xref: /freebsd/sys/vm/vm_pageout.c (revision 6966ac055c3b7a39266fb982493330df7a097997)
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 out of memory 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_page_queue() must be used to
261  * determine whether the page has been logically dequeued by another thread.
262  * Once this check is performed, the page lock guarantees that the page will
263  * not be disassociated from the queue.
264  */
265 static __always_inline void
266 vm_pageout_collect_batch(struct scan_state *ss, const bool dequeue)
267 {
268 	struct vm_pagequeue *pq;
269 	vm_page_t m, marker, n;
270 
271 	marker = ss->marker;
272 	pq = ss->pq;
273 
274 	KASSERT((marker->a.flags & PGA_ENQUEUED) != 0,
275 	    ("marker %p not enqueued", ss->marker));
276 
277 	vm_pagequeue_lock(pq);
278 	for (m = TAILQ_NEXT(marker, plinks.q); m != NULL &&
279 	    ss->scanned < ss->maxscan && ss->bq.bq_cnt < VM_BATCHQUEUE_SIZE;
280 	    m = n, ss->scanned++) {
281 		n = TAILQ_NEXT(m, plinks.q);
282 		if ((m->flags & PG_MARKER) == 0) {
283 			KASSERT((m->a.flags & PGA_ENQUEUED) != 0,
284 			    ("page %p not enqueued", m));
285 			KASSERT((m->flags & PG_FICTITIOUS) == 0,
286 			    ("Fictitious page %p cannot be in page queue", m));
287 			KASSERT((m->oflags & VPO_UNMANAGED) == 0,
288 			    ("Unmanaged page %p cannot be in page queue", m));
289 		} else if (dequeue)
290 			continue;
291 
292 		(void)vm_batchqueue_insert(&ss->bq, m);
293 		if (dequeue) {
294 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
295 			vm_page_aflag_clear(m, PGA_ENQUEUED);
296 		}
297 	}
298 	TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
299 	if (__predict_true(m != NULL))
300 		TAILQ_INSERT_BEFORE(m, marker, plinks.q);
301 	else
302 		TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
303 	if (dequeue)
304 		vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
305 	vm_pagequeue_unlock(pq);
306 }
307 
308 /*
309  * Return the next page to be scanned, or NULL if the scan is complete.
310  */
311 static __always_inline vm_page_t
312 vm_pageout_next(struct scan_state *ss, const bool dequeue)
313 {
314 
315 	if (ss->bq.bq_cnt == 0)
316 		vm_pageout_collect_batch(ss, dequeue);
317 	return (vm_batchqueue_pop(&ss->bq));
318 }
319 
320 /*
321  * Determine whether processing of a page should be deferred and ensure that any
322  * outstanding queue operations are processed.
323  */
324 static __always_inline bool
325 vm_pageout_defer(vm_page_t m, const uint8_t queue, const bool enqueued)
326 {
327 	vm_page_astate_t as;
328 
329 	as = vm_page_astate_load(m);
330 	if (__predict_false(as.queue != queue ||
331 	    ((as.flags & PGA_ENQUEUED) != 0) != enqueued))
332 		return (true);
333 	if ((as.flags & PGA_QUEUE_OP_MASK) != 0) {
334 		vm_page_pqbatch_submit(m, queue);
335 		return (true);
336 	}
337 	return (false);
338 }
339 
340 /*
341  * Scan for pages at adjacent offsets within the given page's object that are
342  * eligible for laundering, form a cluster of these pages and the given page,
343  * and launder that cluster.
344  */
345 static int
346 vm_pageout_cluster(vm_page_t m)
347 {
348 	vm_object_t object;
349 	vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
350 	vm_pindex_t pindex;
351 	int ib, is, page_base, pageout_count;
352 
353 	object = m->object;
354 	VM_OBJECT_ASSERT_WLOCKED(object);
355 	pindex = m->pindex;
356 
357 	vm_page_assert_xbusied(m);
358 
359 	mc[vm_pageout_page_count] = pb = ps = m;
360 	pageout_count = 1;
361 	page_base = vm_pageout_page_count;
362 	ib = 1;
363 	is = 1;
364 
365 	/*
366 	 * We can cluster only if the page is not clean, busy, or held, and
367 	 * the page is in the laundry queue.
368 	 *
369 	 * During heavy mmap/modification loads the pageout
370 	 * daemon can really fragment the underlying file
371 	 * due to flushing pages out of order and not trying to
372 	 * align the clusters (which leaves sporadic out-of-order
373 	 * holes).  To solve this problem we do the reverse scan
374 	 * first and attempt to align our cluster, then do a
375 	 * forward scan if room remains.
376 	 */
377 more:
378 	while (ib != 0 && pageout_count < vm_pageout_page_count) {
379 		if (ib > pindex) {
380 			ib = 0;
381 			break;
382 		}
383 		if ((p = vm_page_prev(pb)) == NULL ||
384 		    vm_page_tryxbusy(p) == 0) {
385 			ib = 0;
386 			break;
387 		}
388 		if (vm_page_wired(p)) {
389 			ib = 0;
390 			vm_page_xunbusy(p);
391 			break;
392 		}
393 		vm_page_test_dirty(p);
394 		if (p->dirty == 0) {
395 			ib = 0;
396 			vm_page_xunbusy(p);
397 			break;
398 		}
399 		if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
400 			vm_page_xunbusy(p);
401 			ib = 0;
402 			break;
403 		}
404 		mc[--page_base] = pb = p;
405 		++pageout_count;
406 		++ib;
407 
408 		/*
409 		 * We are at an alignment boundary.  Stop here, and switch
410 		 * directions.  Do not clear ib.
411 		 */
412 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
413 			break;
414 	}
415 	while (pageout_count < vm_pageout_page_count &&
416 	    pindex + is < object->size) {
417 		if ((p = vm_page_next(ps)) == NULL ||
418 		    vm_page_tryxbusy(p) == 0)
419 			break;
420 		if (vm_page_wired(p)) {
421 			vm_page_xunbusy(p);
422 			break;
423 		}
424 		vm_page_test_dirty(p);
425 		if (p->dirty == 0) {
426 			vm_page_xunbusy(p);
427 			break;
428 		}
429 		if (!vm_page_in_laundry(p) || !vm_page_try_remove_write(p)) {
430 			vm_page_xunbusy(p);
431 			break;
432 		}
433 		mc[page_base + pageout_count] = ps = p;
434 		++pageout_count;
435 		++is;
436 	}
437 
438 	/*
439 	 * If we exhausted our forward scan, continue with the reverse scan
440 	 * when possible, even past an alignment boundary.  This catches
441 	 * boundary conditions.
442 	 */
443 	if (ib != 0 && pageout_count < vm_pageout_page_count)
444 		goto more;
445 
446 	return (vm_pageout_flush(&mc[page_base], pageout_count,
447 	    VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
448 }
449 
450 /*
451  * vm_pageout_flush() - launder the given pages
452  *
453  *	The given pages are laundered.  Note that we setup for the start of
454  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
455  *	reference count all in here rather then in the parent.  If we want
456  *	the parent to do more sophisticated things we may have to change
457  *	the ordering.
458  *
459  *	Returned runlen is the count of pages between mreq and first
460  *	page after mreq with status VM_PAGER_AGAIN.
461  *	*eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
462  *	for any page in runlen set.
463  */
464 int
465 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
466     boolean_t *eio)
467 {
468 	vm_object_t object = mc[0]->object;
469 	int pageout_status[count];
470 	int numpagedout = 0;
471 	int i, runlen;
472 
473 	VM_OBJECT_ASSERT_WLOCKED(object);
474 
475 	/*
476 	 * Initiate I/O.  Mark the pages shared busy and verify that they're
477 	 * valid and read-only.
478 	 *
479 	 * We do not have to fixup the clean/dirty bits here... we can
480 	 * allow the pager to do it after the I/O completes.
481 	 *
482 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
483 	 * edge case with file fragments.
484 	 */
485 	for (i = 0; i < count; i++) {
486 		KASSERT(vm_page_all_valid(mc[i]),
487 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
488 			mc[i], i, count));
489 		KASSERT((mc[i]->a.flags & PGA_WRITEABLE) == 0,
490 		    ("vm_pageout_flush: writeable page %p", mc[i]));
491 		vm_page_busy_downgrade(mc[i]);
492 	}
493 	vm_object_pip_add(object, count);
494 
495 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
496 
497 	runlen = count - mreq;
498 	if (eio != NULL)
499 		*eio = FALSE;
500 	for (i = 0; i < count; i++) {
501 		vm_page_t mt = mc[i];
502 
503 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
504 		    !pmap_page_is_write_mapped(mt),
505 		    ("vm_pageout_flush: page %p is not write protected", mt));
506 		switch (pageout_status[i]) {
507 		case VM_PAGER_OK:
508 			/*
509 			 * The page may have moved since laundering started, in
510 			 * which case it should be left alone.
511 			 */
512 			if (vm_page_in_laundry(mt))
513 				vm_page_deactivate_noreuse(mt);
514 			/* FALLTHROUGH */
515 		case VM_PAGER_PEND:
516 			numpagedout++;
517 			break;
518 		case VM_PAGER_BAD:
519 			/*
520 			 * The page is outside the object's range.  We pretend
521 			 * that the page out worked and clean the page, so the
522 			 * changes will be lost if the page is reclaimed by
523 			 * the page daemon.
524 			 */
525 			vm_page_undirty(mt);
526 			if (vm_page_in_laundry(mt))
527 				vm_page_deactivate_noreuse(mt);
528 			break;
529 		case VM_PAGER_ERROR:
530 		case VM_PAGER_FAIL:
531 			/*
532 			 * If the page couldn't be paged out to swap because the
533 			 * pager wasn't able to find space, place the page in
534 			 * the PQ_UNSWAPPABLE holding queue.  This is an
535 			 * optimization that prevents the page daemon from
536 			 * wasting CPU cycles on pages that cannot be reclaimed
537 			 * becase no swap device is configured.
538 			 *
539 			 * Otherwise, reactivate the page so that it doesn't
540 			 * clog the laundry and inactive queues.  (We will try
541 			 * paging it out again later.)
542 			 */
543 			if (object->type == OBJT_SWAP &&
544 			    pageout_status[i] == VM_PAGER_FAIL) {
545 				vm_page_unswappable(mt);
546 				numpagedout++;
547 			} else
548 				vm_page_activate(mt);
549 			if (eio != NULL && i >= mreq && i - mreq < runlen)
550 				*eio = TRUE;
551 			break;
552 		case VM_PAGER_AGAIN:
553 			if (i >= mreq && i - mreq < runlen)
554 				runlen = i - mreq;
555 			break;
556 		}
557 
558 		/*
559 		 * If the operation is still going, leave the page busy to
560 		 * block all other accesses. Also, leave the paging in
561 		 * progress indicator set so that we don't attempt an object
562 		 * collapse.
563 		 */
564 		if (pageout_status[i] != VM_PAGER_PEND) {
565 			vm_object_pip_wakeup(object);
566 			vm_page_sunbusy(mt);
567 		}
568 	}
569 	if (prunlen != NULL)
570 		*prunlen = runlen;
571 	return (numpagedout);
572 }
573 
574 static void
575 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
576 {
577 
578 	atomic_store_rel_int(&swapdev_enabled, 1);
579 }
580 
581 static void
582 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
583 {
584 
585 	if (swap_pager_nswapdev() == 1)
586 		atomic_store_rel_int(&swapdev_enabled, 0);
587 }
588 
589 /*
590  * Attempt to acquire all of the necessary locks to launder a page and
591  * then call through the clustering layer to PUTPAGES.  Wait a short
592  * time for a vnode lock.
593  *
594  * Requires the page and object lock on entry, releases both before return.
595  * Returns 0 on success and an errno otherwise.
596  */
597 static int
598 vm_pageout_clean(vm_page_t m, int *numpagedout)
599 {
600 	struct vnode *vp;
601 	struct mount *mp;
602 	vm_object_t object;
603 	vm_pindex_t pindex;
604 	int error, lockmode;
605 
606 	object = m->object;
607 	VM_OBJECT_ASSERT_WLOCKED(object);
608 	error = 0;
609 	vp = NULL;
610 	mp = NULL;
611 
612 	/*
613 	 * The object is already known NOT to be dead.   It
614 	 * is possible for the vget() to block the whole
615 	 * pageout daemon, but the new low-memory handling
616 	 * code should prevent it.
617 	 *
618 	 * We can't wait forever for the vnode lock, we might
619 	 * deadlock due to a vn_read() getting stuck in
620 	 * vm_wait while holding this vnode.  We skip the
621 	 * vnode if we can't get it in a reasonable amount
622 	 * of time.
623 	 */
624 	if (object->type == OBJT_VNODE) {
625 		vm_page_xunbusy(m);
626 		vp = object->handle;
627 		if (vp->v_type == VREG &&
628 		    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
629 			mp = NULL;
630 			error = EDEADLK;
631 			goto unlock_all;
632 		}
633 		KASSERT(mp != NULL,
634 		    ("vp %p with NULL v_mount", vp));
635 		vm_object_reference_locked(object);
636 		pindex = m->pindex;
637 		VM_OBJECT_WUNLOCK(object);
638 		lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
639 		    LK_SHARED : LK_EXCLUSIVE;
640 		if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
641 			vp = NULL;
642 			error = EDEADLK;
643 			goto unlock_mp;
644 		}
645 		VM_OBJECT_WLOCK(object);
646 
647 		/*
648 		 * Ensure that the object and vnode were not disassociated
649 		 * while locks were dropped.
650 		 */
651 		if (vp->v_object != object) {
652 			error = ENOENT;
653 			goto unlock_all;
654 		}
655 
656 		/*
657 		 * While the object was unlocked, the page may have been:
658 		 * (1) moved to a different queue,
659 		 * (2) reallocated to a different object,
660 		 * (3) reallocated to a different offset, or
661 		 * (4) cleaned.
662 		 */
663 		if (!vm_page_in_laundry(m) || m->object != object ||
664 		    m->pindex != pindex || m->dirty == 0) {
665 			error = ENXIO;
666 			goto unlock_all;
667 		}
668 
669 		/*
670 		 * The page may have been busied while the object lock was
671 		 * released.
672 		 */
673 		if (vm_page_tryxbusy(m) == 0) {
674 			error = EBUSY;
675 			goto unlock_all;
676 		}
677 	}
678 
679 	/*
680 	 * Remove all writeable mappings, failing if the page is wired.
681 	 */
682 	if (!vm_page_try_remove_write(m)) {
683 		vm_page_xunbusy(m);
684 		error = EBUSY;
685 		goto unlock_all;
686 	}
687 
688 	/*
689 	 * If a page is dirty, then it is either being washed
690 	 * (but not yet cleaned) or it is still in the
691 	 * laundry.  If it is still in the laundry, then we
692 	 * start the cleaning operation.
693 	 */
694 	if ((*numpagedout = vm_pageout_cluster(m)) == 0)
695 		error = EIO;
696 
697 unlock_all:
698 	VM_OBJECT_WUNLOCK(object);
699 
700 unlock_mp:
701 	if (mp != NULL) {
702 		if (vp != NULL)
703 			vput(vp);
704 		vm_object_deallocate(object);
705 		vn_finished_write(mp);
706 	}
707 
708 	return (error);
709 }
710 
711 /*
712  * Attempt to launder the specified number of pages.
713  *
714  * Returns the number of pages successfully laundered.
715  */
716 static int
717 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
718 {
719 	struct scan_state ss;
720 	struct vm_pagequeue *pq;
721 	vm_object_t object;
722 	vm_page_t m, marker;
723 	vm_page_astate_t new, old;
724 	int act_delta, error, numpagedout, queue, refs, starting_target;
725 	int vnodes_skipped;
726 	bool pageout_ok;
727 
728 	object = NULL;
729 	starting_target = launder;
730 	vnodes_skipped = 0;
731 
732 	/*
733 	 * Scan the laundry queues for pages eligible to be laundered.  We stop
734 	 * once the target number of dirty pages have been laundered, or once
735 	 * we've reached the end of the queue.  A single iteration of this loop
736 	 * may cause more than one page to be laundered because of clustering.
737 	 *
738 	 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
739 	 * swap devices are configured.
740 	 */
741 	if (atomic_load_acq_int(&swapdev_enabled))
742 		queue = PQ_UNSWAPPABLE;
743 	else
744 		queue = PQ_LAUNDRY;
745 
746 scan:
747 	marker = &vmd->vmd_markers[queue];
748 	pq = &vmd->vmd_pagequeues[queue];
749 	vm_pagequeue_lock(pq);
750 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
751 	while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
752 		if (__predict_false((m->flags & PG_MARKER) != 0))
753 			continue;
754 
755 		/*
756 		 * Don't touch a page that was removed from the queue after the
757 		 * page queue lock was released.  Otherwise, ensure that any
758 		 * pending queue operations, such as dequeues for wired pages,
759 		 * are handled.
760 		 */
761 		if (vm_pageout_defer(m, queue, true))
762 			continue;
763 
764 		/*
765 		 * Lock the page's object.
766 		 */
767 		if (object == NULL || object != m->object) {
768 			if (object != NULL)
769 				VM_OBJECT_WUNLOCK(object);
770 			object = (vm_object_t)atomic_load_ptr(&m->object);
771 			if (__predict_false(object == NULL))
772 				/* The page is being freed by another thread. */
773 				continue;
774 
775 			/* Depends on type-stability. */
776 			VM_OBJECT_WLOCK(object);
777 			if (__predict_false(m->object != object)) {
778 				VM_OBJECT_WUNLOCK(object);
779 				object = NULL;
780 				continue;
781 			}
782 		}
783 
784 		if (vm_page_tryxbusy(m) == 0)
785 			continue;
786 
787 		/*
788 		 * Check for wirings now that we hold the object lock and have
789 		 * exclusively busied the page.  If the page is mapped, it may
790 		 * still be wired by pmap lookups.  The call to
791 		 * vm_page_try_remove_all() below atomically checks for such
792 		 * wirings and removes mappings.  If the page is unmapped, the
793 		 * wire count is guaranteed not to increase after this check.
794 		 */
795 		if (__predict_false(vm_page_wired(m)))
796 			goto skip_page;
797 
798 		/*
799 		 * Invalid pages can be easily freed.  They cannot be
800 		 * mapped; vm_page_free() asserts this.
801 		 */
802 		if (vm_page_none_valid(m))
803 			goto free_page;
804 
805 		refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
806 
807 		for (old = vm_page_astate_load(m);;) {
808 			/*
809 			 * Check to see if the page has been removed from the
810 			 * queue since the first such check.  Leave it alone if
811 			 * so, discarding any references collected by
812 			 * pmap_ts_referenced().
813 			 */
814 			if (__predict_false(_vm_page_queue(old) == PQ_NONE))
815 				goto skip_page;
816 
817 			new = old;
818 			act_delta = refs;
819 			if ((old.flags & PGA_REFERENCED) != 0) {
820 				new.flags &= ~PGA_REFERENCED;
821 				act_delta++;
822 			}
823 			if (act_delta == 0) {
824 				;
825 			} else if (object->ref_count != 0) {
826 				/*
827 				 * Increase the activation count if the page was
828 				 * referenced while in the laundry queue.  This
829 				 * makes it less likely that the page will be
830 				 * returned prematurely to the laundry queue.
831 				 */
832 				new.act_count += ACT_ADVANCE +
833 				    act_delta;
834 				if (new.act_count > ACT_MAX)
835 					new.act_count = ACT_MAX;
836 
837 				new.flags &= ~PGA_QUEUE_OP_MASK;
838 				new.flags |= PGA_REQUEUE;
839 				new.queue = PQ_ACTIVE;
840 				if (!vm_page_pqstate_commit(m, &old, new))
841 					continue;
842 
843 				/*
844 				 * If this was a background laundering, count
845 				 * activated pages towards our target.  The
846 				 * purpose of background laundering is to ensure
847 				 * that pages are eventually cycled through the
848 				 * laundry queue, and an activation is a valid
849 				 * way out.
850 				 */
851 				if (!in_shortfall)
852 					launder--;
853 				VM_CNT_INC(v_reactivated);
854 				goto skip_page;
855 			} else if ((object->flags & OBJ_DEAD) == 0) {
856 				new.flags |= PGA_REQUEUE;
857 				if (!vm_page_pqstate_commit(m, &old, new))
858 					continue;
859 				goto skip_page;
860 			}
861 			break;
862 		}
863 
864 		/*
865 		 * If the page appears to be clean at the machine-independent
866 		 * layer, then remove all of its mappings from the pmap in
867 		 * anticipation of freeing it.  If, however, any of the page's
868 		 * mappings allow write access, then the page may still be
869 		 * modified until the last of those mappings are removed.
870 		 */
871 		if (object->ref_count != 0) {
872 			vm_page_test_dirty(m);
873 			if (m->dirty == 0 && !vm_page_try_remove_all(m))
874 				goto skip_page;
875 		}
876 
877 		/*
878 		 * Clean pages are freed, and dirty pages are paged out unless
879 		 * they belong to a dead object.  Requeueing dirty pages from
880 		 * dead objects is pointless, as they are being paged out and
881 		 * freed by the thread that destroyed the object.
882 		 */
883 		if (m->dirty == 0) {
884 free_page:
885 			/*
886 			 * Now we are guaranteed that no other threads are
887 			 * manipulating the page, check for a last-second
888 			 * reference.
889 			 */
890 			if (vm_pageout_defer(m, queue, true))
891 				goto skip_page;
892 			vm_page_free(m);
893 			VM_CNT_INC(v_dfree);
894 		} else if ((object->flags & OBJ_DEAD) == 0) {
895 			if (object->type != OBJT_SWAP &&
896 			    object->type != OBJT_DEFAULT)
897 				pageout_ok = true;
898 			else if (disable_swap_pageouts)
899 				pageout_ok = false;
900 			else
901 				pageout_ok = true;
902 			if (!pageout_ok) {
903 				vm_page_launder(m);
904 				goto skip_page;
905 			}
906 
907 			/*
908 			 * Form a cluster with adjacent, dirty pages from the
909 			 * same object, and page out that entire cluster.
910 			 *
911 			 * The adjacent, dirty pages must also be in the
912 			 * laundry.  However, their mappings are not checked
913 			 * for new references.  Consequently, a recently
914 			 * referenced page may be paged out.  However, that
915 			 * page will not be prematurely reclaimed.  After page
916 			 * out, the page will be placed in the inactive queue,
917 			 * where any new references will be detected and the
918 			 * page reactivated.
919 			 */
920 			error = vm_pageout_clean(m, &numpagedout);
921 			if (error == 0) {
922 				launder -= numpagedout;
923 				ss.scanned += numpagedout;
924 			} else if (error == EDEADLK) {
925 				pageout_lock_miss++;
926 				vnodes_skipped++;
927 			}
928 			object = NULL;
929 		} else {
930 skip_page:
931 			vm_page_xunbusy(m);
932 		}
933 	}
934 	if (object != NULL) {
935 		VM_OBJECT_WUNLOCK(object);
936 		object = NULL;
937 	}
938 	vm_pagequeue_lock(pq);
939 	vm_pageout_end_scan(&ss);
940 	vm_pagequeue_unlock(pq);
941 
942 	if (launder > 0 && queue == PQ_UNSWAPPABLE) {
943 		queue = PQ_LAUNDRY;
944 		goto scan;
945 	}
946 
947 	/*
948 	 * Wakeup the sync daemon if we skipped a vnode in a writeable object
949 	 * and we didn't launder enough pages.
950 	 */
951 	if (vnodes_skipped > 0 && launder > 0)
952 		(void)speedup_syncer();
953 
954 	return (starting_target - launder);
955 }
956 
957 /*
958  * Compute the integer square root.
959  */
960 static u_int
961 isqrt(u_int num)
962 {
963 	u_int bit, root, tmp;
964 
965 	bit = num != 0 ? (1u << ((fls(num) - 1) & ~1)) : 0;
966 	root = 0;
967 	while (bit != 0) {
968 		tmp = root + bit;
969 		root >>= 1;
970 		if (num >= tmp) {
971 			num -= tmp;
972 			root += bit;
973 		}
974 		bit >>= 2;
975 	}
976 	return (root);
977 }
978 
979 /*
980  * Perform the work of the laundry thread: periodically wake up and determine
981  * whether any pages need to be laundered.  If so, determine the number of pages
982  * that need to be laundered, and launder them.
983  */
984 static void
985 vm_pageout_laundry_worker(void *arg)
986 {
987 	struct vm_domain *vmd;
988 	struct vm_pagequeue *pq;
989 	uint64_t nclean, ndirty, nfreed;
990 	int domain, last_target, launder, shortfall, shortfall_cycle, target;
991 	bool in_shortfall;
992 
993 	domain = (uintptr_t)arg;
994 	vmd = VM_DOMAIN(domain);
995 	pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
996 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
997 
998 	shortfall = 0;
999 	in_shortfall = false;
1000 	shortfall_cycle = 0;
1001 	last_target = target = 0;
1002 	nfreed = 0;
1003 
1004 	/*
1005 	 * Calls to these handlers are serialized by the swap syscall lock.
1006 	 */
1007 	(void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
1008 	    EVENTHANDLER_PRI_ANY);
1009 	(void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
1010 	    EVENTHANDLER_PRI_ANY);
1011 
1012 	/*
1013 	 * The pageout laundry worker is never done, so loop forever.
1014 	 */
1015 	for (;;) {
1016 		KASSERT(target >= 0, ("negative target %d", target));
1017 		KASSERT(shortfall_cycle >= 0,
1018 		    ("negative cycle %d", shortfall_cycle));
1019 		launder = 0;
1020 
1021 		/*
1022 		 * First determine whether we need to launder pages to meet a
1023 		 * shortage of free pages.
1024 		 */
1025 		if (shortfall > 0) {
1026 			in_shortfall = true;
1027 			shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1028 			target = shortfall;
1029 		} else if (!in_shortfall)
1030 			goto trybackground;
1031 		else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1032 			/*
1033 			 * We recently entered shortfall and began laundering
1034 			 * pages.  If we have completed that laundering run
1035 			 * (and we are no longer in shortfall) or we have met
1036 			 * our laundry target through other activity, then we
1037 			 * can stop laundering pages.
1038 			 */
1039 			in_shortfall = false;
1040 			target = 0;
1041 			goto trybackground;
1042 		}
1043 		launder = target / shortfall_cycle--;
1044 		goto dolaundry;
1045 
1046 		/*
1047 		 * There's no immediate need to launder any pages; see if we
1048 		 * meet the conditions to perform background laundering:
1049 		 *
1050 		 * 1. The ratio of dirty to clean inactive pages exceeds the
1051 		 *    background laundering threshold, or
1052 		 * 2. we haven't yet reached the target of the current
1053 		 *    background laundering run.
1054 		 *
1055 		 * The background laundering threshold is not a constant.
1056 		 * Instead, it is a slowly growing function of the number of
1057 		 * clean pages freed by the page daemon since the last
1058 		 * background laundering.  Thus, as the ratio of dirty to
1059 		 * clean inactive pages grows, the amount of memory pressure
1060 		 * required to trigger laundering decreases.  We ensure
1061 		 * that the threshold is non-zero after an inactive queue
1062 		 * scan, even if that scan failed to free a single clean page.
1063 		 */
1064 trybackground:
1065 		nclean = vmd->vmd_free_count +
1066 		    vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1067 		ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1068 		if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1069 		    vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1070 			target = vmd->vmd_background_launder_target;
1071 		}
1072 
1073 		/*
1074 		 * We have a non-zero background laundering target.  If we've
1075 		 * laundered up to our maximum without observing a page daemon
1076 		 * request, just stop.  This is a safety belt that ensures we
1077 		 * don't launder an excessive amount if memory pressure is low
1078 		 * and the ratio of dirty to clean pages is large.  Otherwise,
1079 		 * proceed at the background laundering rate.
1080 		 */
1081 		if (target > 0) {
1082 			if (nfreed > 0) {
1083 				nfreed = 0;
1084 				last_target = target;
1085 			} else if (last_target - target >=
1086 			    vm_background_launder_max * PAGE_SIZE / 1024) {
1087 				target = 0;
1088 			}
1089 			launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1090 			launder /= VM_LAUNDER_RATE;
1091 			if (launder > target)
1092 				launder = target;
1093 		}
1094 
1095 dolaundry:
1096 		if (launder > 0) {
1097 			/*
1098 			 * Because of I/O clustering, the number of laundered
1099 			 * pages could exceed "target" by the maximum size of
1100 			 * a cluster minus one.
1101 			 */
1102 			target -= min(vm_pageout_launder(vmd, launder,
1103 			    in_shortfall), target);
1104 			pause("laundp", hz / VM_LAUNDER_RATE);
1105 		}
1106 
1107 		/*
1108 		 * If we're not currently laundering pages and the page daemon
1109 		 * hasn't posted a new request, sleep until the page daemon
1110 		 * kicks us.
1111 		 */
1112 		vm_pagequeue_lock(pq);
1113 		if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1114 			(void)mtx_sleep(&vmd->vmd_laundry_request,
1115 			    vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1116 
1117 		/*
1118 		 * If the pagedaemon has indicated that it's in shortfall, start
1119 		 * a shortfall laundering unless we're already in the middle of
1120 		 * one.  This may preempt a background laundering.
1121 		 */
1122 		if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1123 		    (!in_shortfall || shortfall_cycle == 0)) {
1124 			shortfall = vm_laundry_target(vmd) +
1125 			    vmd->vmd_pageout_deficit;
1126 			target = 0;
1127 		} else
1128 			shortfall = 0;
1129 
1130 		if (target == 0)
1131 			vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1132 		nfreed += vmd->vmd_clean_pages_freed;
1133 		vmd->vmd_clean_pages_freed = 0;
1134 		vm_pagequeue_unlock(pq);
1135 	}
1136 }
1137 
1138 /*
1139  * Compute the number of pages we want to try to move from the
1140  * active queue to either the inactive or laundry queue.
1141  *
1142  * When scanning active pages during a shortage, we make clean pages
1143  * count more heavily towards the page shortage than dirty pages.
1144  * This is because dirty pages must be laundered before they can be
1145  * reused and thus have less utility when attempting to quickly
1146  * alleviate a free page shortage.  However, this weighting also
1147  * causes the scan to deactivate dirty pages more aggressively,
1148  * improving the effectiveness of clustering.
1149  */
1150 static int
1151 vm_pageout_active_target(struct vm_domain *vmd)
1152 {
1153 	int shortage;
1154 
1155 	shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1156 	    (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1157 	    vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1158 	shortage *= act_scan_laundry_weight;
1159 	return (shortage);
1160 }
1161 
1162 /*
1163  * Scan the active queue.  If there is no shortage of inactive pages, scan a
1164  * small portion of the queue in order to maintain quasi-LRU.
1165  */
1166 static void
1167 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1168 {
1169 	struct scan_state ss;
1170 	vm_object_t object;
1171 	vm_page_t m, marker;
1172 	struct vm_pagequeue *pq;
1173 	vm_page_astate_t old, new;
1174 	long min_scan;
1175 	int act_delta, max_scan, ps_delta, refs, scan_tick;
1176 	uint8_t nqueue;
1177 
1178 	marker = &vmd->vmd_markers[PQ_ACTIVE];
1179 	pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1180 	vm_pagequeue_lock(pq);
1181 
1182 	/*
1183 	 * If we're just idle polling attempt to visit every
1184 	 * active page within 'update_period' seconds.
1185 	 */
1186 	scan_tick = ticks;
1187 	if (vm_pageout_update_period != 0) {
1188 		min_scan = pq->pq_cnt;
1189 		min_scan *= scan_tick - vmd->vmd_last_active_scan;
1190 		min_scan /= hz * vm_pageout_update_period;
1191 	} else
1192 		min_scan = 0;
1193 	if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1194 		vmd->vmd_last_active_scan = scan_tick;
1195 
1196 	/*
1197 	 * Scan the active queue for pages that can be deactivated.  Update
1198 	 * the per-page activity counter and use it to identify deactivation
1199 	 * candidates.  Held pages may be deactivated.
1200 	 *
1201 	 * To avoid requeuing each page that remains in the active queue, we
1202 	 * implement the CLOCK algorithm.  To keep the implementation of the
1203 	 * enqueue operation consistent for all page queues, we use two hands,
1204 	 * represented by marker pages. Scans begin at the first hand, which
1205 	 * precedes the second hand in the queue.  When the two hands meet,
1206 	 * they are moved back to the head and tail of the queue, respectively,
1207 	 * and scanning resumes.
1208 	 */
1209 	max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1210 act_scan:
1211 	vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1212 	while ((m = vm_pageout_next(&ss, false)) != NULL) {
1213 		if (__predict_false(m == &vmd->vmd_clock[1])) {
1214 			vm_pagequeue_lock(pq);
1215 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1216 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1217 			TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1218 			    plinks.q);
1219 			TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1220 			    plinks.q);
1221 			max_scan -= ss.scanned;
1222 			vm_pageout_end_scan(&ss);
1223 			goto act_scan;
1224 		}
1225 		if (__predict_false((m->flags & PG_MARKER) != 0))
1226 			continue;
1227 
1228 		/*
1229 		 * Don't touch a page that was removed from the queue after the
1230 		 * page queue lock was released.  Otherwise, ensure that any
1231 		 * pending queue operations, such as dequeues for wired pages,
1232 		 * are handled.
1233 		 */
1234 		if (vm_pageout_defer(m, PQ_ACTIVE, true))
1235 			continue;
1236 
1237 		/*
1238 		 * A page's object pointer may be set to NULL before
1239 		 * the object lock is acquired.
1240 		 */
1241 		object = (vm_object_t)atomic_load_ptr(&m->object);
1242 		if (__predict_false(object == NULL))
1243 			/*
1244 			 * The page has been removed from its object.
1245 			 */
1246 			continue;
1247 
1248 		/* Deferred free of swap space. */
1249 		if ((m->a.flags & PGA_SWAP_FREE) != 0 &&
1250 		    VM_OBJECT_TRYWLOCK(object)) {
1251 			if (m->object == object)
1252 				vm_pager_page_unswapped(m);
1253 			VM_OBJECT_WUNLOCK(object);
1254 		}
1255 
1256 		/*
1257 		 * Check to see "how much" the page has been used.
1258 		 *
1259 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1260 		 * that a reference from a concurrently destroyed mapping is
1261 		 * observed here and now.
1262 		 *
1263 		 * Perform an unsynchronized object ref count check.  While
1264 		 * the page lock ensures that the page is not reallocated to
1265 		 * another object, in particular, one with unmanaged mappings
1266 		 * that cannot support pmap_ts_referenced(), two races are,
1267 		 * nonetheless, possible:
1268 		 * 1) The count was transitioning to zero, but we saw a non-
1269 		 *    zero value.  pmap_ts_referenced() will return zero
1270 		 *    because the page is not mapped.
1271 		 * 2) The count was transitioning to one, but we saw zero.
1272 		 *    This race delays the detection of a new reference.  At
1273 		 *    worst, we will deactivate and reactivate the page.
1274 		 */
1275 		refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1276 
1277 		old = vm_page_astate_load(m);
1278 		do {
1279 			/*
1280 			 * Check to see if the page has been removed from the
1281 			 * queue since the first such check.  Leave it alone if
1282 			 * so, discarding any references collected by
1283 			 * pmap_ts_referenced().
1284 			 */
1285 			if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1286 				break;
1287 
1288 			/*
1289 			 * Advance or decay the act_count based on recent usage.
1290 			 */
1291 			new = old;
1292 			act_delta = refs;
1293 			if ((old.flags & PGA_REFERENCED) != 0) {
1294 				new.flags &= ~PGA_REFERENCED;
1295 				act_delta++;
1296 			}
1297 			if (act_delta != 0) {
1298 				new.act_count += ACT_ADVANCE + act_delta;
1299 				if (new.act_count > ACT_MAX)
1300 					new.act_count = ACT_MAX;
1301 			} else {
1302 				new.act_count -= min(new.act_count,
1303 				    ACT_DECLINE);
1304 			}
1305 
1306 			if (new.act_count > 0) {
1307 				/*
1308 				 * Adjust the activation count and keep the page
1309 				 * in the active queue.  The count might be left
1310 				 * unchanged if it is saturated.  The page may
1311 				 * have been moved to a different queue since we
1312 				 * started the scan, in which case we move it
1313 				 * back.
1314 				 */
1315 				ps_delta = 0;
1316 				if (old.queue != PQ_ACTIVE) {
1317 					new.flags &= ~PGA_QUEUE_OP_MASK;
1318 					new.flags |= PGA_REQUEUE;
1319 					new.queue = PQ_ACTIVE;
1320 				}
1321 			} else {
1322 				/*
1323 				 * When not short for inactive pages, let dirty
1324 				 * pages go through the inactive queue before
1325 				 * moving to the laundry queue.  This gives them
1326 				 * some extra time to be reactivated,
1327 				 * potentially avoiding an expensive pageout.
1328 				 * However, during a page shortage, the inactive
1329 				 * queue is necessarily small, and so dirty
1330 				 * pages would only spend a trivial amount of
1331 				 * time in the inactive queue.  Therefore, we
1332 				 * might as well place them directly in the
1333 				 * laundry queue to reduce queuing overhead.
1334 				 *
1335 				 * Calling vm_page_test_dirty() here would
1336 				 * require acquisition of the object's write
1337 				 * lock.  However, during a page shortage,
1338 				 * directing dirty pages into the laundry queue
1339 				 * is only an optimization and not a
1340 				 * requirement.  Therefore, we simply rely on
1341 				 * the opportunistic updates to the page's dirty
1342 				 * field by the pmap.
1343 				 */
1344 				if (page_shortage <= 0) {
1345 					nqueue = PQ_INACTIVE;
1346 					ps_delta = 0;
1347 				} else if (m->dirty == 0) {
1348 					nqueue = PQ_INACTIVE;
1349 					ps_delta = act_scan_laundry_weight;
1350 				} else {
1351 					nqueue = PQ_LAUNDRY;
1352 					ps_delta = 1;
1353 				}
1354 
1355 				new.flags &= ~PGA_QUEUE_OP_MASK;
1356 				new.flags |= PGA_REQUEUE;
1357 				new.queue = nqueue;
1358 			}
1359 		} while (!vm_page_pqstate_commit(m, &old, new));
1360 
1361 		page_shortage -= ps_delta;
1362 	}
1363 	vm_pagequeue_lock(pq);
1364 	TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1365 	TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1366 	vm_pageout_end_scan(&ss);
1367 	vm_pagequeue_unlock(pq);
1368 }
1369 
1370 static int
1371 vm_pageout_reinsert_inactive_page(struct vm_pagequeue *pq, vm_page_t marker,
1372     vm_page_t m)
1373 {
1374 	vm_page_astate_t as;
1375 
1376 	vm_pagequeue_assert_locked(pq);
1377 
1378 	as = vm_page_astate_load(m);
1379 	if (as.queue != PQ_INACTIVE || (as.flags & PGA_ENQUEUED) != 0)
1380 		return (0);
1381 	vm_page_aflag_set(m, PGA_ENQUEUED);
1382 	TAILQ_INSERT_BEFORE(marker, m, plinks.q);
1383 	return (1);
1384 }
1385 
1386 /*
1387  * Re-add stuck pages to the inactive queue.  We will examine them again
1388  * during the next scan.  If the queue state of a page has changed since
1389  * it was physically removed from the page queue in
1390  * vm_pageout_collect_batch(), don't do anything with that page.
1391  */
1392 static void
1393 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1394     vm_page_t m)
1395 {
1396 	struct vm_pagequeue *pq;
1397 	vm_page_t marker;
1398 	int delta;
1399 
1400 	delta = 0;
1401 	marker = ss->marker;
1402 	pq = ss->pq;
1403 
1404 	if (m != NULL) {
1405 		if (vm_batchqueue_insert(bq, m))
1406 			return;
1407 		vm_pagequeue_lock(pq);
1408 		delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1409 	} else
1410 		vm_pagequeue_lock(pq);
1411 	while ((m = vm_batchqueue_pop(bq)) != NULL)
1412 		delta += vm_pageout_reinsert_inactive_page(pq, marker, m);
1413 	vm_pagequeue_cnt_add(pq, delta);
1414 	vm_pagequeue_unlock(pq);
1415 	vm_batchqueue_init(bq);
1416 }
1417 
1418 /*
1419  * Attempt to reclaim the requested number of pages from the inactive queue.
1420  * Returns true if the shortage was addressed.
1421  */
1422 static int
1423 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1424     int *addl_shortage)
1425 {
1426 	struct scan_state ss;
1427 	struct vm_batchqueue rq;
1428 	vm_page_t m, marker;
1429 	struct vm_pagequeue *pq;
1430 	vm_object_t object;
1431 	vm_page_astate_t old, new;
1432 	int act_delta, addl_page_shortage, deficit, page_shortage, refs;
1433 	int starting_page_shortage;
1434 
1435 	/*
1436 	 * The addl_page_shortage is an estimate of the number of temporarily
1437 	 * stuck pages in the inactive queue.  In other words, the
1438 	 * number of pages from the inactive count that should be
1439 	 * discounted in setting the target for the active queue scan.
1440 	 */
1441 	addl_page_shortage = 0;
1442 
1443 	/*
1444 	 * vmd_pageout_deficit counts the number of pages requested in
1445 	 * allocations that failed because of a free page shortage.  We assume
1446 	 * that the allocations will be reattempted and thus include the deficit
1447 	 * in our scan target.
1448 	 */
1449 	deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1450 	starting_page_shortage = page_shortage = shortage + deficit;
1451 
1452 	object = NULL;
1453 	vm_batchqueue_init(&rq);
1454 
1455 	/*
1456 	 * Start scanning the inactive queue for pages that we can free.  The
1457 	 * scan will stop when we reach the target or we have scanned the
1458 	 * entire queue.  (Note that m->a.act_count is not used to make
1459 	 * decisions for the inactive queue, only for the active queue.)
1460 	 */
1461 	marker = &vmd->vmd_markers[PQ_INACTIVE];
1462 	pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1463 	vm_pagequeue_lock(pq);
1464 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1465 	while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1466 		KASSERT((m->flags & PG_MARKER) == 0,
1467 		    ("marker page %p was dequeued", m));
1468 
1469 		/*
1470 		 * Don't touch a page that was removed from the queue after the
1471 		 * page queue lock was released.  Otherwise, ensure that any
1472 		 * pending queue operations, such as dequeues for wired pages,
1473 		 * are handled.
1474 		 */
1475 		if (vm_pageout_defer(m, PQ_INACTIVE, false))
1476 			continue;
1477 
1478 		/*
1479 		 * Lock the page's object.
1480 		 */
1481 		if (object == NULL || object != m->object) {
1482 			if (object != NULL)
1483 				VM_OBJECT_WUNLOCK(object);
1484 			object = (vm_object_t)atomic_load_ptr(&m->object);
1485 			if (__predict_false(object == NULL))
1486 				/* The page is being freed by another thread. */
1487 				continue;
1488 
1489 			/* Depends on type-stability. */
1490 			VM_OBJECT_WLOCK(object);
1491 			if (__predict_false(m->object != object)) {
1492 				VM_OBJECT_WUNLOCK(object);
1493 				object = NULL;
1494 				goto reinsert;
1495 			}
1496 		}
1497 
1498 		if (vm_page_tryxbusy(m) == 0) {
1499 			/*
1500 			 * Don't mess with busy pages.  Leave them at
1501 			 * the front of the queue.  Most likely, they
1502 			 * are being paged out and will leave the
1503 			 * queue shortly after the scan finishes.  So,
1504 			 * they ought to be discounted from the
1505 			 * inactive count.
1506 			 */
1507 			addl_page_shortage++;
1508 			goto reinsert;
1509 		}
1510 
1511 		/* Deferred free of swap space. */
1512 		if ((m->a.flags & PGA_SWAP_FREE) != 0)
1513 			vm_pager_page_unswapped(m);
1514 
1515 		/*
1516 		 * Check for wirings now that we hold the object lock and have
1517 		 * exclusively busied the page.  If the page is mapped, it may
1518 		 * still be wired by pmap lookups.  The call to
1519 		 * vm_page_try_remove_all() below atomically checks for such
1520 		 * wirings and removes mappings.  If the page is unmapped, the
1521 		 * wire count is guaranteed not to increase after this check.
1522 		 */
1523 		if (__predict_false(vm_page_wired(m)))
1524 			goto skip_page;
1525 
1526 		/*
1527 		 * Invalid pages can be easily freed. They cannot be
1528 		 * mapped, vm_page_free() asserts this.
1529 		 */
1530 		if (vm_page_none_valid(m))
1531 			goto free_page;
1532 
1533 		refs = object->ref_count != 0 ? pmap_ts_referenced(m) : 0;
1534 
1535 		for (old = vm_page_astate_load(m);;) {
1536 			/*
1537 			 * Check to see if the page has been removed from the
1538 			 * queue since the first such check.  Leave it alone if
1539 			 * so, discarding any references collected by
1540 			 * pmap_ts_referenced().
1541 			 */
1542 			if (__predict_false(_vm_page_queue(old) == PQ_NONE))
1543 				goto skip_page;
1544 
1545 			new = old;
1546 			act_delta = refs;
1547 			if ((old.flags & PGA_REFERENCED) != 0) {
1548 				new.flags &= ~PGA_REFERENCED;
1549 				act_delta++;
1550 			}
1551 			if (act_delta == 0) {
1552 				;
1553 			} else if (object->ref_count != 0) {
1554 				/*
1555 				 * Increase the activation count if the
1556 				 * page was referenced while in the
1557 				 * inactive queue.  This makes it less
1558 				 * likely that the page will be returned
1559 				 * prematurely to the inactive queue.
1560 				 */
1561 				new.act_count += ACT_ADVANCE +
1562 				    act_delta;
1563 				if (new.act_count > ACT_MAX)
1564 					new.act_count = ACT_MAX;
1565 
1566 				new.flags &= ~PGA_QUEUE_OP_MASK;
1567 				new.flags |= PGA_REQUEUE;
1568 				new.queue = PQ_ACTIVE;
1569 				if (!vm_page_pqstate_commit(m, &old, new))
1570 					continue;
1571 
1572 				VM_CNT_INC(v_reactivated);
1573 				goto skip_page;
1574 			} else if ((object->flags & OBJ_DEAD) == 0) {
1575 				new.queue = PQ_INACTIVE;
1576 				new.flags |= PGA_REQUEUE;
1577 				if (!vm_page_pqstate_commit(m, &old, new))
1578 					continue;
1579 				goto skip_page;
1580 			}
1581 			break;
1582 		}
1583 
1584 		/*
1585 		 * If the page appears to be clean at the machine-independent
1586 		 * layer, then remove all of its mappings from the pmap in
1587 		 * anticipation of freeing it.  If, however, any of the page's
1588 		 * mappings allow write access, then the page may still be
1589 		 * modified until the last of those mappings are removed.
1590 		 */
1591 		if (object->ref_count != 0) {
1592 			vm_page_test_dirty(m);
1593 			if (m->dirty == 0 && !vm_page_try_remove_all(m))
1594 				goto skip_page;
1595 		}
1596 
1597 		/*
1598 		 * Clean pages can be freed, but dirty pages must be sent back
1599 		 * to the laundry, unless they belong to a dead object.
1600 		 * Requeueing dirty pages from dead objects is pointless, as
1601 		 * they are being paged out and freed by the thread that
1602 		 * destroyed the object.
1603 		 */
1604 		if (m->dirty == 0) {
1605 free_page:
1606 			/*
1607 			 * Now we are guaranteed that no other threads are
1608 			 * manipulating the page, check for a last-second
1609 			 * reference that would save it from doom.
1610 			 */
1611 			if (vm_pageout_defer(m, PQ_INACTIVE, false))
1612 				goto skip_page;
1613 
1614 			/*
1615 			 * Because we dequeued the page and have already checked
1616 			 * for pending dequeue and enqueue requests, we can
1617 			 * safely disassociate the page from the inactive queue
1618 			 * without holding the queue lock.
1619 			 */
1620 			m->a.queue = PQ_NONE;
1621 			vm_page_free(m);
1622 			page_shortage--;
1623 			continue;
1624 		}
1625 		if ((object->flags & OBJ_DEAD) == 0)
1626 			vm_page_launder(m);
1627 skip_page:
1628 		vm_page_xunbusy(m);
1629 		continue;
1630 reinsert:
1631 		vm_pageout_reinsert_inactive(&ss, &rq, m);
1632 	}
1633 	if (object != NULL)
1634 		VM_OBJECT_WUNLOCK(object);
1635 	vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1636 	vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1637 	vm_pagequeue_lock(pq);
1638 	vm_pageout_end_scan(&ss);
1639 	vm_pagequeue_unlock(pq);
1640 
1641 	VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1642 
1643 	/*
1644 	 * Wake up the laundry thread so that it can perform any needed
1645 	 * laundering.  If we didn't meet our target, we're in shortfall and
1646 	 * need to launder more aggressively.  If PQ_LAUNDRY is empty and no
1647 	 * swap devices are configured, the laundry thread has no work to do, so
1648 	 * don't bother waking it up.
1649 	 *
1650 	 * The laundry thread uses the number of inactive queue scans elapsed
1651 	 * since the last laundering to determine whether to launder again, so
1652 	 * keep count.
1653 	 */
1654 	if (starting_page_shortage > 0) {
1655 		pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1656 		vm_pagequeue_lock(pq);
1657 		if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1658 		    (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1659 			if (page_shortage > 0) {
1660 				vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1661 				VM_CNT_INC(v_pdshortfalls);
1662 			} else if (vmd->vmd_laundry_request !=
1663 			    VM_LAUNDRY_SHORTFALL)
1664 				vmd->vmd_laundry_request =
1665 				    VM_LAUNDRY_BACKGROUND;
1666 			wakeup(&vmd->vmd_laundry_request);
1667 		}
1668 		vmd->vmd_clean_pages_freed +=
1669 		    starting_page_shortage - page_shortage;
1670 		vm_pagequeue_unlock(pq);
1671 	}
1672 
1673 	/*
1674 	 * Wakeup the swapout daemon if we didn't free the targeted number of
1675 	 * pages.
1676 	 */
1677 	if (page_shortage > 0)
1678 		vm_swapout_run();
1679 
1680 	/*
1681 	 * If the inactive queue scan fails repeatedly to meet its
1682 	 * target, kill the largest process.
1683 	 */
1684 	vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1685 
1686 	/*
1687 	 * Reclaim pages by swapping out idle processes, if configured to do so.
1688 	 */
1689 	vm_swapout_run_idle();
1690 
1691 	/*
1692 	 * See the description of addl_page_shortage above.
1693 	 */
1694 	*addl_shortage = addl_page_shortage + deficit;
1695 
1696 	return (page_shortage <= 0);
1697 }
1698 
1699 static int vm_pageout_oom_vote;
1700 
1701 /*
1702  * The pagedaemon threads randlomly select one to perform the
1703  * OOM.  Trying to kill processes before all pagedaemons
1704  * failed to reach free target is premature.
1705  */
1706 static void
1707 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1708     int starting_page_shortage)
1709 {
1710 	int old_vote;
1711 
1712 	if (starting_page_shortage <= 0 || starting_page_shortage !=
1713 	    page_shortage)
1714 		vmd->vmd_oom_seq = 0;
1715 	else
1716 		vmd->vmd_oom_seq++;
1717 	if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1718 		if (vmd->vmd_oom) {
1719 			vmd->vmd_oom = FALSE;
1720 			atomic_subtract_int(&vm_pageout_oom_vote, 1);
1721 		}
1722 		return;
1723 	}
1724 
1725 	/*
1726 	 * Do not follow the call sequence until OOM condition is
1727 	 * cleared.
1728 	 */
1729 	vmd->vmd_oom_seq = 0;
1730 
1731 	if (vmd->vmd_oom)
1732 		return;
1733 
1734 	vmd->vmd_oom = TRUE;
1735 	old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1736 	if (old_vote != vm_ndomains - 1)
1737 		return;
1738 
1739 	/*
1740 	 * The current pagedaemon thread is the last in the quorum to
1741 	 * start OOM.  Initiate the selection and signaling of the
1742 	 * victim.
1743 	 */
1744 	vm_pageout_oom(VM_OOM_MEM);
1745 
1746 	/*
1747 	 * After one round of OOM terror, recall our vote.  On the
1748 	 * next pass, current pagedaemon would vote again if the low
1749 	 * memory condition is still there, due to vmd_oom being
1750 	 * false.
1751 	 */
1752 	vmd->vmd_oom = FALSE;
1753 	atomic_subtract_int(&vm_pageout_oom_vote, 1);
1754 }
1755 
1756 /*
1757  * The OOM killer is the page daemon's action of last resort when
1758  * memory allocation requests have been stalled for a prolonged period
1759  * of time because it cannot reclaim memory.  This function computes
1760  * the approximate number of physical pages that could be reclaimed if
1761  * the specified address space is destroyed.
1762  *
1763  * Private, anonymous memory owned by the address space is the
1764  * principal resource that we expect to recover after an OOM kill.
1765  * Since the physical pages mapped by the address space's COW entries
1766  * are typically shared pages, they are unlikely to be released and so
1767  * they are not counted.
1768  *
1769  * To get to the point where the page daemon runs the OOM killer, its
1770  * efforts to write-back vnode-backed pages may have stalled.  This
1771  * could be caused by a memory allocation deadlock in the write path
1772  * that might be resolved by an OOM kill.  Therefore, physical pages
1773  * belonging to vnode-backed objects are counted, because they might
1774  * be freed without being written out first if the address space holds
1775  * the last reference to an unlinked vnode.
1776  *
1777  * Similarly, physical pages belonging to OBJT_PHYS objects are
1778  * counted because the address space might hold the last reference to
1779  * the object.
1780  */
1781 static long
1782 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1783 {
1784 	vm_map_t map;
1785 	vm_map_entry_t entry;
1786 	vm_object_t obj;
1787 	long res;
1788 
1789 	map = &vmspace->vm_map;
1790 	KASSERT(!map->system_map, ("system map"));
1791 	sx_assert(&map->lock, SA_LOCKED);
1792 	res = 0;
1793 	VM_MAP_ENTRY_FOREACH(entry, map) {
1794 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1795 			continue;
1796 		obj = entry->object.vm_object;
1797 		if (obj == NULL)
1798 			continue;
1799 		if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1800 		    obj->ref_count != 1)
1801 			continue;
1802 		switch (obj->type) {
1803 		case OBJT_DEFAULT:
1804 		case OBJT_SWAP:
1805 		case OBJT_PHYS:
1806 		case OBJT_VNODE:
1807 			res += obj->resident_page_count;
1808 			break;
1809 		}
1810 	}
1811 	return (res);
1812 }
1813 
1814 static int vm_oom_ratelim_last;
1815 static int vm_oom_pf_secs = 10;
1816 SYSCTL_INT(_vm, OID_AUTO, oom_pf_secs, CTLFLAG_RWTUN, &vm_oom_pf_secs, 0,
1817     "");
1818 static struct mtx vm_oom_ratelim_mtx;
1819 
1820 void
1821 vm_pageout_oom(int shortage)
1822 {
1823 	struct proc *p, *bigproc;
1824 	vm_offset_t size, bigsize;
1825 	struct thread *td;
1826 	struct vmspace *vm;
1827 	int now;
1828 	bool breakout;
1829 
1830 	/*
1831 	 * For OOM requests originating from vm_fault(), there is a high
1832 	 * chance that a single large process faults simultaneously in
1833 	 * several threads.  Also, on an active system running many
1834 	 * processes of middle-size, like buildworld, all of them
1835 	 * could fault almost simultaneously as well.
1836 	 *
1837 	 * To avoid killing too many processes, rate-limit OOMs
1838 	 * initiated by vm_fault() time-outs on the waits for free
1839 	 * pages.
1840 	 */
1841 	mtx_lock(&vm_oom_ratelim_mtx);
1842 	now = ticks;
1843 	if (shortage == VM_OOM_MEM_PF &&
1844 	    (u_int)(now - vm_oom_ratelim_last) < hz * vm_oom_pf_secs) {
1845 		mtx_unlock(&vm_oom_ratelim_mtx);
1846 		return;
1847 	}
1848 	vm_oom_ratelim_last = now;
1849 	mtx_unlock(&vm_oom_ratelim_mtx);
1850 
1851 	/*
1852 	 * We keep the process bigproc locked once we find it to keep anyone
1853 	 * from messing with it; however, there is a possibility of
1854 	 * deadlock if process B is bigproc and one of its child processes
1855 	 * attempts to propagate a signal to B while we are waiting for A's
1856 	 * lock while walking this list.  To avoid this, we don't block on
1857 	 * the process lock but just skip a process if it is already locked.
1858 	 */
1859 	bigproc = NULL;
1860 	bigsize = 0;
1861 	sx_slock(&allproc_lock);
1862 	FOREACH_PROC_IN_SYSTEM(p) {
1863 		PROC_LOCK(p);
1864 
1865 		/*
1866 		 * If this is a system, protected or killed process, skip it.
1867 		 */
1868 		if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1869 		    P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1870 		    p->p_pid == 1 || P_KILLED(p) ||
1871 		    (p->p_pid < 48 && swap_pager_avail != 0)) {
1872 			PROC_UNLOCK(p);
1873 			continue;
1874 		}
1875 		/*
1876 		 * If the process is in a non-running type state,
1877 		 * don't touch it.  Check all the threads individually.
1878 		 */
1879 		breakout = false;
1880 		FOREACH_THREAD_IN_PROC(p, td) {
1881 			thread_lock(td);
1882 			if (!TD_ON_RUNQ(td) &&
1883 			    !TD_IS_RUNNING(td) &&
1884 			    !TD_IS_SLEEPING(td) &&
1885 			    !TD_IS_SUSPENDED(td) &&
1886 			    !TD_IS_SWAPPED(td)) {
1887 				thread_unlock(td);
1888 				breakout = true;
1889 				break;
1890 			}
1891 			thread_unlock(td);
1892 		}
1893 		if (breakout) {
1894 			PROC_UNLOCK(p);
1895 			continue;
1896 		}
1897 		/*
1898 		 * get the process size
1899 		 */
1900 		vm = vmspace_acquire_ref(p);
1901 		if (vm == NULL) {
1902 			PROC_UNLOCK(p);
1903 			continue;
1904 		}
1905 		_PHOLD_LITE(p);
1906 		PROC_UNLOCK(p);
1907 		sx_sunlock(&allproc_lock);
1908 		if (!vm_map_trylock_read(&vm->vm_map)) {
1909 			vmspace_free(vm);
1910 			sx_slock(&allproc_lock);
1911 			PRELE(p);
1912 			continue;
1913 		}
1914 		size = vmspace_swap_count(vm);
1915 		if (shortage == VM_OOM_MEM || shortage == VM_OOM_MEM_PF)
1916 			size += vm_pageout_oom_pagecount(vm);
1917 		vm_map_unlock_read(&vm->vm_map);
1918 		vmspace_free(vm);
1919 		sx_slock(&allproc_lock);
1920 
1921 		/*
1922 		 * If this process is bigger than the biggest one,
1923 		 * remember it.
1924 		 */
1925 		if (size > bigsize) {
1926 			if (bigproc != NULL)
1927 				PRELE(bigproc);
1928 			bigproc = p;
1929 			bigsize = size;
1930 		} else {
1931 			PRELE(p);
1932 		}
1933 	}
1934 	sx_sunlock(&allproc_lock);
1935 	if (bigproc != NULL) {
1936 		if (vm_panic_on_oom != 0)
1937 			panic("out of swap space");
1938 		PROC_LOCK(bigproc);
1939 		killproc(bigproc, "out of swap space");
1940 		sched_nice(bigproc, PRIO_MIN);
1941 		_PRELE(bigproc);
1942 		PROC_UNLOCK(bigproc);
1943 	}
1944 }
1945 
1946 /*
1947  * Signal a free page shortage to subsystems that have registered an event
1948  * handler.  Reclaim memory from UMA in the event of a severe shortage.
1949  * Return true if the free page count should be re-evaluated.
1950  */
1951 static bool
1952 vm_pageout_lowmem(void)
1953 {
1954 	static int lowmem_ticks = 0;
1955 	int last;
1956 	bool ret;
1957 
1958 	ret = false;
1959 
1960 	last = atomic_load_int(&lowmem_ticks);
1961 	while ((u_int)(ticks - last) / hz >= lowmem_period) {
1962 		if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
1963 			continue;
1964 
1965 		/*
1966 		 * Decrease registered cache sizes.
1967 		 */
1968 		SDT_PROBE0(vm, , , vm__lowmem_scan);
1969 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1970 
1971 		/*
1972 		 * We do this explicitly after the caches have been
1973 		 * drained above.
1974 		 */
1975 		uma_reclaim(UMA_RECLAIM_TRIM);
1976 		ret = true;
1977 		break;
1978 	}
1979 
1980 	/*
1981 	 * Kick off an asynchronous reclaim of cached memory if one of the
1982 	 * page daemons is failing to keep up with demand.  Use the "severe"
1983 	 * threshold instead of "min" to ensure that we do not blow away the
1984 	 * caches if a subset of the NUMA domains are depleted by kernel memory
1985 	 * allocations; the domainset iterators automatically skip domains
1986 	 * below the "min" threshold on the first pass.
1987 	 *
1988 	 * UMA reclaim worker has its own rate-limiting mechanism, so don't
1989 	 * worry about kicking it too often.
1990 	 */
1991 	if (vm_page_count_severe())
1992 		uma_reclaim_wakeup();
1993 
1994 	return (ret);
1995 }
1996 
1997 static void
1998 vm_pageout_worker(void *arg)
1999 {
2000 	struct vm_domain *vmd;
2001 	u_int ofree;
2002 	int addl_shortage, domain, shortage;
2003 	bool target_met;
2004 
2005 	domain = (uintptr_t)arg;
2006 	vmd = VM_DOMAIN(domain);
2007 	shortage = 0;
2008 	target_met = true;
2009 
2010 	/*
2011 	 * XXXKIB It could be useful to bind pageout daemon threads to
2012 	 * the cores belonging to the domain, from which vm_page_array
2013 	 * is allocated.
2014 	 */
2015 
2016 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
2017 	vmd->vmd_last_active_scan = ticks;
2018 
2019 	/*
2020 	 * The pageout daemon worker is never done, so loop forever.
2021 	 */
2022 	while (TRUE) {
2023 		vm_domain_pageout_lock(vmd);
2024 
2025 		/*
2026 		 * We need to clear wanted before we check the limits.  This
2027 		 * prevents races with wakers who will check wanted after they
2028 		 * reach the limit.
2029 		 */
2030 		atomic_store_int(&vmd->vmd_pageout_wanted, 0);
2031 
2032 		/*
2033 		 * Might the page daemon need to run again?
2034 		 */
2035 		if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
2036 			/*
2037 			 * Yes.  If the scan failed to produce enough free
2038 			 * pages, sleep uninterruptibly for some time in the
2039 			 * hope that the laundry thread will clean some pages.
2040 			 */
2041 			vm_domain_pageout_unlock(vmd);
2042 			if (!target_met)
2043 				pause("pwait", hz / VM_INACT_SCAN_RATE);
2044 		} else {
2045 			/*
2046 			 * No, sleep until the next wakeup or until pages
2047 			 * need to have their reference stats updated.
2048 			 */
2049 			if (mtx_sleep(&vmd->vmd_pageout_wanted,
2050 			    vm_domain_pageout_lockptr(vmd), PDROP | PVM,
2051 			    "psleep", hz / VM_INACT_SCAN_RATE) == 0)
2052 				VM_CNT_INC(v_pdwakeups);
2053 		}
2054 
2055 		/* Prevent spurious wakeups by ensuring that wanted is set. */
2056 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2057 
2058 		/*
2059 		 * Use the controller to calculate how many pages to free in
2060 		 * this interval, and scan the inactive queue.  If the lowmem
2061 		 * handlers appear to have freed up some pages, subtract the
2062 		 * difference from the inactive queue scan target.
2063 		 */
2064 		shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
2065 		if (shortage > 0) {
2066 			ofree = vmd->vmd_free_count;
2067 			if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
2068 				shortage -= min(vmd->vmd_free_count - ofree,
2069 				    (u_int)shortage);
2070 			target_met = vm_pageout_scan_inactive(vmd, shortage,
2071 			    &addl_shortage);
2072 		} else
2073 			addl_shortage = 0;
2074 
2075 		/*
2076 		 * Scan the active queue.  A positive value for shortage
2077 		 * indicates that we must aggressively deactivate pages to avoid
2078 		 * a shortfall.
2079 		 */
2080 		shortage = vm_pageout_active_target(vmd) + addl_shortage;
2081 		vm_pageout_scan_active(vmd, shortage);
2082 	}
2083 }
2084 
2085 /*
2086  * Initialize basic pageout daemon settings.  See the comment above the
2087  * definition of vm_domain for some explanation of how these thresholds are
2088  * used.
2089  */
2090 static void
2091 vm_pageout_init_domain(int domain)
2092 {
2093 	struct vm_domain *vmd;
2094 	struct sysctl_oid *oid;
2095 
2096 	vmd = VM_DOMAIN(domain);
2097 	vmd->vmd_interrupt_free_min = 2;
2098 
2099 	/*
2100 	 * v_free_reserved needs to include enough for the largest
2101 	 * swap pager structures plus enough for any pv_entry structs
2102 	 * when paging.
2103 	 */
2104 	vmd->vmd_pageout_free_min = 2 * MAXBSIZE / PAGE_SIZE +
2105 	    vmd->vmd_interrupt_free_min;
2106 	vmd->vmd_free_reserved = vm_pageout_page_count +
2107 	    vmd->vmd_pageout_free_min + vmd->vmd_page_count / 768;
2108 	vmd->vmd_free_min = vmd->vmd_page_count / 200;
2109 	vmd->vmd_free_severe = vmd->vmd_free_min / 2;
2110 	vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2111 	vmd->vmd_free_min += vmd->vmd_free_reserved;
2112 	vmd->vmd_free_severe += vmd->vmd_free_reserved;
2113 	vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2114 	if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2115 		vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2116 
2117 	/*
2118 	 * Set the default wakeup threshold to be 10% below the paging
2119 	 * target.  This keeps the steady state out of shortfall.
2120 	 */
2121 	vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2122 
2123 	/*
2124 	 * Target amount of memory to move out of the laundry queue during a
2125 	 * background laundering.  This is proportional to the amount of system
2126 	 * memory.
2127 	 */
2128 	vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2129 	    vmd->vmd_free_min) / 10;
2130 
2131 	/* Initialize the pageout daemon pid controller. */
2132 	pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2133 	    vmd->vmd_free_target, PIDCTRL_BOUND,
2134 	    PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2135 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2136 	    "pidctrl", CTLFLAG_RD, NULL, "");
2137 	pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2138 }
2139 
2140 static void
2141 vm_pageout_init(void)
2142 {
2143 	u_int freecount;
2144 	int i;
2145 
2146 	/*
2147 	 * Initialize some paging parameters.
2148 	 */
2149 	if (vm_cnt.v_page_count < 2000)
2150 		vm_pageout_page_count = 8;
2151 
2152 	freecount = 0;
2153 	for (i = 0; i < vm_ndomains; i++) {
2154 		struct vm_domain *vmd;
2155 
2156 		vm_pageout_init_domain(i);
2157 		vmd = VM_DOMAIN(i);
2158 		vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2159 		vm_cnt.v_free_target += vmd->vmd_free_target;
2160 		vm_cnt.v_free_min += vmd->vmd_free_min;
2161 		vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2162 		vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2163 		vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2164 		vm_cnt.v_free_severe += vmd->vmd_free_severe;
2165 		freecount += vmd->vmd_free_count;
2166 	}
2167 
2168 	/*
2169 	 * Set interval in seconds for active scan.  We want to visit each
2170 	 * page at least once every ten minutes.  This is to prevent worst
2171 	 * case paging behaviors with stale active LRU.
2172 	 */
2173 	if (vm_pageout_update_period == 0)
2174 		vm_pageout_update_period = 600;
2175 
2176 	if (vm_page_max_user_wired == 0)
2177 		vm_page_max_user_wired = freecount / 3;
2178 }
2179 
2180 /*
2181  *     vm_pageout is the high level pageout daemon.
2182  */
2183 static void
2184 vm_pageout(void)
2185 {
2186 	struct proc *p;
2187 	struct thread *td;
2188 	int error, first, i;
2189 
2190 	p = curproc;
2191 	td = curthread;
2192 
2193 	mtx_init(&vm_oom_ratelim_mtx, "vmoomr", NULL, MTX_DEF);
2194 	swap_pager_swap_init();
2195 	for (first = -1, i = 0; i < vm_ndomains; i++) {
2196 		if (VM_DOMAIN_EMPTY(i)) {
2197 			if (bootverbose)
2198 				printf("domain %d empty; skipping pageout\n",
2199 				    i);
2200 			continue;
2201 		}
2202 		if (first == -1)
2203 			first = i;
2204 		else {
2205 			error = kthread_add(vm_pageout_worker,
2206 			    (void *)(uintptr_t)i, p, NULL, 0, 0, "dom%d", i);
2207 			if (error != 0)
2208 				panic("starting pageout for domain %d: %d\n",
2209 				    i, error);
2210 		}
2211 		error = kthread_add(vm_pageout_laundry_worker,
2212 		    (void *)(uintptr_t)i, p, NULL, 0, 0, "laundry: dom%d", i);
2213 		if (error != 0)
2214 			panic("starting laundry for domain %d: %d", i, error);
2215 	}
2216 	error = kthread_add(uma_reclaim_worker, NULL, p, NULL, 0, 0, "uma");
2217 	if (error != 0)
2218 		panic("starting uma_reclaim helper, error %d\n", error);
2219 
2220 	snprintf(td->td_name, sizeof(td->td_name), "dom%d", first);
2221 	vm_pageout_worker((void *)(uintptr_t)first);
2222 }
2223 
2224 /*
2225  * Perform an advisory wakeup of the page daemon.
2226  */
2227 void
2228 pagedaemon_wakeup(int domain)
2229 {
2230 	struct vm_domain *vmd;
2231 
2232 	vmd = VM_DOMAIN(domain);
2233 	vm_domain_pageout_assert_unlocked(vmd);
2234 	if (curproc == pageproc)
2235 		return;
2236 
2237 	if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2238 		vm_domain_pageout_lock(vmd);
2239 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2240 		wakeup(&vmd->vmd_pageout_wanted);
2241 		vm_domain_pageout_unlock(vmd);
2242 	}
2243 }
2244