xref: /freebsd/sys/vm/vm_pageout.c (revision 999c1fd64b489eda8c04f1e1529f828ebe5c7794)
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 
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
83 #include <sys/eventhandler.h>
84 #include <sys/lock.h>
85 #include <sys/mutex.h>
86 #include <sys/proc.h>
87 #include <sys/kthread.h>
88 #include <sys/ktr.h>
89 #include <sys/mount.h>
90 #include <sys/racct.h>
91 #include <sys/resourcevar.h>
92 #include <sys/sched.h>
93 #include <sys/sdt.h>
94 #include <sys/signalvar.h>
95 #include <sys/smp.h>
96 #include <sys/time.h>
97 #include <sys/vnode.h>
98 #include <sys/vmmeter.h>
99 #include <sys/rwlock.h>
100 #include <sys/sx.h>
101 #include <sys/sysctl.h>
102 
103 #include <vm/vm.h>
104 #include <vm/vm_param.h>
105 #include <vm/vm_object.h>
106 #include <vm/vm_page.h>
107 #include <vm/vm_map.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_pager.h>
110 #include <vm/vm_phys.h>
111 #include <vm/swap_pager.h>
112 #include <vm/vm_extern.h>
113 #include <vm/uma.h>
114 
115 /*
116  * System initialization
117  */
118 
119 /* the kernel process "vm_pageout"*/
120 static void vm_pageout(void);
121 static void vm_pageout_init(void);
122 static int vm_pageout_clean(vm_page_t m);
123 static int vm_pageout_cluster(vm_page_t m);
124 static void vm_pageout_scan(struct vm_domain *vmd, int pass);
125 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
126     int starting_page_shortage);
127 
128 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
129     NULL);
130 
131 struct proc *pageproc;
132 
133 static struct kproc_desc page_kp = {
134 	"pagedaemon",
135 	vm_pageout,
136 	&pageproc
137 };
138 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
139     &page_kp);
140 
141 SDT_PROVIDER_DEFINE(vm);
142 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
143 
144 #if !defined(NO_SWAPPING)
145 /* the kernel process "vm_daemon"*/
146 static void vm_daemon(void);
147 static struct	proc *vmproc;
148 
149 static struct kproc_desc vm_kp = {
150 	"vmdaemon",
151 	vm_daemon,
152 	&vmproc
153 };
154 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
155 #endif
156 
157 
158 int vm_pageout_deficit;		/* Estimated number of pages deficit */
159 u_int vm_pageout_wakeup_thresh;
160 static int vm_pageout_oom_seq = 12;
161 bool vm_pageout_wanted;		/* Event on which pageout daemon sleeps */
162 bool vm_pages_needed;		/* Are threads waiting for free pages? */
163 
164 #if !defined(NO_SWAPPING)
165 static int vm_pageout_req_swapout;	/* XXX */
166 static int vm_daemon_needed;
167 static struct mtx vm_daemon_mtx;
168 /* Allow for use by vm_pageout before vm_daemon is initialized. */
169 MTX_SYSINIT(vm_daemon, &vm_daemon_mtx, "vm daemon", MTX_DEF);
170 #endif
171 static int vm_max_launder = 32;
172 static int vm_pageout_update_period;
173 static int defer_swap_pageouts;
174 static int disable_swap_pageouts;
175 static int lowmem_period = 10;
176 static time_t lowmem_uptime;
177 
178 #if defined(NO_SWAPPING)
179 static int vm_swap_enabled = 0;
180 static int vm_swap_idle_enabled = 0;
181 #else
182 static int vm_swap_enabled = 1;
183 static int vm_swap_idle_enabled = 0;
184 #endif
185 
186 static int vm_panic_on_oom = 0;
187 
188 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
189 	CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
190 	"panic on out of memory instead of killing the largest process");
191 
192 SYSCTL_INT(_vm, OID_AUTO, pageout_wakeup_thresh,
193 	CTLFLAG_RW, &vm_pageout_wakeup_thresh, 0,
194 	"free page threshold for waking up the pageout daemon");
195 
196 SYSCTL_INT(_vm, OID_AUTO, max_launder,
197 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
198 
199 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
200 	CTLFLAG_RW, &vm_pageout_update_period, 0,
201 	"Maximum active LRU update period");
202 
203 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RW, &lowmem_period, 0,
204 	"Low memory callback period");
205 
206 #if defined(NO_SWAPPING)
207 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
208 	CTLFLAG_RD, &vm_swap_enabled, 0, "Enable entire process swapout");
209 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
210 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
211 #else
212 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
213 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
214 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
215 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
216 #endif
217 
218 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
219 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
220 
221 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
222 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
223 
224 static int pageout_lock_miss;
225 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
226 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
227 
228 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
229 	CTLFLAG_RW, &vm_pageout_oom_seq, 0,
230 	"back-to-back calls to oom detector to start OOM");
231 
232 #define VM_PAGEOUT_PAGE_COUNT 16
233 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
234 
235 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
236 SYSCTL_INT(_vm, OID_AUTO, max_wired,
237 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
238 
239 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
240 #if !defined(NO_SWAPPING)
241 static void vm_pageout_map_deactivate_pages(vm_map_t, long);
242 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long);
243 static void vm_req_vmdaemon(int req);
244 #endif
245 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
246 
247 /*
248  * Initialize a dummy page for marking the caller's place in the specified
249  * paging queue.  In principle, this function only needs to set the flag
250  * PG_MARKER.  Nonetheless, it wirte busies and initializes the hold count
251  * to one as safety precautions.
252  */
253 static void
254 vm_pageout_init_marker(vm_page_t marker, u_short queue)
255 {
256 
257 	bzero(marker, sizeof(*marker));
258 	marker->flags = PG_MARKER;
259 	marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
260 	marker->queue = queue;
261 	marker->hold_count = 1;
262 }
263 
264 /*
265  * vm_pageout_fallback_object_lock:
266  *
267  * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
268  * known to have failed and page queue must be either PQ_ACTIVE or
269  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queue
270  * while locking the vm object.  Use marker page to detect page queue
271  * changes and maintain notion of next page on page queue.  Return
272  * TRUE if no changes were detected, FALSE otherwise.  vm object is
273  * locked on return.
274  *
275  * This function depends on both the lock portion of struct vm_object
276  * and normal struct vm_page being type stable.
277  */
278 static boolean_t
279 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
280 {
281 	struct vm_page marker;
282 	struct vm_pagequeue *pq;
283 	boolean_t unchanged;
284 	u_short queue;
285 	vm_object_t object;
286 
287 	queue = m->queue;
288 	vm_pageout_init_marker(&marker, queue);
289 	pq = vm_page_pagequeue(m);
290 	object = m->object;
291 
292 	TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
293 	vm_pagequeue_unlock(pq);
294 	vm_page_unlock(m);
295 	VM_OBJECT_WLOCK(object);
296 	vm_page_lock(m);
297 	vm_pagequeue_lock(pq);
298 
299 	/*
300 	 * The page's object might have changed, and/or the page might
301 	 * have moved from its original position in the queue.  If the
302 	 * page's object has changed, then the caller should abandon
303 	 * processing the page because the wrong object lock was
304 	 * acquired.  Use the marker's plinks.q, not the page's, to
305 	 * determine if the page has been moved.  The state of the
306 	 * page's plinks.q can be indeterminate; whereas, the marker's
307 	 * plinks.q must be valid.
308 	 */
309 	*next = TAILQ_NEXT(&marker, plinks.q);
310 	unchanged = m->object == object &&
311 	    m == TAILQ_PREV(&marker, pglist, plinks.q);
312 	KASSERT(!unchanged || m->queue == queue,
313 	    ("page %p queue %d %d", m, queue, m->queue));
314 	TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
315 	return (unchanged);
316 }
317 
318 /*
319  * Lock the page while holding the page queue lock.  Use marker page
320  * to detect page queue changes and maintain notion of next page on
321  * page queue.  Return TRUE if no changes were detected, FALSE
322  * otherwise.  The page is locked on return. The page queue lock might
323  * be dropped and reacquired.
324  *
325  * This function depends on normal struct vm_page being type stable.
326  */
327 static boolean_t
328 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
329 {
330 	struct vm_page marker;
331 	struct vm_pagequeue *pq;
332 	boolean_t unchanged;
333 	u_short queue;
334 
335 	vm_page_lock_assert(m, MA_NOTOWNED);
336 	if (vm_page_trylock(m))
337 		return (TRUE);
338 
339 	queue = m->queue;
340 	vm_pageout_init_marker(&marker, queue);
341 	pq = vm_page_pagequeue(m);
342 
343 	TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
344 	vm_pagequeue_unlock(pq);
345 	vm_page_lock(m);
346 	vm_pagequeue_lock(pq);
347 
348 	/* Page queue might have changed. */
349 	*next = TAILQ_NEXT(&marker, plinks.q);
350 	unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
351 	KASSERT(!unchanged || m->queue == queue,
352 	    ("page %p queue %d %d", m, queue, m->queue));
353 	TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
354 	return (unchanged);
355 }
356 
357 /*
358  * vm_pageout_clean:
359  *
360  * Clean the page and remove it from the laundry.
361  *
362  * We set the busy bit to cause potential page faults on this page to
363  * block.  Note the careful timing, however, the busy bit isn't set till
364  * late and we cannot do anything that will mess with the page.
365  */
366 static int
367 vm_pageout_cluster(vm_page_t m)
368 {
369 	vm_object_t object;
370 	vm_page_t mc[2*vm_pageout_page_count], pb, ps;
371 	int pageout_count;
372 	int ib, is, page_base;
373 	vm_pindex_t pindex = m->pindex;
374 
375 	vm_page_lock_assert(m, MA_OWNED);
376 	object = m->object;
377 	VM_OBJECT_ASSERT_WLOCKED(object);
378 
379 	/*
380 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
381 	 * with the new swapper, but we could have serious problems paging
382 	 * out other object types if there is insufficient memory.
383 	 *
384 	 * Unfortunately, checking free memory here is far too late, so the
385 	 * check has been moved up a procedural level.
386 	 */
387 
388 	/*
389 	 * Can't clean the page if it's busy or held.
390 	 */
391 	vm_page_assert_unbusied(m);
392 	KASSERT(m->hold_count == 0, ("vm_pageout_clean: page %p is held", m));
393 	vm_page_unlock(m);
394 
395 	mc[vm_pageout_page_count] = pb = ps = m;
396 	pageout_count = 1;
397 	page_base = vm_pageout_page_count;
398 	ib = 1;
399 	is = 1;
400 
401 	/*
402 	 * Scan object for clusterable pages.
403 	 *
404 	 * We can cluster ONLY if: ->> the page is NOT
405 	 * clean, wired, busy, held, or mapped into a
406 	 * buffer, and one of the following:
407 	 * 1) The page is inactive, or a seldom used
408 	 *    active page.
409 	 * -or-
410 	 * 2) we force the issue.
411 	 *
412 	 * During heavy mmap/modification loads the pageout
413 	 * daemon can really fragment the underlying file
414 	 * due to flushing pages out of order and not trying
415 	 * align the clusters (which leave sporatic out-of-order
416 	 * holes).  To solve this problem we do the reverse scan
417 	 * first and attempt to align our cluster, then do a
418 	 * forward scan if room remains.
419 	 */
420 more:
421 	while (ib && pageout_count < vm_pageout_page_count) {
422 		vm_page_t p;
423 
424 		if (ib > pindex) {
425 			ib = 0;
426 			break;
427 		}
428 
429 		if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
430 			ib = 0;
431 			break;
432 		}
433 		vm_page_test_dirty(p);
434 		if (p->dirty == 0) {
435 			ib = 0;
436 			break;
437 		}
438 		vm_page_lock(p);
439 		if (p->queue != PQ_INACTIVE ||
440 		    p->hold_count != 0) {	/* may be undergoing I/O */
441 			vm_page_unlock(p);
442 			ib = 0;
443 			break;
444 		}
445 		vm_page_unlock(p);
446 		mc[--page_base] = pb = p;
447 		++pageout_count;
448 		++ib;
449 		/*
450 		 * alignment boundary, stop here and switch directions.  Do
451 		 * not clear ib.
452 		 */
453 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
454 			break;
455 	}
456 
457 	while (pageout_count < vm_pageout_page_count &&
458 	    pindex + is < object->size) {
459 		vm_page_t p;
460 
461 		if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
462 			break;
463 		vm_page_test_dirty(p);
464 		if (p->dirty == 0)
465 			break;
466 		vm_page_lock(p);
467 		if (p->queue != PQ_INACTIVE ||
468 		    p->hold_count != 0) {	/* may be undergoing I/O */
469 			vm_page_unlock(p);
470 			break;
471 		}
472 		vm_page_unlock(p);
473 		mc[page_base + pageout_count] = ps = p;
474 		++pageout_count;
475 		++is;
476 	}
477 
478 	/*
479 	 * If we exhausted our forward scan, continue with the reverse scan
480 	 * when possible, even past a page boundary.  This catches boundary
481 	 * conditions.
482 	 */
483 	if (ib && pageout_count < vm_pageout_page_count)
484 		goto more;
485 
486 	/*
487 	 * we allow reads during pageouts...
488 	 */
489 	return (vm_pageout_flush(&mc[page_base], pageout_count, 0, 0, NULL,
490 	    NULL));
491 }
492 
493 /*
494  * vm_pageout_flush() - launder the given pages
495  *
496  *	The given pages are laundered.  Note that we setup for the start of
497  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
498  *	reference count all in here rather then in the parent.  If we want
499  *	the parent to do more sophisticated things we may have to change
500  *	the ordering.
501  *
502  *	Returned runlen is the count of pages between mreq and first
503  *	page after mreq with status VM_PAGER_AGAIN.
504  *	*eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
505  *	for any page in runlen set.
506  */
507 int
508 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
509     boolean_t *eio)
510 {
511 	vm_object_t object = mc[0]->object;
512 	int pageout_status[count];
513 	int numpagedout = 0;
514 	int i, runlen;
515 
516 	VM_OBJECT_ASSERT_WLOCKED(object);
517 
518 	/*
519 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
520 	 * mark the pages read-only.
521 	 *
522 	 * We do not have to fixup the clean/dirty bits here... we can
523 	 * allow the pager to do it after the I/O completes.
524 	 *
525 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
526 	 * edge case with file fragments.
527 	 */
528 	for (i = 0; i < count; i++) {
529 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
530 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
531 			mc[i], i, count));
532 		vm_page_sbusy(mc[i]);
533 		pmap_remove_write(mc[i]);
534 	}
535 	vm_object_pip_add(object, count);
536 
537 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
538 
539 	runlen = count - mreq;
540 	if (eio != NULL)
541 		*eio = FALSE;
542 	for (i = 0; i < count; i++) {
543 		vm_page_t mt = mc[i];
544 
545 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
546 		    !pmap_page_is_write_mapped(mt),
547 		    ("vm_pageout_flush: page %p is not write protected", mt));
548 		switch (pageout_status[i]) {
549 		case VM_PAGER_OK:
550 		case VM_PAGER_PEND:
551 			numpagedout++;
552 			break;
553 		case VM_PAGER_BAD:
554 			/*
555 			 * Page outside of range of object. Right now we
556 			 * essentially lose the changes by pretending it
557 			 * worked.
558 			 */
559 			vm_page_undirty(mt);
560 			break;
561 		case VM_PAGER_ERROR:
562 		case VM_PAGER_FAIL:
563 			/*
564 			 * If page couldn't be paged out, then reactivate the
565 			 * page so it doesn't clog the inactive list.  (We
566 			 * will try paging out it again later).
567 			 */
568 			vm_page_lock(mt);
569 			vm_page_activate(mt);
570 			vm_page_unlock(mt);
571 			if (eio != NULL && i >= mreq && i - mreq < runlen)
572 				*eio = TRUE;
573 			break;
574 		case VM_PAGER_AGAIN:
575 			if (i >= mreq && i - mreq < runlen)
576 				runlen = i - mreq;
577 			break;
578 		}
579 
580 		/*
581 		 * If the operation is still going, leave the page busy to
582 		 * block all other accesses. Also, leave the paging in
583 		 * progress indicator set so that we don't attempt an object
584 		 * collapse.
585 		 */
586 		if (pageout_status[i] != VM_PAGER_PEND) {
587 			vm_object_pip_wakeup(object);
588 			vm_page_sunbusy(mt);
589 		}
590 	}
591 	if (prunlen != NULL)
592 		*prunlen = runlen;
593 	return (numpagedout);
594 }
595 
596 #if !defined(NO_SWAPPING)
597 /*
598  *	vm_pageout_object_deactivate_pages
599  *
600  *	Deactivate enough pages to satisfy the inactive target
601  *	requirements.
602  *
603  *	The object and map must be locked.
604  */
605 static void
606 vm_pageout_object_deactivate_pages(pmap_t pmap, vm_object_t first_object,
607     long desired)
608 {
609 	vm_object_t backing_object, object;
610 	vm_page_t p;
611 	int act_delta, remove_mode;
612 
613 	VM_OBJECT_ASSERT_LOCKED(first_object);
614 	if ((first_object->flags & OBJ_FICTITIOUS) != 0)
615 		return;
616 	for (object = first_object;; object = backing_object) {
617 		if (pmap_resident_count(pmap) <= desired)
618 			goto unlock_return;
619 		VM_OBJECT_ASSERT_LOCKED(object);
620 		if ((object->flags & OBJ_UNMANAGED) != 0 ||
621 		    object->paging_in_progress != 0)
622 			goto unlock_return;
623 
624 		remove_mode = 0;
625 		if (object->shadow_count > 1)
626 			remove_mode = 1;
627 		/*
628 		 * Scan the object's entire memory queue.
629 		 */
630 		TAILQ_FOREACH(p, &object->memq, listq) {
631 			if (pmap_resident_count(pmap) <= desired)
632 				goto unlock_return;
633 			if (vm_page_busied(p))
634 				continue;
635 			PCPU_INC(cnt.v_pdpages);
636 			vm_page_lock(p);
637 			if (p->wire_count != 0 || p->hold_count != 0 ||
638 			    !pmap_page_exists_quick(pmap, p)) {
639 				vm_page_unlock(p);
640 				continue;
641 			}
642 			act_delta = pmap_ts_referenced(p);
643 			if ((p->aflags & PGA_REFERENCED) != 0) {
644 				if (act_delta == 0)
645 					act_delta = 1;
646 				vm_page_aflag_clear(p, PGA_REFERENCED);
647 			}
648 			if (p->queue != PQ_ACTIVE && act_delta != 0) {
649 				vm_page_activate(p);
650 				p->act_count += act_delta;
651 			} else if (p->queue == PQ_ACTIVE) {
652 				if (act_delta == 0) {
653 					p->act_count -= min(p->act_count,
654 					    ACT_DECLINE);
655 					if (!remove_mode && p->act_count == 0) {
656 						pmap_remove_all(p);
657 						vm_page_deactivate(p);
658 					} else
659 						vm_page_requeue(p);
660 				} else {
661 					vm_page_activate(p);
662 					if (p->act_count < ACT_MAX -
663 					    ACT_ADVANCE)
664 						p->act_count += ACT_ADVANCE;
665 					vm_page_requeue(p);
666 				}
667 			} else if (p->queue == PQ_INACTIVE)
668 				pmap_remove_all(p);
669 			vm_page_unlock(p);
670 		}
671 		if ((backing_object = object->backing_object) == NULL)
672 			goto unlock_return;
673 		VM_OBJECT_RLOCK(backing_object);
674 		if (object != first_object)
675 			VM_OBJECT_RUNLOCK(object);
676 	}
677 unlock_return:
678 	if (object != first_object)
679 		VM_OBJECT_RUNLOCK(object);
680 }
681 
682 /*
683  * deactivate some number of pages in a map, try to do it fairly, but
684  * that is really hard to do.
685  */
686 static void
687 vm_pageout_map_deactivate_pages(map, desired)
688 	vm_map_t map;
689 	long desired;
690 {
691 	vm_map_entry_t tmpe;
692 	vm_object_t obj, bigobj;
693 	int nothingwired;
694 
695 	if (!vm_map_trylock(map))
696 		return;
697 
698 	bigobj = NULL;
699 	nothingwired = TRUE;
700 
701 	/*
702 	 * first, search out the biggest object, and try to free pages from
703 	 * that.
704 	 */
705 	tmpe = map->header.next;
706 	while (tmpe != &map->header) {
707 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
708 			obj = tmpe->object.vm_object;
709 			if (obj != NULL && VM_OBJECT_TRYRLOCK(obj)) {
710 				if (obj->shadow_count <= 1 &&
711 				    (bigobj == NULL ||
712 				     bigobj->resident_page_count < obj->resident_page_count)) {
713 					if (bigobj != NULL)
714 						VM_OBJECT_RUNLOCK(bigobj);
715 					bigobj = obj;
716 				} else
717 					VM_OBJECT_RUNLOCK(obj);
718 			}
719 		}
720 		if (tmpe->wired_count > 0)
721 			nothingwired = FALSE;
722 		tmpe = tmpe->next;
723 	}
724 
725 	if (bigobj != NULL) {
726 		vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired);
727 		VM_OBJECT_RUNLOCK(bigobj);
728 	}
729 	/*
730 	 * Next, hunt around for other pages to deactivate.  We actually
731 	 * do this search sort of wrong -- .text first is not the best idea.
732 	 */
733 	tmpe = map->header.next;
734 	while (tmpe != &map->header) {
735 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
736 			break;
737 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
738 			obj = tmpe->object.vm_object;
739 			if (obj != NULL) {
740 				VM_OBJECT_RLOCK(obj);
741 				vm_pageout_object_deactivate_pages(map->pmap, obj, desired);
742 				VM_OBJECT_RUNLOCK(obj);
743 			}
744 		}
745 		tmpe = tmpe->next;
746 	}
747 
748 	/*
749 	 * Remove all mappings if a process is swapped out, this will free page
750 	 * table pages.
751 	 */
752 	if (desired == 0 && nothingwired) {
753 		pmap_remove(vm_map_pmap(map), vm_map_min(map),
754 		    vm_map_max(map));
755 	}
756 
757 	vm_map_unlock(map);
758 }
759 #endif		/* !defined(NO_SWAPPING) */
760 
761 /*
762  * Attempt to acquire all of the necessary locks to launder a page and
763  * then call through the clustering layer to PUTPAGES.  Wait a short
764  * time for a vnode lock.
765  *
766  * Requires the page and object lock on entry, releases both before return.
767  * Returns 0 on success and an errno otherwise.
768  */
769 static int
770 vm_pageout_clean(vm_page_t m)
771 {
772 	struct vnode *vp;
773 	struct mount *mp;
774 	vm_object_t object;
775 	vm_pindex_t pindex;
776 	int error, lockmode;
777 
778 	vm_page_assert_locked(m);
779 	object = m->object;
780 	VM_OBJECT_ASSERT_WLOCKED(object);
781 	error = 0;
782 	vp = NULL;
783 	mp = NULL;
784 
785 	/*
786 	 * The object is already known NOT to be dead.   It
787 	 * is possible for the vget() to block the whole
788 	 * pageout daemon, but the new low-memory handling
789 	 * code should prevent it.
790 	 *
791 	 * We can't wait forever for the vnode lock, we might
792 	 * deadlock due to a vn_read() getting stuck in
793 	 * vm_wait while holding this vnode.  We skip the
794 	 * vnode if we can't get it in a reasonable amount
795 	 * of time.
796 	 */
797 	if (object->type == OBJT_VNODE) {
798 		vm_page_unlock(m);
799 		vp = object->handle;
800 		if (vp->v_type == VREG &&
801 		    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
802 			mp = NULL;
803 			error = EDEADLK;
804 			goto unlock_all;
805 		}
806 		KASSERT(mp != NULL,
807 		    ("vp %p with NULL v_mount", vp));
808 		vm_object_reference_locked(object);
809 		pindex = m->pindex;
810 		VM_OBJECT_WUNLOCK(object);
811 		lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
812 		    LK_SHARED : LK_EXCLUSIVE;
813 		if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
814 			vp = NULL;
815 			error = EDEADLK;
816 			goto unlock_mp;
817 		}
818 		VM_OBJECT_WLOCK(object);
819 		vm_page_lock(m);
820 		/*
821 		 * While the object and page were unlocked, the page
822 		 * may have been:
823 		 * (1) moved to a different queue,
824 		 * (2) reallocated to a different object,
825 		 * (3) reallocated to a different offset, or
826 		 * (4) cleaned.
827 		 */
828 		if (m->queue != PQ_INACTIVE || m->object != object ||
829 		    m->pindex != pindex || m->dirty == 0) {
830 			vm_page_unlock(m);
831 			error = ENXIO;
832 			goto unlock_all;
833 		}
834 
835 		/*
836 		 * The page may have been busied or held while the object
837 		 * and page locks were released.
838 		 */
839 		if (vm_page_busied(m) || m->hold_count != 0) {
840 			vm_page_unlock(m);
841 			error = EBUSY;
842 			goto unlock_all;
843 		}
844 	}
845 
846 	/*
847 	 * If a page is dirty, then it is either being washed
848 	 * (but not yet cleaned) or it is still in the
849 	 * laundry.  If it is still in the laundry, then we
850 	 * start the cleaning operation.
851 	 */
852 	if (vm_pageout_cluster(m) == 0)
853 		error = EIO;
854 
855 unlock_all:
856 	VM_OBJECT_WUNLOCK(object);
857 
858 unlock_mp:
859 	vm_page_lock_assert(m, MA_NOTOWNED);
860 	if (mp != NULL) {
861 		if (vp != NULL)
862 			vput(vp);
863 		vm_object_deallocate(object);
864 		vn_finished_write(mp);
865 	}
866 
867 	return (error);
868 }
869 
870 /*
871  *	vm_pageout_scan does the dirty work for the pageout daemon.
872  *
873  *	pass 0 - Update active LRU/deactivate pages
874  *	pass 1 - Free inactive pages
875  *	pass 2 - Launder dirty pages
876  */
877 static void
878 vm_pageout_scan(struct vm_domain *vmd, int pass)
879 {
880 	vm_page_t m, next;
881 	struct vm_pagequeue *pq;
882 	vm_object_t object;
883 	long min_scan;
884 	int act_delta, addl_page_shortage, deficit, error, maxlaunder, maxscan;
885 	int page_shortage, scan_tick, scanned, starting_page_shortage;
886 	int vnodes_skipped;
887 	boolean_t pageout_ok, queue_locked;
888 
889 	/*
890 	 * If we need to reclaim memory ask kernel caches to return
891 	 * some.  We rate limit to avoid thrashing.
892 	 */
893 	if (vmd == &vm_dom[0] && pass > 0 &&
894 	    (time_uptime - lowmem_uptime) >= lowmem_period) {
895 		/*
896 		 * Decrease registered cache sizes.
897 		 */
898 		SDT_PROBE0(vm, , , vm__lowmem_scan);
899 		EVENTHANDLER_INVOKE(vm_lowmem, 0);
900 		/*
901 		 * We do this explicitly after the caches have been
902 		 * drained above.
903 		 */
904 		uma_reclaim();
905 		lowmem_uptime = time_uptime;
906 	}
907 
908 	/*
909 	 * The addl_page_shortage is the number of temporarily
910 	 * stuck pages in the inactive queue.  In other words, the
911 	 * number of pages from the inactive count that should be
912 	 * discounted in setting the target for the active queue scan.
913 	 */
914 	addl_page_shortage = 0;
915 
916 	/*
917 	 * Calculate the number of pages that we want to free.
918 	 */
919 	if (pass > 0) {
920 		deficit = atomic_readandclear_int(&vm_pageout_deficit);
921 		page_shortage = vm_paging_target() + deficit;
922 	} else
923 		page_shortage = deficit = 0;
924 	starting_page_shortage = page_shortage;
925 
926 	/*
927 	 * maxlaunder limits the number of dirty pages we flush per scan.
928 	 * For most systems a smaller value (16 or 32) is more robust under
929 	 * extreme memory and disk pressure because any unnecessary writes
930 	 * to disk can result in extreme performance degredation.  However,
931 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
932 	 * used) will die horribly with limited laundering.  If the pageout
933 	 * daemon cannot clean enough pages in the first pass, we let it go
934 	 * all out in succeeding passes.
935 	 */
936 	if ((maxlaunder = vm_max_launder) <= 1)
937 		maxlaunder = 1;
938 	if (pass > 1)
939 		maxlaunder = 10000;
940 
941 	vnodes_skipped = 0;
942 
943 	/*
944 	 * Start scanning the inactive queue for pages that we can free.  The
945 	 * scan will stop when we reach the target or we have scanned the
946 	 * entire queue.  (Note that m->act_count is not used to make
947 	 * decisions for the inactive queue, only for the active queue.)
948 	 */
949 	pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
950 	maxscan = pq->pq_cnt;
951 	vm_pagequeue_lock(pq);
952 	queue_locked = TRUE;
953 	for (m = TAILQ_FIRST(&pq->pq_pl);
954 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
955 	     m = next) {
956 		vm_pagequeue_assert_locked(pq);
957 		KASSERT(queue_locked, ("unlocked inactive queue"));
958 		KASSERT(m->queue == PQ_INACTIVE, ("Inactive queue %p", m));
959 
960 		PCPU_INC(cnt.v_pdpages);
961 		next = TAILQ_NEXT(m, plinks.q);
962 
963 		/*
964 		 * skip marker pages
965 		 */
966 		if (m->flags & PG_MARKER)
967 			continue;
968 
969 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
970 		    ("Fictitious page %p cannot be in inactive queue", m));
971 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
972 		    ("Unmanaged page %p cannot be in inactive queue", m));
973 
974 		/*
975 		 * The page or object lock acquisitions fail if the
976 		 * page was removed from the queue or moved to a
977 		 * different position within the queue.  In either
978 		 * case, addl_page_shortage should not be incremented.
979 		 */
980 		if (!vm_pageout_page_lock(m, &next))
981 			goto unlock_page;
982 		else if (m->hold_count != 0) {
983 			/*
984 			 * Held pages are essentially stuck in the
985 			 * queue.  So, they ought to be discounted
986 			 * from the inactive count.  See the
987 			 * calculation of the page_shortage for the
988 			 * loop over the active queue below.
989 			 */
990 			addl_page_shortage++;
991 			goto unlock_page;
992 		}
993 		object = m->object;
994 		if (!VM_OBJECT_TRYWLOCK(object)) {
995 			if (!vm_pageout_fallback_object_lock(m, &next))
996 				goto unlock_object;
997 			else if (m->hold_count != 0) {
998 				addl_page_shortage++;
999 				goto unlock_object;
1000 			}
1001 		}
1002 		if (vm_page_busied(m)) {
1003 			/*
1004 			 * Don't mess with busy pages.  Leave them at
1005 			 * the front of the queue.  Most likely, they
1006 			 * are being paged out and will leave the
1007 			 * queue shortly after the scan finishes.  So,
1008 			 * they ought to be discounted from the
1009 			 * inactive count.
1010 			 */
1011 			addl_page_shortage++;
1012 unlock_object:
1013 			VM_OBJECT_WUNLOCK(object);
1014 unlock_page:
1015 			vm_page_unlock(m);
1016 			continue;
1017 		}
1018 		KASSERT(m->hold_count == 0, ("Held page %p", m));
1019 
1020 		/*
1021 		 * We unlock the inactive page queue, invalidating the
1022 		 * 'next' pointer.  Use our marker to remember our
1023 		 * place.
1024 		 */
1025 		TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1026 		vm_pagequeue_unlock(pq);
1027 		queue_locked = FALSE;
1028 
1029 		/*
1030 		 * Invalid pages can be easily freed. They cannot be
1031 		 * mapped, vm_page_free() asserts this.
1032 		 */
1033 		if (m->valid == 0)
1034 			goto free_page;
1035 
1036 		/*
1037 		 * If the page has been referenced and the object is not dead,
1038 		 * reactivate or requeue the page depending on whether the
1039 		 * object is mapped.
1040 		 */
1041 		if ((m->aflags & PGA_REFERENCED) != 0) {
1042 			vm_page_aflag_clear(m, PGA_REFERENCED);
1043 			act_delta = 1;
1044 		} else
1045 			act_delta = 0;
1046 		if (object->ref_count != 0) {
1047 			act_delta += pmap_ts_referenced(m);
1048 		} else {
1049 			KASSERT(!pmap_page_is_mapped(m),
1050 			    ("vm_pageout_scan: page %p is mapped", m));
1051 		}
1052 		if (act_delta != 0) {
1053 			if (object->ref_count != 0) {
1054 				vm_page_activate(m);
1055 
1056 				/*
1057 				 * Increase the activation count if the page
1058 				 * was referenced while in the inactive queue.
1059 				 * This makes it less likely that the page will
1060 				 * be returned prematurely to the inactive
1061 				 * queue.
1062  				 */
1063 				m->act_count += act_delta + ACT_ADVANCE;
1064 				goto drop_page;
1065 			} else if ((object->flags & OBJ_DEAD) == 0)
1066 				goto requeue_page;
1067 		}
1068 
1069 		/*
1070 		 * If the page appears to be clean at the machine-independent
1071 		 * layer, then remove all of its mappings from the pmap in
1072 		 * anticipation of freeing it.  If, however, any of the page's
1073 		 * mappings allow write access, then the page may still be
1074 		 * modified until the last of those mappings are removed.
1075 		 */
1076 		if (object->ref_count != 0) {
1077 			vm_page_test_dirty(m);
1078 			if (m->dirty == 0)
1079 				pmap_remove_all(m);
1080 		}
1081 
1082 		if (m->dirty == 0) {
1083 			/*
1084 			 * Clean pages can be freed.
1085 			 */
1086 free_page:
1087 			vm_page_free(m);
1088 			PCPU_INC(cnt.v_dfree);
1089 			--page_shortage;
1090 		} else if ((object->flags & OBJ_DEAD) != 0) {
1091 			/*
1092 			 * Leave dirty pages from dead objects at the front of
1093 			 * the queue.  They are being paged out and freed by
1094 			 * the thread that destroyed the object.  They will
1095 			 * leave the queue shortly after the scan finishes, so
1096 			 * they should be discounted from the inactive count.
1097 			 */
1098 			addl_page_shortage++;
1099 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass < 2) {
1100 			/*
1101 			 * Dirty pages need to be paged out, but flushing
1102 			 * a page is extremely expensive versus freeing
1103 			 * a clean page.  Rather then artificially limiting
1104 			 * the number of pages we can flush, we instead give
1105 			 * dirty pages extra priority on the inactive queue
1106 			 * by forcing them to be cycled through the queue
1107 			 * twice before being flushed, after which the
1108 			 * (now clean) page will cycle through once more
1109 			 * before being freed.  This significantly extends
1110 			 * the thrash point for a heavily loaded machine.
1111 			 */
1112 			m->flags |= PG_WINATCFLS;
1113 requeue_page:
1114 			vm_pagequeue_lock(pq);
1115 			queue_locked = TRUE;
1116 			vm_page_requeue_locked(m);
1117 		} else if (maxlaunder > 0) {
1118 			/*
1119 			 * We always want to try to flush some dirty pages if
1120 			 * we encounter them, to keep the system stable.
1121 			 * Normally this number is small, but under extreme
1122 			 * pressure where there are insufficient clean pages
1123 			 * on the inactive queue, we may have to go all out.
1124 			 */
1125 
1126 			if (object->type != OBJT_SWAP &&
1127 			    object->type != OBJT_DEFAULT)
1128 				pageout_ok = TRUE;
1129 			else if (disable_swap_pageouts)
1130 				pageout_ok = FALSE;
1131 			else if (defer_swap_pageouts)
1132 				pageout_ok = vm_page_count_min();
1133 			else
1134 				pageout_ok = TRUE;
1135 			if (!pageout_ok)
1136 				goto requeue_page;
1137 			error = vm_pageout_clean(m);
1138 			/*
1139 			 * Decrement page_shortage on success to account for
1140 			 * the (future) cleaned page.  Otherwise we could wind
1141 			 * up laundering or cleaning too many pages.
1142 			 */
1143 			if (error == 0) {
1144 				page_shortage--;
1145 				maxlaunder--;
1146 			} else if (error == EDEADLK) {
1147 				pageout_lock_miss++;
1148 				vnodes_skipped++;
1149 			} else if (error == EBUSY) {
1150 				addl_page_shortage++;
1151 			}
1152 			vm_page_lock_assert(m, MA_NOTOWNED);
1153 			goto relock_queue;
1154 		}
1155 drop_page:
1156 		vm_page_unlock(m);
1157 		VM_OBJECT_WUNLOCK(object);
1158 relock_queue:
1159 		if (!queue_locked) {
1160 			vm_pagequeue_lock(pq);
1161 			queue_locked = TRUE;
1162 		}
1163 		next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1164 		TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1165 	}
1166 	vm_pagequeue_unlock(pq);
1167 
1168 #if !defined(NO_SWAPPING)
1169 	/*
1170 	 * Wakeup the swapout daemon if we didn't free the targeted number of
1171 	 * pages.
1172 	 */
1173 	if (vm_swap_enabled && page_shortage > 0)
1174 		vm_req_vmdaemon(VM_SWAP_NORMAL);
1175 #endif
1176 
1177 	/*
1178 	 * Wakeup the sync daemon if we skipped a vnode in a writeable object
1179 	 * and we didn't free enough pages.
1180 	 */
1181 	if (vnodes_skipped > 0 && page_shortage > vm_cnt.v_free_target -
1182 	    vm_cnt.v_free_min)
1183 		(void)speedup_syncer();
1184 
1185 	/*
1186 	 * If the inactive queue scan fails repeatedly to meet its
1187 	 * target, kill the largest process.
1188 	 */
1189 	vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1190 
1191 	/*
1192 	 * Compute the number of pages we want to try to move from the
1193 	 * active queue to the inactive queue.
1194 	 */
1195 	page_shortage = vm_cnt.v_inactive_target - vm_cnt.v_inactive_count +
1196 	    vm_paging_target() + deficit + addl_page_shortage;
1197 
1198 	pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1199 	vm_pagequeue_lock(pq);
1200 	maxscan = pq->pq_cnt;
1201 
1202 	/*
1203 	 * If we're just idle polling attempt to visit every
1204 	 * active page within 'update_period' seconds.
1205 	 */
1206 	scan_tick = ticks;
1207 	if (vm_pageout_update_period != 0) {
1208 		min_scan = pq->pq_cnt;
1209 		min_scan *= scan_tick - vmd->vmd_last_active_scan;
1210 		min_scan /= hz * vm_pageout_update_period;
1211 	} else
1212 		min_scan = 0;
1213 	if (min_scan > 0 || (page_shortage > 0 && maxscan > 0))
1214 		vmd->vmd_last_active_scan = scan_tick;
1215 
1216 	/*
1217 	 * Scan the active queue for pages that can be deactivated.  Update
1218 	 * the per-page activity counter and use it to identify deactivation
1219 	 * candidates.
1220 	 */
1221 	for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1222 	    min_scan || (page_shortage > 0 && scanned < maxscan)); m = next,
1223 	    scanned++) {
1224 
1225 		KASSERT(m->queue == PQ_ACTIVE,
1226 		    ("vm_pageout_scan: page %p isn't active", m));
1227 
1228 		next = TAILQ_NEXT(m, plinks.q);
1229 		if ((m->flags & PG_MARKER) != 0)
1230 			continue;
1231 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
1232 		    ("Fictitious page %p cannot be in active queue", m));
1233 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1234 		    ("Unmanaged page %p cannot be in active queue", m));
1235 		if (!vm_pageout_page_lock(m, &next)) {
1236 			vm_page_unlock(m);
1237 			continue;
1238 		}
1239 
1240 		/*
1241 		 * The count for pagedaemon pages is done after checking the
1242 		 * page for eligibility...
1243 		 */
1244 		PCPU_INC(cnt.v_pdpages);
1245 
1246 		/*
1247 		 * Check to see "how much" the page has been used.
1248 		 */
1249 		if ((m->aflags & PGA_REFERENCED) != 0) {
1250 			vm_page_aflag_clear(m, PGA_REFERENCED);
1251 			act_delta = 1;
1252 		} else
1253 			act_delta = 0;
1254 
1255 		/*
1256 		 * Unlocked object ref count check.  Two races are possible.
1257 		 * 1) The ref was transitioning to zero and we saw non-zero,
1258 		 *    the pmap bits will be checked unnecessarily.
1259 		 * 2) The ref was transitioning to one and we saw zero.
1260 		 *    The page lock prevents a new reference to this page so
1261 		 *    we need not check the reference bits.
1262 		 */
1263 		if (m->object->ref_count != 0)
1264 			act_delta += pmap_ts_referenced(m);
1265 
1266 		/*
1267 		 * Advance or decay the act_count based on recent usage.
1268 		 */
1269 		if (act_delta != 0) {
1270 			m->act_count += ACT_ADVANCE + act_delta;
1271 			if (m->act_count > ACT_MAX)
1272 				m->act_count = ACT_MAX;
1273 		} else
1274 			m->act_count -= min(m->act_count, ACT_DECLINE);
1275 
1276 		/*
1277 		 * Move this page to the tail of the active or inactive
1278 		 * queue depending on usage.
1279 		 */
1280 		if (m->act_count == 0) {
1281 			/* Dequeue to avoid later lock recursion. */
1282 			vm_page_dequeue_locked(m);
1283 			vm_page_deactivate(m);
1284 			page_shortage--;
1285 		} else
1286 			vm_page_requeue_locked(m);
1287 		vm_page_unlock(m);
1288 	}
1289 	vm_pagequeue_unlock(pq);
1290 #if !defined(NO_SWAPPING)
1291 	/*
1292 	 * Idle process swapout -- run once per second.
1293 	 */
1294 	if (vm_swap_idle_enabled) {
1295 		static long lsec;
1296 		if (time_second != lsec) {
1297 			vm_req_vmdaemon(VM_SWAP_IDLE);
1298 			lsec = time_second;
1299 		}
1300 	}
1301 #endif
1302 }
1303 
1304 static int vm_pageout_oom_vote;
1305 
1306 /*
1307  * The pagedaemon threads randlomly select one to perform the
1308  * OOM.  Trying to kill processes before all pagedaemons
1309  * failed to reach free target is premature.
1310  */
1311 static void
1312 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1313     int starting_page_shortage)
1314 {
1315 	int old_vote;
1316 
1317 	if (starting_page_shortage <= 0 || starting_page_shortage !=
1318 	    page_shortage)
1319 		vmd->vmd_oom_seq = 0;
1320 	else
1321 		vmd->vmd_oom_seq++;
1322 	if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1323 		if (vmd->vmd_oom) {
1324 			vmd->vmd_oom = FALSE;
1325 			atomic_subtract_int(&vm_pageout_oom_vote, 1);
1326 		}
1327 		return;
1328 	}
1329 
1330 	/*
1331 	 * Do not follow the call sequence until OOM condition is
1332 	 * cleared.
1333 	 */
1334 	vmd->vmd_oom_seq = 0;
1335 
1336 	if (vmd->vmd_oom)
1337 		return;
1338 
1339 	vmd->vmd_oom = TRUE;
1340 	old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1341 	if (old_vote != vm_ndomains - 1)
1342 		return;
1343 
1344 	/*
1345 	 * The current pagedaemon thread is the last in the quorum to
1346 	 * start OOM.  Initiate the selection and signaling of the
1347 	 * victim.
1348 	 */
1349 	vm_pageout_oom(VM_OOM_MEM);
1350 
1351 	/*
1352 	 * After one round of OOM terror, recall our vote.  On the
1353 	 * next pass, current pagedaemon would vote again if the low
1354 	 * memory condition is still there, due to vmd_oom being
1355 	 * false.
1356 	 */
1357 	vmd->vmd_oom = FALSE;
1358 	atomic_subtract_int(&vm_pageout_oom_vote, 1);
1359 }
1360 
1361 /*
1362  * The OOM killer is the page daemon's action of last resort when
1363  * memory allocation requests have been stalled for a prolonged period
1364  * of time because it cannot reclaim memory.  This function computes
1365  * the approximate number of physical pages that could be reclaimed if
1366  * the specified address space is destroyed.
1367  *
1368  * Private, anonymous memory owned by the address space is the
1369  * principal resource that we expect to recover after an OOM kill.
1370  * Since the physical pages mapped by the address space's COW entries
1371  * are typically shared pages, they are unlikely to be released and so
1372  * they are not counted.
1373  *
1374  * To get to the point where the page daemon runs the OOM killer, its
1375  * efforts to write-back vnode-backed pages may have stalled.  This
1376  * could be caused by a memory allocation deadlock in the write path
1377  * that might be resolved by an OOM kill.  Therefore, physical pages
1378  * belonging to vnode-backed objects are counted, because they might
1379  * be freed without being written out first if the address space holds
1380  * the last reference to an unlinked vnode.
1381  *
1382  * Similarly, physical pages belonging to OBJT_PHYS objects are
1383  * counted because the address space might hold the last reference to
1384  * the object.
1385  */
1386 static long
1387 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1388 {
1389 	vm_map_t map;
1390 	vm_map_entry_t entry;
1391 	vm_object_t obj;
1392 	long res;
1393 
1394 	map = &vmspace->vm_map;
1395 	KASSERT(!map->system_map, ("system map"));
1396 	sx_assert(&map->lock, SA_LOCKED);
1397 	res = 0;
1398 	for (entry = map->header.next; entry != &map->header;
1399 	    entry = entry->next) {
1400 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1401 			continue;
1402 		obj = entry->object.vm_object;
1403 		if (obj == NULL)
1404 			continue;
1405 		if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1406 		    obj->ref_count != 1)
1407 			continue;
1408 		switch (obj->type) {
1409 		case OBJT_DEFAULT:
1410 		case OBJT_SWAP:
1411 		case OBJT_PHYS:
1412 		case OBJT_VNODE:
1413 			res += obj->resident_page_count;
1414 			break;
1415 		}
1416 	}
1417 	return (res);
1418 }
1419 
1420 void
1421 vm_pageout_oom(int shortage)
1422 {
1423 	struct proc *p, *bigproc;
1424 	vm_offset_t size, bigsize;
1425 	struct thread *td;
1426 	struct vmspace *vm;
1427 
1428 	/*
1429 	 * We keep the process bigproc locked once we find it to keep anyone
1430 	 * from messing with it; however, there is a possibility of
1431 	 * deadlock if process B is bigproc and one of it's child processes
1432 	 * attempts to propagate a signal to B while we are waiting for A's
1433 	 * lock while walking this list.  To avoid this, we don't block on
1434 	 * the process lock but just skip a process if it is already locked.
1435 	 */
1436 	bigproc = NULL;
1437 	bigsize = 0;
1438 	sx_slock(&allproc_lock);
1439 	FOREACH_PROC_IN_SYSTEM(p) {
1440 		int breakout;
1441 
1442 		PROC_LOCK(p);
1443 
1444 		/*
1445 		 * If this is a system, protected or killed process, skip it.
1446 		 */
1447 		if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1448 		    P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1449 		    p->p_pid == 1 || P_KILLED(p) ||
1450 		    (p->p_pid < 48 && swap_pager_avail != 0)) {
1451 			PROC_UNLOCK(p);
1452 			continue;
1453 		}
1454 		/*
1455 		 * If the process is in a non-running type state,
1456 		 * don't touch it.  Check all the threads individually.
1457 		 */
1458 		breakout = 0;
1459 		FOREACH_THREAD_IN_PROC(p, td) {
1460 			thread_lock(td);
1461 			if (!TD_ON_RUNQ(td) &&
1462 			    !TD_IS_RUNNING(td) &&
1463 			    !TD_IS_SLEEPING(td) &&
1464 			    !TD_IS_SUSPENDED(td) &&
1465 			    !TD_IS_SWAPPED(td)) {
1466 				thread_unlock(td);
1467 				breakout = 1;
1468 				break;
1469 			}
1470 			thread_unlock(td);
1471 		}
1472 		if (breakout) {
1473 			PROC_UNLOCK(p);
1474 			continue;
1475 		}
1476 		/*
1477 		 * get the process size
1478 		 */
1479 		vm = vmspace_acquire_ref(p);
1480 		if (vm == NULL) {
1481 			PROC_UNLOCK(p);
1482 			continue;
1483 		}
1484 		_PHOLD_LITE(p);
1485 		PROC_UNLOCK(p);
1486 		sx_sunlock(&allproc_lock);
1487 		if (!vm_map_trylock_read(&vm->vm_map)) {
1488 			vmspace_free(vm);
1489 			sx_slock(&allproc_lock);
1490 			PRELE(p);
1491 			continue;
1492 		}
1493 		size = vmspace_swap_count(vm);
1494 		if (shortage == VM_OOM_MEM)
1495 			size += vm_pageout_oom_pagecount(vm);
1496 		vm_map_unlock_read(&vm->vm_map);
1497 		vmspace_free(vm);
1498 		sx_slock(&allproc_lock);
1499 
1500 		/*
1501 		 * If this process is bigger than the biggest one,
1502 		 * remember it.
1503 		 */
1504 		if (size > bigsize) {
1505 			if (bigproc != NULL)
1506 				PRELE(bigproc);
1507 			bigproc = p;
1508 			bigsize = size;
1509 		} else {
1510 			PRELE(p);
1511 		}
1512 	}
1513 	sx_sunlock(&allproc_lock);
1514 	if (bigproc != NULL) {
1515 		if (vm_panic_on_oom != 0)
1516 			panic("out of swap space");
1517 		PROC_LOCK(bigproc);
1518 		killproc(bigproc, "out of swap space");
1519 		sched_nice(bigproc, PRIO_MIN);
1520 		_PRELE(bigproc);
1521 		PROC_UNLOCK(bigproc);
1522 		wakeup(&vm_cnt.v_free_count);
1523 	}
1524 }
1525 
1526 static void
1527 vm_pageout_worker(void *arg)
1528 {
1529 	struct vm_domain *domain;
1530 	int domidx;
1531 
1532 	domidx = (uintptr_t)arg;
1533 	domain = &vm_dom[domidx];
1534 
1535 	/*
1536 	 * XXXKIB It could be useful to bind pageout daemon threads to
1537 	 * the cores belonging to the domain, from which vm_page_array
1538 	 * is allocated.
1539 	 */
1540 
1541 	KASSERT(domain->vmd_segs != 0, ("domain without segments"));
1542 	domain->vmd_last_active_scan = ticks;
1543 	vm_pageout_init_marker(&domain->vmd_marker, PQ_INACTIVE);
1544 	vm_pageout_init_marker(&domain->vmd_inacthead, PQ_INACTIVE);
1545 	TAILQ_INSERT_HEAD(&domain->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1546 	    &domain->vmd_inacthead, plinks.q);
1547 
1548 	/*
1549 	 * The pageout daemon worker is never done, so loop forever.
1550 	 */
1551 	while (TRUE) {
1552 		mtx_lock(&vm_page_queue_free_mtx);
1553 
1554 		/*
1555 		 * Generally, after a level >= 1 scan, if there are enough
1556 		 * free pages to wakeup the waiters, then they are already
1557 		 * awake.  A call to vm_page_free() during the scan awakened
1558 		 * them.  However, in the following case, this wakeup serves
1559 		 * to bound the amount of time that a thread might wait.
1560 		 * Suppose a thread's call to vm_page_alloc() fails, but
1561 		 * before that thread calls VM_WAIT, enough pages are freed by
1562 		 * other threads to alleviate the free page shortage.  The
1563 		 * thread will, nonetheless, wait until another page is freed
1564 		 * or this wakeup is performed.
1565 		 */
1566 		if (vm_pages_needed && !vm_page_count_min()) {
1567 			vm_pages_needed = false;
1568 			wakeup(&vm_cnt.v_free_count);
1569 		}
1570 
1571 		/*
1572 		 * Do not clear vm_pageout_wanted until we reach our target.
1573 		 * Otherwise, we may be awakened over and over again, wasting
1574 		 * CPU time.
1575 		 */
1576 		if (vm_pageout_wanted && !vm_paging_needed())
1577 			vm_pageout_wanted = false;
1578 
1579 		/*
1580 		 * Might the page daemon receive a wakeup call?
1581 		 */
1582 		if (vm_pageout_wanted) {
1583 			/*
1584 			 * No.  Either vm_pageout_wanted was set by another
1585 			 * thread during the previous scan, which must have
1586 			 * been a level 0 scan, or vm_pageout_wanted was
1587 			 * already set and the scan failed to free enough
1588 			 * pages.  If we haven't yet performed a level >= 2
1589 			 * scan (unlimited dirty cleaning), then upgrade the
1590 			 * level and scan again now.  Otherwise, sleep a bit
1591 			 * and try again later.
1592 			 */
1593 			mtx_unlock(&vm_page_queue_free_mtx);
1594 			if (domain->vmd_pass > 1)
1595 				pause("psleep", hz / 2);
1596 			domain->vmd_pass++;
1597 		} else {
1598 			/*
1599 			 * Yes.  Sleep until pages need to be reclaimed or
1600 			 * have their reference stats updated.
1601 			 */
1602 			if (mtx_sleep(&vm_pageout_wanted,
1603 			    &vm_page_queue_free_mtx, PDROP | PVM, "psleep",
1604 			    hz) == 0) {
1605 				PCPU_INC(cnt.v_pdwakeups);
1606 				domain->vmd_pass = 1;
1607 			} else
1608 				domain->vmd_pass = 0;
1609 		}
1610 
1611 		vm_pageout_scan(domain, domain->vmd_pass);
1612 	}
1613 }
1614 
1615 /*
1616  *	vm_pageout_init initialises basic pageout daemon settings.
1617  */
1618 static void
1619 vm_pageout_init(void)
1620 {
1621 	/*
1622 	 * Initialize some paging parameters.
1623 	 */
1624 	vm_cnt.v_interrupt_free_min = 2;
1625 	if (vm_cnt.v_page_count < 2000)
1626 		vm_pageout_page_count = 8;
1627 
1628 	/*
1629 	 * v_free_reserved needs to include enough for the largest
1630 	 * swap pager structures plus enough for any pv_entry structs
1631 	 * when paging.
1632 	 */
1633 	if (vm_cnt.v_page_count > 1024)
1634 		vm_cnt.v_free_min = 4 + (vm_cnt.v_page_count - 1024) / 200;
1635 	else
1636 		vm_cnt.v_free_min = 4;
1637 	vm_cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1638 	    vm_cnt.v_interrupt_free_min;
1639 	vm_cnt.v_free_reserved = vm_pageout_page_count +
1640 	    vm_cnt.v_pageout_free_min + (vm_cnt.v_page_count / 768);
1641 	vm_cnt.v_free_severe = vm_cnt.v_free_min / 2;
1642 	vm_cnt.v_free_target = 4 * vm_cnt.v_free_min + vm_cnt.v_free_reserved;
1643 	vm_cnt.v_free_min += vm_cnt.v_free_reserved;
1644 	vm_cnt.v_free_severe += vm_cnt.v_free_reserved;
1645 	vm_cnt.v_inactive_target = (3 * vm_cnt.v_free_target) / 2;
1646 	if (vm_cnt.v_inactive_target > vm_cnt.v_free_count / 3)
1647 		vm_cnt.v_inactive_target = vm_cnt.v_free_count / 3;
1648 
1649 	/*
1650 	 * Set the default wakeup threshold to be 10% above the minimum
1651 	 * page limit.  This keeps the steady state out of shortfall.
1652 	 */
1653 	vm_pageout_wakeup_thresh = (vm_cnt.v_free_min / 10) * 11;
1654 
1655 	/*
1656 	 * Set interval in seconds for active scan.  We want to visit each
1657 	 * page at least once every ten minutes.  This is to prevent worst
1658 	 * case paging behaviors with stale active LRU.
1659 	 */
1660 	if (vm_pageout_update_period == 0)
1661 		vm_pageout_update_period = 600;
1662 
1663 	/* XXX does not really belong here */
1664 	if (vm_page_max_wired == 0)
1665 		vm_page_max_wired = vm_cnt.v_free_count / 3;
1666 }
1667 
1668 /*
1669  *     vm_pageout is the high level pageout daemon.
1670  */
1671 static void
1672 vm_pageout(void)
1673 {
1674 	int error;
1675 #ifdef VM_NUMA_ALLOC
1676 	int i;
1677 #endif
1678 
1679 	swap_pager_swap_init();
1680 #ifdef VM_NUMA_ALLOC
1681 	for (i = 1; i < vm_ndomains; i++) {
1682 		error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1683 		    curproc, NULL, 0, 0, "dom%d", i);
1684 		if (error != 0) {
1685 			panic("starting pageout for domain %d, error %d\n",
1686 			    i, error);
1687 		}
1688 	}
1689 #endif
1690 	error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1691 	    0, 0, "uma");
1692 	if (error != 0)
1693 		panic("starting uma_reclaim helper, error %d\n", error);
1694 	vm_pageout_worker((void *)(uintptr_t)0);
1695 }
1696 
1697 /*
1698  * Unless the free page queue lock is held by the caller, this function
1699  * should be regarded as advisory.  Specifically, the caller should
1700  * not msleep() on &vm_cnt.v_free_count following this function unless
1701  * the free page queue lock is held until the msleep() is performed.
1702  */
1703 void
1704 pagedaemon_wakeup(void)
1705 {
1706 
1707 	if (!vm_pageout_wanted && curthread->td_proc != pageproc) {
1708 		vm_pageout_wanted = true;
1709 		wakeup(&vm_pageout_wanted);
1710 	}
1711 }
1712 
1713 #if !defined(NO_SWAPPING)
1714 static void
1715 vm_req_vmdaemon(int req)
1716 {
1717 	static int lastrun = 0;
1718 
1719 	mtx_lock(&vm_daemon_mtx);
1720 	vm_pageout_req_swapout |= req;
1721 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1722 		wakeup(&vm_daemon_needed);
1723 		lastrun = ticks;
1724 	}
1725 	mtx_unlock(&vm_daemon_mtx);
1726 }
1727 
1728 static void
1729 vm_daemon(void)
1730 {
1731 	struct rlimit rsslim;
1732 	struct proc *p;
1733 	struct thread *td;
1734 	struct vmspace *vm;
1735 	int breakout, swapout_flags, tryagain, attempts;
1736 #ifdef RACCT
1737 	uint64_t rsize, ravailable;
1738 #endif
1739 
1740 	while (TRUE) {
1741 		mtx_lock(&vm_daemon_mtx);
1742 		msleep(&vm_daemon_needed, &vm_daemon_mtx, PPAUSE, "psleep",
1743 #ifdef RACCT
1744 		    racct_enable ? hz : 0
1745 #else
1746 		    0
1747 #endif
1748 		);
1749 		swapout_flags = vm_pageout_req_swapout;
1750 		vm_pageout_req_swapout = 0;
1751 		mtx_unlock(&vm_daemon_mtx);
1752 		if (swapout_flags)
1753 			swapout_procs(swapout_flags);
1754 
1755 		/*
1756 		 * scan the processes for exceeding their rlimits or if
1757 		 * process is swapped out -- deactivate pages
1758 		 */
1759 		tryagain = 0;
1760 		attempts = 0;
1761 again:
1762 		attempts++;
1763 		sx_slock(&allproc_lock);
1764 		FOREACH_PROC_IN_SYSTEM(p) {
1765 			vm_pindex_t limit, size;
1766 
1767 			/*
1768 			 * if this is a system process or if we have already
1769 			 * looked at this process, skip it.
1770 			 */
1771 			PROC_LOCK(p);
1772 			if (p->p_state != PRS_NORMAL ||
1773 			    p->p_flag & (P_INEXEC | P_SYSTEM | P_WEXIT)) {
1774 				PROC_UNLOCK(p);
1775 				continue;
1776 			}
1777 			/*
1778 			 * if the process is in a non-running type state,
1779 			 * don't touch it.
1780 			 */
1781 			breakout = 0;
1782 			FOREACH_THREAD_IN_PROC(p, td) {
1783 				thread_lock(td);
1784 				if (!TD_ON_RUNQ(td) &&
1785 				    !TD_IS_RUNNING(td) &&
1786 				    !TD_IS_SLEEPING(td) &&
1787 				    !TD_IS_SUSPENDED(td)) {
1788 					thread_unlock(td);
1789 					breakout = 1;
1790 					break;
1791 				}
1792 				thread_unlock(td);
1793 			}
1794 			if (breakout) {
1795 				PROC_UNLOCK(p);
1796 				continue;
1797 			}
1798 			/*
1799 			 * get a limit
1800 			 */
1801 			lim_rlimit_proc(p, RLIMIT_RSS, &rsslim);
1802 			limit = OFF_TO_IDX(
1803 			    qmin(rsslim.rlim_cur, rsslim.rlim_max));
1804 
1805 			/*
1806 			 * let processes that are swapped out really be
1807 			 * swapped out set the limit to nothing (will force a
1808 			 * swap-out.)
1809 			 */
1810 			if ((p->p_flag & P_INMEM) == 0)
1811 				limit = 0;	/* XXX */
1812 			vm = vmspace_acquire_ref(p);
1813 			_PHOLD_LITE(p);
1814 			PROC_UNLOCK(p);
1815 			if (vm == NULL) {
1816 				PRELE(p);
1817 				continue;
1818 			}
1819 			sx_sunlock(&allproc_lock);
1820 
1821 			size = vmspace_resident_count(vm);
1822 			if (size >= limit) {
1823 				vm_pageout_map_deactivate_pages(
1824 				    &vm->vm_map, limit);
1825 			}
1826 #ifdef RACCT
1827 			if (racct_enable) {
1828 				rsize = IDX_TO_OFF(size);
1829 				PROC_LOCK(p);
1830 				racct_set(p, RACCT_RSS, rsize);
1831 				ravailable = racct_get_available(p, RACCT_RSS);
1832 				PROC_UNLOCK(p);
1833 				if (rsize > ravailable) {
1834 					/*
1835 					 * Don't be overly aggressive; this
1836 					 * might be an innocent process,
1837 					 * and the limit could've been exceeded
1838 					 * by some memory hog.  Don't try
1839 					 * to deactivate more than 1/4th
1840 					 * of process' resident set size.
1841 					 */
1842 					if (attempts <= 8) {
1843 						if (ravailable < rsize -
1844 						    (rsize / 4)) {
1845 							ravailable = rsize -
1846 							    (rsize / 4);
1847 						}
1848 					}
1849 					vm_pageout_map_deactivate_pages(
1850 					    &vm->vm_map,
1851 					    OFF_TO_IDX(ravailable));
1852 					/* Update RSS usage after paging out. */
1853 					size = vmspace_resident_count(vm);
1854 					rsize = IDX_TO_OFF(size);
1855 					PROC_LOCK(p);
1856 					racct_set(p, RACCT_RSS, rsize);
1857 					PROC_UNLOCK(p);
1858 					if (rsize > ravailable)
1859 						tryagain = 1;
1860 				}
1861 			}
1862 #endif
1863 			vmspace_free(vm);
1864 			sx_slock(&allproc_lock);
1865 			PRELE(p);
1866 		}
1867 		sx_sunlock(&allproc_lock);
1868 		if (tryagain != 0 && attempts <= 10)
1869 			goto again;
1870 	}
1871 }
1872 #endif			/* !defined(NO_SWAPPING) */
1873