
626 			}
627 		} else if (maxlaunder > 0) {
628 			int written;
629 			struct vnode *vp = NULL;
630 
631 			object = m->object;
632 			if (object->flags & OBJ_DEAD) {
633 				m = next;
634 				continue;
635 			}
636 
637 			if (object->type == OBJT_VNODE) {
638 				vp = object->handle;
639 				if (VOP_ISLOCKED(vp) || vget(vp, 1)) {
640 					if (object->flags & OBJ_MIGHTBEDIRTY)
641 						++vnodes_skipped;
642 					m = next;
643 					continue;
644 				}
645 			}
646 
647 			/*
648 			 * If a page is dirty, then it is either being washed
649 			 * (but not yet cleaned) or it is still in the
650 			 * laundry.  If it is still in the laundry, then we
651 			 * start the cleaning operation.
652 			 */
653 			written = vm_pageout_clean(m, 0);
654 
655 			if (vp)
656 				vput(vp);
657 
658 			if (!next) {
659 				break;
660 			}
661 			maxlaunder -= written;
662 			/*
663 			 * if the next page has been re-activated, start
664 			 * scanning again
665 			 */
666 			if ((next->flags & PG_INACTIVE) == 0) {
667 				goto rescan1;
668 			}
669 		}
670 		m = next;
671 	}
672 
673 	/*
674 	 * Compute the page shortage.  If we are still very low on memory be
675 	 * sure that we will move a minimal amount of pages from active to
676 	 * inactive.
677 	 */
678 
679 	page_shortage = cnt.v_inactive_target -
680 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
681 	if (page_shortage <= 0) {
682 		if (pages_freed == 0) {
683 			page_shortage = cnt.v_free_min - cnt.v_free_count;
684 		} else {
685 			page_shortage = 1;
686 		}
687 	}
688 	maxscan = MAXSCAN;
689 	pcount = cnt.v_active_count;
690 	m = vm_page_queue_active.tqh_first;
691 	while ((m != NULL) && (maxscan > 0) && (pcount-- > 0) && (page_shortage > 0)) {
692 
693 		cnt.v_pdpages++;
694 		next = m->pageq.tqe_next;
695 
696 		/*
697 		 * Don't deactivate pages that are busy.
698 		 */
699 		if ((m->busy != 0) ||
700 		    (m->flags & PG_BUSY) ||
701 		    (m->hold_count != 0)) {
702 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
703 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
704 			m = next;
705 			continue;
706 		}
707 		if (m->object->ref_count && ((m->flags & (PG_REFERENCED|PG_WANTED)) ||
708 			pmap_is_referenced(VM_PAGE_TO_PHYS(m)))) {
709 			int s;
710 
711 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
712 			m->flags &= ~PG_REFERENCED;
713 			if (m->act_count < ACT_MAX) {
714 				m->act_count += ACT_ADVANCE;
715 			}
716 			TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
717 			TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
718 		} else {
719 			m->flags &= ~PG_REFERENCED;
720 			pmap_clear_reference(VM_PAGE_TO_PHYS(m));
721 			m->act_count -= min(m->act_count, ACT_DECLINE);
722 
723 			/*
724 			 * if the page act_count is zero -- then we deactivate
725 			 */
726 			if (!m->act_count && (page_shortage > 0)) {
727 				if (m->object->ref_count == 0) {
728 					--page_shortage;
729 					vm_page_test_dirty(m);
730 					if ((m->bmapped == 0) && (m->dirty == 0) ) {
731 						m->act_count = 0;
732 						vm_page_cache(m);
733 					} else {
734 						vm_page_deactivate(m);
735 					}
736 				} else {
737 					vm_page_deactivate(m);
738 					--page_shortage;
739 				}
740 			} else if (m->act_count) {
741 				TAILQ_REMOVE(&vm_page_queue_active, m, pageq);
742 				TAILQ_INSERT_TAIL(&vm_page_queue_active, m, pageq);
743 			}
744 		}
745 		maxscan--;
746 		m = next;
747 	}
748 
749 	/*
750 	 * We try to maintain some *really* free pages, this allows interrupt
751 	 * code to be guaranteed space.
752 	 */
753 	while (cnt.v_free_count < cnt.v_free_reserved) {
754 		m = vm_page_queue_cache.tqh_first;
755 		if (!m)
756 			break;
757 		vm_page_free(m);
758 		cnt.v_dfree++;
759 	}
760 
761 	/*
762 	 * If we didn't get enough free pages, and we have skipped a vnode
763 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
764 	 * if we did not get enough free pages.
765 	 */
766 	if ((cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_target) {
767 		if (vnodes_skipped &&
768 		    (cnt.v_cache_count + cnt.v_free_count) < cnt.v_free_min) {
769 			if (!vfs_update_wakeup) {
770 				vfs_update_wakeup = 1;
771 				wakeup(&vfs_update_wakeup);
772 			}
773 		}
774 		/*
775 		 * now swap processes out if we are in low memory conditions
776 		 */
777 		if (!swap_pager_full && vm_swap_size &&
778 			vm_pageout_req_swapout == 0) {
779 			vm_pageout_req_swapout = 1;
780 			vm_req_vmdaemon();
781 		}
782 	}
783 
784 	if ((cnt.v_inactive_count + cnt.v_free_count + cnt.v_cache_count) <
785 	    (cnt.v_inactive_target + cnt.v_free_min)) {
786 		vm_req_vmdaemon();
787 	}
788 
789 	/*
790 	 * make sure that we have swap space -- if we are low on memory and
791 	 * swap -- then kill the biggest process.
792 	 */
793 	if ((vm_swap_size == 0 || swap_pager_full) &&
794 	    ((cnt.v_free_count + cnt.v_cache_count) < cnt.v_free_min)) {
795 		bigproc = NULL;
796 		bigsize = 0;
797 		for (p = (struct proc *) allproc; p != NULL; p = p->p_next) {
798 			/*
799 			 * if this is a system process, skip it
800 			 */
801 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
802 			    ((p->p_pid < 48) && (vm_swap_size != 0))) {
803 				continue;
804 			}
805 			/*
806 			 * if the process is in a non-running type state,
807 			 * don't touch it.
808 			 */
809 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
810 				continue;
811 			}
812 			/*
813 			 * get the process size
814 			 */
815 			size = p->p_vmspace->vm_pmap.pm_stats.resident_count;
816 			/*
817 			 * if the this process is bigger than the biggest one
818 			 * remember it.
819 			 */
820 			if (size > bigsize) {
821 				bigproc = p;
822 				bigsize = size;
823 			}
824 		}
825 		if (bigproc != NULL) {
826 			printf("Process %lu killed by vm_pageout -- out of swap\n", (u_long) bigproc->p_pid);
827 			psignal(bigproc, SIGKILL);
828 			bigproc->p_estcpu = 0;
829 			bigproc->p_nice = PRIO_MIN;
830 			resetpriority(bigproc);
831 			wakeup(&cnt.v_free_count);
832 		}
833 	}
834 	return force_wakeup;
835 }
836 
837 /*
838  *	vm_pageout is the high level pageout daemon.
839  */
840 static void
841 vm_pageout()
842 {
843 	(void) spl0();
844 
845 	/*
846 	 * Initialize some paging parameters.
847 	 */
848 
849 	cnt.v_interrupt_free_min = 2;
850 
851 	if (cnt.v_page_count > 1024)
852 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
853 	else
854 		cnt.v_free_min = 4;
855 	/*
856 	 * free_reserved needs to include enough for the largest swap pager
857 	 * structures plus enough for any pv_entry structs when paging.
858 	 */
859 	cnt.v_pageout_free_min = 6 + cnt.v_page_count / 1024 +
860 				cnt.v_interrupt_free_min;
861 	cnt.v_free_reserved = cnt.v_pageout_free_min + 6;
862 	cnt.v_free_target = 3 * cnt.v_free_min + cnt.v_free_reserved;
863 	cnt.v_free_min += cnt.v_free_reserved;
864 
865 	if (cnt.v_page_count > 1024) {
866 		cnt.v_cache_max = (cnt.v_free_count - 1024) / 2;
867 		cnt.v_cache_min = (cnt.v_free_count - 1024) / 8;
868 		cnt.v_inactive_target = 2*cnt.v_cache_min + 192;
869 	} else {
870 		cnt.v_cache_min = 0;
871 		cnt.v_cache_max = 0;
872 		cnt.v_inactive_target = cnt.v_free_count / 4;
873 	}
874 
875 	/* XXX does not really belong here */
876 	if (vm_page_max_wired == 0)
877 		vm_page_max_wired = cnt.v_free_count / 3;
878 
879 
880 	swap_pager_swap_init();
881 	/*
882 	 * The pageout daemon is never done, so loop forever.
883 	 */
884 	while (TRUE) {
885 		int s = splhigh();
886 
887 		if (!vm_pages_needed ||
888 			((cnt.v_free_count >= cnt.v_free_reserved) &&
889 			 (cnt.v_free_count + cnt.v_cache_count >= cnt.v_free_min))) {
890 			vm_pages_needed = 0;
891 			tsleep(&vm_pages_needed, PVM, "psleep", 0);
892 		}
893 		vm_pages_needed = 0;
894 		splx(s);
895 		cnt.v_pdwakeups++;
896 		vm_pager_sync();
897 		vm_pageout_scan();
898 		vm_pager_sync();
899 		wakeup(&cnt.v_free_count);
900 		wakeup(kmem_map);
901 	}
902 }
903 
904 static void
905 vm_daemon()
906 {
907 	vm_object_t object;
908 	struct proc *p;
909 
910 	while (TRUE) {
911 		tsleep(&vm_daemon_needed, PUSER, "psleep", 0);
912 		if (vm_pageout_req_swapout) {
913 			swapout_procs();
914 			vm_pageout_req_swapout = 0;
915 		}
916 		/*
917 		 * scan the processes for exceeding their rlimits or if
918 		 * process is swapped out -- deactivate pages
919 		 */
920 
921 		for (p = (struct proc *) allproc; p != NULL; p = p->p_next) {
922 			int overage;
923 			quad_t limit;
924 			vm_offset_t size;
925 
926 			/*
927 			 * if this is a system process or if we have already
928 			 * looked at this process, skip it.
929 			 */
930 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
931 				continue;
932 			}
933 			/*
934 			 * if the process is in a non-running type state,
935 			 * don't touch it.
936 			 */
937 			if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
938 				continue;
939 			}
940 			/*
941 			 * get a limit
942 			 */
943 			limit = qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
944 			    p->p_rlimit[RLIMIT_RSS].rlim_max);
945 
946 			/*
947 			 * let processes that are swapped out really be
948 			 * swapped out set the limit to nothing (will force a
949 			 * swap-out.)
950 			 */
951 			if ((p->p_flag & P_INMEM) == 0)
952 				limit = 0;	/* XXX */
953 
954 			size = p->p_vmspace->vm_pmap.pm_stats.resident_count * PAGE_SIZE;
955 			if (limit >= 0 && size >= limit) {
956 				overage = (size - limit) >> PAGE_SHIFT;
957 				vm_pageout_map_deactivate_pages(&p->p_vmspace->vm_map,
958 				    (vm_map_entry_t) 0, &overage, vm_pageout_object_deactivate_pages);
959 			}
960 		}
961 
962 		/*
963 		 * we remove cached objects that have no RSS...
964 		 */
965 restart:
966 		object = vm_object_cached_list.tqh_first;
967 		while (object) {
968 			/*
969 			 * if there are no resident pages -- get rid of the object
970 			 */
971 			if (object->resident_page_count == 0) {
972 				vm_object_reference(object);
973 				pager_cache(object, FALSE);
974 				goto restart;
975 			}
976 			object = object->cached_list.tqe_next;
977 		}
978 	}
979 }
980