xref: /freebsd/sys/vm/vm_pageout.c (revision 9162f64b58d01ec01481d60b6cdc06ffd8e8c7fc)
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  * Copyright (c) 2005 Yahoo! Technologies Norway AS
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
77 
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/eventhandler.h>
83 #include <sys/lock.h>
84 #include <sys/mutex.h>
85 #include <sys/proc.h>
86 #include <sys/kthread.h>
87 #include <sys/ktr.h>
88 #include <sys/mount.h>
89 #include <sys/resourcevar.h>
90 #include <sys/sched.h>
91 #include <sys/signalvar.h>
92 #include <sys/vnode.h>
93 #include <sys/vmmeter.h>
94 #include <sys/sx.h>
95 #include <sys/sysctl.h>
96 
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_map.h>
102 #include <vm/vm_pageout.h>
103 #include <vm/vm_pager.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_extern.h>
106 #include <vm/uma.h>
107 
108 #include <machine/mutex.h>
109 
110 /*
111  * System initialization
112  */
113 
114 /* the kernel process "vm_pageout"*/
115 static void vm_pageout(void);
116 static int vm_pageout_clean(vm_page_t);
117 static void vm_pageout_scan(int pass);
118 
119 struct proc *pageproc;
120 
121 static struct kproc_desc page_kp = {
122 	"pagedaemon",
123 	vm_pageout,
124 	&pageproc
125 };
126 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start,
127     &page_kp);
128 
129 #if !defined(NO_SWAPPING)
130 /* the kernel process "vm_daemon"*/
131 static void vm_daemon(void);
132 static struct	proc *vmproc;
133 
134 static struct kproc_desc vm_kp = {
135 	"vmdaemon",
136 	vm_daemon,
137 	&vmproc
138 };
139 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
140 #endif
141 
142 
143 int vm_pages_needed;		/* Event on which pageout daemon sleeps */
144 int vm_pageout_deficit;		/* Estimated number of pages deficit */
145 int vm_pageout_pages_needed;	/* flag saying that the pageout daemon needs pages */
146 
147 #if !defined(NO_SWAPPING)
148 static int vm_pageout_req_swapout;	/* XXX */
149 static int vm_daemon_needed;
150 static struct mtx vm_daemon_mtx;
151 /* Allow for use by vm_pageout before vm_daemon is initialized. */
152 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
153 #endif
154 static int vm_max_launder = 32;
155 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
156 static int vm_pageout_full_stats_interval = 0;
157 static int vm_pageout_algorithm=0;
158 static int defer_swap_pageouts=0;
159 static int disable_swap_pageouts=0;
160 
161 #if defined(NO_SWAPPING)
162 static int vm_swap_enabled=0;
163 static int vm_swap_idle_enabled=0;
164 #else
165 static int vm_swap_enabled=1;
166 static int vm_swap_idle_enabled=0;
167 #endif
168 
169 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
170 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
171 
172 SYSCTL_INT(_vm, OID_AUTO, max_launder,
173 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
174 
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
176 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
177 
178 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
179 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
180 
181 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
182 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
183 
184 #if defined(NO_SWAPPING)
185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
186 	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
188 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
189 #else
190 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
191 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
192 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
193 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
194 #endif
195 
196 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
197 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
198 
199 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
200 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
201 
202 static int pageout_lock_miss;
203 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
204 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
205 
206 #define VM_PAGEOUT_PAGE_COUNT 16
207 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
208 
209 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
210 SYSCTL_INT(_vm, OID_AUTO, max_wired,
211 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
212 
213 #if !defined(NO_SWAPPING)
214 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
215 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
216 static void vm_req_vmdaemon(int req);
217 #endif
218 static void vm_pageout_page_stats(void);
219 
220 /*
221  * vm_pageout_fallback_object_lock:
222  *
223  * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is
224  * known to have failed and page queue must be either PQ_ACTIVE or
225  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queues
226  * while locking the vm object.  Use marker page to detect page queue
227  * changes and maintain notion of next page on page queue.  Return
228  * TRUE if no changes were detected, FALSE otherwise.  vm object is
229  * locked on return.
230  *
231  * This function depends on both the lock portion of struct vm_object
232  * and normal struct vm_page being type stable.
233  */
234 boolean_t
235 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
236 {
237 	struct vm_page marker;
238 	boolean_t unchanged;
239 	u_short queue;
240 	vm_object_t object;
241 
242 	/*
243 	 * Initialize our marker
244 	 */
245 	bzero(&marker, sizeof(marker));
246 	marker.flags = PG_FICTITIOUS | PG_MARKER;
247 	marker.oflags = VPO_BUSY;
248 	marker.queue = m->queue;
249 	marker.wire_count = 1;
250 
251 	queue = m->queue;
252 	object = m->object;
253 
254 	TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl,
255 			   m, &marker, pageq);
256 	vm_page_unlock_queues();
257 	VM_OBJECT_LOCK(object);
258 	vm_page_lock_queues();
259 
260 	/* Page queue might have changed. */
261 	*next = TAILQ_NEXT(&marker, pageq);
262 	unchanged = (m->queue == queue &&
263 		     m->object == object &&
264 		     &marker == TAILQ_NEXT(m, pageq));
265 	TAILQ_REMOVE(&vm_page_queues[queue].pl,
266 		     &marker, pageq);
267 	return (unchanged);
268 }
269 
270 /*
271  * vm_pageout_clean:
272  *
273  * Clean the page and remove it from the laundry.
274  *
275  * We set the busy bit to cause potential page faults on this page to
276  * block.  Note the careful timing, however, the busy bit isn't set till
277  * late and we cannot do anything that will mess with the page.
278  */
279 static int
280 vm_pageout_clean(m)
281 	vm_page_t m;
282 {
283 	vm_object_t object;
284 	vm_page_t mc[2*vm_pageout_page_count];
285 	int pageout_count;
286 	int ib, is, page_base;
287 	vm_pindex_t pindex = m->pindex;
288 
289 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
290 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
291 
292 	/*
293 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
294 	 * with the new swapper, but we could have serious problems paging
295 	 * out other object types if there is insufficient memory.
296 	 *
297 	 * Unfortunately, checking free memory here is far too late, so the
298 	 * check has been moved up a procedural level.
299 	 */
300 
301 	/*
302 	 * Can't clean the page if it's busy or held.
303 	 */
304 	if ((m->hold_count != 0) ||
305 	    ((m->busy != 0) || (m->oflags & VPO_BUSY))) {
306 		return 0;
307 	}
308 
309 	mc[vm_pageout_page_count] = m;
310 	pageout_count = 1;
311 	page_base = vm_pageout_page_count;
312 	ib = 1;
313 	is = 1;
314 
315 	/*
316 	 * Scan object for clusterable pages.
317 	 *
318 	 * We can cluster ONLY if: ->> the page is NOT
319 	 * clean, wired, busy, held, or mapped into a
320 	 * buffer, and one of the following:
321 	 * 1) The page is inactive, or a seldom used
322 	 *    active page.
323 	 * -or-
324 	 * 2) we force the issue.
325 	 *
326 	 * During heavy mmap/modification loads the pageout
327 	 * daemon can really fragment the underlying file
328 	 * due to flushing pages out of order and not trying
329 	 * align the clusters (which leave sporatic out-of-order
330 	 * holes).  To solve this problem we do the reverse scan
331 	 * first and attempt to align our cluster, then do a
332 	 * forward scan if room remains.
333 	 */
334 	object = m->object;
335 more:
336 	while (ib && pageout_count < vm_pageout_page_count) {
337 		vm_page_t p;
338 
339 		if (ib > pindex) {
340 			ib = 0;
341 			break;
342 		}
343 
344 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
345 			ib = 0;
346 			break;
347 		}
348 		if ((p->oflags & VPO_BUSY) || p->busy) {
349 			ib = 0;
350 			break;
351 		}
352 		vm_page_test_dirty(p);
353 		if ((p->dirty & p->valid) == 0 ||
354 		    p->queue != PQ_INACTIVE ||
355 		    p->wire_count != 0 ||	/* may be held by buf cache */
356 		    p->hold_count != 0) {	/* may be undergoing I/O */
357 			ib = 0;
358 			break;
359 		}
360 		mc[--page_base] = p;
361 		++pageout_count;
362 		++ib;
363 		/*
364 		 * alignment boundry, stop here and switch directions.  Do
365 		 * not clear ib.
366 		 */
367 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
368 			break;
369 	}
370 
371 	while (pageout_count < vm_pageout_page_count &&
372 	    pindex + is < object->size) {
373 		vm_page_t p;
374 
375 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
376 			break;
377 		if ((p->oflags & VPO_BUSY) || p->busy) {
378 			break;
379 		}
380 		vm_page_test_dirty(p);
381 		if ((p->dirty & p->valid) == 0 ||
382 		    p->queue != PQ_INACTIVE ||
383 		    p->wire_count != 0 ||	/* may be held by buf cache */
384 		    p->hold_count != 0) {	/* may be undergoing I/O */
385 			break;
386 		}
387 		mc[page_base + pageout_count] = p;
388 		++pageout_count;
389 		++is;
390 	}
391 
392 	/*
393 	 * If we exhausted our forward scan, continue with the reverse scan
394 	 * when possible, even past a page boundry.  This catches boundry
395 	 * conditions.
396 	 */
397 	if (ib && pageout_count < vm_pageout_page_count)
398 		goto more;
399 
400 	/*
401 	 * we allow reads during pageouts...
402 	 */
403 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0));
404 }
405 
406 /*
407  * vm_pageout_flush() - launder the given pages
408  *
409  *	The given pages are laundered.  Note that we setup for the start of
410  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
411  *	reference count all in here rather then in the parent.  If we want
412  *	the parent to do more sophisticated things we may have to change
413  *	the ordering.
414  */
415 int
416 vm_pageout_flush(vm_page_t *mc, int count, int flags)
417 {
418 	vm_object_t object = mc[0]->object;
419 	int pageout_status[count];
420 	int numpagedout = 0;
421 	int i;
422 
423 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
424 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
425 	/*
426 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
427 	 * mark the pages read-only.
428 	 *
429 	 * We do not have to fixup the clean/dirty bits here... we can
430 	 * allow the pager to do it after the I/O completes.
431 	 *
432 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
433 	 * edge case with file fragments.
434 	 */
435 	for (i = 0; i < count; i++) {
436 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
437 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
438 			mc[i], i, count));
439 		vm_page_io_start(mc[i]);
440 		pmap_remove_write(mc[i]);
441 	}
442 	vm_page_unlock_queues();
443 	vm_object_pip_add(object, count);
444 
445 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
446 
447 	vm_page_lock_queues();
448 	for (i = 0; i < count; i++) {
449 		vm_page_t mt = mc[i];
450 
451 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
452 		    (mt->flags & PG_WRITEABLE) == 0,
453 		    ("vm_pageout_flush: page %p is not write protected", mt));
454 		switch (pageout_status[i]) {
455 		case VM_PAGER_OK:
456 		case VM_PAGER_PEND:
457 			numpagedout++;
458 			break;
459 		case VM_PAGER_BAD:
460 			/*
461 			 * Page outside of range of object. Right now we
462 			 * essentially lose the changes by pretending it
463 			 * worked.
464 			 */
465 			pmap_clear_modify(mt);
466 			vm_page_undirty(mt);
467 			break;
468 		case VM_PAGER_ERROR:
469 		case VM_PAGER_FAIL:
470 			/*
471 			 * If page couldn't be paged out, then reactivate the
472 			 * page so it doesn't clog the inactive list.  (We
473 			 * will try paging out it again later).
474 			 */
475 			vm_page_activate(mt);
476 			break;
477 		case VM_PAGER_AGAIN:
478 			break;
479 		}
480 
481 		/*
482 		 * If the operation is still going, leave the page busy to
483 		 * block all other accesses. Also, leave the paging in
484 		 * progress indicator set so that we don't attempt an object
485 		 * collapse.
486 		 */
487 		if (pageout_status[i] != VM_PAGER_PEND) {
488 			vm_object_pip_wakeup(object);
489 			vm_page_io_finish(mt);
490 			if (vm_page_count_severe())
491 				vm_page_try_to_cache(mt);
492 		}
493 	}
494 	return numpagedout;
495 }
496 
497 #if !defined(NO_SWAPPING)
498 /*
499  *	vm_pageout_object_deactivate_pages
500  *
501  *	deactivate enough pages to satisfy the inactive target
502  *	requirements or if vm_page_proc_limit is set, then
503  *	deactivate all of the pages in the object and its
504  *	backing_objects.
505  *
506  *	The object and map must be locked.
507  */
508 static void
509 vm_pageout_object_deactivate_pages(pmap, first_object, desired)
510 	pmap_t pmap;
511 	vm_object_t first_object;
512 	long desired;
513 {
514 	vm_object_t backing_object, object;
515 	vm_page_t p, next;
516 	int actcount, rcount, remove_mode;
517 
518 	VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED);
519 	if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS)
520 		return;
521 	for (object = first_object;; object = backing_object) {
522 		if (pmap_resident_count(pmap) <= desired)
523 			goto unlock_return;
524 		if (object->paging_in_progress)
525 			goto unlock_return;
526 
527 		remove_mode = 0;
528 		if (object->shadow_count > 1)
529 			remove_mode = 1;
530 		/*
531 		 * scan the objects entire memory queue
532 		 */
533 		rcount = object->resident_page_count;
534 		p = TAILQ_FIRST(&object->memq);
535 		vm_page_lock_queues();
536 		while (p && (rcount-- > 0)) {
537 			if (pmap_resident_count(pmap) <= desired) {
538 				vm_page_unlock_queues();
539 				goto unlock_return;
540 			}
541 			next = TAILQ_NEXT(p, listq);
542 			cnt.v_pdpages++;
543 			if (p->wire_count != 0 ||
544 			    p->hold_count != 0 ||
545 			    p->busy != 0 ||
546 			    (p->oflags & VPO_BUSY) ||
547 			    (p->flags & PG_UNMANAGED) ||
548 			    !pmap_page_exists_quick(pmap, p)) {
549 				p = next;
550 				continue;
551 			}
552 			actcount = pmap_ts_referenced(p);
553 			if (actcount) {
554 				vm_page_flag_set(p, PG_REFERENCED);
555 			} else if (p->flags & PG_REFERENCED) {
556 				actcount = 1;
557 			}
558 			if ((p->queue != PQ_ACTIVE) &&
559 				(p->flags & PG_REFERENCED)) {
560 				vm_page_activate(p);
561 				p->act_count += actcount;
562 				vm_page_flag_clear(p, PG_REFERENCED);
563 			} else if (p->queue == PQ_ACTIVE) {
564 				if ((p->flags & PG_REFERENCED) == 0) {
565 					p->act_count -= min(p->act_count, ACT_DECLINE);
566 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
567 						pmap_remove_all(p);
568 						vm_page_deactivate(p);
569 					} else {
570 						vm_page_requeue(p);
571 					}
572 				} else {
573 					vm_page_activate(p);
574 					vm_page_flag_clear(p, PG_REFERENCED);
575 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
576 						p->act_count += ACT_ADVANCE;
577 					vm_page_requeue(p);
578 				}
579 			} else if (p->queue == PQ_INACTIVE) {
580 				pmap_remove_all(p);
581 			}
582 			p = next;
583 		}
584 		vm_page_unlock_queues();
585 		if ((backing_object = object->backing_object) == NULL)
586 			goto unlock_return;
587 		VM_OBJECT_LOCK(backing_object);
588 		if (object != first_object)
589 			VM_OBJECT_UNLOCK(object);
590 	}
591 unlock_return:
592 	if (object != first_object)
593 		VM_OBJECT_UNLOCK(object);
594 }
595 
596 /*
597  * deactivate some number of pages in a map, try to do it fairly, but
598  * that is really hard to do.
599  */
600 static void
601 vm_pageout_map_deactivate_pages(map, desired)
602 	vm_map_t map;
603 	long desired;
604 {
605 	vm_map_entry_t tmpe;
606 	vm_object_t obj, bigobj;
607 	int nothingwired;
608 
609 	if (!vm_map_trylock(map))
610 		return;
611 
612 	bigobj = NULL;
613 	nothingwired = TRUE;
614 
615 	/*
616 	 * first, search out the biggest object, and try to free pages from
617 	 * that.
618 	 */
619 	tmpe = map->header.next;
620 	while (tmpe != &map->header) {
621 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
622 			obj = tmpe->object.vm_object;
623 			if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) {
624 				if (obj->shadow_count <= 1 &&
625 				    (bigobj == NULL ||
626 				     bigobj->resident_page_count < obj->resident_page_count)) {
627 					if (bigobj != NULL)
628 						VM_OBJECT_UNLOCK(bigobj);
629 					bigobj = obj;
630 				} else
631 					VM_OBJECT_UNLOCK(obj);
632 			}
633 		}
634 		if (tmpe->wired_count > 0)
635 			nothingwired = FALSE;
636 		tmpe = tmpe->next;
637 	}
638 
639 	if (bigobj != NULL) {
640 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
641 		VM_OBJECT_UNLOCK(bigobj);
642 	}
643 	/*
644 	 * Next, hunt around for other pages to deactivate.  We actually
645 	 * do this search sort of wrong -- .text first is not the best idea.
646 	 */
647 	tmpe = map->header.next;
648 	while (tmpe != &map->header) {
649 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
650 			break;
651 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
652 			obj = tmpe->object.vm_object;
653 			if (obj != NULL) {
654 				VM_OBJECT_LOCK(obj);
655 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
656 				VM_OBJECT_UNLOCK(obj);
657 			}
658 		}
659 		tmpe = tmpe->next;
660 	}
661 
662 	/*
663 	 * Remove all mappings if a process is swapped out, this will free page
664 	 * table pages.
665 	 */
666 	if (desired == 0 && nothingwired) {
667 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
668 		    vm_map_max(map));
669 	}
670 	vm_map_unlock(map);
671 }
672 #endif		/* !defined(NO_SWAPPING) */
673 
674 /*
675  *	vm_pageout_scan does the dirty work for the pageout daemon.
676  */
677 static void
678 vm_pageout_scan(int pass)
679 {
680 	vm_page_t m, next;
681 	struct vm_page marker;
682 	int page_shortage, maxscan, pcount;
683 	int addl_page_shortage, addl_page_shortage_init;
684 	vm_object_t object;
685 	int actcount;
686 	int vnodes_skipped = 0;
687 	int maxlaunder;
688 
689 	/*
690 	 * Decrease registered cache sizes.
691 	 */
692 	EVENTHANDLER_INVOKE(vm_lowmem, 0);
693 	/*
694 	 * We do this explicitly after the caches have been drained above.
695 	 */
696 	uma_reclaim();
697 
698 	addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit);
699 
700 	/*
701 	 * Calculate the number of pages we want to either free or move
702 	 * to the cache.
703 	 */
704 	page_shortage = vm_paging_target() + addl_page_shortage_init;
705 
706 	/*
707 	 * Initialize our marker
708 	 */
709 	bzero(&marker, sizeof(marker));
710 	marker.flags = PG_FICTITIOUS | PG_MARKER;
711 	marker.oflags = VPO_BUSY;
712 	marker.queue = PQ_INACTIVE;
713 	marker.wire_count = 1;
714 
715 	/*
716 	 * Start scanning the inactive queue for pages we can move to the
717 	 * cache or free.  The scan will stop when the target is reached or
718 	 * we have scanned the entire inactive queue.  Note that m->act_count
719 	 * is not used to form decisions for the inactive queue, only for the
720 	 * active queue.
721 	 *
722 	 * maxlaunder limits the number of dirty pages we flush per scan.
723 	 * For most systems a smaller value (16 or 32) is more robust under
724 	 * extreme memory and disk pressure because any unnecessary writes
725 	 * to disk can result in extreme performance degredation.  However,
726 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
727 	 * used) will die horribly with limited laundering.  If the pageout
728 	 * daemon cannot clean enough pages in the first pass, we let it go
729 	 * all out in succeeding passes.
730 	 */
731 	if ((maxlaunder = vm_max_launder) <= 1)
732 		maxlaunder = 1;
733 	if (pass)
734 		maxlaunder = 10000;
735 	vm_page_lock_queues();
736 rescan0:
737 	addl_page_shortage = addl_page_shortage_init;
738 	maxscan = cnt.v_inactive_count;
739 
740 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
741 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
742 	     m = next) {
743 
744 		cnt.v_pdpages++;
745 
746 		if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) {
747 			goto rescan0;
748 		}
749 
750 		next = TAILQ_NEXT(m, pageq);
751 		object = m->object;
752 
753 		/*
754 		 * skip marker pages
755 		 */
756 		if (m->flags & PG_MARKER)
757 			continue;
758 
759 		/*
760 		 * A held page may be undergoing I/O, so skip it.
761 		 */
762 		if (m->hold_count) {
763 			vm_page_requeue(m);
764 			addl_page_shortage++;
765 			continue;
766 		}
767 		/*
768 		 * Don't mess with busy pages, keep in the front of the
769 		 * queue, most likely are being paged out.
770 		 */
771 		if (!VM_OBJECT_TRYLOCK(object) &&
772 		    (!vm_pageout_fallback_object_lock(m, &next) ||
773 		     m->hold_count != 0)) {
774 			VM_OBJECT_UNLOCK(object);
775 			addl_page_shortage++;
776 			continue;
777 		}
778 		if (m->busy || (m->oflags & VPO_BUSY)) {
779 			VM_OBJECT_UNLOCK(object);
780 			addl_page_shortage++;
781 			continue;
782 		}
783 
784 		/*
785 		 * If the object is not being used, we ignore previous
786 		 * references.
787 		 */
788 		if (object->ref_count == 0) {
789 			vm_page_flag_clear(m, PG_REFERENCED);
790 			pmap_clear_reference(m);
791 
792 		/*
793 		 * Otherwise, if the page has been referenced while in the
794 		 * inactive queue, we bump the "activation count" upwards,
795 		 * making it less likely that the page will be added back to
796 		 * the inactive queue prematurely again.  Here we check the
797 		 * page tables (or emulated bits, if any), given the upper
798 		 * level VM system not knowing anything about existing
799 		 * references.
800 		 */
801 		} else if (((m->flags & PG_REFERENCED) == 0) &&
802 			(actcount = pmap_ts_referenced(m))) {
803 			vm_page_activate(m);
804 			VM_OBJECT_UNLOCK(object);
805 			m->act_count += (actcount + ACT_ADVANCE);
806 			continue;
807 		}
808 
809 		/*
810 		 * If the upper level VM system knows about any page
811 		 * references, we activate the page.  We also set the
812 		 * "activation count" higher than normal so that we will less
813 		 * likely place pages back onto the inactive queue again.
814 		 */
815 		if ((m->flags & PG_REFERENCED) != 0) {
816 			vm_page_flag_clear(m, PG_REFERENCED);
817 			actcount = pmap_ts_referenced(m);
818 			vm_page_activate(m);
819 			VM_OBJECT_UNLOCK(object);
820 			m->act_count += (actcount + ACT_ADVANCE + 1);
821 			continue;
822 		}
823 
824 		/*
825 		 * If the upper level VM system doesn't know anything about
826 		 * the page being dirty, we have to check for it again.  As
827 		 * far as the VM code knows, any partially dirty pages are
828 		 * fully dirty.
829 		 */
830 		if (m->dirty == 0 && !pmap_is_modified(m)) {
831 			/*
832 			 * Avoid a race condition: Unless write access is
833 			 * removed from the page, another processor could
834 			 * modify it before all access is removed by the call
835 			 * to vm_page_cache() below.  If vm_page_cache() finds
836 			 * that the page has been modified when it removes all
837 			 * access, it panics because it cannot cache dirty
838 			 * pages.  In principle, we could eliminate just write
839 			 * access here rather than all access.  In the expected
840 			 * case, when there are no last instant modifications
841 			 * to the page, removing all access will be cheaper
842 			 * overall.
843 			 */
844 			if ((m->flags & PG_WRITEABLE) != 0)
845 				pmap_remove_all(m);
846 		} else {
847 			vm_page_dirty(m);
848 		}
849 
850 		if (m->valid == 0) {
851 			/*
852 			 * Invalid pages can be easily freed
853 			 */
854 			vm_page_free(m);
855 			cnt.v_dfree++;
856 			--page_shortage;
857 		} else if (m->dirty == 0) {
858 			/*
859 			 * Clean pages can be placed onto the cache queue.
860 			 * This effectively frees them.
861 			 */
862 			vm_page_cache(m);
863 			--page_shortage;
864 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
865 			/*
866 			 * Dirty pages need to be paged out, but flushing
867 			 * a page is extremely expensive verses freeing
868 			 * a clean page.  Rather then artificially limiting
869 			 * the number of pages we can flush, we instead give
870 			 * dirty pages extra priority on the inactive queue
871 			 * by forcing them to be cycled through the queue
872 			 * twice before being flushed, after which the
873 			 * (now clean) page will cycle through once more
874 			 * before being freed.  This significantly extends
875 			 * the thrash point for a heavily loaded machine.
876 			 */
877 			vm_page_flag_set(m, PG_WINATCFLS);
878 			vm_page_requeue(m);
879 		} else if (maxlaunder > 0) {
880 			/*
881 			 * We always want to try to flush some dirty pages if
882 			 * we encounter them, to keep the system stable.
883 			 * Normally this number is small, but under extreme
884 			 * pressure where there are insufficient clean pages
885 			 * on the inactive queue, we may have to go all out.
886 			 */
887 			int swap_pageouts_ok, vfslocked = 0;
888 			struct vnode *vp = NULL;
889 			struct mount *mp = NULL;
890 
891 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
892 				swap_pageouts_ok = 1;
893 			} else {
894 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
895 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
896 				vm_page_count_min());
897 
898 			}
899 
900 			/*
901 			 * We don't bother paging objects that are "dead".
902 			 * Those objects are in a "rundown" state.
903 			 */
904 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
905 				VM_OBJECT_UNLOCK(object);
906 				vm_page_requeue(m);
907 				continue;
908 			}
909 
910 			/*
911 			 * Following operations may unlock
912 			 * vm_page_queue_mtx, invalidating the 'next'
913 			 * pointer.  To prevent an inordinate number
914 			 * of restarts we use our marker to remember
915 			 * our place.
916 			 *
917 			 */
918 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
919 					   m, &marker, pageq);
920 			/*
921 			 * The object is already known NOT to be dead.   It
922 			 * is possible for the vget() to block the whole
923 			 * pageout daemon, but the new low-memory handling
924 			 * code should prevent it.
925 			 *
926 			 * The previous code skipped locked vnodes and, worse,
927 			 * reordered pages in the queue.  This results in
928 			 * completely non-deterministic operation and, on a
929 			 * busy system, can lead to extremely non-optimal
930 			 * pageouts.  For example, it can cause clean pages
931 			 * to be freed and dirty pages to be moved to the end
932 			 * of the queue.  Since dirty pages are also moved to
933 			 * the end of the queue once-cleaned, this gives
934 			 * way too large a weighting to defering the freeing
935 			 * of dirty pages.
936 			 *
937 			 * We can't wait forever for the vnode lock, we might
938 			 * deadlock due to a vn_read() getting stuck in
939 			 * vm_wait while holding this vnode.  We skip the
940 			 * vnode if we can't get it in a reasonable amount
941 			 * of time.
942 			 */
943 			if (object->type == OBJT_VNODE) {
944 				vp = object->handle;
945 				if (vp->v_type == VREG &&
946 				    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
947 					mp = NULL;
948 					++pageout_lock_miss;
949 					if (object->flags & OBJ_MIGHTBEDIRTY)
950 						vnodes_skipped++;
951 					goto unlock_and_continue;
952 				}
953 				vm_page_unlock_queues();
954 				vm_object_reference_locked(object);
955 				VM_OBJECT_UNLOCK(object);
956 				vfslocked = VFS_LOCK_GIANT(vp->v_mount);
957 				if (vget(vp, LK_EXCLUSIVE | LK_TIMELOCK,
958 				    curthread)) {
959 					VM_OBJECT_LOCK(object);
960 					vm_page_lock_queues();
961 					++pageout_lock_miss;
962 					if (object->flags & OBJ_MIGHTBEDIRTY)
963 						vnodes_skipped++;
964 					vp = NULL;
965 					goto unlock_and_continue;
966 				}
967 				VM_OBJECT_LOCK(object);
968 				vm_page_lock_queues();
969 				/*
970 				 * The page might have been moved to another
971 				 * queue during potential blocking in vget()
972 				 * above.  The page might have been freed and
973 				 * reused for another vnode.
974 				 */
975 				if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE ||
976 				    m->object != object ||
977 				    TAILQ_NEXT(m, pageq) != &marker) {
978 					if (object->flags & OBJ_MIGHTBEDIRTY)
979 						vnodes_skipped++;
980 					goto unlock_and_continue;
981 				}
982 
983 				/*
984 				 * The page may have been busied during the
985 				 * blocking in vget().  We don't move the
986 				 * page back onto the end of the queue so that
987 				 * statistics are more correct if we don't.
988 				 */
989 				if (m->busy || (m->oflags & VPO_BUSY)) {
990 					goto unlock_and_continue;
991 				}
992 
993 				/*
994 				 * If the page has become held it might
995 				 * be undergoing I/O, so skip it
996 				 */
997 				if (m->hold_count) {
998 					vm_page_requeue(m);
999 					if (object->flags & OBJ_MIGHTBEDIRTY)
1000 						vnodes_skipped++;
1001 					goto unlock_and_continue;
1002 				}
1003 			}
1004 
1005 			/*
1006 			 * If a page is dirty, then it is either being washed
1007 			 * (but not yet cleaned) or it is still in the
1008 			 * laundry.  If it is still in the laundry, then we
1009 			 * start the cleaning operation.
1010 			 *
1011 			 * decrement page_shortage on success to account for
1012 			 * the (future) cleaned page.  Otherwise we could wind
1013 			 * up laundering or cleaning too many pages.
1014 			 */
1015 			if (vm_pageout_clean(m) != 0) {
1016 				--page_shortage;
1017 				--maxlaunder;
1018 			}
1019 unlock_and_continue:
1020 			VM_OBJECT_UNLOCK(object);
1021 			if (mp != NULL) {
1022 				vm_page_unlock_queues();
1023 				if (vp != NULL)
1024 					vput(vp);
1025 				VFS_UNLOCK_GIANT(vfslocked);
1026 				vm_object_deallocate(object);
1027 				vn_finished_write(mp);
1028 				vm_page_lock_queues();
1029 			}
1030 			next = TAILQ_NEXT(&marker, pageq);
1031 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
1032 				     &marker, pageq);
1033 			continue;
1034 		}
1035 		VM_OBJECT_UNLOCK(object);
1036 	}
1037 
1038 	/*
1039 	 * Compute the number of pages we want to try to move from the
1040 	 * active queue to the inactive queue.
1041 	 */
1042 	page_shortage = vm_paging_target() +
1043 		cnt.v_inactive_target - cnt.v_inactive_count;
1044 	page_shortage += addl_page_shortage;
1045 
1046 	/*
1047 	 * Scan the active queue for things we can deactivate. We nominally
1048 	 * track the per-page activity counter and use it to locate
1049 	 * deactivation candidates.
1050 	 */
1051 	pcount = cnt.v_active_count;
1052 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1053 
1054 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
1055 
1056 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1057 		    ("vm_pageout_scan: page %p isn't active", m));
1058 
1059 		next = TAILQ_NEXT(m, pageq);
1060 		object = m->object;
1061 		if ((m->flags & PG_MARKER) != 0) {
1062 			m = next;
1063 			continue;
1064 		}
1065 		if (!VM_OBJECT_TRYLOCK(object) &&
1066 		    !vm_pageout_fallback_object_lock(m, &next)) {
1067 			VM_OBJECT_UNLOCK(object);
1068 			m = next;
1069 			continue;
1070 		}
1071 
1072 		/*
1073 		 * Don't deactivate pages that are busy.
1074 		 */
1075 		if ((m->busy != 0) ||
1076 		    (m->oflags & VPO_BUSY) ||
1077 		    (m->hold_count != 0)) {
1078 			VM_OBJECT_UNLOCK(object);
1079 			vm_page_requeue(m);
1080 			m = next;
1081 			continue;
1082 		}
1083 
1084 		/*
1085 		 * The count for pagedaemon pages is done after checking the
1086 		 * page for eligibility...
1087 		 */
1088 		cnt.v_pdpages++;
1089 
1090 		/*
1091 		 * Check to see "how much" the page has been used.
1092 		 */
1093 		actcount = 0;
1094 		if (object->ref_count != 0) {
1095 			if (m->flags & PG_REFERENCED) {
1096 				actcount += 1;
1097 			}
1098 			actcount += pmap_ts_referenced(m);
1099 			if (actcount) {
1100 				m->act_count += ACT_ADVANCE + actcount;
1101 				if (m->act_count > ACT_MAX)
1102 					m->act_count = ACT_MAX;
1103 			}
1104 		}
1105 
1106 		/*
1107 		 * Since we have "tested" this bit, we need to clear it now.
1108 		 */
1109 		vm_page_flag_clear(m, PG_REFERENCED);
1110 
1111 		/*
1112 		 * Only if an object is currently being used, do we use the
1113 		 * page activation count stats.
1114 		 */
1115 		if (actcount && (object->ref_count != 0)) {
1116 			vm_page_requeue(m);
1117 		} else {
1118 			m->act_count -= min(m->act_count, ACT_DECLINE);
1119 			if (vm_pageout_algorithm ||
1120 			    object->ref_count == 0 ||
1121 			    m->act_count == 0) {
1122 				page_shortage--;
1123 				if (object->ref_count == 0) {
1124 					pmap_remove_all(m);
1125 					if (m->dirty == 0)
1126 						vm_page_cache(m);
1127 					else
1128 						vm_page_deactivate(m);
1129 				} else {
1130 					vm_page_deactivate(m);
1131 				}
1132 			} else {
1133 				vm_page_requeue(m);
1134 			}
1135 		}
1136 		VM_OBJECT_UNLOCK(object);
1137 		m = next;
1138 	}
1139 	vm_page_unlock_queues();
1140 #if !defined(NO_SWAPPING)
1141 	/*
1142 	 * Idle process swapout -- run once per second.
1143 	 */
1144 	if (vm_swap_idle_enabled) {
1145 		static long lsec;
1146 		if (time_second != lsec) {
1147 			vm_req_vmdaemon(VM_SWAP_IDLE);
1148 			lsec = time_second;
1149 		}
1150 	}
1151 #endif
1152 
1153 	/*
1154 	 * If we didn't get enough free pages, and we have skipped a vnode
1155 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1156 	 * if we did not get enough free pages.
1157 	 */
1158 	if (vm_paging_target() > 0) {
1159 		if (vnodes_skipped && vm_page_count_min())
1160 			(void) speedup_syncer();
1161 #if !defined(NO_SWAPPING)
1162 		if (vm_swap_enabled && vm_page_count_target())
1163 			vm_req_vmdaemon(VM_SWAP_NORMAL);
1164 #endif
1165 	}
1166 
1167 	/*
1168 	 * If we are critically low on one of RAM or swap and low on
1169 	 * the other, kill the largest process.  However, we avoid
1170 	 * doing this on the first pass in order to give ourselves a
1171 	 * chance to flush out dirty vnode-backed pages and to allow
1172 	 * active pages to be moved to the inactive queue and reclaimed.
1173 	 */
1174 	if (pass != 0 &&
1175 	    ((swap_pager_avail < 64 && vm_page_count_min()) ||
1176 	     (swap_pager_full && vm_paging_target() > 0)))
1177 		vm_pageout_oom(VM_OOM_MEM);
1178 }
1179 
1180 
1181 void
1182 vm_pageout_oom(int shortage)
1183 {
1184 	struct proc *p, *bigproc;
1185 	vm_offset_t size, bigsize;
1186 	struct thread *td;
1187 
1188 	/*
1189 	 * We keep the process bigproc locked once we find it to keep anyone
1190 	 * from messing with it; however, there is a possibility of
1191 	 * deadlock if process B is bigproc and one of it's child processes
1192 	 * attempts to propagate a signal to B while we are waiting for A's
1193 	 * lock while walking this list.  To avoid this, we don't block on
1194 	 * the process lock but just skip a process if it is already locked.
1195 	 */
1196 	bigproc = NULL;
1197 	bigsize = 0;
1198 	sx_slock(&allproc_lock);
1199 	FOREACH_PROC_IN_SYSTEM(p) {
1200 		int breakout;
1201 
1202 		if (PROC_TRYLOCK(p) == 0)
1203 			continue;
1204 		/*
1205 		 * If this is a system or protected process, skip it.
1206 		 */
1207 		if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1208 		    (p->p_flag & P_PROTECTED) ||
1209 		    ((p->p_pid < 48) && (swap_pager_avail != 0))) {
1210 			PROC_UNLOCK(p);
1211 			continue;
1212 		}
1213 		/*
1214 		 * If the process is in a non-running type state,
1215 		 * don't touch it.  Check all the threads individually.
1216 		 */
1217 		breakout = 0;
1218 		FOREACH_THREAD_IN_PROC(p, td) {
1219 			thread_lock(td);
1220 			if (!TD_ON_RUNQ(td) &&
1221 			    !TD_IS_RUNNING(td) &&
1222 			    !TD_IS_SLEEPING(td)) {
1223 				thread_unlock(td);
1224 				breakout = 1;
1225 				break;
1226 			}
1227 			thread_unlock(td);
1228 		}
1229 		if (breakout) {
1230 			PROC_UNLOCK(p);
1231 			continue;
1232 		}
1233 		/*
1234 		 * get the process size
1235 		 */
1236 		if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) {
1237 			PROC_UNLOCK(p);
1238 			continue;
1239 		}
1240 		size = vmspace_swap_count(p->p_vmspace);
1241 		vm_map_unlock_read(&p->p_vmspace->vm_map);
1242 		if (shortage == VM_OOM_MEM)
1243 			size += vmspace_resident_count(p->p_vmspace);
1244 		/*
1245 		 * if the this process is bigger than the biggest one
1246 		 * remember it.
1247 		 */
1248 		if (size > bigsize) {
1249 			if (bigproc != NULL)
1250 				PROC_UNLOCK(bigproc);
1251 			bigproc = p;
1252 			bigsize = size;
1253 		} else
1254 			PROC_UNLOCK(p);
1255 	}
1256 	sx_sunlock(&allproc_lock);
1257 	if (bigproc != NULL) {
1258 		killproc(bigproc, "out of swap space");
1259 		sched_nice(bigproc, PRIO_MIN);
1260 		PROC_UNLOCK(bigproc);
1261 		wakeup(&cnt.v_free_count);
1262 	}
1263 }
1264 
1265 /*
1266  * This routine tries to maintain the pseudo LRU active queue,
1267  * so that during long periods of time where there is no paging,
1268  * that some statistic accumulation still occurs.  This code
1269  * helps the situation where paging just starts to occur.
1270  */
1271 static void
1272 vm_pageout_page_stats()
1273 {
1274 	vm_object_t object;
1275 	vm_page_t m,next;
1276 	int pcount,tpcount;		/* Number of pages to check */
1277 	static int fullintervalcount = 0;
1278 	int page_shortage;
1279 
1280 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1281 	page_shortage =
1282 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1283 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1284 
1285 	if (page_shortage <= 0)
1286 		return;
1287 
1288 	pcount = cnt.v_active_count;
1289 	fullintervalcount += vm_pageout_stats_interval;
1290 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1291 		tpcount = (int64_t)vm_pageout_stats_max * cnt.v_active_count /
1292 		    cnt.v_page_count;
1293 		if (pcount > tpcount)
1294 			pcount = tpcount;
1295 	} else {
1296 		fullintervalcount = 0;
1297 	}
1298 
1299 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1300 	while ((m != NULL) && (pcount-- > 0)) {
1301 		int actcount;
1302 
1303 		KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE),
1304 		    ("vm_pageout_page_stats: page %p isn't active", m));
1305 
1306 		next = TAILQ_NEXT(m, pageq);
1307 		object = m->object;
1308 
1309 		if ((m->flags & PG_MARKER) != 0) {
1310 			m = next;
1311 			continue;
1312 		}
1313 		if (!VM_OBJECT_TRYLOCK(object) &&
1314 		    !vm_pageout_fallback_object_lock(m, &next)) {
1315 			VM_OBJECT_UNLOCK(object);
1316 			m = next;
1317 			continue;
1318 		}
1319 
1320 		/*
1321 		 * Don't deactivate pages that are busy.
1322 		 */
1323 		if ((m->busy != 0) ||
1324 		    (m->oflags & VPO_BUSY) ||
1325 		    (m->hold_count != 0)) {
1326 			VM_OBJECT_UNLOCK(object);
1327 			vm_page_requeue(m);
1328 			m = next;
1329 			continue;
1330 		}
1331 
1332 		actcount = 0;
1333 		if (m->flags & PG_REFERENCED) {
1334 			vm_page_flag_clear(m, PG_REFERENCED);
1335 			actcount += 1;
1336 		}
1337 
1338 		actcount += pmap_ts_referenced(m);
1339 		if (actcount) {
1340 			m->act_count += ACT_ADVANCE + actcount;
1341 			if (m->act_count > ACT_MAX)
1342 				m->act_count = ACT_MAX;
1343 			vm_page_requeue(m);
1344 		} else {
1345 			if (m->act_count == 0) {
1346 				/*
1347 				 * We turn off page access, so that we have
1348 				 * more accurate RSS stats.  We don't do this
1349 				 * in the normal page deactivation when the
1350 				 * system is loaded VM wise, because the
1351 				 * cost of the large number of page protect
1352 				 * operations would be higher than the value
1353 				 * of doing the operation.
1354 				 */
1355 				pmap_remove_all(m);
1356 				vm_page_deactivate(m);
1357 			} else {
1358 				m->act_count -= min(m->act_count, ACT_DECLINE);
1359 				vm_page_requeue(m);
1360 			}
1361 		}
1362 		VM_OBJECT_UNLOCK(object);
1363 		m = next;
1364 	}
1365 }
1366 
1367 /*
1368  *	vm_pageout is the high level pageout daemon.
1369  */
1370 static void
1371 vm_pageout()
1372 {
1373 	int error, pass;
1374 
1375 	/*
1376 	 * Initialize some paging parameters.
1377 	 */
1378 	cnt.v_interrupt_free_min = 2;
1379 	if (cnt.v_page_count < 2000)
1380 		vm_pageout_page_count = 8;
1381 
1382 	/*
1383 	 * v_free_reserved needs to include enough for the largest
1384 	 * swap pager structures plus enough for any pv_entry structs
1385 	 * when paging.
1386 	 */
1387 	if (cnt.v_page_count > 1024)
1388 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1389 	else
1390 		cnt.v_free_min = 4;
1391 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1392 	    cnt.v_interrupt_free_min;
1393 	cnt.v_free_reserved = vm_pageout_page_count +
1394 	    cnt.v_pageout_free_min + (cnt.v_page_count / 768);
1395 	cnt.v_free_severe = cnt.v_free_min / 2;
1396 	cnt.v_free_min += cnt.v_free_reserved;
1397 	cnt.v_free_severe += cnt.v_free_reserved;
1398 
1399 	/*
1400 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1401 	 * that these are more a measure of the VM cache queue hysteresis
1402 	 * then the VM free queue.  Specifically, v_free_target is the
1403 	 * high water mark (free+cache pages).
1404 	 *
1405 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1406 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1407 	 * be big enough to handle memory needs while the pageout daemon
1408 	 * is signalled and run to free more pages.
1409 	 */
1410 	if (cnt.v_free_count > 6144)
1411 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1412 	else
1413 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1414 
1415 	if (cnt.v_free_count > 2048) {
1416 		cnt.v_cache_min = cnt.v_free_target;
1417 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1418 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1419 	} else {
1420 		cnt.v_cache_min = 0;
1421 		cnt.v_cache_max = 0;
1422 		cnt.v_inactive_target = cnt.v_free_count / 4;
1423 	}
1424 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1425 		cnt.v_inactive_target = cnt.v_free_count / 3;
1426 
1427 	/* XXX does not really belong here */
1428 	if (vm_page_max_wired == 0)
1429 		vm_page_max_wired = cnt.v_free_count / 3;
1430 
1431 	if (vm_pageout_stats_max == 0)
1432 		vm_pageout_stats_max = cnt.v_free_target;
1433 
1434 	/*
1435 	 * Set interval in seconds for stats scan.
1436 	 */
1437 	if (vm_pageout_stats_interval == 0)
1438 		vm_pageout_stats_interval = 5;
1439 	if (vm_pageout_full_stats_interval == 0)
1440 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1441 
1442 	swap_pager_swap_init();
1443 	pass = 0;
1444 	/*
1445 	 * The pageout daemon is never done, so loop forever.
1446 	 */
1447 	while (TRUE) {
1448 		/*
1449 		 * If we have enough free memory, wakeup waiters.  Do
1450 		 * not clear vm_pages_needed until we reach our target,
1451 		 * otherwise we may be woken up over and over again and
1452 		 * waste a lot of cpu.
1453 		 */
1454 		mtx_lock(&vm_page_queue_free_mtx);
1455 		if (vm_pages_needed && !vm_page_count_min()) {
1456 			if (!vm_paging_needed())
1457 				vm_pages_needed = 0;
1458 			wakeup(&cnt.v_free_count);
1459 		}
1460 		if (vm_pages_needed) {
1461 			/*
1462 			 * Still not done, take a second pass without waiting
1463 			 * (unlimited dirty cleaning), otherwise sleep a bit
1464 			 * and try again.
1465 			 */
1466 			++pass;
1467 			if (pass > 1)
1468 				msleep(&vm_pages_needed,
1469 				    &vm_page_queue_free_mtx, PVM, "psleep",
1470 				    hz / 2);
1471 		} else {
1472 			/*
1473 			 * Good enough, sleep & handle stats.  Prime the pass
1474 			 * for the next run.
1475 			 */
1476 			if (pass > 1)
1477 				pass = 1;
1478 			else
1479 				pass = 0;
1480 			error = msleep(&vm_pages_needed,
1481 			    &vm_page_queue_free_mtx, PVM, "psleep",
1482 			    vm_pageout_stats_interval * hz);
1483 			if (error && !vm_pages_needed) {
1484 				mtx_unlock(&vm_page_queue_free_mtx);
1485 				pass = 0;
1486 				vm_page_lock_queues();
1487 				vm_pageout_page_stats();
1488 				vm_page_unlock_queues();
1489 				continue;
1490 			}
1491 		}
1492 		if (vm_pages_needed)
1493 			cnt.v_pdwakeups++;
1494 		mtx_unlock(&vm_page_queue_free_mtx);
1495 		vm_pageout_scan(pass);
1496 	}
1497 }
1498 
1499 /*
1500  * Unless the free page queue lock is held by the caller, this function
1501  * should be regarded as advisory.  Specifically, the caller should
1502  * not msleep() on &cnt.v_free_count following this function unless
1503  * the free page queue lock is held until the msleep() is performed.
1504  */
1505 void
1506 pagedaemon_wakeup()
1507 {
1508 
1509 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1510 		vm_pages_needed = 1;
1511 		wakeup(&vm_pages_needed);
1512 	}
1513 }
1514 
1515 #if !defined(NO_SWAPPING)
1516 static void
1517 vm_req_vmdaemon(int req)
1518 {
1519 	static int lastrun = 0;
1520 
1521 	mtx_lock(&vm_daemon_mtx);
1522 	vm_pageout_req_swapout |= req;
1523 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1524 		wakeup(&vm_daemon_needed);
1525 		lastrun = ticks;
1526 	}
1527 	mtx_unlock(&vm_daemon_mtx);
1528 }
1529 
1530 static void
1531 vm_daemon()
1532 {
1533 	struct rlimit rsslim;
1534 	struct proc *p;
1535 	struct thread *td;
1536 	int breakout, swapout_flags;
1537 
1538 	while (TRUE) {
1539 		mtx_lock(&vm_daemon_mtx);
1540 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep", 0);
1541 		swapout_flags = vm_pageout_req_swapout;
1542 		vm_pageout_req_swapout = 0;
1543 		mtx_unlock(&vm_daemon_mtx);
1544 		if (swapout_flags)
1545 			swapout_procs(swapout_flags);
1546 
1547 		/*
1548 		 * scan the processes for exceeding their rlimits or if
1549 		 * process is swapped out -- deactivate pages
1550 		 */
1551 		sx_slock(&allproc_lock);
1552 		FOREACH_PROC_IN_SYSTEM(p) {
1553 			vm_pindex_t limit, size;
1554 
1555 			/*
1556 			 * if this is a system process or if we have already
1557 			 * looked at this process, skip it.
1558 			 */
1559 			PROC_LOCK(p);
1560 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1561 				PROC_UNLOCK(p);
1562 				continue;
1563 			}
1564 			/*
1565 			 * if the process is in a non-running type state,
1566 			 * don't touch it.
1567 			 */
1568 			breakout = 0;
1569 			FOREACH_THREAD_IN_PROC(p, td) {
1570 				thread_lock(td);
1571 				if (!TD_ON_RUNQ(td) &&
1572 				    !TD_IS_RUNNING(td) &&
1573 				    !TD_IS_SLEEPING(td)) {
1574 					thread_unlock(td);
1575 					breakout = 1;
1576 					break;
1577 				}
1578 				thread_unlock(td);
1579 			}
1580 			if (breakout) {
1581 				PROC_UNLOCK(p);
1582 				continue;
1583 			}
1584 			/*
1585 			 * get a limit
1586 			 */
1587 			lim_rlimit(p, RLIMIT_RSS, &rsslim);
1588 			limit = OFF_TO_IDX(
1589 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1590 
1591 			/*
1592 			 * let processes that are swapped out really be
1593 			 * swapped out set the limit to nothing (will force a
1594 			 * swap-out.)
1595 			 */
1596 			if ((p->p_flag & P_INMEM) == 0)
1597 				limit = 0;	/* XXX */
1598 			PROC_UNLOCK(p);
1599 
1600 			size = vmspace_resident_count(p->p_vmspace);
1601 			if (limit >= 0 && size >= limit) {
1602 				vm_pageout_map_deactivate_pages(
1603 				    &p->p_vmspace->vm_map, limit);
1604 			}
1605 		}
1606 		sx_sunlock(&allproc_lock);
1607 	}
1608 }
1609 #endif			/* !defined(NO_SWAPPING) */
1610