xref: /freebsd/sys/vm/vm_pageout.c (revision a134ebd6e63f658f2d3d04ac0c60d23bcaa86dd7)
1 /*-
2  * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1994 John S. Dyson
7  * All rights reserved.
8  * Copyright (c) 1994 David Greenman
9  * All rights reserved.
10  * Copyright (c) 2005 Yahoo! Technologies Norway AS
11  * All rights reserved.
12  *
13  * This code is derived from software contributed to Berkeley by
14  * The Mach Operating System project at Carnegie-Mellon University.
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  * 1. Redistributions of source code must retain the above copyright
20  *    notice, this list of conditions and the following disclaimer.
21  * 2. Redistributions in binary form must reproduce the above copyright
22  *    notice, this list of conditions and the following disclaimer in the
23  *    documentation and/or other materials provided with the distribution.
24  * 3. All advertising materials mentioning features or use of this software
25  *    must display the following acknowledgement:
26  *	This product includes software developed by the University of
27  *	California, Berkeley and its contributors.
28  * 4. Neither the name of the University nor the names of its contributors
29  *    may be used to endorse or promote products derived from this software
30  *    without specific prior written permission.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42  * SUCH DAMAGE.
43  *
44  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
45  *
46  *
47  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
48  * All rights reserved.
49  *
50  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
51  *
52  * Permission to use, copy, modify and distribute this software and
53  * its documentation is hereby granted, provided that both the copyright
54  * notice and this permission notice appear in all copies of the
55  * software, derivative works or modified versions, and any portions
56  * thereof, and that both notices appear in supporting documentation.
57  *
58  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
59  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
60  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
61  *
62  * Carnegie Mellon requests users of this software to return to
63  *
64  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
65  *  School of Computer Science
66  *  Carnegie Mellon University
67  *  Pittsburgh PA 15213-3890
68  *
69  * any improvements or extensions that they make and grant Carnegie the
70  * rights to redistribute these changes.
71  */
72 
73 /*
74  *	The proverbial page-out daemon.
75  */
76 
77 #include <sys/cdefs.h>
78 __FBSDID("$FreeBSD$");
79 
80 #include "opt_vm.h"
81 
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
85 #include <sys/eventhandler.h>
86 #include <sys/lock.h>
87 #include <sys/mutex.h>
88 #include <sys/proc.h>
89 #include <sys/kthread.h>
90 #include <sys/ktr.h>
91 #include <sys/mount.h>
92 #include <sys/racct.h>
93 #include <sys/resourcevar.h>
94 #include <sys/sched.h>
95 #include <sys/sdt.h>
96 #include <sys/signalvar.h>
97 #include <sys/smp.h>
98 #include <sys/time.h>
99 #include <sys/vnode.h>
100 #include <sys/vmmeter.h>
101 #include <sys/rwlock.h>
102 #include <sys/sx.h>
103 #include <sys/sysctl.h>
104 
105 #include <vm/vm.h>
106 #include <vm/vm_param.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_map.h>
110 #include <vm/vm_pageout.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_phys.h>
113 #include <vm/vm_pagequeue.h>
114 #include <vm/swap_pager.h>
115 #include <vm/vm_extern.h>
116 #include <vm/uma.h>
117 
118 /*
119  * System initialization
120  */
121 
122 /* the kernel process "vm_pageout"*/
123 static void vm_pageout(void);
124 static void vm_pageout_init(void);
125 static int vm_pageout_clean(vm_page_t m, int *numpagedout);
126 static int vm_pageout_cluster(vm_page_t m);
127 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
128     int starting_page_shortage);
129 
130 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
131     NULL);
132 
133 struct proc *pageproc;
134 
135 static struct kproc_desc page_kp = {
136 	"pagedaemon",
137 	vm_pageout,
138 	&pageproc
139 };
140 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
141     &page_kp);
142 
143 SDT_PROVIDER_DEFINE(vm);
144 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
145 
146 /* Pagedaemon activity rates, in subdivisions of one second. */
147 #define	VM_LAUNDER_RATE		10
148 #define	VM_INACT_SCAN_RATE	10
149 
150 static int vm_pageout_oom_seq = 12;
151 
152 static int vm_pageout_update_period;
153 static int disable_swap_pageouts;
154 static int lowmem_period = 10;
155 static int swapdev_enabled;
156 
157 static int vm_panic_on_oom = 0;
158 
159 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
160 	CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
161 	"Panic on the given number of out-of-memory errors instead of killing the largest process");
162 
163 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
164 	CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
165 	"Maximum active LRU update period");
166 
167 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
168 	"Low memory callback period");
169 
170 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
171 	CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
172 
173 static int pageout_lock_miss;
174 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
175 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
176 
177 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
178 	CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
179 	"back-to-back calls to oom detector to start OOM");
180 
181 static int act_scan_laundry_weight = 3;
182 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
183     &act_scan_laundry_weight, 0,
184     "weight given to clean vs. dirty pages in active queue scans");
185 
186 static u_int vm_background_launder_rate = 4096;
187 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
188     &vm_background_launder_rate, 0,
189     "background laundering rate, in kilobytes per second");
190 
191 static u_int vm_background_launder_max = 20 * 1024;
192 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
193     &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
194 
195 int vm_pageout_page_count = 32;
196 
197 u_long vm_page_max_user_wired;
198 SYSCTL_ULONG(_vm, OID_AUTO, max_user_wired, CTLFLAG_RW,
199     &vm_page_max_user_wired, 0,
200     "system-wide limit to user-wired page count");
201 
202 static u_int isqrt(u_int num);
203 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
204     bool in_shortfall);
205 static void vm_pageout_laundry_worker(void *arg);
206 
207 struct scan_state {
208 	struct vm_batchqueue bq;
209 	struct vm_pagequeue *pq;
210 	vm_page_t	marker;
211 	int		maxscan;
212 	int		scanned;
213 };
214 
215 static void
216 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
217     vm_page_t marker, vm_page_t after, int maxscan)
218 {
219 
220 	vm_pagequeue_assert_locked(pq);
221 	KASSERT((marker->a.flags & PGA_ENQUEUED) == 0,
222 	    ("marker %p already enqueued", marker));
223 
224 	if (after == NULL)
225 		TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
226 	else
227 		TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
228 	vm_page_aflag_set(marker, PGA_ENQUEUED);
229 
230 	vm_batchqueue_init(&ss->bq);
231 	ss->pq = pq;
232 	ss->marker = marker;
233 	ss->maxscan = maxscan;
234 	ss->scanned = 0;
235 	vm_pagequeue_unlock(pq);
236 }
237 
238 static void
239 vm_pageout_end_scan(struct scan_state *ss)
240 {
241 	struct vm_pagequeue *pq;
242 
243 	pq = ss->pq;
244 	vm_pagequeue_assert_locked(pq);
245 	KASSERT((ss->marker->a.flags & PGA_ENQUEUED) != 0,
246 	    ("marker %p not enqueued", ss->marker));
247 
248 	TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
249 	vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
250 	pq->pq_pdpages += ss->scanned;
251 }
252 
253 /*
254  * Add a small number of queued pages to a batch queue for later processing
255  * without the corresponding queue lock held.  The caller must have enqueued a
256  * marker page at the desired start point for the scan.  Pages will be
257  * physically dequeued if the caller so requests.  Otherwise, the returned
258  * batch may contain marker pages, and it is up to the caller to handle them.
259  *
260  * When processing the batch queue, vm_pageout_defer() must be used to
261  * determine whether the page has been logically dequeued since the batch was
262  * collected.
263  */
264 static __always_inline void
265 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
266 {
267 	struct vm_pagequeue *pq;
268 	vm_page_t m, marker, n;
269 
270 	marker = ss->marker;
271 	pq = ss->pq;
272 
273 	KASSERT((marker->a.flags & PGA_ENQUEUED) != 0,
274 	    ("marker %p not enqueued", ss->marker));
275 
276 	vm_pagequeue_lock(pq);
277 	for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
278 	    ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
279 	    m = n, ss->scanned++) {
280 		n = TAILQ_NEXT(m, plinks.q);
281 		if ((m->flags & PG_MARKER) == 0) {
282 			KASSERT((m->a.flags & PGA_ENQUEUED) != 0,
283 			    ("page %p not enqueued", m));
284 			KASSERT((m->flags & PG_FICTITIOUS) == 0,
285 			    ("Fictitious page %p cannot be in page queue", m));
286 			KASSERT((m->oflags & VPO_UNMANAGED) == 0,
287 			    ("Unmanaged page %p cannot be in page queue", m));
288 		} else if (dequeue)
289 			continue;
290 
291 		(void)vm_batchqueue_insert(&ss->bq, m);
292 		if (dequeue) {
293 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
294 			vm_page_aflag_clear(m, PGA_ENQUEUED);
295 		}
296 	}
297 	TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
298 	if (__predict_true(m != NULL))
299 		TAILQ_INSERT_BEFORE(m, marker, plinks.q);
300 	else
301 		TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
302 	if (dequeue)
303 		vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
304 	vm_pagequeue_unlock(pq);
305 }
306 
307 /*
308  * Return the next page to be scanned, or NULL if the scan is complete.
309  */
310 static __always_inline vm_page_t
311 vm_pageout_next(struct scan_state *ss, const bool dequeue)
312 {
313 
314 	if (ss->bq.bq_cnt == 0)
315 		vm_pageout_collect_batch(ss, dequeue);
316 	return (vm_batchqueue_pop(&ss->bq));
317 }
318 
319 /*
320  * Determine whether processing of a page should be deferred and ensure that any
321  * outstanding queue operations are processed.
322  */
323 static __always_inline bool
324 vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued)
325 {
326 	vm_page_astate_t as;
327 
328 	as = vm_page_astate_load(m);
329 	if (__predict_false(as.queue != queue ||
330 	    ((as.flags & PGA_ENQUEUED) != 0) != enqueued))
331 		return (true);
332 	if ((as.flags & PGA_QUEUE_OP_MASK) != 0) {
333 		vm_page_pqbatch_submit(m, queue);
334 		return (true);
335 	}
336 	return (false);
337 }
338 
339 /*
340  * Scan for pages at adjacent offsets within the given page's object that are
341  * eligible for laundering, form a cluster of these pages and the given page,
342  * and launder that cluster.
343  */
344 static int
345 vm_pageout_cluster(vm_page_t m)
346 {
347 	vm_object_t object;
348 	vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
349 	vm_pindex_t pindex;
350 	int ib, is, page_base, pageout_count;
351 
352 	object = m->object;
353 	VM_OBJECT_ASSERT_WLOCKED(object);
354 	pindex = m->pindex;
355 
356 	vm_page_assert_xbusied(m);
357 
358 	mc[vm_pageout_page_count] = pb = ps = m;
359 	pageout_count = 1;
360 	page_base = vm_pageout_page_count;
361 	ib = 1;
362 	is = 1;
363 
364 	/*
365 	 * We can cluster only if the page is not clean, busy, or held, and
366 	 * the page is in the laundry queue.
367 	 *
368 	 * During heavy mmap/modification loads the pageout
369 	 * daemon can really fragment the underlying file
370 	 * due to flushing pages out of order and not trying to
371 	 * align the clusters (which leaves sporadic out-of-order
372 	 * holes).  To solve this problem we do the reverse scan
373 	 * first and attempt to align our cluster, then do a
374 	 * forward scan if room remains.
375 	 */
376 more:
377 	while (ib != 0 && pageout_count < vm_pageout_page_count) {
378 		if (ib > pindex) {
379 			ib = 0;
380 			break;
381 		}
382 		if ((p = vm_page_prev(pb)) == NULL ||
383 		    vm_page_tryxbusy(p) == 0) {
384 			ib = 0;
385 			break;
386 		}
387 		if (vm_page_wired(p)) {
388 			ib = 0;
389 			vm_page_xunbusy(p);
390 			break;
391 		}
392 		vm_page_test_dirty(p);
393 		if (p->dirty == 0) {
394 			ib = 0;
395 			vm_page_xunbusy(p);
396 			break;
397 		}
398 		if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
399 			vm_page_xunbusy(p);
400 			ib = 0;
401 			break;
402 		}
403 		mc[--page_base] = pb = p;
404 		++pageout_count;
405 		++ib;
406 
407 		/*
408 		 * We are at an alignment boundary.  Stop here, and switch
409 		 * directions.  Do not clear ib.
410 		 */
411 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
412 			break;
413 	}
414 	while (pageout_count < vm_pageout_page_count &&
415 	    pindex + is < object->size) {
416 		if ((p = vm_page_next(ps)) == NULL ||
417 		    vm_page_tryxbusy(p) == 0)
418 			break;
419 		if (vm_page_wired(p)) {
420 			vm_page_xunbusy(p);
421 			break;
422 		}
423 		vm_page_test_dirty(p);
424 		if (p->dirty == 0) {
425 			vm_page_xunbusy(p);
426 			break;
427 		}
428 		if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
429 			vm_page_xunbusy(p);
430 			break;
431 		}
432 		mc[page_base + pageout_count] = ps = p;
433 		++pageout_count;
434 		++is;
435 	}
436 
437 	/*
438 	 * If we exhausted our forward scan, continue with the reverse scan
439 	 * when possible, even past an alignment boundary.  This catches
440 	 * boundary conditions.
441 	 */
442 	if (ib != 0 && pageout_count < vm_pageout_page_count)
443 		goto more;
444 
445 	return (vm_pageout_flush(&mc[page_base], pageout_count,
446 	    VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
447 }
448 
449 /*
450  * vm_pageout_flush() - launder the given pages
451  *
452  *	The given pages are laundered.  Note that we setup for the start of
453  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
454  *	reference count all in here rather then in the parent.  If we want
455  *	the parent to do more sophisticated things we may have to change
456  *	the ordering.
457  *
458  *	Returned runlen is the count of pages between mreq and first
459  *	page after mreq with status VM_PAGER_AGAIN.
460  *	*eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
461  *	for any page in runlen set.
462  */
463 int
464 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
465     boolean_t *eio)
466 {
467 	vm_object_t object = mc[0]->object;
468 	int pageout_status[count];
469 	int numpagedout = 0;
470 	int i, runlen;
471 
472 	VM_OBJECT_ASSERT_WLOCKED(object);
473 
474 	/*
475 	 * Initiate I/O.  Mark the pages shared busy and verify that they're
476 	 * valid and read-only.
477 	 *
478 	 * We do not have to fixup the clean/dirty bits here... we can
479 	 * allow the pager to do it after the I/O completes.
480 	 *
481 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
482 	 * edge case with file fragments.
483 	 */
484 	for (i = 0; i < count; i++) {
485 		KASSERT(vm_page_all_valid(mc[i]),
486 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
487 			mc[i], i, count));
488 		KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
489 		    ("vm_pageout_flush: writeable page %p", mc[i]));
490 		vm_page_busy_downgrade(mc[i]);
491 	}
492 	vm_object_pip_add(object, count);
493 
494 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
495 
496 	runlen = count - mreq;
497 	if (eio != NULL)
498 		*eio = FALSE;
499 	for (i = 0; i < count; i++) {
500 		vm_page_t mt = mc[i];
501 
502 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
503 		    !pmap_page_is_write_mapped(mt),
504 		    ("vm_pageout_flush: page %p is not write protected", mt));
505 		switch (pageout_status[i]) {
506 		case VM_PAGER_OK:
507 			/*
508 			 * The page may have moved since laundering started, in
509 			 * which case it should be left alone.
510 			 */
511 			if (vm_page_in_laundry(mt))
512 				vm_page_deactivate_noreuse(mt);
513 			/* FALLTHROUGH */
514 		case VM_PAGER_PEND:
515 			numpagedout++;
516 			break;
517 		case VM_PAGER_BAD:
518 			/*
519 			 * The page is outside the object's range.  We pretend
520 			 * that the page out worked and clean the page, so the
521 			 * changes will be lost if the page is reclaimed by
522 			 * the page daemon.
523 			 */
524 			vm_page_undirty(mt);
525 			if (vm_page_in_laundry(mt))
526 				vm_page_deactivate_noreuse(mt);
527 			break;
528 		case VM_PAGER_ERROR:
529 		case VM_PAGER_FAIL:
530 			/*
531 			 * If the page couldn't be paged out to swap because the
532 			 * pager wasn't able to find space, place the page in
533 			 * the PQ_UNSWAPPABLE holding queue.  This is an
534 			 * optimization that prevents the page daemon from
535 			 * wasting CPU cycles on pages that cannot be reclaimed
536 			 * becase no swap device is configured.
537 			 *
538 			 * Otherwise, reactivate the page so that it doesn't
539 			 * clog the laundry and inactive queues.  (We will try
540 			 * paging it out again later.)
541 			 */
542 			if (object->type == OBJT_SWAP &&
543 			    pageout_status[i] == VM_PAGER_FAIL) {
544 				vm_page_unswappable(mt);
545 				numpagedout++;
546 			} else
547 				vm_page_activate(mt);
548 			if (eio != NULL && i >= mreq && i - mreq < runlen)
549 				*eio = TRUE;
550 			break;
551 		case VM_PAGER_AGAIN:
552 			if (i >= mreq && i - mreq < runlen)
553 				runlen = i - mreq;
554 			break;
555 		}
556 
557 		/*
558 		 * If the operation is still going, leave the page busy to
559 		 * block all other accesses. Also, leave the paging in
560 		 * progress indicator set so that we don't attempt an object
561 		 * collapse.
562 		 */
563 		if (pageout_status[i] != VM_PAGER_PEND) {
564 			vm_object_pip_wakeup(object);
565 			vm_page_sunbusy(mt);
566 		}
567 	}
568 	if (prunlen != NULL)
569 		*prunlen = runlen;
570 	return (numpagedout);
571 }
572 
573 static void
574 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
575 {
576 
577 	atomic_store_rel_int(&swapdev_enabled, 1);
578 }
579 
580 static void
581 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
582 {
583 
584 	if (swap_pager_nswapdev() == 1)
585 		atomic_store_rel_int(&swapdev_enabled, 0);
586 }
587 
588 /*
589  * Attempt to acquire all of the necessary locks to launder a page and
590  * then call through the clustering layer to PUTPAGES.  Wait a short
591  * time for a vnode lock.
592  *
593  * Requires the page and object lock on entry, releases both before return.
594  * Returns 0 on success and an errno otherwise.
595  */
596 static int
597 vm_pageout_clean(vm_page_t m, int *numpagedout)
598 {
599 	struct vnode *vp;
600 	struct mount *mp;
601 	vm_object_t object;
602 	vm_pindex_t pindex;
603 	int error, lockmode;
604 
605 	object = m->object;
606 	VM_OBJECT_ASSERT_WLOCKED(object);
607 	error = 0;
608 	vp = NULL;
609 	mp = NULL;
610 
611 	/*
612 	 * The object is already known NOT to be dead.   It
613 	 * is possible for the vget() to block the whole
614 	 * pageout daemon, but the new low-memory handling
615 	 * code should prevent it.
616 	 *
617 	 * We can't wait forever for the vnode lock, we might
618 	 * deadlock due to a vn_read() getting stuck in
619 	 * vm_wait while holding this vnode.  We skip the
620 	 * vnode if we can't get it in a reasonable amount
621 	 * of time.
622 	 */
623 	if (object->type == OBJT_VNODE) {
624 		vm_page_xunbusy(m);
625 		vp = object->handle;
626 		if (vp->v_type == VREG &&
627 		    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
628 			mp = NULL;
629 			error = EDEADLK;
630 			goto unlock_all;
631 		}
632 		KASSERT(mp != NULL,
633 		    ("vp %p with NULL v_mount", vp));
634 		vm_object_reference_locked(object);
635 		pindex = m->pindex;
636 		VM_OBJECT_WUNLOCK(object);
637 		lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
638 		    LK_SHARED : LK_EXCLUSIVE;
639 		if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
640 			vp = NULL;
641 			error = EDEADLK;
642 			goto unlock_mp;
643 		}
644 		VM_OBJECT_WLOCK(object);
645 
646 		/*
647 		 * Ensure that the object and vnode were not disassociated
648 		 * while locks were dropped.
649 		 */
650 		if (vp->v_object != object) {
651 			error = ENOENT;
652 			goto unlock_all;
653 		}
654 
655 		/*
656 		 * While the object was unlocked, the page may have been:
657 		 * (1) moved to a different queue,
658 		 * (2) reallocated to a different object,
659 		 * (3) reallocated to a different offset, or
660 		 * (4) cleaned.
661 		 */
662 		if (!vm_page_in_laundry(m) || m->object != object ||
663 		    m->pindex != pindex || m->dirty == 0) {
664 			error = ENXIO;
665 			goto unlock_all;
666 		}
667 
668 		/*
669 		 * The page may have been busied while the object lock was
670 		 * released.
671 		 */
672 		if (vm_page_tryxbusy(m) == 0) {
673 			error = EBUSY;
674 			goto unlock_all;
675 		}
676 	}
677 
678 	/*
679 	 * Remove all writeable mappings, failing if the page is wired.
680 	 */
681 	if (!vm_page_try_remove_write(m)) {
682 		vm_page_xunbusy(m);
683 		error = EBUSY;
684 		goto unlock_all;
685 	}
686 
687 	/*
688 	 * If a page is dirty, then it is either being washed
689 	 * (but not yet cleaned) or it is still in the
690 	 * laundry.  If it is still in the laundry, then we
691 	 * start the cleaning operation.
692 	 */
693 	if ((*numpagedout = vm_pageout_cluster(m)) == 0)
694 		error = EIO;
695 
696 unlock_all:
697 	VM_OBJECT_WUNLOCK(object);
698 
699 unlock_mp:
700 	if (mp != NULL) {
701 		if (vp != NULL)
702 			vput(vp);
703 		vm_object_deallocate(object);
704 		vn_finished_write(mp);
705 	}
706 
707 	return (error);
708 }
709 
710 /*
711  * Attempt to launder the specified number of pages.
712  *
713  * Returns the number of pages successfully laundered.
714  */
715 static int
716 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
717 {
718 	struct scan_state ss;
719 	struct vm_pagequeue *pq;
720 	vm_object_t object;
721 	vm_page_t m, marker;
722 	vm_page_astate_t new, old;
723 	int act_delta, error, numpagedout, queue, refs, starting_target;
724 	int vnodes_skipped;
725 	bool pageout_ok;
726 
727 	object = NULL;
728 	starting_target = launder;
729 	vnodes_skipped = 0;
730 
731 	/*
732 	 * Scan the laundry queues for pages eligible to be laundered.  We stop
733 	 * once the target number of dirty pages have been laundered, or once
734 	 * we've reached the end of the queue.  A single iteration of this loop
735 	 * may cause more than one page to be laundered because of clustering.
736 	 *
737 	 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
738 	 * swap devices are configured.
739 	 */
740 	if (atomic_load_acq_int(&swapdev_enabled))
741 		queue = PQ_UNSWAPPABLE;
742 	else
743 		queue = PQ_LAUNDRY;
744 
745 scan:
746 	marker = &vmd->vmd_markers[queue];
747 	pq = &vmd->vmd_pagequeues[queue];
748 	vm_pagequeue_lock(pq);
749 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
750 	while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
751 		if (__predict_false((m->flags & PG_MARKER) != 0))
752 			continue;
753 
754 		/*
755 		 * Don't touch a page that was removed from the queue after the
756 		 * page queue lock was released.  Otherwise, ensure that any
757 		 * pending queue operations, such as dequeues for wired pages,
758 		 * are handled.
759 		 */
760 		if (vm_pageout_defer(m, queue, true))
761 			continue;
762 
763 		/*
764 		 * Lock the page's object.
765 		 */
766 		if (object == NULL || object != m->object) {
767 			if (object != NULL)
768 				VM_OBJECT_WUNLOCK(object);
769 			object = atomic_load_ptr(&m->object);
770 			if (__predict_false(object == NULL))
771 				/* The page is being freed by another thread. */
772 				continue;
773 
774 			/* Depends on type-stability. */
775 			VM_OBJECT_WLOCK(object);
776 			if (__predict_false(m->object != object)) {
777 				VM_OBJECT_WUNLOCK(object);
778 				object = NULL;
779 				continue;
780 			}
781 		}
782 
783 		if (vm_page_tryxbusy(m) == 0)
784 			continue;
785 
786 		/*
787 		 * Check for wirings now that we hold the object lock and have
788 		 * exclusively busied the page.  If the page is mapped, it may
789 		 * still be wired by pmap lookups.  The call to
790 		 * vm_page_try_remove_all() below atomically checks for such
791 		 * wirings and removes mappings.  If the page is unmapped, the
792 		 * wire count is guaranteed not to increase after this check.
793 		 */
794 		if (__predict_false(vm_page_wired(m)))
795 			goto skip_page;
796 
797 		/*
798 		 * Invalid pages can be easily freed.  They cannot be
799 		 * mapped; vm_page_free() asserts this.
800 		 */
801 		if (vm_page_none_valid(m))
802 			goto free_page;
803 
804 		refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
805 
806 		for (old = vm_page_astate_load(m);;) {
807 			/*
808 			 * Check to see if the page has been removed from the
809 			 * queue since the first such check.  Leave it alone if
810 			 * so, discarding any references collected by
811 			 * pmap_ts_referenced().
812 			 */
813 			if (__predict_false(_vm_page_queue(old) == PQ_NONE))
814 				goto skip_page;
815 
816 			new = old;
817 			act_delta = refs;
818 			if ((old.flags & PGA_REFERENCED) != 0) {
819 				new.flags &= ~PGA_REFERENCED;
820 				act_delta++;
821 			}
822 			if (act_delta == 0) {
823 				;
824 			} else if (object->ref_count != 0) {
825 				/*
826 				 * Increase the activation count if the page was
827 				 * referenced while in the laundry queue.  This
828 				 * makes it less likely that the page will be
829 				 * returned prematurely to the laundry queue.
830 				 */
831 				new.act_count += ACT_ADVANCE +
832 				    act_delta;
833 				if (new.act_count > ACT_MAX)
834 					new.act_count = ACT_MAX;
835 
836 				new.flags &= ~PGA_QUEUE_OP_MASK;
837 				new.flags |= PGA_REQUEUE;
838 				new.queue = PQ_ACTIVE;
839 				if (!vm_page_pqstate_commit(m, &old, new))
840 					continue;
841 
842 				/*
843 				 * If this was a background laundering, count
844 				 * activated pages towards our target.  The
845 				 * purpose of background laundering is to ensure
846 				 * that pages are eventually cycled through the
847 				 * laundry queue, and an activation is a valid
848 				 * way out.
849 				 */
850 				if (!in_shortfall)
851 					launder--;
852 				VM_CNT_INC(v_reactivated);
853 				goto skip_page;
854 			} else if ((object->flags & OBJ_DEAD) == 0) {
855 				new.flags |= PGA_REQUEUE;
856 				if (!vm_page_pqstate_commit(m, &old, new))
857 					continue;
858 				goto skip_page;
859 			}
860 			break;
861 		}
862 
863 		/*
864 		 * If the page appears to be clean at the machine-independent
865 		 * layer, then remove all of its mappings from the pmap in
866 		 * anticipation of freeing it.  If, however, any of the page's
867 		 * mappings allow write access, then the page may still be
868 		 * modified until the last of those mappings are removed.
869 		 */
870 		if (object->ref_count != 0) {
871 			vm_page_test_dirty(m);
872 			if (m->dirty == 0 && !vm_page_try_remove_all(m))
873 				goto skip_page;
874 		}
875 
876 		/*
877 		 * Clean pages are freed, and dirty pages are paged out unless
878 		 * they belong to a dead object.  Requeueing dirty pages from
879 		 * dead objects is pointless, as they are being paged out and
880 		 * freed by the thread that destroyed the object.
881 		 */
882 		if (m->dirty == 0) {
883 free_page:
884 			/*
885 			 * Now we are guaranteed that no other threads are
886 			 * manipulating the page, check for a last-second
887 			 * reference.
888 			 */
889 			if (vm_pageout_defer(m, queue, true))
890 				goto skip_page;
891 			vm_page_free(m);
892 			VM_CNT_INC(v_dfree);
893 		} else if ((object->flags & OBJ_DEAD) == 0) {
894 			if (object->type != OBJT_SWAP &&
895 			    object->type != OBJT_DEFAULT)
896 				pageout_ok = true;
897 			else if (disable_swap_pageouts)
898 				pageout_ok = false;
899 			else
900 				pageout_ok = true;
901 			if (!pageout_ok) {
902 				vm_page_launder(m);
903 				goto skip_page;
904 			}
905 
906 			/*
907 			 * Form a cluster with adjacent, dirty pages from the
908 			 * same object, and page out that entire cluster.
909 			 *
910 			 * The adjacent, dirty pages must also be in the
911 			 * laundry.  However, their mappings are not checked
912 			 * for new references.  Consequently, a recently
913 			 * referenced page may be paged out.  However, that
914 			 * page will not be prematurely reclaimed.  After page
915 			 * out, the page will be placed in the inactive queue,
916 			 * where any new references will be detected and the
917 			 * page reactivated.
918 			 */
919 			error = vm_pageout_clean(m, &numpagedout);
920 			if (error == 0) {
921 				launder -= numpagedout;
922 				ss.scanned += numpagedout;
923 			} else if (error == EDEADLK) {
924 				pageout_lock_miss++;
925 				vnodes_skipped++;
926 			}
927 			object = NULL;
928 		} else {
929 skip_page:
930 			vm_page_xunbusy(m);
931 		}
932 	}
933 	if (object != NULL) {
934 		VM_OBJECT_WUNLOCK(object);
935 		object = NULL;
936 	}
937 	vm_pagequeue_lock(pq);
938 	vm_pageout_end_scan(&ss);
939 	vm_pagequeue_unlock(pq);
940 
941 	if (launder > 0 && queue == PQ_UNSWAPPABLE) {
942 		queue = PQ_LAUNDRY;
943 		goto scan;
944 	}
945 
946 	/*
947 	 * Wakeup the sync daemon if we skipped a vnode in a writeable object
948 	 * and we didn't launder enough pages.
949 	 */
950 	if (vnodes_skipped > 0 && launder > 0)
951 		(void)speedup_syncer();
952 
953 	return (starting_target - launder);
954 }
955 
956 /*
957  * Compute the integer square root.
958  */
959 static u_int
960 isqrt(u_int num)
961 {
962 	u_int bit, root, tmp;
963 
964 	bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
965 	root = 0;
966 	while (bit != 0) {
967 		tmp = root + bit;
968 		root >>= 1;
969 		if (num >= tmp) {
970 			num -= tmp;
971 			root += bit;
972 		}
973 		bit >>= 2;
974 	}
975 	return (root);
976 }
977 
978 /*
979  * Perform the work of the laundry thread: periodically wake up and determine
980  * whether any pages need to be laundered.  If so, determine the number of pages
981  * that need to be laundered, and launder them.
982  */
983 static void
984 vm_pageout_laundry_worker(void *arg)
985 {
986 	struct vm_domain *vmd;
987 	struct vm_pagequeue *pq;
988 	uint64_t nclean, ndirty, nfreed;
989 	int domain, last_target, launder, shortfall, shortfall_cycle, target;
990 	bool in_shortfall;
991 
992 	domain = (uintptr_t)arg;
993 	vmd = VM_DOMAIN(domain);
994 	pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
995 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
996 
997 	shortfall = 0;
998 	in_shortfall = false;
999 	shortfall_cycle = 0;
1000 	last_target = target = 0;
1001 	nfreed = 0;
1002 
1003 	/*
1004 	 * Calls to these handlers are serialized by the swap syscall lock.
1005 	 */
1006 	(void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
1007 	    EVENTHANDLER_PRI_ANY);
1008 	(void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
1009 	    EVENTHANDLER_PRI_ANY);
1010 
1011 	/*
1012 	 * The pageout laundry worker is never done, so loop forever.
1013 	 */
1014 	for (;;) {
1015 		KASSERT(target >= 0, ("negative target %d", target));
1016 		KASSERT(shortfall_cycle >= 0,
1017 		    ("negative cycle %d", shortfall_cycle));
1018 		launder = 0;
1019 
1020 		/*
1021 		 * First determine whether we need to launder pages to meet a
1022 		 * shortage of free pages.
1023 		 */
1024 		if (shortfall > 0) {
1025 			in_shortfall = true;
1026 			shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1027 			target = shortfall;
1028 		} else if (!in_shortfall)
1029 			goto trybackground;
1030 		else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1031 			/*
1032 			 * We recently entered shortfall and began laundering
1033 			 * pages.  If we have completed that laundering run
1034 			 * (and we are no longer in shortfall) or we have met
1035 			 * our laundry target through other activity, then we
1036 			 * can stop laundering pages.
1037 			 */
1038 			in_shortfall = false;
1039 			target = 0;
1040 			goto trybackground;
1041 		}
1042 		launder = target / shortfall_cycle--;
1043 		goto dolaundry;
1044 
1045 		/*
1046 		 * There's no immediate need to launder any pages; see if we
1047 		 * meet the conditions to perform background laundering:
1048 		 *
1049 		 * 1. The ratio of dirty to clean inactive pages exceeds the
1050 		 *    background laundering threshold, or
1051 		 * 2. we haven't yet reached the target of the current
1052 		 *    background laundering run.
1053 		 *
1054 		 * The background laundering threshold is not a constant.
1055 		 * Instead, it is a slowly growing function of the number of
1056 		 * clean pages freed by the page daemon since the last
1057 		 * background laundering.  Thus, as the ratio of dirty to
1058 		 * clean inactive pages grows, the amount of memory pressure
1059 		 * required to trigger laundering decreases.  We ensure
1060 		 * that the threshold is non-zero after an inactive queue
1061 		 * scan, even if that scan failed to free a single clean page.
1062 		 */
1063 trybackground:
1064 		nclean = vmd->vmd_free_count +
1065 		    vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1066 		ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1067 		if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1068 		    vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1069 			target = vmd->vmd_background_launder_target;
1070 		}
1071 
1072 		/*
1073 		 * We have a non-zero background laundering target.  If we've
1074 		 * laundered up to our maximum without observing a page daemon
1075 		 * request, just stop.  This is a safety belt that ensures we
1076 		 * don't launder an excessive amount if memory pressure is low
1077 		 * and the ratio of dirty to clean pages is large.  Otherwise,
1078 		 * proceed at the background laundering rate.
1079 		 */
1080 		if (target > 0) {
1081 			if (nfreed > 0) {
1082 				nfreed = 0;
1083 				last_target = target;
1084 			} else if (last_target - target >=
1085 			    vm_background_launder_max * PAGE_SIZE / 1024) {
1086 				target = 0;
1087 			}
1088 			launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1089 			launder /= VM_LAUNDER_RATE;
1090 			if (launder > target)
1091 				launder = target;
1092 		}
1093 
1094 dolaundry:
1095 		if (launder > 0) {
1096 			/*
1097 			 * Because of I/O clustering, the number of laundered
1098 			 * pages could exceed "target" by the maximum size of
1099 			 * a cluster minus one.
1100 			 */
1101 			target -= min(vm_pageout_launder(vmd, launder,
1102 			    in_shortfall), target);
1103 			pause("laundp", hz / VM_LAUNDER_RATE);
1104 		}
1105 
1106 		/*
1107 		 * If we're not currently laundering pages and the page daemon
1108 		 * hasn't posted a new request, sleep until the page daemon
1109 		 * kicks us.
1110 		 */
1111 		vm_pagequeue_lock(pq);
1112 		if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1113 			(void)mtx_sleep(&vmd->vmd_laundry_request,
1114 			    vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1115 
1116 		/*
1117 		 * If the pagedaemon has indicated that it's in shortfall, start
1118 		 * a shortfall laundering unless we're already in the middle of
1119 		 * one.  This may preempt a background laundering.
1120 		 */
1121 		if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1122 		    (!in_shortfall || shortfall_cycle == 0)) {
1123 			shortfall = vm_laundry_target(vmd) +
1124 			    vmd->vmd_pageout_deficit;
1125 			target = 0;
1126 		} else
1127 			shortfall = 0;
1128 
1129 		if (target == 0)
1130 			vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1131 		nfreed += vmd->vmd_clean_pages_freed;
1132 		vmd->vmd_clean_pages_freed = 0;
1133 		vm_pagequeue_unlock(pq);
1134 	}
1135 }
1136 
1137 /*
1138  * Compute the number of pages we want to try to move from the
1139  * active queue to either the inactive or laundry queue.
1140  *
1141  * When scanning active pages during a shortage, we make clean pages
1142  * count more heavily towards the page shortage than dirty pages.
1143  * This is because dirty pages must be laundered before they can be
1144  * reused and thus have less utility when attempting to quickly
1145  * alleviate a free page shortage.  However, this weighting also
1146  * causes the scan to deactivate dirty pages more aggressively,
1147  * improving the effectiveness of clustering.
1148  */
1149 static int
1150 vm_pageout_active_target(struct vm_domain *vmd)
1151 {
1152 	int shortage;
1153 
1154 	shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1155 	    (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1156 	    vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1157 	shortage *= act_scan_laundry_weight;
1158 	return (shortage);
1159 }
1160 
1161 /*
1162  * Scan the active queue.  If there is no shortage of inactive pages, scan a
1163  * small portion of the queue in order to maintain quasi-LRU.
1164  */
1165 static void
1166 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1167 {
1168 	struct scan_state ss;
1169 	vm_object_t object;
1170 	vm_page_t m, marker;
1171 	struct vm_pagequeue *pq;
1172 	vm_page_astate_t old, new;
1173 	long min_scan;
1174 	int act_delta, max_scan, ps_delta, refs, scan_tick;
1175 	uint8_t nqueue;
1176 
1177 	marker = &vmd->vmd_markers[PQ_ACTIVE];
1178 	pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1179 	vm_pagequeue_lock(pq);
1180 
1181 	/*
1182 	 * If we're just idle polling attempt to visit every
1183 	 * active page within 'update_period' seconds.
1184 	 */
1185 	scan_tick = ticks;
1186 	if (vm_pageout_update_period != 0) {
1187 		min_scan = pq->pq_cnt;
1188 		min_scan *= scan_tick - vmd->vmd_last_active_scan;
1189 		min_scan /= hz * vm_pageout_update_period;
1190 	} else
1191 		min_scan = 0;
1192 	if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1193 		vmd->vmd_last_active_scan = scan_tick;
1194 
1195 	/*
1196 	 * Scan the active queue for pages that can be deactivated.  Update
1197 	 * the per-page activity counter and use it to identify deactivation
1198 	 * candidates.  Held pages may be deactivated.
1199 	 *
1200 	 * To avoid requeuing each page that remains in the active queue, we
1201 	 * implement the CLOCK algorithm.  To keep the implementation of the
1202 	 * enqueue operation consistent for all page queues, we use two hands,
1203 	 * represented by marker pages. Scans begin at the first hand, which
1204 	 * precedes the second hand in the queue.  When the two hands meet,
1205 	 * they are moved back to the head and tail of the queue, respectively,
1206 	 * and scanning resumes.
1207 	 */
1208 	max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1209 act_scan:
1210 	vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1211 	while ((m = vm_pageout_next(&ss, false)) != NULL) {
1212 		if (__predict_false(m == &vmd->vmd_clock[1])) {
1213 			vm_pagequeue_lock(pq);
1214 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1215 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1216 			TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1217 			    plinks.q);
1218 			TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1219 			    plinks.q);
1220 			max_scan -= ss.scanned;
1221 			vm_pageout_end_scan(&ss);
1222 			goto act_scan;
1223 		}
1224 		if (__predict_false((m->flags & PG_MARKER) != 0))
1225 			continue;
1226 
1227 		/*
1228 		 * Don't touch a page that was removed from the queue after the
1229 		 * page queue lock was released.  Otherwise, ensure that any
1230 		 * pending queue operations, such as dequeues for wired pages,
1231 		 * are handled.
1232 		 */
1233 		if (vm_pageout_defer(m, PQ_ACTIVE, true))
1234 			continue;
1235 
1236 		/*
1237 		 * A page's object pointer may be set to NULL before
1238 		 * the object lock is acquired.
1239 		 */
1240 		object = atomic_load_ptr(&m->object);
1241 		if (__predict_false(object == NULL))
1242 			/*
1243 			 * The page has been removed from its object.
1244 			 */
1245 			continue;
1246 
1247 		/* Deferred free of swap space. */
1248 		if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
1249 		    VM_OBJECT_TRYWLOCK(object)) {
1250 			if (m->object == object)
1251 				vm_pager_page_unswapped(m);
1252 			VM_OBJECT_WUNLOCK(object);
1253 		}
1254 
1255 		/*
1256 		 * Check to see "how much" the page has been used.
1257 		 *
1258 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1259 		 * that a reference from a concurrently destroyed mapping is
1260 		 * observed here and now.
1261 		 *
1262 		 * Perform an unsynchronized object ref count check.  While
1263 		 * the page lock ensures that the page is not reallocated to
1264 		 * another object, in particular, one with unmanaged mappings
1265 		 * that cannot support pmap_ts_referenced(), two races are,
1266 		 * nonetheless, possible:
1267 		 * 1) The count was transitioning to zero, but we saw a non-
1268 		 *    zero value.  pmap_ts_referenced() will return zero
1269 		 *    because the page is not mapped.
1270 		 * 2) The count was transitioning to one, but we saw zero.
1271 		 *    This race delays the detection of a new reference.  At
1272 		 *    worst, we will deactivate and reactivate the page.
1273 		 */
1274 		refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1275 
1276 		old = vm_page_astate_load(m);
1277 		do {
1278 			/*
1279 			 * Check to see if the page has been removed from the
1280 			 * queue since the first such check.  Leave it alone if
1281 			 * so, discarding any references collected by
1282 			 * pmap_ts_referenced().
1283 			 */
1284 			if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1285 				break;
1286 
1287 			/*
1288 			 * Advance or decay the act_count based on recent usage.
1289 			 */
1290 			new = old;
1291 			act_delta = refs;
1292 			if ((old.flags & PGA_REFERENCED) != 0) {
1293 				new.flags &= ~PGA_REFERENCED;
1294 				act_delta++;
1295 			}
1296 			if (act_delta != 0) {
1297 				new.act_count += ACT_ADVANCE + act_delta;
1298 				if (new.act_count > ACT_MAX)
1299 					new.act_count = ACT_MAX;
1300 			} else {
1301 				new.act_count -= min(new.act_count,
1302 				    ACT_DECLINE);
1303 			}
1304 
1305 			if (new.act_count > 0) {
1306 				/*
1307 				 * Adjust the activation count and keep the page
1308 				 * in the active queue.  The count might be left
1309 				 * unchanged if it is saturated.  The page may
1310 				 * have been moved to a different queue since we
1311 				 * started the scan, in which case we move it
1312 				 * back.
1313 				 */
1314 				ps_delta = 0;
1315 				if (old.queue != PQ_ACTIVE) {
1316 					new.flags &= ~PGA_QUEUE_OP_MASK;
1317 					new.flags |= PGA_REQUEUE;
1318 					new.queue = PQ_ACTIVE;
1319 				}
1320 			} else {
1321 				/*
1322 				 * When not short for inactive pages, let dirty
1323 				 * pages go through the inactive queue before
1324 				 * moving to the laundry queue.  This gives them
1325 				 * some extra time to be reactivated,
1326 				 * potentially avoiding an expensive pageout.
1327 				 * However, during a page shortage, the inactive
1328 				 * queue is necessarily small, and so dirty
1329 				 * pages would only spend a trivial amount of
1330 				 * time in the inactive queue.  Therefore, we
1331 				 * might as well place them directly in the
1332 				 * laundry queue to reduce queuing overhead.
1333 				 *
1334 				 * Calling vm_page_test_dirty() here would
1335 				 * require acquisition of the object's write
1336 				 * lock.  However, during a page shortage,
1337 				 * directing dirty pages into the laundry queue
1338 				 * is only an optimization and not a
1339 				 * requirement.  Therefore, we simply rely on
1340 				 * the opportunistic updates to the page's dirty
1341 				 * field by the pmap.
1342 				 */
1343 				if (page_shortage <= 0) {
1344 					nqueue = PQ_INACTIVE;
1345 					ps_delta = 0;
1346 				} else if (m->dirty == 0) {
1347 					nqueue = PQ_INACTIVE;
1348 					ps_delta = act_scan_laundry_weight;
1349 				} else {
1350 					nqueue = PQ_LAUNDRY;
1351 					ps_delta = 1;
1352 				}
1353 
1354 				new.flags &= ~PGA_QUEUE_OP_MASK;
1355 				new.flags |= PGA_REQUEUE;
1356 				new.queue = nqueue;
1357 			}
1358 		} while (!vm_page_pqstate_commit(m, &old, new));
1359 
1360 		page_shortage -= ps_delta;
1361 	}
1362 	vm_pagequeue_lock(pq);
1363 	TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1364 	TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1365 	vm_pageout_end_scan(&ss);
1366 	vm_pagequeue_unlock(pq);
1367 }
1368 
1369 static int
1370 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker,
1371     vm_page_t m)
1372 {
1373 	vm_page_astate_t as;
1374 
1375 	vm_pagequeue_assert_locked(pq);
1376 
1377 	as = vm_page_astate_load(m);
1378 	if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
1379 		return (0);
1380 	vm_page_aflag_set(m, PGA_ENQUEUED);
1381 	TAILQ_INSERT_BEFORE(marker, m, plinks.q);
1382 	return (1);
1383 }
1384 
1385 /*
1386  * Re-add stuck pages to the inactive queue.  We will examine them again
1387  * during the next scan.  If the queue state of a page has changed since
1388  * it was physically removed from the page queue in
1389  * vm_pageout_collect_batch(), don't do anything with that page.
1390  */
1391 static void
1392 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1393     vm_page_t m)
1394 {
1395 	struct vm_pagequeue *pq;
1396 	vm_page_t marker;
1397 	int delta;
1398 
1399 	delta = 0;
1400 	marker = ss->marker;
1401 	pq = ss->pq;
1402 
1403 	if (m != NULL) {
1404 		if (vm_batchqueue_insert(bq, m))
1405 			return;
1406 		vm_pagequeue_lock(pq);
1407 		delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1408 	} else
1409 		vm_pagequeue_lock(pq);
1410 	while ((m = vm_batchqueue_pop(bq)) != NULL)
1411 		delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1412 	vm_pagequeue_cnt_add(pq, delta);
1413 	vm_pagequeue_unlock(pq);
1414 	vm_batchqueue_init(bq);
1415 }
1416 
1417 /*
1418  * Attempt to reclaim the requested number of pages from the inactive queue.
1419  * Returns true if the shortage was addressed.
1420  */
1421 static int
1422 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1423     int *addl_shortage)
1424 {
1425 	struct scan_state ss;
1426 	struct vm_batchqueue rq;
1427 	vm_page_t m, marker;
1428 	struct vm_pagequeue *pq;
1429 	vm_object_t object;
1430 	vm_page_astate_t old, new;
1431 	int act_delta, addl_page_shortage, deficit, page_shortage, refs;
1432 	int starting_page_shortage;
1433 
1434 	/*
1435 	 * The addl_page_shortage is an estimate of the number of temporarily
1436 	 * stuck pages in the inactive queue.  In other words, the
1437 	 * number of pages from the inactive count that should be
1438 	 * discounted in setting the target for the active queue scan.
1439 	 */
1440 	addl_page_shortage = 0;
1441 
1442 	/*
1443 	 * vmd_pageout_deficit counts the number of pages requested in
1444 	 * allocations that failed because of a free page shortage.  We assume
1445 	 * that the allocations will be reattempted and thus include the deficit
1446 	 * in our scan target.
1447 	 */
1448 	deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1449 	starting_page_shortage = page_shortage = shortage + deficit;
1450 
1451 	object = NULL;
1452 	vm_batchqueue_init(&rq);
1453 
1454 	/*
1455 	 * Start scanning the inactive queue for pages that we can free.  The
1456 	 * scan will stop when we reach the target or we have scanned the
1457 	 * entire queue.  (Note that m->a.act_count is not used to make
1458 	 * decisions for the inactive queue, only for the active queue.)
1459 	 */
1460 	marker = &vmd->vmd_markers[PQ_INACTIVE];
1461 	pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1462 	vm_pagequeue_lock(pq);
1463 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1464 	while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1465 		KASSERT((m->flags & PG_MARKER) == 0,
1466 		    ("marker page %p was dequeued", m));
1467 
1468 		/*
1469 		 * Don't touch a page that was removed from the queue after the
1470 		 * page queue lock was released.  Otherwise, ensure that any
1471 		 * pending queue operations, such as dequeues for wired pages,
1472 		 * are handled.
1473 		 */
1474 		if (vm_pageout_defer(m, PQ_INACTIVE, false))
1475 			continue;
1476 
1477 		/*
1478 		 * Lock the page's object.
1479 		 */
1480 		if (object == NULL || object != m->object) {
1481 			if (object != NULL)
1482 				VM_OBJECT_WUNLOCK(object);
1483 			object = atomic_load_ptr(&m->object);
1484 			if (__predict_false(object == NULL))
1485 				/* The page is being freed by another thread. */
1486 				continue;
1487 
1488 			/* Depends on type-stability. */
1489 			VM_OBJECT_WLOCK(object);
1490 			if (__predict_false(m->object != object)) {
1491 				VM_OBJECT_WUNLOCK(object);
1492 				object = NULL;
1493 				goto reinsert;
1494 			}
1495 		}
1496 
1497 		if (vm_page_tryxbusy(m) == 0) {
1498 			/*
1499 			 * Don't mess with busy pages.  Leave them at
1500 			 * the front of the queue.  Most likely, they
1501 			 * are being paged out and will leave the
1502 			 * queue shortly after the scan finishes.  So,
1503 			 * they ought to be discounted from the
1504 			 * inactive count.
1505 			 */
1506 			addl_page_shortage++;
1507 			goto reinsert;
1508 		}
1509 
1510 		/* Deferred free of swap space. */
1511 		if ((m->a.flags & PGA_SWAP_FREE) != 0)
1512 			vm_pager_page_unswapped(m);
1513 
1514 		/*
1515 		 * Check for wirings now that we hold the object lock and have
1516 		 * exclusively busied the page.  If the page is mapped, it may
1517 		 * still be wired by pmap lookups.  The call to
1518 		 * vm_page_try_remove_all() below atomically checks for such
1519 		 * wirings and removes mappings.  If the page is unmapped, the
1520 		 * wire count is guaranteed not to increase after this check.
1521 		 */
1522 		if (__predict_false(vm_page_wired(m)))
1523 			goto skip_page;
1524 
1525 		/*
1526 		 * Invalid pages can be easily freed. They cannot be
1527 		 * mapped, vm_page_free() asserts this.
1528 		 */
1529 		if (vm_page_none_valid(m))
1530 			goto free_page;
1531 
1532 		refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1533 
1534 		for (old = vm_page_astate_load(m);;) {
1535 			/*
1536 			 * Check to see if the page has been removed from the
1537 			 * queue since the first such check.  Leave it alone if
1538 			 * so, discarding any references collected by
1539 			 * pmap_ts_referenced().
1540 			 */
1541 			if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1542 				goto skip_page;
1543 
1544 			new = old;
1545 			act_delta = refs;
1546 			if ((old.flags & PGA_REFERENCED) != 0) {
1547 				new.flags &= ~PGA_REFERENCED;
1548 				act_delta++;
1549 			}
1550 			if (act_delta == 0) {
1551 				;
1552 			} else if (object->ref_count != 0) {
1553 				/*
1554 				 * Increase the activation count if the
1555 				 * page was referenced while in the
1556 				 * inactive queue.  This makes it less
1557 				 * likely that the page will be returned
1558 				 * prematurely to the inactive queue.
1559 				 */
1560 				new.act_count += ACT_ADVANCE +
1561 				    act_delta;
1562 				if (new.act_count > ACT_MAX)
1563 					new.act_count = ACT_MAX;
1564 
1565 				new.flags &= ~PGA_QUEUE_OP_MASK;
1566 				new.flags |= PGA_REQUEUE;
1567 				new.queue = PQ_ACTIVE;
1568 				if (!vm_page_pqstate_commit(m, &old, new))
1569 					continue;
1570 
1571 				VM_CNT_INC(v_reactivated);
1572 				goto skip_page;
1573 			} else if ((object->flags & OBJ_DEAD) == 0) {
1574 				new.queue = PQ_INACTIVE;
1575 				new.flags |= PGA_REQUEUE;
1576 				if (!vm_page_pqstate_commit(m, &old, new))
1577 					continue;
1578 				goto skip_page;
1579 			}
1580 			break;
1581 		}
1582 
1583 		/*
1584 		 * If the page appears to be clean at the machine-independent
1585 		 * layer, then remove all of its mappings from the pmap in
1586 		 * anticipation of freeing it.  If, however, any of the page's
1587 		 * mappings allow write access, then the page may still be
1588 		 * modified until the last of those mappings are removed.
1589 		 */
1590 		if (object->ref_count != 0) {
1591 			vm_page_test_dirty(m);
1592 			if (m->dirty == 0 && !vm_page_try_remove_all(m))
1593 				goto skip_page;
1594 		}
1595 
1596 		/*
1597 		 * Clean pages can be freed, but dirty pages must be sent back
1598 		 * to the laundry, unless they belong to a dead object.
1599 		 * Requeueing dirty pages from dead objects is pointless, as
1600 		 * they are being paged out and freed by the thread that
1601 		 * destroyed the object.
1602 		 */
1603 		if (m->dirty == 0) {
1604 free_page:
1605 			/*
1606 			 * Now we are guaranteed that no other threads are
1607 			 * manipulating the page, check for a last-second
1608 			 * reference that would save it from doom.
1609 			 */
1610 			if (vm_pageout_defer(m, PQ_INACTIVE, false))
1611 				goto skip_page;
1612 
1613 			/*
1614 			 * Because we dequeued the page and have already checked
1615 			 * for pending dequeue and enqueue requests, we can
1616 			 * safely disassociate the page from the inactive queue
1617 			 * without holding the queue lock.
1618 			 */
1619 			m->a.queue = PQ_NONE;
1620 			vm_page_free(m);
1621 			page_shortage--;
1622 			continue;
1623 		}
1624 		if ((object->flags & OBJ_DEAD) == 0)
1625 			vm_page_launder(m);
1626 skip_page:
1627 		vm_page_xunbusy(m);
1628 		continue;
1629 reinsert:
1630 		vm_pageout_reinsert_inactive(&ss, &rq, m);
1631 	}
1632 	if (object != NULL)
1633 		VM_OBJECT_WUNLOCK(object);
1634 	vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1635 	vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1636 	vm_pagequeue_lock(pq);
1637 	vm_pageout_end_scan(&ss);
1638 	vm_pagequeue_unlock(pq);
1639 
1640 	VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1641 
1642 	/*
1643 	 * Wake up the laundry thread so that it can perform any needed
1644 	 * laundering.  If we didn't meet our target, we're in shortfall and
1645 	 * need to launder more aggressively.  If PQ_LAUNDRY is empty and no
1646 	 * swap devices are configured, the laundry thread has no work to do, so
1647 	 * don't bother waking it up.
1648 	 *
1649 	 * The laundry thread uses the number of inactive queue scans elapsed
1650 	 * since the last laundering to determine whether to launder again, so
1651 	 * keep count.
1652 	 */
1653 	if (starting_page_shortage > 0) {
1654 		pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1655 		vm_pagequeue_lock(pq);
1656 		if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1657 		    (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1658 			if (page_shortage > 0) {
1659 				vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1660 				VM_CNT_INC(v_pdshortfalls);
1661 			} else if (vmd->vmd_laundry_request !=
1662 			    VM_LAUNDRY_SHORTFALL)
1663 				vmd->vmd_laundry_request =
1664 				    VM_LAUNDRY_BACKGROUND;
1665 			wakeup(&vmd->vmd_laundry_request);
1666 		}
1667 		vmd->vmd_clean_pages_freed +=
1668 		    starting_page_shortage - page_shortage;
1669 		vm_pagequeue_unlock(pq);
1670 	}
1671 
1672 	/*
1673 	 * Wakeup the swapout daemon if we didn't free the targeted number of
1674 	 * pages.
1675 	 */
1676 	if (page_shortage > 0)
1677 		vm_swapout_run();
1678 
1679 	/*
1680 	 * If the inactive queue scan fails repeatedly to meet its
1681 	 * target, kill the largest process.
1682 	 */
1683 	vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1684 
1685 	/*
1686 	 * Reclaim pages by swapping out idle processes, if configured to do so.
1687 	 */
1688 	vm_swapout_run_idle();
1689 
1690 	/*
1691 	 * See the description of addl_page_shortage above.
1692 	 */
1693 	*addl_shortage = addl_page_shortage + deficit;
1694 
1695 	return (page_shortage <= 0);
1696 }
1697 
1698 static int vm_pageout_oom_vote;
1699 
1700 /*
1701  * The pagedaemon threads randlomly select one to perform the
1702  * OOM.  Trying to kill processes before all pagedaemons
1703  * failed to reach free target is premature.
1704  */
1705 static void
1706 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1707     int starting_page_shortage)
1708 {
1709 	int old_vote;
1710 
1711 	if (starting_page_shortage <= 0 || starting_page_shortage !=
1712 	    page_shortage)
1713 		vmd->vmd_oom_seq = 0;
1714 	else
1715 		vmd->vmd_oom_seq++;
1716 	if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1717 		if (vmd->vmd_oom) {
1718 			vmd->vmd_oom = FALSE;
1719 			atomic_subtract_int(&vm_pageout_oom_vote, 1);
1720 		}
1721 		return;
1722 	}
1723 
1724 	/*
1725 	 * Do not follow the call sequence until OOM condition is
1726 	 * cleared.
1727 	 */
1728 	vmd->vmd_oom_seq = 0;
1729 
1730 	if (vmd->vmd_oom)
1731 		return;
1732 
1733 	vmd->vmd_oom = TRUE;
1734 	old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1735 	if (old_vote != vm_ndomains - 1)
1736 		return;
1737 
1738 	/*
1739 	 * The current pagedaemon thread is the last in the quorum to
1740 	 * start OOM.  Initiate the selection and signaling of the
1741 	 * victim.
1742 	 */
1743 	vm_pageout_oom(VM_OOM_MEM);
1744 
1745 	/*
1746 	 * After one round of OOM terror, recall our vote.  On the
1747 	 * next pass, current pagedaemon would vote again if the low
1748 	 * memory condition is still there, due to vmd_oom being
1749 	 * false.
1750 	 */
1751 	vmd->vmd_oom = FALSE;
1752 	atomic_subtract_int(&vm_pageout_oom_vote, 1);
1753 }
1754 
1755 /*
1756  * The OOM killer is the page daemon's action of last resort when
1757  * memory allocation requests have been stalled for a prolonged period
1758  * of time because it cannot reclaim memory.  This function computes
1759  * the approximate number of physical pages that could be reclaimed if
1760  * the specified address space is destroyed.
1761  *
1762  * Private, anonymous memory owned by the address space is the
1763  * principal resource that we expect to recover after an OOM kill.
1764  * Since the physical pages mapped by the address space's COW entries
1765  * are typically shared pages, they are unlikely to be released and so
1766  * they are not counted.
1767  *
1768  * To get to the point where the page daemon runs the OOM killer, its
1769  * efforts to write-back vnode-backed pages may have stalled.  This
1770  * could be caused by a memory allocation deadlock in the write path
1771  * that might be resolved by an OOM kill.  Therefore, physical pages
1772  * belonging to vnode-backed objects are counted, because they might
1773  * be freed without being written out first if the address space holds
1774  * the last reference to an unlinked vnode.
1775  *
1776  * Similarly, physical pages belonging to OBJT_PHYS objects are
1777  * counted because the address space might hold the last reference to
1778  * the object.
1779  */
1780 static long
1781 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1782 {
1783 	vm_map_t map;
1784 	vm_map_entry_t entry;
1785 	vm_object_t obj;
1786 	long res;
1787 
1788 	map = &vmspace->vm_map;
1789 	KASSERT(!map->system_map, ("system map"));
1790 	sx_assert(&map->lock, SA_LOCKED);
1791 	res = 0;
1792 	VM_MAP_ENTRY_FOREACH(entry, map) {
1793 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1794 			continue;
1795 		obj = entry->object.vm_object;
1796 		if (obj == NULL)
1797 			continue;
1798 		if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1799 		    obj->ref_count != 1)
1800 			continue;
1801 		switch (obj->type) {
1802 		case OBJT_DEFAULT:
1803 		case OBJT_SWAP:
1804 		case OBJT_PHYS:
1805 		case OBJT_VNODE:
1806 			res += obj->resident_page_count;
1807 			break;
1808 		}
1809 	}
1810 	return (res);
1811 }
1812 
1813 static int vm_oom_ratelim_last;
1814 static int vm_oom_pf_secs = 10;
1815 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1816     "");
1817 static struct mtx vm_oom_ratelim_mtx;
1818 
1819 void
1820 vm_pageout_oom(int shortage)
1821 {
1822 	struct proc *p, *bigproc;
1823 	vm_offset_t size, bigsize;
1824 	struct thread *td;
1825 	struct vmspace *vm;
1826 	int now;
1827 	bool breakout;
1828 
1829 	/*
1830 	 * For OOM requests originating from vm_fault(), there is a high
1831 	 * chance that a single large process faults simultaneously in
1832 	 * several threads.  Also, on an active system running many
1833 	 * processes of middle-size, like buildworld, all of them
1834 	 * could fault almost simultaneously as well.
1835 	 *
1836 	 * To avoid killing too many processes, rate-limit OOMs
1837 	 * initiated by vm_fault() time-outs on the waits for free
1838 	 * pages.
1839 	 */
1840 	mtx_lock(&vm_oom_ratelim_mtx);
1841 	now = ticks;
1842 	if (shortage == VM_OOM_MEM_PF &&
1843 	    (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1844 		mtx_unlock(&vm_oom_ratelim_mtx);
1845 		return;
1846 	}
1847 	vm_oom_ratelim_last = now;
1848 	mtx_unlock(&vm_oom_ratelim_mtx);
1849 
1850 	/*
1851 	 * We keep the process bigproc locked once we find it to keep anyone
1852 	 * from messing with it; however, there is a possibility of
1853 	 * deadlock if process B is bigproc and one of its child processes
1854 	 * attempts to propagate a signal to B while we are waiting for A's
1855 	 * lock while walking this list.  To avoid this, we don't block on
1856 	 * the process lock but just skip a process if it is already locked.
1857 	 */
1858 	bigproc = NULL;
1859 	bigsize = 0;
1860 	sx_slock(&allproc_lock);
1861 	FOREACH_PROC_IN_SYSTEM(p) {
1862 		PROC_LOCK(p);
1863 
1864 		/*
1865 		 * If this is a system, protected or killed process, skip it.
1866 		 */
1867 		if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1868 		    P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1869 		    p->p_pid == 1 || P_KILLED(p) ||
1870 		    (p->p_pid < 48 && swap_pager_avail != 0)) {
1871 			PROC_UNLOCK(p);
1872 			continue;
1873 		}
1874 		/*
1875 		 * If the process is in a non-running type state,
1876 		 * don't touch it.  Check all the threads individually.
1877 		 */
1878 		breakout = false;
1879 		FOREACH_THREAD_IN_PROC(p, td) {
1880 			thread_lock(td);
1881 			if (!TD_ON_RUNQ(td) &&
1882 			    !TD_IS_RUNNING(td) &&
1883 			    !TD_IS_SLEEPING(td) &&
1884 			    !TD_IS_SUSPENDED(td) &&
1885 			    !TD_IS_SWAPPED(td)) {
1886 				thread_unlock(td);
1887 				breakout = true;
1888 				break;
1889 			}
1890 			thread_unlock(td);
1891 		}
1892 		if (breakout) {
1893 			PROC_UNLOCK(p);
1894 			continue;
1895 		}
1896 		/*
1897 		 * get the process size
1898 		 */
1899 		vm = vmspace_acquire_ref(p);
1900 		if (vm == NULL) {
1901 			PROC_UNLOCK(p);
1902 			continue;
1903 		}
1904 		_PHOLD_LITE(p);
1905 		PROC_UNLOCK(p);
1906 		sx_sunlock(&allproc_lock);
1907 		if (!vm_map_trylock_read(&vm->vm_map)) {
1908 			vmspace_free(vm);
1909 			sx_slock(&allproc_lock);
1910 			PRELE(p);
1911 			continue;
1912 		}
1913 		size = vmspace_swap_count(vm);
1914 		if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
1915 			size += vm_pageout_oom_pagecount(vm);
1916 		vm_map_unlock_read(&vm->vm_map);
1917 		vmspace_free(vm);
1918 		sx_slock(&allproc_lock);
1919 
1920 		/*
1921 		 * If this process is bigger than the biggest one,
1922 		 * remember it.
1923 		 */
1924 		if (size > bigsize) {
1925 			if (bigproc != NULL)
1926 				PRELE(bigproc);
1927 			bigproc = p;
1928 			bigsize = size;
1929 		} else {
1930 			PRELE(p);
1931 		}
1932 	}
1933 	sx_sunlock(&allproc_lock);
1934 	if (bigproc != NULL) {
1935 		if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0)
1936 			panic("out of swap space");
1937 		PROC_LOCK(bigproc);
1938 		killproc(bigproc, "out of swap space");
1939 		sched_nice(bigproc, PRIO_MIN);
1940 		_PRELE(bigproc);
1941 		PROC_UNLOCK(bigproc);
1942 	}
1943 }
1944 
1945 /*
1946  * Signal a free page shortage to subsystems that have registered an event
1947  * handler.  Reclaim memory from UMA in the event of a severe shortage.
1948  * Return true if the free page count should be re-evaluated.
1949  */
1950 static bool
1951 vm_pageout_lowmem(void)
1952 {
1953 	static int lowmem_ticks = 0;
1954 	int last;
1955 	bool ret;
1956 
1957 	ret = false;
1958 
1959 	last = atomic_load_int(&lowmem_ticks);
1960 	while ((u_int)(ticks - last) / hz >= lowmem_period) {
1961 		if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
1962 			continue;
1963 
1964 		/*
1965 		 * Decrease registered cache sizes.
1966 		 */
1967 		SDT_PROBE0(vm, , , vm__lowmem_scan);
1968 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1969 
1970 		/*
1971 		 * We do this explicitly after the caches have been
1972 		 * drained above.
1973 		 */
1974 		uma_reclaim(UMA_RECLAIM_TRIM);
1975 		ret = true;
1976 		break;
1977 	}
1978 
1979 	/*
1980 	 * Kick off an asynchronous reclaim of cached memory if one of the
1981 	 * page daemons is failing to keep up with demand.  Use the "severe"
1982 	 * threshold instead of "min" to ensure that we do not blow away the
1983 	 * caches if a subset of the NUMA domains are depleted by kernel memory
1984 	 * allocations; the domainset iterators automatically skip domains
1985 	 * below the "min" threshold on the first pass.
1986 	 *
1987 	 * UMA reclaim worker has its own rate-limiting mechanism, so don't
1988 	 * worry about kicking it too often.
1989 	 */
1990 	if (vm_page_count_severe())
1991 		uma_reclaim_wakeup();
1992 
1993 	return (ret);
1994 }
1995 
1996 static void
1997 vm_pageout_worker(void *arg)
1998 {
1999 	struct vm_domain *vmd;
2000 	u_int ofree;
2001 	int addl_shortage, domain, shortage;
2002 	bool target_met;
2003 
2004 	domain = (uintptr_t)arg;
2005 	vmd = VM_DOMAIN(domain);
2006 	shortage = 0;
2007 	target_met = true;
2008 
2009 	/*
2010 	 * XXXKIB It could be useful to bind pageout daemon threads to
2011 	 * the cores belonging to the domain, from which vm_page_array
2012 	 * is allocated.
2013 	 */
2014 
2015 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
2016 	vmd->vmd_last_active_scan = ticks;
2017 
2018 	/*
2019 	 * The pageout daemon worker is never done, so loop forever.
2020 	 */
2021 	while (TRUE) {
2022 		vm_domain_pageout_lock(vmd);
2023 
2024 		/*
2025 		 * We need to clear wanted before we check the limits.  This
2026 		 * prevents races with wakers who will check wanted after they
2027 		 * reach the limit.
2028 		 */
2029 		atomic_store_int(&vmd->vmd_pageout_wanted, 0);
2030 
2031 		/*
2032 		 * Might the page daemon need to run again?
2033 		 */
2034 		if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
2035 			/*
2036 			 * Yes.  If the scan failed to produce enough free
2037 			 * pages, sleep uninterruptibly for some time in the
2038 			 * hope that the laundry thread will clean some pages.
2039 			 */
2040 			vm_domain_pageout_unlock(vmd);
2041 			if (!target_met)
2042 				pause("pwait", hz / VM_INACT_SCAN_RATE);
2043 		} else {
2044 			/*
2045 			 * No, sleep until the next wakeup or until pages
2046 			 * need to have their reference stats updated.
2047 			 */
2048 			if (mtx_sleep(&vmd->vmd_pageout_wanted,
2049 			    vm_domain_pageout_lockptr(vmd), PDROP | PVM,
2050 			    "psleep", hz / VM_INACT_SCAN_RATE) == 0)
2051 				VM_CNT_INC(v_pdwakeups);
2052 		}
2053 
2054 		/* Prevent spurious wakeups by ensuring that wanted is set. */
2055 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2056 
2057 		/*
2058 		 * Use the controller to calculate how many pages to free in
2059 		 * this interval, and scan the inactive queue.  If the lowmem
2060 		 * handlers appear to have freed up some pages, subtract the
2061 		 * difference from the inactive queue scan target.
2062 		 */
2063 		shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
2064 		if (shortage > 0) {
2065 			ofree = vmd->vmd_free_count;
2066 			if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
2067 				shortage -= min(vmd->vmd_free_count - ofree,
2068 				    (u_int)shortage);
2069 			target_met = vm_pageout_scan_inactive(vmd, shortage,
2070 			    &addl_shortage);
2071 		} else
2072 			addl_shortage = 0;
2073 
2074 		/*
2075 		 * Scan the active queue.  A positive value for shortage
2076 		 * indicates that we must aggressively deactivate pages to avoid
2077 		 * a shortfall.
2078 		 */
2079 		shortage = vm_pageout_active_target(vmd) + addl_shortage;
2080 		vm_pageout_scan_active(vmd, shortage);
2081 	}
2082 }
2083 
2084 /*
2085  * Initialize basic pageout daemon settings.  See the comment above the
2086  * definition of vm_domain for some explanation of how these thresholds are
2087  * used.
2088  */
2089 static void
2090 vm_pageout_init_domain(int domain)
2091 {
2092 	struct vm_domain *vmd;
2093 	struct sysctl_oid *oid;
2094 
2095 	vmd = VM_DOMAIN(domain);
2096 	vmd->vmd_interrupt_free_min = 2;
2097 
2098 	/*
2099 	 * v_free_reserved needs to include enough for the largest
2100 	 * swap pager structures plus enough for any pv_entry structs
2101 	 * when paging.
2102 	 */
2103 	vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE +
2104 	    vmd->vmd_interrupt_free_min;
2105 	vmd->vmd_free_reserved = vm_pageout_page_count +
2106 	    vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768;
2107 	vmd->vmd_free_min = vmd->vmd_page_count / 200;
2108 	vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2109 	vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2110 	vmd->vmd_free_min += vmd->vmd_free_reserved;
2111 	vmd->vmd_free_severe += vmd->vmd_free_reserved;
2112 	vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2113 	if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2114 		vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2115 
2116 	/*
2117 	 * Set the default wakeup threshold to be 10% below the paging
2118 	 * target.  This keeps the steady state out of shortfall.
2119 	 */
2120 	vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2121 
2122 	/*
2123 	 * Target amount of memory to move out of the laundry queue during a
2124 	 * background laundering.  This is proportional to the amount of system
2125 	 * memory.
2126 	 */
2127 	vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2128 	    vmd->vmd_free_min) / 10;
2129 
2130 	/* Initialize the pageout daemon pid controller. */
2131 	pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2132 	    vmd->vmd_free_target, PIDCTRL_BOUND,
2133 	    PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2134 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2135 	    "pidctrl", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "");
2136 	pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2137 }
2138 
2139 static void
2140 vm_pageout_init(void)
2141 {
2142 	u_int freecount;
2143 	int i;
2144 
2145 	/*
2146 	 * Initialize some paging parameters.
2147 	 */
2148 	if (vm_cnt.v_page_count < 2000)
2149 		vm_pageout_page_count = 8;
2150 
2151 	freecount = 0;
2152 	for (i = 0; i < vm_ndomains; i++) {
2153 		struct vm_domain *vmd;
2154 
2155 		vm_pageout_init_domain(i);
2156 		vmd = VM_DOMAIN(i);
2157 		vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2158 		vm_cnt.v_free_target += vmd->vmd_free_target;
2159 		vm_cnt.v_free_min += vmd->vmd_free_min;
2160 		vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2161 		vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2162 		vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2163 		vm_cnt.v_free_severe += vmd->vmd_free_severe;
2164 		freecount += vmd->vmd_free_count;
2165 	}
2166 
2167 	/*
2168 	 * Set interval in seconds for active scan.  We want to visit each
2169 	 * page at least once every ten minutes.  This is to prevent worst
2170 	 * case paging behaviors with stale active LRU.
2171 	 */
2172 	if (vm_pageout_update_period == 0)
2173 		vm_pageout_update_period = 600;
2174 
2175 	if (vm_page_max_user_wired == 0)
2176 		vm_page_max_user_wired = freecount / 3;
2177 }
2178 
2179 /*
2180  *     vm_pageout is the high level pageout daemon.
2181  */
2182 static void
2183 vm_pageout(void)
2184 {
2185 	struct proc *p;
2186 	struct thread *td;
2187 	int error, first, i;
2188 
2189 	p = curproc;
2190 	td = curthread;
2191 
2192 	mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2193 	swap_pager_swap_init();
2194 	for (first = -1, i = 0; i < vm_ndomains; i++) {
2195 		if (VM_DOMAIN_EMPTY(i)) {
2196 			if (bootverbose)
2197 				printf("domain %d empty; skipping pageout\n",
2198 				    i);
2199 			continue;
2200 		}
2201 		if (first == -1)
2202 			first = i;
2203 		else {
2204 			error = kthread_add(vm_pageout_worker,
2205 			    (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2206 			if (error != 0)
2207 				panic("starting pageout for domain %d: %d\n",
2208 				    i, error);
2209 		}
2210 		error = kthread_add(vm_pageout_laundry_worker,
2211 		    (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2212 		if (error != 0)
2213 			panic("starting laundry for domain %d: %d", i, error);
2214 	}
2215 	error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2216 	if (error != 0)
2217 		panic("starting uma_reclaim helper, error %d\n", error);
2218 
2219 	snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2220 	vm_pageout_worker((void *)(uintptr_t)first);
2221 }
2222 
2223 /*
2224  * Perform an advisory wakeup of the page daemon.
2225  */
2226 void
2227 pagedaemon_wakeup(int domain)
2228 {
2229 	struct vm_domain *vmd;
2230 
2231 	vmd = VM_DOMAIN(domain);
2232 	vm_domain_pageout_assert_unlocked(vmd);
2233 	if (curproc == pageproc)
2234 		return;
2235 
2236 	if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2237 		vm_domain_pageout_lock(vmd);
2238 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2239 		wakeup(&vmd->vmd_pageout_wanted);
2240 		vm_domain_pageout_unlock(vmd);
2241 	}
2242 }
2243