xref: /freebsd/sys/vm/vm_pageout.c (revision d3d381b2b194b4d24853e92eecef55f262688d1a)
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 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
198 SYSCTL_INT(_vm, OID_AUTO, max_wired,
199 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
200 
201 static u_int isqrt(u_int num);
202 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
203     bool in_shortfall);
204 static void vm_pageout_laundry_worker(void *arg);
205 
206 struct scan_state {
207 	struct vm_batchqueue bq;
208 	struct vm_pagequeue *pq;
209 	vm_page_t	marker;
210 	int		maxscan;
211 	int		scanned;
212 };
213 
214 static void
215 vm_pageout_init_scan(struct scan_state *ss, struct vm_pagequeue *pq,
216     vm_page_t marker, vm_page_t after, int maxscan)
217 {
218 
219 	vm_pagequeue_assert_locked(pq);
220 	KASSERT((marker->aflags & PGA_ENQUEUED) == 0,
221 	    ("marker %p already enqueued", marker));
222 
223 	if (after == NULL)
224 		TAILQ_INSERT_HEAD(&pq->pq_pl, marker, plinks.q);
225 	else
226 		TAILQ_INSERT_AFTER(&pq->pq_pl, after, marker, plinks.q);
227 	vm_page_aflag_set(marker, PGA_ENQUEUED);
228 
229 	vm_batchqueue_init(&ss->bq);
230 	ss->pq = pq;
231 	ss->marker = marker;
232 	ss->maxscan = maxscan;
233 	ss->scanned = 0;
234 	vm_pagequeue_unlock(pq);
235 }
236 
237 static void
238 vm_pageout_end_scan(struct scan_state *ss)
239 {
240 	struct vm_pagequeue *pq;
241 
242 	pq = ss->pq;
243 	vm_pagequeue_assert_locked(pq);
244 	KASSERT((ss->marker->aflags & PGA_ENQUEUED) != 0,
245 	    ("marker %p not enqueued", ss->marker));
246 
247 	TAILQ_REMOVE(&pq->pq_pl, ss->marker, plinks.q);
248 	vm_page_aflag_clear(ss->marker, PGA_ENQUEUED);
249 	pq->pq_pdpages += ss->scanned;
250 }
251 
252 /*
253  * Add a small number of queued pages to a batch queue for later processing
254  * without the corresponding queue lock held.  The caller must have enqueued a
255  * marker page at the desired start point for the scan.  Pages will be
256  * physically dequeued if the caller so requests.  Otherwise, the returned
257  * batch may contain marker pages, and it is up to the caller to handle them.
258  *
259  * When processing the batch queue, vm_page_queue() must be used to
260  * determine whether the page has been logically dequeued by another thread.
261  * Once this check is performed, the page lock guarantees that the page will
262  * not be disassociated from the queue.
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;
269 
270 	marker = ss->marker;
271 	pq = ss->pq;
272 
273 	KASSERT((marker->aflags & 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 = TAILQ_NEXT(m, plinks.q), ss->scanned++) {
280 		if ((m->flags & PG_MARKER) == 0) {
281 			KASSERT((m->aflags & PGA_ENQUEUED) != 0,
282 			    ("page %p not enqueued", m));
283 			KASSERT((m->flags & PG_FICTITIOUS) == 0,
284 			    ("Fictitious page %p cannot be in page queue", m));
285 			KASSERT((m->oflags & VPO_UNMANAGED) == 0,
286 			    ("Unmanaged page %p cannot be in page queue", m));
287 		} else if (dequeue)
288 			continue;
289 
290 		(void)vm_batchqueue_insert(&ss->bq, m);
291 		if (dequeue) {
292 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
293 			vm_page_aflag_clear(m, PGA_ENQUEUED);
294 		}
295 	}
296 	TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
297 	if (__predict_true(m != NULL))
298 		TAILQ_INSERT_BEFORE(m, marker, plinks.q);
299 	else
300 		TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
301 	if (dequeue)
302 		vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
303 	vm_pagequeue_unlock(pq);
304 }
305 
306 /* Return the next page to be scanned, or NULL if the scan is complete. */
307 static __always_inline vm_page_t
308 vm_pageout_next(struct scan_state *ss, const bool dequeue)
309 {
310 
311 	if (ss->bq.bq_cnt == 0)
312 		vm_pageout_collect_batch(ss, dequeue);
313 	return (vm_batchqueue_pop(&ss->bq));
314 }
315 
316 /*
317  * Scan for pages at adjacent offsets within the given page's object that are
318  * eligible for laundering, form a cluster of these pages and the given page,
319  * and launder that cluster.
320  */
321 static int
322 vm_pageout_cluster(vm_page_t m)
323 {
324 	vm_object_t object;
325 	vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
326 	vm_pindex_t pindex;
327 	int ib, is, page_base, pageout_count;
328 
329 	vm_page_assert_locked(m);
330 	object = m->object;
331 	VM_OBJECT_ASSERT_WLOCKED(object);
332 	pindex = m->pindex;
333 
334 	vm_page_assert_unbusied(m);
335 	KASSERT(!vm_page_held(m), ("page %p is held", m));
336 
337 	pmap_remove_write(m);
338 	vm_page_unlock(m);
339 
340 	mc[vm_pageout_page_count] = pb = ps = m;
341 	pageout_count = 1;
342 	page_base = vm_pageout_page_count;
343 	ib = 1;
344 	is = 1;
345 
346 	/*
347 	 * We can cluster only if the page is not clean, busy, or held, and
348 	 * the page is in the laundry queue.
349 	 *
350 	 * During heavy mmap/modification loads the pageout
351 	 * daemon can really fragment the underlying file
352 	 * due to flushing pages out of order and not trying to
353 	 * align the clusters (which leaves sporadic out-of-order
354 	 * holes).  To solve this problem we do the reverse scan
355 	 * first and attempt to align our cluster, then do a
356 	 * forward scan if room remains.
357 	 */
358 more:
359 	while (ib != 0 && pageout_count < vm_pageout_page_count) {
360 		if (ib > pindex) {
361 			ib = 0;
362 			break;
363 		}
364 		if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
365 			ib = 0;
366 			break;
367 		}
368 		vm_page_test_dirty(p);
369 		if (p->dirty == 0) {
370 			ib = 0;
371 			break;
372 		}
373 		vm_page_lock(p);
374 		if (vm_page_held(p) || !vm_page_in_laundry(p)) {
375 			vm_page_unlock(p);
376 			ib = 0;
377 			break;
378 		}
379 		pmap_remove_write(p);
380 		vm_page_unlock(p);
381 		mc[--page_base] = pb = p;
382 		++pageout_count;
383 		++ib;
384 
385 		/*
386 		 * We are at an alignment boundary.  Stop here, and switch
387 		 * directions.  Do not clear ib.
388 		 */
389 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
390 			break;
391 	}
392 	while (pageout_count < vm_pageout_page_count &&
393 	    pindex + is < object->size) {
394 		if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
395 			break;
396 		vm_page_test_dirty(p);
397 		if (p->dirty == 0)
398 			break;
399 		vm_page_lock(p);
400 		if (vm_page_held(p) || !vm_page_in_laundry(p)) {
401 			vm_page_unlock(p);
402 			break;
403 		}
404 		pmap_remove_write(p);
405 		vm_page_unlock(p);
406 		mc[page_base + pageout_count] = ps = p;
407 		++pageout_count;
408 		++is;
409 	}
410 
411 	/*
412 	 * If we exhausted our forward scan, continue with the reverse scan
413 	 * when possible, even past an alignment boundary.  This catches
414 	 * boundary conditions.
415 	 */
416 	if (ib != 0 && pageout_count < vm_pageout_page_count)
417 		goto more;
418 
419 	return (vm_pageout_flush(&mc[page_base], pageout_count,
420 	    VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
421 }
422 
423 /*
424  * vm_pageout_flush() - launder the given pages
425  *
426  *	The given pages are laundered.  Note that we setup for the start of
427  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
428  *	reference count all in here rather then in the parent.  If we want
429  *	the parent to do more sophisticated things we may have to change
430  *	the ordering.
431  *
432  *	Returned runlen is the count of pages between mreq and first
433  *	page after mreq with status VM_PAGER_AGAIN.
434  *	*eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
435  *	for any page in runlen set.
436  */
437 int
438 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
439     boolean_t *eio)
440 {
441 	vm_object_t object = mc[0]->object;
442 	int pageout_status[count];
443 	int numpagedout = 0;
444 	int i, runlen;
445 
446 	VM_OBJECT_ASSERT_WLOCKED(object);
447 
448 	/*
449 	 * Initiate I/O.  Mark the pages busy and verify that they're valid
450 	 * and read-only.
451 	 *
452 	 * We do not have to fixup the clean/dirty bits here... we can
453 	 * allow the pager to do it after the I/O completes.
454 	 *
455 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
456 	 * edge case with file fragments.
457 	 */
458 	for (i = 0; i < count; i++) {
459 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
460 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
461 			mc[i], i, count));
462 		KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
463 		    ("vm_pageout_flush: writeable page %p", mc[i]));
464 		vm_page_sbusy(mc[i]);
465 	}
466 	vm_object_pip_add(object, count);
467 
468 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
469 
470 	runlen = count - mreq;
471 	if (eio != NULL)
472 		*eio = FALSE;
473 	for (i = 0; i < count; i++) {
474 		vm_page_t mt = mc[i];
475 
476 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
477 		    !pmap_page_is_write_mapped(mt),
478 		    ("vm_pageout_flush: page %p is not write protected", mt));
479 		switch (pageout_status[i]) {
480 		case VM_PAGER_OK:
481 			vm_page_lock(mt);
482 			if (vm_page_in_laundry(mt))
483 				vm_page_deactivate_noreuse(mt);
484 			vm_page_unlock(mt);
485 			/* FALLTHROUGH */
486 		case VM_PAGER_PEND:
487 			numpagedout++;
488 			break;
489 		case VM_PAGER_BAD:
490 			/*
491 			 * The page is outside the object's range.  We pretend
492 			 * that the page out worked and clean the page, so the
493 			 * changes will be lost if the page is reclaimed by
494 			 * the page daemon.
495 			 */
496 			vm_page_undirty(mt);
497 			vm_page_lock(mt);
498 			if (vm_page_in_laundry(mt))
499 				vm_page_deactivate_noreuse(mt);
500 			vm_page_unlock(mt);
501 			break;
502 		case VM_PAGER_ERROR:
503 		case VM_PAGER_FAIL:
504 			/*
505 			 * If the page couldn't be paged out to swap because the
506 			 * pager wasn't able to find space, place the page in
507 			 * the PQ_UNSWAPPABLE holding queue.  This is an
508 			 * optimization that prevents the page daemon from
509 			 * wasting CPU cycles on pages that cannot be reclaimed
510 			 * becase no swap device is configured.
511 			 *
512 			 * Otherwise, reactivate the page so that it doesn't
513 			 * clog the laundry and inactive queues.  (We will try
514 			 * paging it out again later.)
515 			 */
516 			vm_page_lock(mt);
517 			if (object->type == OBJT_SWAP &&
518 			    pageout_status[i] == VM_PAGER_FAIL) {
519 				vm_page_unswappable(mt);
520 				numpagedout++;
521 			} else
522 				vm_page_activate(mt);
523 			vm_page_unlock(mt);
524 			if (eio != NULL && i >= mreq && i - mreq < runlen)
525 				*eio = TRUE;
526 			break;
527 		case VM_PAGER_AGAIN:
528 			if (i >= mreq && i - mreq < runlen)
529 				runlen = i - mreq;
530 			break;
531 		}
532 
533 		/*
534 		 * If the operation is still going, leave the page busy to
535 		 * block all other accesses. Also, leave the paging in
536 		 * progress indicator set so that we don't attempt an object
537 		 * collapse.
538 		 */
539 		if (pageout_status[i] != VM_PAGER_PEND) {
540 			vm_object_pip_wakeup(object);
541 			vm_page_sunbusy(mt);
542 		}
543 	}
544 	if (prunlen != NULL)
545 		*prunlen = runlen;
546 	return (numpagedout);
547 }
548 
549 static void
550 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
551 {
552 
553 	atomic_store_rel_int(&swapdev_enabled, 1);
554 }
555 
556 static void
557 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
558 {
559 
560 	if (swap_pager_nswapdev() == 1)
561 		atomic_store_rel_int(&swapdev_enabled, 0);
562 }
563 
564 /*
565  * Attempt to acquire all of the necessary locks to launder a page and
566  * then call through the clustering layer to PUTPAGES.  Wait a short
567  * time for a vnode lock.
568  *
569  * Requires the page and object lock on entry, releases both before return.
570  * Returns 0 on success and an errno otherwise.
571  */
572 static int
573 vm_pageout_clean(vm_page_t m, int *numpagedout)
574 {
575 	struct vnode *vp;
576 	struct mount *mp;
577 	vm_object_t object;
578 	vm_pindex_t pindex;
579 	int error, lockmode;
580 
581 	vm_page_assert_locked(m);
582 	object = m->object;
583 	VM_OBJECT_ASSERT_WLOCKED(object);
584 	error = 0;
585 	vp = NULL;
586 	mp = NULL;
587 
588 	/*
589 	 * The object is already known NOT to be dead.   It
590 	 * is possible for the vget() to block the whole
591 	 * pageout daemon, but the new low-memory handling
592 	 * code should prevent it.
593 	 *
594 	 * We can't wait forever for the vnode lock, we might
595 	 * deadlock due to a vn_read() getting stuck in
596 	 * vm_wait while holding this vnode.  We skip the
597 	 * vnode if we can't get it in a reasonable amount
598 	 * of time.
599 	 */
600 	if (object->type == OBJT_VNODE) {
601 		vm_page_unlock(m);
602 		vp = object->handle;
603 		if (vp->v_type == VREG &&
604 		    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
605 			mp = NULL;
606 			error = EDEADLK;
607 			goto unlock_all;
608 		}
609 		KASSERT(mp != NULL,
610 		    ("vp %p with NULL v_mount", vp));
611 		vm_object_reference_locked(object);
612 		pindex = m->pindex;
613 		VM_OBJECT_WUNLOCK(object);
614 		lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
615 		    LK_SHARED : LK_EXCLUSIVE;
616 		if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
617 			vp = NULL;
618 			error = EDEADLK;
619 			goto unlock_mp;
620 		}
621 		VM_OBJECT_WLOCK(object);
622 
623 		/*
624 		 * Ensure that the object and vnode were not disassociated
625 		 * while locks were dropped.
626 		 */
627 		if (vp->v_object != object) {
628 			error = ENOENT;
629 			goto unlock_all;
630 		}
631 		vm_page_lock(m);
632 
633 		/*
634 		 * While the object and page were unlocked, the page
635 		 * may have been:
636 		 * (1) moved to a different queue,
637 		 * (2) reallocated to a different object,
638 		 * (3) reallocated to a different offset, or
639 		 * (4) cleaned.
640 		 */
641 		if (!vm_page_in_laundry(m) || m->object != object ||
642 		    m->pindex != pindex || m->dirty == 0) {
643 			vm_page_unlock(m);
644 			error = ENXIO;
645 			goto unlock_all;
646 		}
647 
648 		/*
649 		 * The page may have been busied or referenced while the object
650 		 * and page locks were released.
651 		 */
652 		if (vm_page_busied(m) || vm_page_held(m)) {
653 			vm_page_unlock(m);
654 			error = EBUSY;
655 			goto unlock_all;
656 		}
657 	}
658 
659 	/*
660 	 * If a page is dirty, then it is either being washed
661 	 * (but not yet cleaned) or it is still in the
662 	 * laundry.  If it is still in the laundry, then we
663 	 * start the cleaning operation.
664 	 */
665 	if ((*numpagedout = vm_pageout_cluster(m)) == 0)
666 		error = EIO;
667 
668 unlock_all:
669 	VM_OBJECT_WUNLOCK(object);
670 
671 unlock_mp:
672 	vm_page_lock_assert(m, MA_NOTOWNED);
673 	if (mp != NULL) {
674 		if (vp != NULL)
675 			vput(vp);
676 		vm_object_deallocate(object);
677 		vn_finished_write(mp);
678 	}
679 
680 	return (error);
681 }
682 
683 /*
684  * Attempt to launder the specified number of pages.
685  *
686  * Returns the number of pages successfully laundered.
687  */
688 static int
689 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
690 {
691 	struct scan_state ss;
692 	struct vm_pagequeue *pq;
693 	struct mtx *mtx;
694 	vm_object_t object;
695 	vm_page_t m, marker;
696 	int act_delta, error, numpagedout, queue, starting_target;
697 	int vnodes_skipped;
698 	bool obj_locked, pageout_ok;
699 
700 	mtx = NULL;
701 	obj_locked = false;
702 	object = NULL;
703 	starting_target = launder;
704 	vnodes_skipped = 0;
705 
706 	/*
707 	 * Scan the laundry queues for pages eligible to be laundered.  We stop
708 	 * once the target number of dirty pages have been laundered, or once
709 	 * we've reached the end of the queue.  A single iteration of this loop
710 	 * may cause more than one page to be laundered because of clustering.
711 	 *
712 	 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
713 	 * swap devices are configured.
714 	 */
715 	if (atomic_load_acq_int(&swapdev_enabled))
716 		queue = PQ_UNSWAPPABLE;
717 	else
718 		queue = PQ_LAUNDRY;
719 
720 scan:
721 	marker = &vmd->vmd_markers[queue];
722 	pq = &vmd->vmd_pagequeues[queue];
723 	vm_pagequeue_lock(pq);
724 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
725 	while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
726 		if (__predict_false((m->flags & PG_MARKER) != 0))
727 			continue;
728 
729 		vm_page_change_lock(m, &mtx);
730 
731 recheck:
732 		/*
733 		 * The page may have been disassociated from the queue
734 		 * while locks were dropped.
735 		 */
736 		if (vm_page_queue(m) != queue)
737 			continue;
738 
739 		/*
740 		 * A requeue was requested, so this page gets a second
741 		 * chance.
742 		 */
743 		if ((m->aflags & PGA_REQUEUE) != 0) {
744 			vm_page_requeue(m);
745 			continue;
746 		}
747 
748 		/*
749 		 * Held pages are essentially stuck in the queue.
750 		 *
751 		 * Wired pages may not be freed.  Complete their removal
752 		 * from the queue now to avoid needless revisits during
753 		 * future scans.
754 		 */
755 		if (m->hold_count != 0)
756 			continue;
757 		if (m->wire_count != 0) {
758 			vm_page_dequeue_deferred(m);
759 			continue;
760 		}
761 
762 		if (object != m->object) {
763 			if (obj_locked) {
764 				VM_OBJECT_WUNLOCK(object);
765 				obj_locked = false;
766 			}
767 			object = m->object;
768 		}
769 		if (!obj_locked) {
770 			if (!VM_OBJECT_TRYWLOCK(object)) {
771 				mtx_unlock(mtx);
772 				/* Depends on type-stability. */
773 				VM_OBJECT_WLOCK(object);
774 				obj_locked = true;
775 				mtx_lock(mtx);
776 				goto recheck;
777 			} else
778 				obj_locked = true;
779 		}
780 
781 		if (vm_page_busied(m))
782 			continue;
783 
784 		/*
785 		 * Invalid pages can be easily freed.  They cannot be
786 		 * mapped; vm_page_free() asserts this.
787 		 */
788 		if (m->valid == 0)
789 			goto free_page;
790 
791 		/*
792 		 * If the page has been referenced and the object is not dead,
793 		 * reactivate or requeue the page depending on whether the
794 		 * object is mapped.
795 		 *
796 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
797 		 * that a reference from a concurrently destroyed mapping is
798 		 * observed here and now.
799 		 */
800 		if (object->ref_count != 0)
801 			act_delta = pmap_ts_referenced(m);
802 		else {
803 			KASSERT(!pmap_page_is_mapped(m),
804 			    ("page %p is mapped", m));
805 			act_delta = 0;
806 		}
807 		if ((m->aflags & PGA_REFERENCED) != 0) {
808 			vm_page_aflag_clear(m, PGA_REFERENCED);
809 			act_delta++;
810 		}
811 		if (act_delta != 0) {
812 			if (object->ref_count != 0) {
813 				VM_CNT_INC(v_reactivated);
814 				vm_page_activate(m);
815 
816 				/*
817 				 * Increase the activation count if the page
818 				 * was referenced while in the laundry queue.
819 				 * This makes it less likely that the page will
820 				 * be returned prematurely to the inactive
821 				 * queue.
822  				 */
823 				m->act_count += act_delta + ACT_ADVANCE;
824 
825 				/*
826 				 * If this was a background laundering, count
827 				 * activated pages towards our target.  The
828 				 * purpose of background laundering is to ensure
829 				 * that pages are eventually cycled through the
830 				 * laundry queue, and an activation is a valid
831 				 * way out.
832 				 */
833 				if (!in_shortfall)
834 					launder--;
835 				continue;
836 			} else if ((object->flags & OBJ_DEAD) == 0) {
837 				vm_page_requeue(m);
838 				continue;
839 			}
840 		}
841 
842 		/*
843 		 * If the page appears to be clean at the machine-independent
844 		 * layer, then remove all of its mappings from the pmap in
845 		 * anticipation of freeing it.  If, however, any of the page's
846 		 * mappings allow write access, then the page may still be
847 		 * modified until the last of those mappings are removed.
848 		 */
849 		if (object->ref_count != 0) {
850 			vm_page_test_dirty(m);
851 			if (m->dirty == 0)
852 				pmap_remove_all(m);
853 		}
854 
855 		/*
856 		 * Clean pages are freed, and dirty pages are paged out unless
857 		 * they belong to a dead object.  Requeueing dirty pages from
858 		 * dead objects is pointless, as they are being paged out and
859 		 * freed by the thread that destroyed the object.
860 		 */
861 		if (m->dirty == 0) {
862 free_page:
863 			vm_page_free(m);
864 			VM_CNT_INC(v_dfree);
865 		} else if ((object->flags & OBJ_DEAD) == 0) {
866 			if (object->type != OBJT_SWAP &&
867 			    object->type != OBJT_DEFAULT)
868 				pageout_ok = true;
869 			else if (disable_swap_pageouts)
870 				pageout_ok = false;
871 			else
872 				pageout_ok = true;
873 			if (!pageout_ok) {
874 				vm_page_requeue(m);
875 				continue;
876 			}
877 
878 			/*
879 			 * Form a cluster with adjacent, dirty pages from the
880 			 * same object, and page out that entire cluster.
881 			 *
882 			 * The adjacent, dirty pages must also be in the
883 			 * laundry.  However, their mappings are not checked
884 			 * for new references.  Consequently, a recently
885 			 * referenced page may be paged out.  However, that
886 			 * page will not be prematurely reclaimed.  After page
887 			 * out, the page will be placed in the inactive queue,
888 			 * where any new references will be detected and the
889 			 * page reactivated.
890 			 */
891 			error = vm_pageout_clean(m, &numpagedout);
892 			if (error == 0) {
893 				launder -= numpagedout;
894 				ss.scanned += numpagedout;
895 			} else if (error == EDEADLK) {
896 				pageout_lock_miss++;
897 				vnodes_skipped++;
898 			}
899 			mtx = NULL;
900 			obj_locked = false;
901 		}
902 	}
903 	if (mtx != NULL) {
904 		mtx_unlock(mtx);
905 		mtx = NULL;
906 	}
907 	if (obj_locked) {
908 		VM_OBJECT_WUNLOCK(object);
909 		obj_locked = false;
910 	}
911 	vm_pagequeue_lock(pq);
912 	vm_pageout_end_scan(&ss);
913 	vm_pagequeue_unlock(pq);
914 
915 	if (launder > 0 && queue == PQ_UNSWAPPABLE) {
916 		queue = PQ_LAUNDRY;
917 		goto scan;
918 	}
919 
920 	/*
921 	 * Wakeup the sync daemon if we skipped a vnode in a writeable object
922 	 * and we didn't launder enough pages.
923 	 */
924 	if (vnodes_skipped > 0 && launder > 0)
925 		(void)speedup_syncer();
926 
927 	return (starting_target - launder);
928 }
929 
930 /*
931  * Compute the integer square root.
932  */
933 static u_int
934 isqrt(u_int num)
935 {
936 	u_int bit, root, tmp;
937 
938 	bit = 1u << ((NBBY * sizeof(u_int)) - 2);
939 	while (bit > num)
940 		bit >>= 2;
941 	root = 0;
942 	while (bit != 0) {
943 		tmp = root + bit;
944 		root >>= 1;
945 		if (num >= tmp) {
946 			num -= tmp;
947 			root += bit;
948 		}
949 		bit >>= 2;
950 	}
951 	return (root);
952 }
953 
954 /*
955  * Perform the work of the laundry thread: periodically wake up and determine
956  * whether any pages need to be laundered.  If so, determine the number of pages
957  * that need to be laundered, and launder them.
958  */
959 static void
960 vm_pageout_laundry_worker(void *arg)
961 {
962 	struct vm_domain *vmd;
963 	struct vm_pagequeue *pq;
964 	uint64_t nclean, ndirty, nfreed;
965 	int domain, last_target, launder, shortfall, shortfall_cycle, target;
966 	bool in_shortfall;
967 
968 	domain = (uintptr_t)arg;
969 	vmd = VM_DOMAIN(domain);
970 	pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
971 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
972 
973 	shortfall = 0;
974 	in_shortfall = false;
975 	shortfall_cycle = 0;
976 	target = 0;
977 	nfreed = 0;
978 
979 	/*
980 	 * Calls to these handlers are serialized by the swap syscall lock.
981 	 */
982 	(void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
983 	    EVENTHANDLER_PRI_ANY);
984 	(void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
985 	    EVENTHANDLER_PRI_ANY);
986 
987 	/*
988 	 * The pageout laundry worker is never done, so loop forever.
989 	 */
990 	for (;;) {
991 		KASSERT(target >= 0, ("negative target %d", target));
992 		KASSERT(shortfall_cycle >= 0,
993 		    ("negative cycle %d", shortfall_cycle));
994 		launder = 0;
995 
996 		/*
997 		 * First determine whether we need to launder pages to meet a
998 		 * shortage of free pages.
999 		 */
1000 		if (shortfall > 0) {
1001 			in_shortfall = true;
1002 			shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1003 			target = shortfall;
1004 		} else if (!in_shortfall)
1005 			goto trybackground;
1006 		else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1007 			/*
1008 			 * We recently entered shortfall and began laundering
1009 			 * pages.  If we have completed that laundering run
1010 			 * (and we are no longer in shortfall) or we have met
1011 			 * our laundry target through other activity, then we
1012 			 * can stop laundering pages.
1013 			 */
1014 			in_shortfall = false;
1015 			target = 0;
1016 			goto trybackground;
1017 		}
1018 		launder = target / shortfall_cycle--;
1019 		goto dolaundry;
1020 
1021 		/*
1022 		 * There's no immediate need to launder any pages; see if we
1023 		 * meet the conditions to perform background laundering:
1024 		 *
1025 		 * 1. The ratio of dirty to clean inactive pages exceeds the
1026 		 *    background laundering threshold, or
1027 		 * 2. we haven't yet reached the target of the current
1028 		 *    background laundering run.
1029 		 *
1030 		 * The background laundering threshold is not a constant.
1031 		 * Instead, it is a slowly growing function of the number of
1032 		 * clean pages freed by the page daemon since the last
1033 		 * background laundering.  Thus, as the ratio of dirty to
1034 		 * clean inactive pages grows, the amount of memory pressure
1035 		 * required to trigger laundering decreases.  We ensure
1036 		 * that the threshold is non-zero after an inactive queue
1037 		 * scan, even if that scan failed to free a single clean page.
1038 		 */
1039 trybackground:
1040 		nclean = vmd->vmd_free_count +
1041 		    vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1042 		ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1043 		if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1044 		    vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1045 			target = vmd->vmd_background_launder_target;
1046 		}
1047 
1048 		/*
1049 		 * We have a non-zero background laundering target.  If we've
1050 		 * laundered up to our maximum without observing a page daemon
1051 		 * request, just stop.  This is a safety belt that ensures we
1052 		 * don't launder an excessive amount if memory pressure is low
1053 		 * and the ratio of dirty to clean pages is large.  Otherwise,
1054 		 * proceed at the background laundering rate.
1055 		 */
1056 		if (target > 0) {
1057 			if (nfreed > 0) {
1058 				nfreed = 0;
1059 				last_target = target;
1060 			} else if (last_target - target >=
1061 			    vm_background_launder_max * PAGE_SIZE / 1024) {
1062 				target = 0;
1063 			}
1064 			launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1065 			launder /= VM_LAUNDER_RATE;
1066 			if (launder > target)
1067 				launder = target;
1068 		}
1069 
1070 dolaundry:
1071 		if (launder > 0) {
1072 			/*
1073 			 * Because of I/O clustering, the number of laundered
1074 			 * pages could exceed "target" by the maximum size of
1075 			 * a cluster minus one.
1076 			 */
1077 			target -= min(vm_pageout_launder(vmd, launder,
1078 			    in_shortfall), target);
1079 			pause("laundp", hz / VM_LAUNDER_RATE);
1080 		}
1081 
1082 		/*
1083 		 * If we're not currently laundering pages and the page daemon
1084 		 * hasn't posted a new request, sleep until the page daemon
1085 		 * kicks us.
1086 		 */
1087 		vm_pagequeue_lock(pq);
1088 		if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1089 			(void)mtx_sleep(&vmd->vmd_laundry_request,
1090 			    vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1091 
1092 		/*
1093 		 * If the pagedaemon has indicated that it's in shortfall, start
1094 		 * a shortfall laundering unless we're already in the middle of
1095 		 * one.  This may preempt a background laundering.
1096 		 */
1097 		if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1098 		    (!in_shortfall || shortfall_cycle == 0)) {
1099 			shortfall = vm_laundry_target(vmd) +
1100 			    vmd->vmd_pageout_deficit;
1101 			target = 0;
1102 		} else
1103 			shortfall = 0;
1104 
1105 		if (target == 0)
1106 			vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1107 		nfreed += vmd->vmd_clean_pages_freed;
1108 		vmd->vmd_clean_pages_freed = 0;
1109 		vm_pagequeue_unlock(pq);
1110 	}
1111 }
1112 
1113 /*
1114  * Compute the number of pages we want to try to move from the
1115  * active queue to either the inactive or laundry queue.
1116  *
1117  * When scanning active pages during a shortage, we make clean pages
1118  * count more heavily towards the page shortage than dirty pages.
1119  * This is because dirty pages must be laundered before they can be
1120  * reused and thus have less utility when attempting to quickly
1121  * alleviate a free page shortage.  However, this weighting also
1122  * causes the scan to deactivate dirty pages more aggressively,
1123  * improving the effectiveness of clustering.
1124  */
1125 static int
1126 vm_pageout_active_target(struct vm_domain *vmd)
1127 {
1128 	int shortage;
1129 
1130 	shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1131 	    (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1132 	    vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1133 	shortage *= act_scan_laundry_weight;
1134 	return (shortage);
1135 }
1136 
1137 /*
1138  * Scan the active queue.  If there is no shortage of inactive pages, scan a
1139  * small portion of the queue in order to maintain quasi-LRU.
1140  */
1141 static void
1142 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1143 {
1144 	struct scan_state ss;
1145 	struct mtx *mtx;
1146 	vm_page_t m, marker;
1147 	struct vm_pagequeue *pq;
1148 	long min_scan;
1149 	int act_delta, max_scan, scan_tick;
1150 
1151 	marker = &vmd->vmd_markers[PQ_ACTIVE];
1152 	pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1153 	vm_pagequeue_lock(pq);
1154 
1155 	/*
1156 	 * If we're just idle polling attempt to visit every
1157 	 * active page within 'update_period' seconds.
1158 	 */
1159 	scan_tick = ticks;
1160 	if (vm_pageout_update_period != 0) {
1161 		min_scan = pq->pq_cnt;
1162 		min_scan *= scan_tick - vmd->vmd_last_active_scan;
1163 		min_scan /= hz * vm_pageout_update_period;
1164 	} else
1165 		min_scan = 0;
1166 	if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1167 		vmd->vmd_last_active_scan = scan_tick;
1168 
1169 	/*
1170 	 * Scan the active queue for pages that can be deactivated.  Update
1171 	 * the per-page activity counter and use it to identify deactivation
1172 	 * candidates.  Held pages may be deactivated.
1173 	 *
1174 	 * To avoid requeuing each page that remains in the active queue, we
1175 	 * implement the CLOCK algorithm.  To keep the implementation of the
1176 	 * enqueue operation consistent for all page queues, we use two hands,
1177 	 * represented by marker pages. Scans begin at the first hand, which
1178 	 * precedes the second hand in the queue.  When the two hands meet,
1179 	 * they are moved back to the head and tail of the queue, respectively,
1180 	 * and scanning resumes.
1181 	 */
1182 	max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1183 	mtx = NULL;
1184 act_scan:
1185 	vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1186 	while ((m = vm_pageout_next(&ss, false)) != NULL) {
1187 		if (__predict_false(m == &vmd->vmd_clock[1])) {
1188 			vm_pagequeue_lock(pq);
1189 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1190 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1191 			TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1192 			    plinks.q);
1193 			TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1194 			    plinks.q);
1195 			max_scan -= ss.scanned;
1196 			vm_pageout_end_scan(&ss);
1197 			goto act_scan;
1198 		}
1199 		if (__predict_false((m->flags & PG_MARKER) != 0))
1200 			continue;
1201 
1202 		vm_page_change_lock(m, &mtx);
1203 
1204 		/*
1205 		 * The page may have been disassociated from the queue
1206 		 * while locks were dropped.
1207 		 */
1208 		if (vm_page_queue(m) != PQ_ACTIVE)
1209 			continue;
1210 
1211 		/*
1212 		 * Wired pages are dequeued lazily.
1213 		 */
1214 		if (m->wire_count != 0) {
1215 			vm_page_dequeue_deferred(m);
1216 			continue;
1217 		}
1218 
1219 		/*
1220 		 * Check to see "how much" the page has been used.
1221 		 *
1222 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1223 		 * that a reference from a concurrently destroyed mapping is
1224 		 * observed here and now.
1225 		 *
1226 		 * Perform an unsynchronized object ref count check.  While
1227 		 * the page lock ensures that the page is not reallocated to
1228 		 * another object, in particular, one with unmanaged mappings
1229 		 * that cannot support pmap_ts_referenced(), two races are,
1230 		 * nonetheless, possible:
1231 		 * 1) The count was transitioning to zero, but we saw a non-
1232 		 *    zero value.  pmap_ts_referenced() will return zero
1233 		 *    because the page is not mapped.
1234 		 * 2) The count was transitioning to one, but we saw zero.
1235 		 *    This race delays the detection of a new reference.  At
1236 		 *    worst, we will deactivate and reactivate the page.
1237 		 */
1238 		if (m->object->ref_count != 0)
1239 			act_delta = pmap_ts_referenced(m);
1240 		else
1241 			act_delta = 0;
1242 		if ((m->aflags & PGA_REFERENCED) != 0) {
1243 			vm_page_aflag_clear(m, PGA_REFERENCED);
1244 			act_delta++;
1245 		}
1246 
1247 		/*
1248 		 * Advance or decay the act_count based on recent usage.
1249 		 */
1250 		if (act_delta != 0) {
1251 			m->act_count += ACT_ADVANCE + act_delta;
1252 			if (m->act_count > ACT_MAX)
1253 				m->act_count = ACT_MAX;
1254 		} else
1255 			m->act_count -= min(m->act_count, ACT_DECLINE);
1256 
1257 		if (m->act_count == 0) {
1258 			/*
1259 			 * When not short for inactive pages, let dirty pages go
1260 			 * through the inactive queue before moving to the
1261 			 * laundry queues.  This gives them some extra time to
1262 			 * be reactivated, potentially avoiding an expensive
1263 			 * pageout.  However, during a page shortage, the
1264 			 * inactive queue is necessarily small, and so dirty
1265 			 * pages would only spend a trivial amount of time in
1266 			 * the inactive queue.  Therefore, we might as well
1267 			 * place them directly in the laundry queue to reduce
1268 			 * queuing overhead.
1269 			 */
1270 			if (page_shortage <= 0)
1271 				vm_page_deactivate(m);
1272 			else {
1273 				/*
1274 				 * Calling vm_page_test_dirty() here would
1275 				 * require acquisition of the object's write
1276 				 * lock.  However, during a page shortage,
1277 				 * directing dirty pages into the laundry
1278 				 * queue is only an optimization and not a
1279 				 * requirement.  Therefore, we simply rely on
1280 				 * the opportunistic updates to the page's
1281 				 * dirty field by the pmap.
1282 				 */
1283 				if (m->dirty == 0) {
1284 					vm_page_deactivate(m);
1285 					page_shortage -=
1286 					    act_scan_laundry_weight;
1287 				} else {
1288 					vm_page_launder(m);
1289 					page_shortage--;
1290 				}
1291 			}
1292 		}
1293 	}
1294 	if (mtx != NULL) {
1295 		mtx_unlock(mtx);
1296 		mtx = NULL;
1297 	}
1298 	vm_pagequeue_lock(pq);
1299 	TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1300 	TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1301 	vm_pageout_end_scan(&ss);
1302 	vm_pagequeue_unlock(pq);
1303 }
1304 
1305 static int
1306 vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m)
1307 {
1308 	struct vm_domain *vmd;
1309 
1310 	if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0)
1311 		return (0);
1312 	vm_page_aflag_set(m, PGA_ENQUEUED);
1313 	if ((m->aflags & PGA_REQUEUE_HEAD) != 0) {
1314 		vmd = vm_pagequeue_domain(m);
1315 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
1316 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1317 	} else if ((m->aflags & PGA_REQUEUE) != 0) {
1318 		TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q);
1319 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1320 	} else
1321 		TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q);
1322 	return (1);
1323 }
1324 
1325 /*
1326  * Re-add stuck pages to the inactive queue.  We will examine them again
1327  * during the next scan.  If the queue state of a page has changed since
1328  * it was physically removed from the page queue in
1329  * vm_pageout_collect_batch(), don't do anything with that page.
1330  */
1331 static void
1332 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1333     vm_page_t m)
1334 {
1335 	struct vm_pagequeue *pq;
1336 	int delta;
1337 
1338 	delta = 0;
1339 	pq = ss->pq;
1340 
1341 	if (m != NULL) {
1342 		if (vm_batchqueue_insert(bq, m))
1343 			return;
1344 		vm_pagequeue_lock(pq);
1345 		delta += vm_pageout_reinsert_inactive_page(ss, m);
1346 	} else
1347 		vm_pagequeue_lock(pq);
1348 	while ((m = vm_batchqueue_pop(bq)) != NULL)
1349 		delta += vm_pageout_reinsert_inactive_page(ss, m);
1350 	vm_pagequeue_cnt_add(pq, delta);
1351 	vm_pagequeue_unlock(pq);
1352 	vm_batchqueue_init(bq);
1353 }
1354 
1355 /*
1356  * Attempt to reclaim the requested number of pages from the inactive queue.
1357  * Returns true if the shortage was addressed.
1358  */
1359 static int
1360 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1361     int *addl_shortage)
1362 {
1363 	struct scan_state ss;
1364 	struct vm_batchqueue rq;
1365 	struct mtx *mtx;
1366 	vm_page_t m, marker;
1367 	struct vm_pagequeue *pq;
1368 	vm_object_t object;
1369 	int act_delta, addl_page_shortage, deficit, page_shortage;
1370 	int starting_page_shortage;
1371 	bool obj_locked;
1372 
1373 	/*
1374 	 * The addl_page_shortage is an estimate of the number of temporarily
1375 	 * stuck pages in the inactive queue.  In other words, the
1376 	 * number of pages from the inactive count that should be
1377 	 * discounted in setting the target for the active queue scan.
1378 	 */
1379 	addl_page_shortage = 0;
1380 
1381 	/*
1382 	 * vmd_pageout_deficit counts the number of pages requested in
1383 	 * allocations that failed because of a free page shortage.  We assume
1384 	 * that the allocations will be reattempted and thus include the deficit
1385 	 * in our scan target.
1386 	 */
1387 	deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1388 	starting_page_shortage = page_shortage = shortage + deficit;
1389 
1390 	mtx = NULL;
1391 	obj_locked = false;
1392 	object = NULL;
1393 	vm_batchqueue_init(&rq);
1394 
1395 	/*
1396 	 * Start scanning the inactive queue for pages that we can free.  The
1397 	 * scan will stop when we reach the target or we have scanned the
1398 	 * entire queue.  (Note that m->act_count is not used to make
1399 	 * decisions for the inactive queue, only for the active queue.)
1400 	 */
1401 	marker = &vmd->vmd_markers[PQ_INACTIVE];
1402 	pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1403 	vm_pagequeue_lock(pq);
1404 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1405 	while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1406 		KASSERT((m->flags & PG_MARKER) == 0,
1407 		    ("marker page %p was dequeued", m));
1408 
1409 		vm_page_change_lock(m, &mtx);
1410 
1411 recheck:
1412 		/*
1413 		 * The page may have been disassociated from the queue
1414 		 * while locks were dropped.
1415 		 */
1416 		if (vm_page_queue(m) != PQ_INACTIVE) {
1417 			addl_page_shortage++;
1418 			continue;
1419 		}
1420 
1421 		/*
1422 		 * The page was re-enqueued after the page queue lock was
1423 		 * dropped, or a requeue was requested.  This page gets a second
1424 		 * chance.
1425 		 */
1426 		if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE |
1427 		    PGA_REQUEUE_HEAD)) != 0)
1428 			goto reinsert;
1429 
1430 		/*
1431 		 * Held pages are essentially stuck in the queue.  So,
1432 		 * they ought to be discounted from the inactive count.
1433 		 * See the description of addl_page_shortage above.
1434 		 *
1435 		 * Wired pages may not be freed.  Complete their removal
1436 		 * from the queue now to avoid needless revisits during
1437 		 * future scans.
1438 		 */
1439 		if (m->hold_count != 0) {
1440 			addl_page_shortage++;
1441 			goto reinsert;
1442 		}
1443 		if (m->wire_count != 0) {
1444 			vm_page_dequeue_deferred(m);
1445 			continue;
1446 		}
1447 
1448 		if (object != m->object) {
1449 			if (obj_locked) {
1450 				VM_OBJECT_WUNLOCK(object);
1451 				obj_locked = false;
1452 			}
1453 			object = m->object;
1454 		}
1455 		if (!obj_locked) {
1456 			if (!VM_OBJECT_TRYWLOCK(object)) {
1457 				mtx_unlock(mtx);
1458 				/* Depends on type-stability. */
1459 				VM_OBJECT_WLOCK(object);
1460 				obj_locked = true;
1461 				mtx_lock(mtx);
1462 				goto recheck;
1463 			} else
1464 				obj_locked = true;
1465 		}
1466 
1467 		if (vm_page_busied(m)) {
1468 			/*
1469 			 * Don't mess with busy pages.  Leave them at
1470 			 * the front of the queue.  Most likely, they
1471 			 * are being paged out and will leave the
1472 			 * queue shortly after the scan finishes.  So,
1473 			 * they ought to be discounted from the
1474 			 * inactive count.
1475 			 */
1476 			addl_page_shortage++;
1477 			goto reinsert;
1478 		}
1479 
1480 		/*
1481 		 * Invalid pages can be easily freed. They cannot be
1482 		 * mapped, vm_page_free() asserts this.
1483 		 */
1484 		if (m->valid == 0)
1485 			goto free_page;
1486 
1487 		/*
1488 		 * If the page has been referenced and the object is not dead,
1489 		 * reactivate or requeue the page depending on whether the
1490 		 * object is mapped.
1491 		 *
1492 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1493 		 * that a reference from a concurrently destroyed mapping is
1494 		 * observed here and now.
1495 		 */
1496 		if (object->ref_count != 0)
1497 			act_delta = pmap_ts_referenced(m);
1498 		else {
1499 			KASSERT(!pmap_page_is_mapped(m),
1500 			    ("page %p is mapped", m));
1501 			act_delta = 0;
1502 		}
1503 		if ((m->aflags & PGA_REFERENCED) != 0) {
1504 			vm_page_aflag_clear(m, PGA_REFERENCED);
1505 			act_delta++;
1506 		}
1507 		if (act_delta != 0) {
1508 			if (object->ref_count != 0) {
1509 				VM_CNT_INC(v_reactivated);
1510 				vm_page_activate(m);
1511 
1512 				/*
1513 				 * Increase the activation count if the page
1514 				 * was referenced while in the inactive queue.
1515 				 * This makes it less likely that the page will
1516 				 * be returned prematurely to the inactive
1517 				 * queue.
1518  				 */
1519 				m->act_count += act_delta + ACT_ADVANCE;
1520 				continue;
1521 			} else if ((object->flags & OBJ_DEAD) == 0) {
1522 				vm_page_aflag_set(m, PGA_REQUEUE);
1523 				goto reinsert;
1524 			}
1525 		}
1526 
1527 		/*
1528 		 * If the page appears to be clean at the machine-independent
1529 		 * layer, then remove all of its mappings from the pmap in
1530 		 * anticipation of freeing it.  If, however, any of the page's
1531 		 * mappings allow write access, then the page may still be
1532 		 * modified until the last of those mappings are removed.
1533 		 */
1534 		if (object->ref_count != 0) {
1535 			vm_page_test_dirty(m);
1536 			if (m->dirty == 0)
1537 				pmap_remove_all(m);
1538 		}
1539 
1540 		/*
1541 		 * Clean pages can be freed, but dirty pages must be sent back
1542 		 * to the laundry, unless they belong to a dead object.
1543 		 * Requeueing dirty pages from dead objects is pointless, as
1544 		 * they are being paged out and freed by the thread that
1545 		 * destroyed the object.
1546 		 */
1547 		if (m->dirty == 0) {
1548 free_page:
1549 			/*
1550 			 * Because we dequeued the page and have already
1551 			 * checked for concurrent dequeue and enqueue
1552 			 * requests, we can safely disassociate the page
1553 			 * from the inactive queue.
1554 			 */
1555 			KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
1556 			    ("page %p has queue state", m));
1557 			m->queue = PQ_NONE;
1558 			vm_page_free(m);
1559 			page_shortage--;
1560 		} else if ((object->flags & OBJ_DEAD) == 0)
1561 			vm_page_launder(m);
1562 		continue;
1563 reinsert:
1564 		vm_pageout_reinsert_inactive(&ss, &rq, m);
1565 	}
1566 	if (mtx != NULL) {
1567 		mtx_unlock(mtx);
1568 		mtx = NULL;
1569 	}
1570 	if (obj_locked) {
1571 		VM_OBJECT_WUNLOCK(object);
1572 		obj_locked = false;
1573 	}
1574 	vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1575 	vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1576 	vm_pagequeue_lock(pq);
1577 	vm_pageout_end_scan(&ss);
1578 	vm_pagequeue_unlock(pq);
1579 
1580 	VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1581 
1582 	/*
1583 	 * Wake up the laundry thread so that it can perform any needed
1584 	 * laundering.  If we didn't meet our target, we're in shortfall and
1585 	 * need to launder more aggressively.  If PQ_LAUNDRY is empty and no
1586 	 * swap devices are configured, the laundry thread has no work to do, so
1587 	 * don't bother waking it up.
1588 	 *
1589 	 * The laundry thread uses the number of inactive queue scans elapsed
1590 	 * since the last laundering to determine whether to launder again, so
1591 	 * keep count.
1592 	 */
1593 	if (starting_page_shortage > 0) {
1594 		pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1595 		vm_pagequeue_lock(pq);
1596 		if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1597 		    (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1598 			if (page_shortage > 0) {
1599 				vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1600 				VM_CNT_INC(v_pdshortfalls);
1601 			} else if (vmd->vmd_laundry_request !=
1602 			    VM_LAUNDRY_SHORTFALL)
1603 				vmd->vmd_laundry_request =
1604 				    VM_LAUNDRY_BACKGROUND;
1605 			wakeup(&vmd->vmd_laundry_request);
1606 		}
1607 		vmd->vmd_clean_pages_freed +=
1608 		    starting_page_shortage - page_shortage;
1609 		vm_pagequeue_unlock(pq);
1610 	}
1611 
1612 	/*
1613 	 * Wakeup the swapout daemon if we didn't free the targeted number of
1614 	 * pages.
1615 	 */
1616 	if (page_shortage > 0)
1617 		vm_swapout_run();
1618 
1619 	/*
1620 	 * If the inactive queue scan fails repeatedly to meet its
1621 	 * target, kill the largest process.
1622 	 */
1623 	vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1624 
1625 	/*
1626 	 * Reclaim pages by swapping out idle processes, if configured to do so.
1627 	 */
1628 	vm_swapout_run_idle();
1629 
1630 	/*
1631 	 * See the description of addl_page_shortage above.
1632 	 */
1633 	*addl_shortage = addl_page_shortage + deficit;
1634 
1635 	return (page_shortage <= 0);
1636 }
1637 
1638 static int vm_pageout_oom_vote;
1639 
1640 /*
1641  * The pagedaemon threads randlomly select one to perform the
1642  * OOM.  Trying to kill processes before all pagedaemons
1643  * failed to reach free target is premature.
1644  */
1645 static void
1646 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1647     int starting_page_shortage)
1648 {
1649 	int old_vote;
1650 
1651 	if (starting_page_shortage <= 0 || starting_page_shortage !=
1652 	    page_shortage)
1653 		vmd->vmd_oom_seq = 0;
1654 	else
1655 		vmd->vmd_oom_seq++;
1656 	if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1657 		if (vmd->vmd_oom) {
1658 			vmd->vmd_oom = FALSE;
1659 			atomic_subtract_int(&vm_pageout_oom_vote, 1);
1660 		}
1661 		return;
1662 	}
1663 
1664 	/*
1665 	 * Do not follow the call sequence until OOM condition is
1666 	 * cleared.
1667 	 */
1668 	vmd->vmd_oom_seq = 0;
1669 
1670 	if (vmd->vmd_oom)
1671 		return;
1672 
1673 	vmd->vmd_oom = TRUE;
1674 	old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1675 	if (old_vote != vm_ndomains - 1)
1676 		return;
1677 
1678 	/*
1679 	 * The current pagedaemon thread is the last in the quorum to
1680 	 * start OOM.  Initiate the selection and signaling of the
1681 	 * victim.
1682 	 */
1683 	vm_pageout_oom(VM_OOM_MEM);
1684 
1685 	/*
1686 	 * After one round of OOM terror, recall our vote.  On the
1687 	 * next pass, current pagedaemon would vote again if the low
1688 	 * memory condition is still there, due to vmd_oom being
1689 	 * false.
1690 	 */
1691 	vmd->vmd_oom = FALSE;
1692 	atomic_subtract_int(&vm_pageout_oom_vote, 1);
1693 }
1694 
1695 /*
1696  * The OOM killer is the page daemon's action of last resort when
1697  * memory allocation requests have been stalled for a prolonged period
1698  * of time because it cannot reclaim memory.  This function computes
1699  * the approximate number of physical pages that could be reclaimed if
1700  * the specified address space is destroyed.
1701  *
1702  * Private, anonymous memory owned by the address space is the
1703  * principal resource that we expect to recover after an OOM kill.
1704  * Since the physical pages mapped by the address space's COW entries
1705  * are typically shared pages, they are unlikely to be released and so
1706  * they are not counted.
1707  *
1708  * To get to the point where the page daemon runs the OOM killer, its
1709  * efforts to write-back vnode-backed pages may have stalled.  This
1710  * could be caused by a memory allocation deadlock in the write path
1711  * that might be resolved by an OOM kill.  Therefore, physical pages
1712  * belonging to vnode-backed objects are counted, because they might
1713  * be freed without being written out first if the address space holds
1714  * the last reference to an unlinked vnode.
1715  *
1716  * Similarly, physical pages belonging to OBJT_PHYS objects are
1717  * counted because the address space might hold the last reference to
1718  * the object.
1719  */
1720 static long
1721 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1722 {
1723 	vm_map_t map;
1724 	vm_map_entry_t entry;
1725 	vm_object_t obj;
1726 	long res;
1727 
1728 	map = &vmspace->vm_map;
1729 	KASSERT(!map->system_map, ("system map"));
1730 	sx_assert(&map->lock, SA_LOCKED);
1731 	res = 0;
1732 	for (entry = map->header.next; entry != &map->header;
1733 	    entry = entry->next) {
1734 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1735 			continue;
1736 		obj = entry->object.vm_object;
1737 		if (obj == NULL)
1738 			continue;
1739 		if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1740 		    obj->ref_count != 1)
1741 			continue;
1742 		switch (obj->type) {
1743 		case OBJT_DEFAULT:
1744 		case OBJT_SWAP:
1745 		case OBJT_PHYS:
1746 		case OBJT_VNODE:
1747 			res += obj->resident_page_count;
1748 			break;
1749 		}
1750 	}
1751 	return (res);
1752 }
1753 
1754 void
1755 vm_pageout_oom(int shortage)
1756 {
1757 	struct proc *p, *bigproc;
1758 	vm_offset_t size, bigsize;
1759 	struct thread *td;
1760 	struct vmspace *vm;
1761 	bool breakout;
1762 
1763 	/*
1764 	 * We keep the process bigproc locked once we find it to keep anyone
1765 	 * from messing with it; however, there is a possibility of
1766 	 * deadlock if process B is bigproc and one of its child processes
1767 	 * attempts to propagate a signal to B while we are waiting for A's
1768 	 * lock while walking this list.  To avoid this, we don't block on
1769 	 * the process lock but just skip a process if it is already locked.
1770 	 */
1771 	bigproc = NULL;
1772 	bigsize = 0;
1773 	sx_slock(&allproc_lock);
1774 	FOREACH_PROC_IN_SYSTEM(p) {
1775 		PROC_LOCK(p);
1776 
1777 		/*
1778 		 * If this is a system, protected or killed process, skip it.
1779 		 */
1780 		if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1781 		    P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1782 		    p->p_pid == 1 || P_KILLED(p) ||
1783 		    (p->p_pid < 48 && swap_pager_avail != 0)) {
1784 			PROC_UNLOCK(p);
1785 			continue;
1786 		}
1787 		/*
1788 		 * If the process is in a non-running type state,
1789 		 * don't touch it.  Check all the threads individually.
1790 		 */
1791 		breakout = false;
1792 		FOREACH_THREAD_IN_PROC(p, td) {
1793 			thread_lock(td);
1794 			if (!TD_ON_RUNQ(td) &&
1795 			    !TD_IS_RUNNING(td) &&
1796 			    !TD_IS_SLEEPING(td) &&
1797 			    !TD_IS_SUSPENDED(td) &&
1798 			    !TD_IS_SWAPPED(td)) {
1799 				thread_unlock(td);
1800 				breakout = true;
1801 				break;
1802 			}
1803 			thread_unlock(td);
1804 		}
1805 		if (breakout) {
1806 			PROC_UNLOCK(p);
1807 			continue;
1808 		}
1809 		/*
1810 		 * get the process size
1811 		 */
1812 		vm = vmspace_acquire_ref(p);
1813 		if (vm == NULL) {
1814 			PROC_UNLOCK(p);
1815 			continue;
1816 		}
1817 		_PHOLD_LITE(p);
1818 		PROC_UNLOCK(p);
1819 		sx_sunlock(&allproc_lock);
1820 		if (!vm_map_trylock_read(&vm->vm_map)) {
1821 			vmspace_free(vm);
1822 			sx_slock(&allproc_lock);
1823 			PRELE(p);
1824 			continue;
1825 		}
1826 		size = vmspace_swap_count(vm);
1827 		if (shortage == VM_OOM_MEM)
1828 			size += vm_pageout_oom_pagecount(vm);
1829 		vm_map_unlock_read(&vm->vm_map);
1830 		vmspace_free(vm);
1831 		sx_slock(&allproc_lock);
1832 
1833 		/*
1834 		 * If this process is bigger than the biggest one,
1835 		 * remember it.
1836 		 */
1837 		if (size > bigsize) {
1838 			if (bigproc != NULL)
1839 				PRELE(bigproc);
1840 			bigproc = p;
1841 			bigsize = size;
1842 		} else {
1843 			PRELE(p);
1844 		}
1845 	}
1846 	sx_sunlock(&allproc_lock);
1847 	if (bigproc != NULL) {
1848 		if (vm_panic_on_oom != 0)
1849 			panic("out of swap space");
1850 		PROC_LOCK(bigproc);
1851 		killproc(bigproc, "out of swap space");
1852 		sched_nice(bigproc, PRIO_MIN);
1853 		_PRELE(bigproc);
1854 		PROC_UNLOCK(bigproc);
1855 	}
1856 }
1857 
1858 static bool
1859 vm_pageout_lowmem(void)
1860 {
1861 	static int lowmem_ticks = 0;
1862 	int last;
1863 
1864 	last = atomic_load_int(&lowmem_ticks);
1865 	while ((u_int)(ticks - last) / hz >= lowmem_period) {
1866 		if (atomic_fcmpset_int(&lowmem_ticks, &last, ticks) == 0)
1867 			continue;
1868 
1869 		/*
1870 		 * Decrease registered cache sizes.
1871 		 */
1872 		SDT_PROBE0(vm, , , vm__lowmem_scan);
1873 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1874 
1875 		/*
1876 		 * We do this explicitly after the caches have been
1877 		 * drained above.
1878 		 */
1879 		uma_reclaim();
1880 		return (true);
1881 	}
1882 	return (false);
1883 }
1884 
1885 static void
1886 vm_pageout_worker(void *arg)
1887 {
1888 	struct vm_domain *vmd;
1889 	u_int ofree;
1890 	int addl_shortage, domain, shortage;
1891 	bool target_met;
1892 
1893 	domain = (uintptr_t)arg;
1894 	vmd = VM_DOMAIN(domain);
1895 	shortage = 0;
1896 	target_met = true;
1897 
1898 	/*
1899 	 * XXXKIB It could be useful to bind pageout daemon threads to
1900 	 * the cores belonging to the domain, from which vm_page_array
1901 	 * is allocated.
1902 	 */
1903 
1904 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1905 	vmd->vmd_last_active_scan = ticks;
1906 
1907 	/*
1908 	 * The pageout daemon worker is never done, so loop forever.
1909 	 */
1910 	while (TRUE) {
1911 		vm_domain_pageout_lock(vmd);
1912 
1913 		/*
1914 		 * We need to clear wanted before we check the limits.  This
1915 		 * prevents races with wakers who will check wanted after they
1916 		 * reach the limit.
1917 		 */
1918 		atomic_store_int(&vmd->vmd_pageout_wanted, 0);
1919 
1920 		/*
1921 		 * Might the page daemon need to run again?
1922 		 */
1923 		if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
1924 			/*
1925 			 * Yes.  If the scan failed to produce enough free
1926 			 * pages, sleep uninterruptibly for some time in the
1927 			 * hope that the laundry thread will clean some pages.
1928 			 */
1929 			vm_domain_pageout_unlock(vmd);
1930 			if (!target_met)
1931 				pause("pwait", hz / VM_INACT_SCAN_RATE);
1932 		} else {
1933 			/*
1934 			 * No, sleep until the next wakeup or until pages
1935 			 * need to have their reference stats updated.
1936 			 */
1937 			if (mtx_sleep(&vmd->vmd_pageout_wanted,
1938 			    vm_domain_pageout_lockptr(vmd), PDROP | PVM,
1939 			    "psleep", hz / VM_INACT_SCAN_RATE) == 0)
1940 				VM_CNT_INC(v_pdwakeups);
1941 		}
1942 
1943 		/* Prevent spurious wakeups by ensuring that wanted is set. */
1944 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1945 
1946 		/*
1947 		 * Use the controller to calculate how many pages to free in
1948 		 * this interval, and scan the inactive queue.  If the lowmem
1949 		 * handlers appear to have freed up some pages, subtract the
1950 		 * difference from the inactive queue scan target.
1951 		 */
1952 		shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
1953 		if (shortage > 0) {
1954 			ofree = vmd->vmd_free_count;
1955 			if (vm_pageout_lowmem() && vmd->vmd_free_count > ofree)
1956 				shortage -= min(vmd->vmd_free_count - ofree,
1957 				    (u_int)shortage);
1958 			target_met = vm_pageout_scan_inactive(vmd, shortage,
1959 			    &addl_shortage);
1960 		} else
1961 			addl_shortage = 0;
1962 
1963 		/*
1964 		 * Scan the active queue.  A positive value for shortage
1965 		 * indicates that we must aggressively deactivate pages to avoid
1966 		 * a shortfall.
1967 		 */
1968 		shortage = vm_pageout_active_target(vmd) + addl_shortage;
1969 		vm_pageout_scan_active(vmd, shortage);
1970 	}
1971 }
1972 
1973 /*
1974  *	vm_pageout_init initialises basic pageout daemon settings.
1975  */
1976 static void
1977 vm_pageout_init_domain(int domain)
1978 {
1979 	struct vm_domain *vmd;
1980 	struct sysctl_oid *oid;
1981 
1982 	vmd = VM_DOMAIN(domain);
1983 	vmd->vmd_interrupt_free_min = 2;
1984 
1985 	/*
1986 	 * v_free_reserved needs to include enough for the largest
1987 	 * swap pager structures plus enough for any pv_entry structs
1988 	 * when paging.
1989 	 */
1990 	if (vmd->vmd_page_count > 1024)
1991 		vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200;
1992 	else
1993 		vmd->vmd_free_min = 4;
1994 	vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1995 	    vmd->vmd_interrupt_free_min;
1996 	vmd->vmd_free_reserved = vm_pageout_page_count +
1997 	    vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768);
1998 	vmd->vmd_free_severe = vmd->vmd_free_min / 2;
1999 	vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
2000 	vmd->vmd_free_min += vmd->vmd_free_reserved;
2001 	vmd->vmd_free_severe += vmd->vmd_free_reserved;
2002 	vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
2003 	if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
2004 		vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
2005 
2006 	/*
2007 	 * Set the default wakeup threshold to be 10% below the paging
2008 	 * target.  This keeps the steady state out of shortfall.
2009 	 */
2010 	vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2011 
2012 	/*
2013 	 * Target amount of memory to move out of the laundry queue during a
2014 	 * background laundering.  This is proportional to the amount of system
2015 	 * memory.
2016 	 */
2017 	vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2018 	    vmd->vmd_free_min) / 10;
2019 
2020 	/* Initialize the pageout daemon pid controller. */
2021 	pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2022 	    vmd->vmd_free_target, PIDCTRL_BOUND,
2023 	    PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2024 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2025 	    "pidctrl", CTLFLAG_RD, NULL, "");
2026 	pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2027 }
2028 
2029 static void
2030 vm_pageout_init(void)
2031 {
2032 	u_int freecount;
2033 	int i;
2034 
2035 	/*
2036 	 * Initialize some paging parameters.
2037 	 */
2038 	if (vm_cnt.v_page_count < 2000)
2039 		vm_pageout_page_count = 8;
2040 
2041 	freecount = 0;
2042 	for (i = 0; i < vm_ndomains; i++) {
2043 		struct vm_domain *vmd;
2044 
2045 		vm_pageout_init_domain(i);
2046 		vmd = VM_DOMAIN(i);
2047 		vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2048 		vm_cnt.v_free_target += vmd->vmd_free_target;
2049 		vm_cnt.v_free_min += vmd->vmd_free_min;
2050 		vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2051 		vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2052 		vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2053 		vm_cnt.v_free_severe += vmd->vmd_free_severe;
2054 		freecount += vmd->vmd_free_count;
2055 	}
2056 
2057 	/*
2058 	 * Set interval in seconds for active scan.  We want to visit each
2059 	 * page at least once every ten minutes.  This is to prevent worst
2060 	 * case paging behaviors with stale active LRU.
2061 	 */
2062 	if (vm_pageout_update_period == 0)
2063 		vm_pageout_update_period = 600;
2064 
2065 	if (vm_page_max_wired == 0)
2066 		vm_page_max_wired = freecount / 3;
2067 }
2068 
2069 /*
2070  *     vm_pageout is the high level pageout daemon.
2071  */
2072 static void
2073 vm_pageout(void)
2074 {
2075 	int error;
2076 	int i;
2077 
2078 	swap_pager_swap_init();
2079 	snprintf(curthread->td_name, sizeof(curthread->td_name), "dom0");
2080 	error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
2081 	    0, 0, "laundry: dom0");
2082 	if (error != 0)
2083 		panic("starting laundry for domain 0, error %d", error);
2084 	for (i = 1; i < vm_ndomains; i++) {
2085 		error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
2086 		    curproc, NULL, 0, 0, "dom%d", i);
2087 		if (error != 0) {
2088 			panic("starting pageout for domain %d, error %d\n",
2089 			    i, error);
2090 		}
2091 		error = kthread_add(vm_pageout_laundry_worker,
2092 		    (void *)(uintptr_t)i, curproc, NULL, 0, 0,
2093 		    "laundry: dom%d", i);
2094 		if (error != 0)
2095 			panic("starting laundry for domain %d, error %d",
2096 			    i, error);
2097 	}
2098 	error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
2099 	    0, 0, "uma");
2100 	if (error != 0)
2101 		panic("starting uma_reclaim helper, error %d\n", error);
2102 	vm_pageout_worker((void *)(uintptr_t)0);
2103 }
2104 
2105 /*
2106  * Perform an advisory wakeup of the page daemon.
2107  */
2108 void
2109 pagedaemon_wakeup(int domain)
2110 {
2111 	struct vm_domain *vmd;
2112 
2113 	vmd = VM_DOMAIN(domain);
2114 	vm_domain_pageout_assert_unlocked(vmd);
2115 	if (curproc == pageproc)
2116 		return;
2117 
2118 	if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2119 		vm_domain_pageout_lock(vmd);
2120 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2121 		wakeup(&vmd->vmd_pageout_wanted);
2122 		vm_domain_pageout_unlock(vmd);
2123 	}
2124 }
2125