xref: /freebsd/sys/vm/vm_pageout.c (revision 2dc4dbb9673c9a3309c2dad59cb588c6f04beaea)
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 time_t lowmem_uptime;
156 static int swapdev_enabled;
157 
158 static int vm_panic_on_oom = 0;
159 
160 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
161 	CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
162 	"panic on out of memory instead of killing the largest process");
163 
164 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
165 	CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
166 	"Maximum active LRU update period");
167 
168 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
169 	"Low memory callback period");
170 
171 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
172 	CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
173 
174 static int pageout_lock_miss;
175 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
176 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
177 
178 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
179 	CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
180 	"back-to-back calls to oom detector to start OOM");
181 
182 static int act_scan_laundry_weight = 3;
183 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
184     &act_scan_laundry_weight, 0,
185     "weight given to clean vs. dirty pages in active queue scans");
186 
187 static u_int vm_background_launder_rate = 4096;
188 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
189     &vm_background_launder_rate, 0,
190     "background laundering rate, in kilobytes per second");
191 
192 static u_int vm_background_launder_max = 20 * 1024;
193 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
194     &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
195 
196 int vm_pageout_page_count = 32;
197 
198 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
199 SYSCTL_INT(_vm, OID_AUTO, max_wired,
200 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to 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->aflags & 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->aflags & 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 	VM_CNT_ADD(v_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;
270 
271 	marker = ss->marker;
272 	pq = ss->pq;
273 
274 	KASSERT((marker->aflags & 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 = TAILQ_NEXT(m, plinks.q), ss->scanned++) {
281 		if ((m->flags & PG_MARKER) == 0) {
282 			KASSERT((m->aflags & PGA_ENQUEUED) != 0,
283 			    ("page %p not enqueued", m));
284 			KASSERT((m->flags & PG_FICTITIOUS) == 0,
285 			    ("Fictitious page %p cannot be in page queue", m));
286 			KASSERT((m->oflags & VPO_UNMANAGED) == 0,
287 			    ("Unmanaged page %p cannot be in page queue", m));
288 		} else if (dequeue)
289 			continue;
290 
291 		(void)vm_batchqueue_insert(&ss->bq, m);
292 		if (dequeue) {
293 			TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
294 			vm_page_aflag_clear(m, PGA_ENQUEUED);
295 		}
296 	}
297 	TAILQ_REMOVE(&pq->pq_pl, marker, plinks.q);
298 	if (__predict_true(m != NULL))
299 		TAILQ_INSERT_BEFORE(m, marker, plinks.q);
300 	else
301 		TAILQ_INSERT_TAIL(&pq->pq_pl, marker, plinks.q);
302 	if (dequeue)
303 		vm_pagequeue_cnt_add(pq, -ss->bq.bq_cnt);
304 	vm_pagequeue_unlock(pq);
305 }
306 
307 /* Return the next page to be scanned, or NULL if the scan is complete. */
308 static __always_inline vm_page_t
309 vm_pageout_next(struct scan_state *ss, const bool dequeue)
310 {
311 
312 	if (ss->bq.bq_cnt == 0)
313 		vm_pageout_collect_batch(ss, dequeue);
314 	return (vm_batchqueue_pop(&ss->bq));
315 }
316 
317 /*
318  * Scan for pages at adjacent offsets within the given page's object that are
319  * eligible for laundering, form a cluster of these pages and the given page,
320  * and launder that cluster.
321  */
322 static int
323 vm_pageout_cluster(vm_page_t m)
324 {
325 	vm_object_t object;
326 	vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
327 	vm_pindex_t pindex;
328 	int ib, is, page_base, pageout_count;
329 
330 	vm_page_assert_locked(m);
331 	object = m->object;
332 	VM_OBJECT_ASSERT_WLOCKED(object);
333 	pindex = m->pindex;
334 
335 	vm_page_assert_unbusied(m);
336 	KASSERT(!vm_page_held(m), ("page %p is held", m));
337 
338 	pmap_remove_write(m);
339 	vm_page_unlock(m);
340 
341 	mc[vm_pageout_page_count] = pb = ps = m;
342 	pageout_count = 1;
343 	page_base = vm_pageout_page_count;
344 	ib = 1;
345 	is = 1;
346 
347 	/*
348 	 * We can cluster only if the page is not clean, busy, or held, and
349 	 * the page is in the laundry queue.
350 	 *
351 	 * During heavy mmap/modification loads the pageout
352 	 * daemon can really fragment the underlying file
353 	 * due to flushing pages out of order and not trying to
354 	 * align the clusters (which leaves sporadic out-of-order
355 	 * holes).  To solve this problem we do the reverse scan
356 	 * first and attempt to align our cluster, then do a
357 	 * forward scan if room remains.
358 	 */
359 more:
360 	while (ib != 0 && pageout_count < vm_pageout_page_count) {
361 		if (ib > pindex) {
362 			ib = 0;
363 			break;
364 		}
365 		if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
366 			ib = 0;
367 			break;
368 		}
369 		vm_page_test_dirty(p);
370 		if (p->dirty == 0) {
371 			ib = 0;
372 			break;
373 		}
374 		vm_page_lock(p);
375 		if (vm_page_held(p) || !vm_page_in_laundry(p)) {
376 			vm_page_unlock(p);
377 			ib = 0;
378 			break;
379 		}
380 		pmap_remove_write(p);
381 		vm_page_unlock(p);
382 		mc[--page_base] = pb = p;
383 		++pageout_count;
384 		++ib;
385 
386 		/*
387 		 * We are at an alignment boundary.  Stop here, and switch
388 		 * directions.  Do not clear ib.
389 		 */
390 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
391 			break;
392 	}
393 	while (pageout_count < vm_pageout_page_count &&
394 	    pindex + is < object->size) {
395 		if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
396 			break;
397 		vm_page_test_dirty(p);
398 		if (p->dirty == 0)
399 			break;
400 		vm_page_lock(p);
401 		if (vm_page_held(p) || !vm_page_in_laundry(p)) {
402 			vm_page_unlock(p);
403 			break;
404 		}
405 		pmap_remove_write(p);
406 		vm_page_unlock(p);
407 		mc[page_base + pageout_count] = ps = p;
408 		++pageout_count;
409 		++is;
410 	}
411 
412 	/*
413 	 * If we exhausted our forward scan, continue with the reverse scan
414 	 * when possible, even past an alignment boundary.  This catches
415 	 * boundary conditions.
416 	 */
417 	if (ib != 0 && pageout_count < vm_pageout_page_count)
418 		goto more;
419 
420 	return (vm_pageout_flush(&mc[page_base], pageout_count,
421 	    VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
422 }
423 
424 /*
425  * vm_pageout_flush() - launder the given pages
426  *
427  *	The given pages are laundered.  Note that we setup for the start of
428  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
429  *	reference count all in here rather then in the parent.  If we want
430  *	the parent to do more sophisticated things we may have to change
431  *	the ordering.
432  *
433  *	Returned runlen is the count of pages between mreq and first
434  *	page after mreq with status VM_PAGER_AGAIN.
435  *	*eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
436  *	for any page in runlen set.
437  */
438 int
439 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
440     boolean_t *eio)
441 {
442 	vm_object_t object = mc[0]->object;
443 	int pageout_status[count];
444 	int numpagedout = 0;
445 	int i, runlen;
446 
447 	VM_OBJECT_ASSERT_WLOCKED(object);
448 
449 	/*
450 	 * Initiate I/O.  Mark the pages busy and verify that they're valid
451 	 * and read-only.
452 	 *
453 	 * We do not have to fixup the clean/dirty bits here... we can
454 	 * allow the pager to do it after the I/O completes.
455 	 *
456 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
457 	 * edge case with file fragments.
458 	 */
459 	for (i = 0; i < count; i++) {
460 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
461 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
462 			mc[i], i, count));
463 		KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
464 		    ("vm_pageout_flush: writeable page %p", mc[i]));
465 		vm_page_sbusy(mc[i]);
466 	}
467 	vm_object_pip_add(object, count);
468 
469 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
470 
471 	runlen = count - mreq;
472 	if (eio != NULL)
473 		*eio = FALSE;
474 	for (i = 0; i < count; i++) {
475 		vm_page_t mt = mc[i];
476 
477 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
478 		    !pmap_page_is_write_mapped(mt),
479 		    ("vm_pageout_flush: page %p is not write protected", mt));
480 		switch (pageout_status[i]) {
481 		case VM_PAGER_OK:
482 			vm_page_lock(mt);
483 			if (vm_page_in_laundry(mt))
484 				vm_page_deactivate_noreuse(mt);
485 			vm_page_unlock(mt);
486 			/* FALLTHROUGH */
487 		case VM_PAGER_PEND:
488 			numpagedout++;
489 			break;
490 		case VM_PAGER_BAD:
491 			/*
492 			 * The page is outside the object's range.  We pretend
493 			 * that the page out worked and clean the page, so the
494 			 * changes will be lost if the page is reclaimed by
495 			 * the page daemon.
496 			 */
497 			vm_page_undirty(mt);
498 			vm_page_lock(mt);
499 			if (vm_page_in_laundry(mt))
500 				vm_page_deactivate_noreuse(mt);
501 			vm_page_unlock(mt);
502 			break;
503 		case VM_PAGER_ERROR:
504 		case VM_PAGER_FAIL:
505 			/*
506 			 * If the page couldn't be paged out to swap because the
507 			 * pager wasn't able to find space, place the page in
508 			 * the PQ_UNSWAPPABLE holding queue.  This is an
509 			 * optimization that prevents the page daemon from
510 			 * wasting CPU cycles on pages that cannot be reclaimed
511 			 * becase no swap device is configured.
512 			 *
513 			 * Otherwise, reactivate the page so that it doesn't
514 			 * clog the laundry and inactive queues.  (We will try
515 			 * paging it out again later.)
516 			 */
517 			vm_page_lock(mt);
518 			if (object->type == OBJT_SWAP &&
519 			    pageout_status[i] == VM_PAGER_FAIL) {
520 				vm_page_unswappable(mt);
521 				numpagedout++;
522 			} else
523 				vm_page_activate(mt);
524 			vm_page_unlock(mt);
525 			if (eio != NULL && i >= mreq && i - mreq < runlen)
526 				*eio = TRUE;
527 			break;
528 		case VM_PAGER_AGAIN:
529 			if (i >= mreq && i - mreq < runlen)
530 				runlen = i - mreq;
531 			break;
532 		}
533 
534 		/*
535 		 * If the operation is still going, leave the page busy to
536 		 * block all other accesses. Also, leave the paging in
537 		 * progress indicator set so that we don't attempt an object
538 		 * collapse.
539 		 */
540 		if (pageout_status[i] != VM_PAGER_PEND) {
541 			vm_object_pip_wakeup(object);
542 			vm_page_sunbusy(mt);
543 		}
544 	}
545 	if (prunlen != NULL)
546 		*prunlen = runlen;
547 	return (numpagedout);
548 }
549 
550 static void
551 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
552 {
553 
554 	atomic_store_rel_int(&swapdev_enabled, 1);
555 }
556 
557 static void
558 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
559 {
560 
561 	if (swap_pager_nswapdev() == 1)
562 		atomic_store_rel_int(&swapdev_enabled, 0);
563 }
564 
565 /*
566  * Attempt to acquire all of the necessary locks to launder a page and
567  * then call through the clustering layer to PUTPAGES.  Wait a short
568  * time for a vnode lock.
569  *
570  * Requires the page and object lock on entry, releases both before return.
571  * Returns 0 on success and an errno otherwise.
572  */
573 static int
574 vm_pageout_clean(vm_page_t m, int *numpagedout)
575 {
576 	struct vnode *vp;
577 	struct mount *mp;
578 	vm_object_t object;
579 	vm_pindex_t pindex;
580 	int error, lockmode;
581 
582 	vm_page_assert_locked(m);
583 	object = m->object;
584 	VM_OBJECT_ASSERT_WLOCKED(object);
585 	error = 0;
586 	vp = NULL;
587 	mp = NULL;
588 
589 	/*
590 	 * The object is already known NOT to be dead.   It
591 	 * is possible for the vget() to block the whole
592 	 * pageout daemon, but the new low-memory handling
593 	 * code should prevent it.
594 	 *
595 	 * We can't wait forever for the vnode lock, we might
596 	 * deadlock due to a vn_read() getting stuck in
597 	 * vm_wait while holding this vnode.  We skip the
598 	 * vnode if we can't get it in a reasonable amount
599 	 * of time.
600 	 */
601 	if (object->type == OBJT_VNODE) {
602 		vm_page_unlock(m);
603 		vp = object->handle;
604 		if (vp->v_type == VREG &&
605 		    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
606 			mp = NULL;
607 			error = EDEADLK;
608 			goto unlock_all;
609 		}
610 		KASSERT(mp != NULL,
611 		    ("vp %p with NULL v_mount", vp));
612 		vm_object_reference_locked(object);
613 		pindex = m->pindex;
614 		VM_OBJECT_WUNLOCK(object);
615 		lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
616 		    LK_SHARED : LK_EXCLUSIVE;
617 		if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
618 			vp = NULL;
619 			error = EDEADLK;
620 			goto unlock_mp;
621 		}
622 		VM_OBJECT_WLOCK(object);
623 
624 		/*
625 		 * Ensure that the object and vnode were not disassociated
626 		 * while locks were dropped.
627 		 */
628 		if (vp->v_object != object) {
629 			error = ENOENT;
630 			goto unlock_all;
631 		}
632 		vm_page_lock(m);
633 
634 		/*
635 		 * While the object and page were unlocked, the page
636 		 * may have been:
637 		 * (1) moved to a different queue,
638 		 * (2) reallocated to a different object,
639 		 * (3) reallocated to a different offset, or
640 		 * (4) cleaned.
641 		 */
642 		if (!vm_page_in_laundry(m) || m->object != object ||
643 		    m->pindex != pindex || m->dirty == 0) {
644 			vm_page_unlock(m);
645 			error = ENXIO;
646 			goto unlock_all;
647 		}
648 
649 		/*
650 		 * The page may have been busied or referenced while the object
651 		 * and page locks were released.
652 		 */
653 		if (vm_page_busied(m) || vm_page_held(m)) {
654 			vm_page_unlock(m);
655 			error = EBUSY;
656 			goto unlock_all;
657 		}
658 	}
659 
660 	/*
661 	 * If a page is dirty, then it is either being washed
662 	 * (but not yet cleaned) or it is still in the
663 	 * laundry.  If it is still in the laundry, then we
664 	 * start the cleaning operation.
665 	 */
666 	if ((*numpagedout = vm_pageout_cluster(m)) == 0)
667 		error = EIO;
668 
669 unlock_all:
670 	VM_OBJECT_WUNLOCK(object);
671 
672 unlock_mp:
673 	vm_page_lock_assert(m, MA_NOTOWNED);
674 	if (mp != NULL) {
675 		if (vp != NULL)
676 			vput(vp);
677 		vm_object_deallocate(object);
678 		vn_finished_write(mp);
679 	}
680 
681 	return (error);
682 }
683 
684 /*
685  * Attempt to launder the specified number of pages.
686  *
687  * Returns the number of pages successfully laundered.
688  */
689 static int
690 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
691 {
692 	struct scan_state ss;
693 	struct vm_pagequeue *pq;
694 	struct mtx *mtx;
695 	vm_object_t object;
696 	vm_page_t m, marker;
697 	int act_delta, error, numpagedout, queue, starting_target;
698 	int vnodes_skipped;
699 	bool obj_locked, pageout_ok;
700 
701 	mtx = NULL;
702 	obj_locked = false;
703 	object = NULL;
704 	starting_target = launder;
705 	vnodes_skipped = 0;
706 
707 	/*
708 	 * Scan the laundry queues for pages eligible to be laundered.  We stop
709 	 * once the target number of dirty pages have been laundered, or once
710 	 * we've reached the end of the queue.  A single iteration of this loop
711 	 * may cause more than one page to be laundered because of clustering.
712 	 *
713 	 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
714 	 * swap devices are configured.
715 	 */
716 	if (atomic_load_acq_int(&swapdev_enabled))
717 		queue = PQ_UNSWAPPABLE;
718 	else
719 		queue = PQ_LAUNDRY;
720 
721 scan:
722 	marker = &vmd->vmd_markers[queue];
723 	pq = &vmd->vmd_pagequeues[queue];
724 	vm_pagequeue_lock(pq);
725 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
726 	while (launder > 0 && (m = vm_pageout_next(&ss, false)) != NULL) {
727 		if (__predict_false((m->flags & PG_MARKER) != 0))
728 			continue;
729 
730 		vm_page_change_lock(m, &mtx);
731 
732 recheck:
733 		/*
734 		 * The page may have been disassociated from the queue
735 		 * while locks were dropped.
736 		 */
737 		if (vm_page_queue(m) != queue)
738 			continue;
739 
740 		/*
741 		 * A requeue was requested, so this page gets a second
742 		 * chance.
743 		 */
744 		if ((m->aflags & PGA_REQUEUE) != 0) {
745 			vm_page_requeue(m);
746 			continue;
747 		}
748 
749 		/*
750 		 * Held pages are essentially stuck in the queue.
751 		 *
752 		 * Wired pages may not be freed.  Complete their removal
753 		 * from the queue now to avoid needless revisits during
754 		 * future scans.
755 		 */
756 		if (m->hold_count != 0)
757 			continue;
758 		if (m->wire_count != 0) {
759 			vm_page_dequeue_deferred(m);
760 			continue;
761 		}
762 
763 		if (object != m->object) {
764 			if (obj_locked) {
765 				VM_OBJECT_WUNLOCK(object);
766 				obj_locked = false;
767 			}
768 			object = m->object;
769 		}
770 		if (!obj_locked) {
771 			if (!VM_OBJECT_TRYWLOCK(object)) {
772 				mtx_unlock(mtx);
773 				/* Depends on type-stability. */
774 				VM_OBJECT_WLOCK(object);
775 				obj_locked = true;
776 				mtx_lock(mtx);
777 				goto recheck;
778 			} else
779 				obj_locked = true;
780 		}
781 
782 		if (vm_page_busied(m))
783 			continue;
784 
785 		/*
786 		 * Invalid pages can be easily freed.  They cannot be
787 		 * mapped; vm_page_free() asserts this.
788 		 */
789 		if (m->valid == 0)
790 			goto free_page;
791 
792 		/*
793 		 * If the page has been referenced and the object is not dead,
794 		 * reactivate or requeue the page depending on whether the
795 		 * object is mapped.
796 		 *
797 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
798 		 * that a reference from a concurrently destroyed mapping is
799 		 * observed here and now.
800 		 */
801 		if (object->ref_count != 0)
802 			act_delta = pmap_ts_referenced(m);
803 		else {
804 			KASSERT(!pmap_page_is_mapped(m),
805 			    ("page %p is mapped", m));
806 			act_delta = 0;
807 		}
808 		if ((m->aflags & PGA_REFERENCED) != 0) {
809 			vm_page_aflag_clear(m, PGA_REFERENCED);
810 			act_delta++;
811 		}
812 		if (act_delta != 0) {
813 			if (object->ref_count != 0) {
814 				VM_CNT_INC(v_reactivated);
815 				vm_page_activate(m);
816 
817 				/*
818 				 * Increase the activation count if the page
819 				 * was referenced while in the laundry queue.
820 				 * This makes it less likely that the page will
821 				 * be returned prematurely to the inactive
822 				 * queue.
823  				 */
824 				m->act_count += act_delta + ACT_ADVANCE;
825 
826 				/*
827 				 * If this was a background laundering, count
828 				 * activated pages towards our target.  The
829 				 * purpose of background laundering is to ensure
830 				 * that pages are eventually cycled through the
831 				 * laundry queue, and an activation is a valid
832 				 * way out.
833 				 */
834 				if (!in_shortfall)
835 					launder--;
836 				continue;
837 			} else if ((object->flags & OBJ_DEAD) == 0) {
838 				vm_page_requeue(m);
839 				continue;
840 			}
841 		}
842 
843 		/*
844 		 * If the page appears to be clean at the machine-independent
845 		 * layer, then remove all of its mappings from the pmap in
846 		 * anticipation of freeing it.  If, however, any of the page's
847 		 * mappings allow write access, then the page may still be
848 		 * modified until the last of those mappings are removed.
849 		 */
850 		if (object->ref_count != 0) {
851 			vm_page_test_dirty(m);
852 			if (m->dirty == 0)
853 				pmap_remove_all(m);
854 		}
855 
856 		/*
857 		 * Clean pages are freed, and dirty pages are paged out unless
858 		 * they belong to a dead object.  Requeueing dirty pages from
859 		 * dead objects is pointless, as they are being paged out and
860 		 * freed by the thread that destroyed the object.
861 		 */
862 		if (m->dirty == 0) {
863 free_page:
864 			vm_page_free(m);
865 			VM_CNT_INC(v_dfree);
866 		} else if ((object->flags & OBJ_DEAD) == 0) {
867 			if (object->type != OBJT_SWAP &&
868 			    object->type != OBJT_DEFAULT)
869 				pageout_ok = true;
870 			else if (disable_swap_pageouts)
871 				pageout_ok = false;
872 			else
873 				pageout_ok = true;
874 			if (!pageout_ok) {
875 				vm_page_requeue(m);
876 				continue;
877 			}
878 
879 			/*
880 			 * Form a cluster with adjacent, dirty pages from the
881 			 * same object, and page out that entire cluster.
882 			 *
883 			 * The adjacent, dirty pages must also be in the
884 			 * laundry.  However, their mappings are not checked
885 			 * for new references.  Consequently, a recently
886 			 * referenced page may be paged out.  However, that
887 			 * page will not be prematurely reclaimed.  After page
888 			 * out, the page will be placed in the inactive queue,
889 			 * where any new references will be detected and the
890 			 * page reactivated.
891 			 */
892 			error = vm_pageout_clean(m, &numpagedout);
893 			if (error == 0) {
894 				launder -= numpagedout;
895 				ss.scanned += numpagedout;
896 			} else if (error == EDEADLK) {
897 				pageout_lock_miss++;
898 				vnodes_skipped++;
899 			}
900 			mtx = NULL;
901 			obj_locked = false;
902 		}
903 	}
904 	if (mtx != NULL) {
905 		mtx_unlock(mtx);
906 		mtx = NULL;
907 	}
908 	if (obj_locked) {
909 		VM_OBJECT_WUNLOCK(object);
910 		obj_locked = false;
911 	}
912 	vm_pagequeue_lock(pq);
913 	vm_pageout_end_scan(&ss);
914 	vm_pagequeue_unlock(pq);
915 
916 	if (launder > 0 && queue == PQ_UNSWAPPABLE) {
917 		queue = PQ_LAUNDRY;
918 		goto scan;
919 	}
920 
921 	/*
922 	 * Wakeup the sync daemon if we skipped a vnode in a writeable object
923 	 * and we didn't launder enough pages.
924 	 */
925 	if (vnodes_skipped > 0 && launder > 0)
926 		(void)speedup_syncer();
927 
928 	return (starting_target - launder);
929 }
930 
931 /*
932  * Compute the integer square root.
933  */
934 static u_int
935 isqrt(u_int num)
936 {
937 	u_int bit, root, tmp;
938 
939 	bit = 1u << ((NBBY * sizeof(u_int)) - 2);
940 	while (bit > num)
941 		bit >>= 2;
942 	root = 0;
943 	while (bit != 0) {
944 		tmp = root + bit;
945 		root >>= 1;
946 		if (num >= tmp) {
947 			num -= tmp;
948 			root += bit;
949 		}
950 		bit >>= 2;
951 	}
952 	return (root);
953 }
954 
955 /*
956  * Perform the work of the laundry thread: periodically wake up and determine
957  * whether any pages need to be laundered.  If so, determine the number of pages
958  * that need to be laundered, and launder them.
959  */
960 static void
961 vm_pageout_laundry_worker(void *arg)
962 {
963 	struct vm_domain *vmd;
964 	struct vm_pagequeue *pq;
965 	uint64_t nclean, ndirty, nfreed;
966 	int domain, last_target, launder, shortfall, shortfall_cycle, target;
967 	bool in_shortfall;
968 
969 	domain = (uintptr_t)arg;
970 	vmd = VM_DOMAIN(domain);
971 	pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
972 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
973 
974 	shortfall = 0;
975 	in_shortfall = false;
976 	shortfall_cycle = 0;
977 	target = 0;
978 	nfreed = 0;
979 
980 	/*
981 	 * Calls to these handlers are serialized by the swap syscall lock.
982 	 */
983 	(void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
984 	    EVENTHANDLER_PRI_ANY);
985 	(void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
986 	    EVENTHANDLER_PRI_ANY);
987 
988 	/*
989 	 * The pageout laundry worker is never done, so loop forever.
990 	 */
991 	for (;;) {
992 		KASSERT(target >= 0, ("negative target %d", target));
993 		KASSERT(shortfall_cycle >= 0,
994 		    ("negative cycle %d", shortfall_cycle));
995 		launder = 0;
996 
997 		/*
998 		 * First determine whether we need to launder pages to meet a
999 		 * shortage of free pages.
1000 		 */
1001 		if (shortfall > 0) {
1002 			in_shortfall = true;
1003 			shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
1004 			target = shortfall;
1005 		} else if (!in_shortfall)
1006 			goto trybackground;
1007 		else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
1008 			/*
1009 			 * We recently entered shortfall and began laundering
1010 			 * pages.  If we have completed that laundering run
1011 			 * (and we are no longer in shortfall) or we have met
1012 			 * our laundry target through other activity, then we
1013 			 * can stop laundering pages.
1014 			 */
1015 			in_shortfall = false;
1016 			target = 0;
1017 			goto trybackground;
1018 		}
1019 		launder = target / shortfall_cycle--;
1020 		goto dolaundry;
1021 
1022 		/*
1023 		 * There's no immediate need to launder any pages; see if we
1024 		 * meet the conditions to perform background laundering:
1025 		 *
1026 		 * 1. The ratio of dirty to clean inactive pages exceeds the
1027 		 *    background laundering threshold, or
1028 		 * 2. we haven't yet reached the target of the current
1029 		 *    background laundering run.
1030 		 *
1031 		 * The background laundering threshold is not a constant.
1032 		 * Instead, it is a slowly growing function of the number of
1033 		 * clean pages freed by the page daemon since the last
1034 		 * background laundering.  Thus, as the ratio of dirty to
1035 		 * clean inactive pages grows, the amount of memory pressure
1036 		 * required to trigger laundering decreases.  We ensure
1037 		 * that the threshold is non-zero after an inactive queue
1038 		 * scan, even if that scan failed to free a single clean page.
1039 		 */
1040 trybackground:
1041 		nclean = vmd->vmd_free_count +
1042 		    vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1043 		ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1044 		if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1045 		    vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1046 			target = vmd->vmd_background_launder_target;
1047 		}
1048 
1049 		/*
1050 		 * We have a non-zero background laundering target.  If we've
1051 		 * laundered up to our maximum without observing a page daemon
1052 		 * request, just stop.  This is a safety belt that ensures we
1053 		 * don't launder an excessive amount if memory pressure is low
1054 		 * and the ratio of dirty to clean pages is large.  Otherwise,
1055 		 * proceed at the background laundering rate.
1056 		 */
1057 		if (target > 0) {
1058 			if (nfreed > 0) {
1059 				nfreed = 0;
1060 				last_target = target;
1061 			} else if (last_target - target >=
1062 			    vm_background_launder_max * PAGE_SIZE / 1024) {
1063 				target = 0;
1064 			}
1065 			launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1066 			launder /= VM_LAUNDER_RATE;
1067 			if (launder > target)
1068 				launder = target;
1069 		}
1070 
1071 dolaundry:
1072 		if (launder > 0) {
1073 			/*
1074 			 * Because of I/O clustering, the number of laundered
1075 			 * pages could exceed "target" by the maximum size of
1076 			 * a cluster minus one.
1077 			 */
1078 			target -= min(vm_pageout_launder(vmd, launder,
1079 			    in_shortfall), target);
1080 			pause("laundp", hz / VM_LAUNDER_RATE);
1081 		}
1082 
1083 		/*
1084 		 * If we're not currently laundering pages and the page daemon
1085 		 * hasn't posted a new request, sleep until the page daemon
1086 		 * kicks us.
1087 		 */
1088 		vm_pagequeue_lock(pq);
1089 		if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1090 			(void)mtx_sleep(&vmd->vmd_laundry_request,
1091 			    vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1092 
1093 		/*
1094 		 * If the pagedaemon has indicated that it's in shortfall, start
1095 		 * a shortfall laundering unless we're already in the middle of
1096 		 * one.  This may preempt a background laundering.
1097 		 */
1098 		if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1099 		    (!in_shortfall || shortfall_cycle == 0)) {
1100 			shortfall = vm_laundry_target(vmd) +
1101 			    vmd->vmd_pageout_deficit;
1102 			target = 0;
1103 		} else
1104 			shortfall = 0;
1105 
1106 		if (target == 0)
1107 			vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1108 		nfreed += vmd->vmd_clean_pages_freed;
1109 		vmd->vmd_clean_pages_freed = 0;
1110 		vm_pagequeue_unlock(pq);
1111 	}
1112 }
1113 
1114 /*
1115  * Compute the number of pages we want to try to move from the
1116  * active queue to either the inactive or laundry queue.
1117  *
1118  * When scanning active pages during a shortage, we make clean pages
1119  * count more heavily towards the page shortage than dirty pages.
1120  * This is because dirty pages must be laundered before they can be
1121  * reused and thus have less utility when attempting to quickly
1122  * alleviate a free page shortage.  However, this weighting also
1123  * causes the scan to deactivate dirty pages more aggressively,
1124  * improving the effectiveness of clustering.
1125  */
1126 static int
1127 vm_pageout_active_target(struct vm_domain *vmd)
1128 {
1129 	int shortage;
1130 
1131 	shortage = vmd->vmd_inactive_target + vm_paging_target(vmd) -
1132 	    (vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt +
1133 	    vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight);
1134 	shortage *= act_scan_laundry_weight;
1135 	return (shortage);
1136 }
1137 
1138 /*
1139  * Scan the active queue.  If there is no shortage of inactive pages, scan a
1140  * small portion of the queue in order to maintain quasi-LRU.
1141  */
1142 static void
1143 vm_pageout_scan_active(struct vm_domain *vmd, int page_shortage)
1144 {
1145 	struct scan_state ss;
1146 	struct mtx *mtx;
1147 	vm_page_t m, marker;
1148 	struct vm_pagequeue *pq;
1149 	long min_scan;
1150 	int act_delta, max_scan, scan_tick;
1151 
1152 	marker = &vmd->vmd_markers[PQ_ACTIVE];
1153 	pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1154 	vm_pagequeue_lock(pq);
1155 
1156 	/*
1157 	 * If we're just idle polling attempt to visit every
1158 	 * active page within 'update_period' seconds.
1159 	 */
1160 	scan_tick = ticks;
1161 	if (vm_pageout_update_period != 0) {
1162 		min_scan = pq->pq_cnt;
1163 		min_scan *= scan_tick - vmd->vmd_last_active_scan;
1164 		min_scan /= hz * vm_pageout_update_period;
1165 	} else
1166 		min_scan = 0;
1167 	if (min_scan > 0 || (page_shortage > 0 && pq->pq_cnt > 0))
1168 		vmd->vmd_last_active_scan = scan_tick;
1169 
1170 	/*
1171 	 * Scan the active queue for pages that can be deactivated.  Update
1172 	 * the per-page activity counter and use it to identify deactivation
1173 	 * candidates.  Held pages may be deactivated.
1174 	 *
1175 	 * To avoid requeuing each page that remains in the active queue, we
1176 	 * implement the CLOCK algorithm.  To keep the implementation of the
1177 	 * enqueue operation consistent for all page queues, we use two hands,
1178 	 * represented by marker pages. Scans begin at the first hand, which
1179 	 * precedes the second hand in the queue.  When the two hands meet,
1180 	 * they are moved back to the head and tail of the queue, respectively,
1181 	 * and scanning resumes.
1182 	 */
1183 	max_scan = page_shortage > 0 ? pq->pq_cnt : min_scan;
1184 	mtx = NULL;
1185 act_scan:
1186 	vm_pageout_init_scan(&ss, pq, marker, &vmd->vmd_clock[0], max_scan);
1187 	while ((m = vm_pageout_next(&ss, false)) != NULL) {
1188 		if (__predict_false(m == &vmd->vmd_clock[1])) {
1189 			vm_pagequeue_lock(pq);
1190 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1191 			TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[1], plinks.q);
1192 			TAILQ_INSERT_HEAD(&pq->pq_pl, &vmd->vmd_clock[0],
1193 			    plinks.q);
1194 			TAILQ_INSERT_TAIL(&pq->pq_pl, &vmd->vmd_clock[1],
1195 			    plinks.q);
1196 			max_scan -= ss.scanned;
1197 			vm_pageout_end_scan(&ss);
1198 			goto act_scan;
1199 		}
1200 		if (__predict_false((m->flags & PG_MARKER) != 0))
1201 			continue;
1202 
1203 		vm_page_change_lock(m, &mtx);
1204 
1205 		/*
1206 		 * The page may have been disassociated from the queue
1207 		 * while locks were dropped.
1208 		 */
1209 		if (vm_page_queue(m) != PQ_ACTIVE)
1210 			continue;
1211 
1212 		/*
1213 		 * Wired pages are dequeued lazily.
1214 		 */
1215 		if (m->wire_count != 0) {
1216 			vm_page_dequeue_deferred(m);
1217 			continue;
1218 		}
1219 
1220 		/*
1221 		 * Check to see "how much" the page has been used.
1222 		 *
1223 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1224 		 * that a reference from a concurrently destroyed mapping is
1225 		 * observed here and now.
1226 		 *
1227 		 * Perform an unsynchronized object ref count check.  While
1228 		 * the page lock ensures that the page is not reallocated to
1229 		 * another object, in particular, one with unmanaged mappings
1230 		 * that cannot support pmap_ts_referenced(), two races are,
1231 		 * nonetheless, possible:
1232 		 * 1) The count was transitioning to zero, but we saw a non-
1233 		 *    zero value.  pmap_ts_referenced() will return zero
1234 		 *    because the page is not mapped.
1235 		 * 2) The count was transitioning to one, but we saw zero.
1236 		 *    This race delays the detection of a new reference.  At
1237 		 *    worst, we will deactivate and reactivate the page.
1238 		 */
1239 		if (m->object->ref_count != 0)
1240 			act_delta = pmap_ts_referenced(m);
1241 		else
1242 			act_delta = 0;
1243 		if ((m->aflags & PGA_REFERENCED) != 0) {
1244 			vm_page_aflag_clear(m, PGA_REFERENCED);
1245 			act_delta++;
1246 		}
1247 
1248 		/*
1249 		 * Advance or decay the act_count based on recent usage.
1250 		 */
1251 		if (act_delta != 0) {
1252 			m->act_count += ACT_ADVANCE + act_delta;
1253 			if (m->act_count > ACT_MAX)
1254 				m->act_count = ACT_MAX;
1255 		} else
1256 			m->act_count -= min(m->act_count, ACT_DECLINE);
1257 
1258 		if (m->act_count == 0) {
1259 			/*
1260 			 * When not short for inactive pages, let dirty pages go
1261 			 * through the inactive queue before moving to the
1262 			 * laundry queues.  This gives them some extra time to
1263 			 * be reactivated, potentially avoiding an expensive
1264 			 * pageout.  However, during a page shortage, the
1265 			 * inactive queue is necessarily small, and so dirty
1266 			 * pages would only spend a trivial amount of time in
1267 			 * the inactive queue.  Therefore, we might as well
1268 			 * place them directly in the laundry queue to reduce
1269 			 * queuing overhead.
1270 			 */
1271 			if (page_shortage <= 0)
1272 				vm_page_deactivate(m);
1273 			else {
1274 				/*
1275 				 * Calling vm_page_test_dirty() here would
1276 				 * require acquisition of the object's write
1277 				 * lock.  However, during a page shortage,
1278 				 * directing dirty pages into the laundry
1279 				 * queue is only an optimization and not a
1280 				 * requirement.  Therefore, we simply rely on
1281 				 * the opportunistic updates to the page's
1282 				 * dirty field by the pmap.
1283 				 */
1284 				if (m->dirty == 0) {
1285 					vm_page_deactivate(m);
1286 					page_shortage -=
1287 					    act_scan_laundry_weight;
1288 				} else {
1289 					vm_page_launder(m);
1290 					page_shortage--;
1291 				}
1292 			}
1293 		}
1294 	}
1295 	if (mtx != NULL) {
1296 		mtx_unlock(mtx);
1297 		mtx = NULL;
1298 	}
1299 	vm_pagequeue_lock(pq);
1300 	TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_clock[0], plinks.q);
1301 	TAILQ_INSERT_AFTER(&pq->pq_pl, marker, &vmd->vmd_clock[0], plinks.q);
1302 	vm_pageout_end_scan(&ss);
1303 	vm_pagequeue_unlock(pq);
1304 }
1305 
1306 static int
1307 vm_pageout_reinsert_inactive_page(struct scan_state *ss, vm_page_t m)
1308 {
1309 	struct vm_domain *vmd;
1310 
1311 	if (m->queue != PQ_INACTIVE || (m->aflags & PGA_ENQUEUED) != 0)
1312 		return (0);
1313 	vm_page_aflag_set(m, PGA_ENQUEUED);
1314 	if ((m->aflags & PGA_REQUEUE_HEAD) != 0) {
1315 		vmd = vm_pagequeue_domain(m);
1316 		TAILQ_INSERT_BEFORE(&vmd->vmd_inacthead, m, plinks.q);
1317 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1318 	} else if ((m->aflags & PGA_REQUEUE) != 0) {
1319 		TAILQ_INSERT_TAIL(&ss->pq->pq_pl, m, plinks.q);
1320 		vm_page_aflag_clear(m, PGA_REQUEUE | PGA_REQUEUE_HEAD);
1321 	} else
1322 		TAILQ_INSERT_BEFORE(ss->marker, m, plinks.q);
1323 	return (1);
1324 }
1325 
1326 /*
1327  * Re-add stuck pages to the inactive queue.  We will examine them again
1328  * during the next scan.  If the queue state of a page has changed since
1329  * it was physically removed from the page queue in
1330  * vm_pageout_collect_batch(), don't do anything with that page.
1331  */
1332 static void
1333 vm_pageout_reinsert_inactive(struct scan_state *ss, struct vm_batchqueue *bq,
1334     vm_page_t m)
1335 {
1336 	struct vm_pagequeue *pq;
1337 	int delta;
1338 
1339 	delta = 0;
1340 	pq = ss->pq;
1341 
1342 	if (m != NULL) {
1343 		if (vm_batchqueue_insert(bq, m))
1344 			return;
1345 		vm_pagequeue_lock(pq);
1346 		delta += vm_pageout_reinsert_inactive_page(ss, m);
1347 	} else
1348 		vm_pagequeue_lock(pq);
1349 	while ((m = vm_batchqueue_pop(bq)) != NULL)
1350 		delta += vm_pageout_reinsert_inactive_page(ss, m);
1351 	vm_pagequeue_cnt_add(pq, delta);
1352 	vm_pagequeue_unlock(pq);
1353 	vm_batchqueue_init(bq);
1354 }
1355 
1356 /*
1357  * Attempt to reclaim the requested number of pages from the inactive queue.
1358  * Returns true if the shortage was addressed.
1359  */
1360 static int
1361 vm_pageout_scan_inactive(struct vm_domain *vmd, int shortage,
1362     int *addl_shortage)
1363 {
1364 	struct scan_state ss;
1365 	struct vm_batchqueue rq;
1366 	struct mtx *mtx;
1367 	vm_page_t m, marker;
1368 	struct vm_pagequeue *pq;
1369 	vm_object_t object;
1370 	int act_delta, addl_page_shortage, deficit, page_shortage;
1371 	int starting_page_shortage;
1372 	bool obj_locked;
1373 
1374 	/*
1375 	 * The addl_page_shortage is an estimate of the number of temporarily
1376 	 * stuck pages in the inactive queue.  In other words, the
1377 	 * number of pages from the inactive count that should be
1378 	 * discounted in setting the target for the active queue scan.
1379 	 */
1380 	addl_page_shortage = 0;
1381 
1382 	/*
1383 	 * vmd_pageout_deficit counts the number of pages requested in
1384 	 * allocations that failed because of a free page shortage.  We assume
1385 	 * that the allocations will be reattempted and thus include the deficit
1386 	 * in our scan target.
1387 	 */
1388 	deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1389 	starting_page_shortage = page_shortage = shortage + deficit;
1390 
1391 	mtx = NULL;
1392 	obj_locked = false;
1393 	object = NULL;
1394 	vm_batchqueue_init(&rq);
1395 
1396 	/*
1397 	 * Start scanning the inactive queue for pages that we can free.  The
1398 	 * scan will stop when we reach the target or we have scanned the
1399 	 * entire queue.  (Note that m->act_count is not used to make
1400 	 * decisions for the inactive queue, only for the active queue.)
1401 	 */
1402 	marker = &vmd->vmd_markers[PQ_INACTIVE];
1403 	pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1404 	vm_pagequeue_lock(pq);
1405 	vm_pageout_init_scan(&ss, pq, marker, NULL, pq->pq_cnt);
1406 	while (page_shortage > 0 && (m = vm_pageout_next(&ss, true)) != NULL) {
1407 		KASSERT((m->flags & PG_MARKER) == 0,
1408 		    ("marker page %p was dequeued", m));
1409 
1410 		vm_page_change_lock(m, &mtx);
1411 
1412 recheck:
1413 		/*
1414 		 * The page may have been disassociated from the queue
1415 		 * while locks were dropped.
1416 		 */
1417 		if (vm_page_queue(m) != PQ_INACTIVE) {
1418 			addl_page_shortage++;
1419 			continue;
1420 		}
1421 
1422 		/*
1423 		 * The page was re-enqueued after the page queue lock was
1424 		 * dropped, or a requeue was requested.  This page gets a second
1425 		 * chance.
1426 		 */
1427 		if ((m->aflags & (PGA_ENQUEUED | PGA_REQUEUE |
1428 		    PGA_REQUEUE_HEAD)) != 0)
1429 			goto reinsert;
1430 
1431 		/*
1432 		 * Held pages are essentially stuck in the queue.  So,
1433 		 * they ought to be discounted from the inactive count.
1434 		 * See the description of addl_page_shortage above.
1435 		 *
1436 		 * Wired pages may not be freed.  Complete their removal
1437 		 * from the queue now to avoid needless revisits during
1438 		 * future scans.
1439 		 */
1440 		if (m->hold_count != 0) {
1441 			addl_page_shortage++;
1442 			goto reinsert;
1443 		}
1444 		if (m->wire_count != 0) {
1445 			vm_page_dequeue_deferred(m);
1446 			continue;
1447 		}
1448 
1449 		if (object != m->object) {
1450 			if (obj_locked) {
1451 				VM_OBJECT_WUNLOCK(object);
1452 				obj_locked = false;
1453 			}
1454 			object = m->object;
1455 		}
1456 		if (!obj_locked) {
1457 			if (!VM_OBJECT_TRYWLOCK(object)) {
1458 				mtx_unlock(mtx);
1459 				/* Depends on type-stability. */
1460 				VM_OBJECT_WLOCK(object);
1461 				obj_locked = true;
1462 				mtx_lock(mtx);
1463 				goto recheck;
1464 			} else
1465 				obj_locked = true;
1466 		}
1467 
1468 		if (vm_page_busied(m)) {
1469 			/*
1470 			 * Don't mess with busy pages.  Leave them at
1471 			 * the front of the queue.  Most likely, they
1472 			 * are being paged out and will leave the
1473 			 * queue shortly after the scan finishes.  So,
1474 			 * they ought to be discounted from the
1475 			 * inactive count.
1476 			 */
1477 			addl_page_shortage++;
1478 			goto reinsert;
1479 		}
1480 
1481 		/*
1482 		 * Invalid pages can be easily freed. They cannot be
1483 		 * mapped, vm_page_free() asserts this.
1484 		 */
1485 		if (m->valid == 0)
1486 			goto free_page;
1487 
1488 		/*
1489 		 * If the page has been referenced and the object is not dead,
1490 		 * reactivate or requeue the page depending on whether the
1491 		 * object is mapped.
1492 		 *
1493 		 * Test PGA_REFERENCED after calling pmap_ts_referenced() so
1494 		 * that a reference from a concurrently destroyed mapping is
1495 		 * observed here and now.
1496 		 */
1497 		if (object->ref_count != 0)
1498 			act_delta = pmap_ts_referenced(m);
1499 		else {
1500 			KASSERT(!pmap_page_is_mapped(m),
1501 			    ("page %p is mapped", m));
1502 			act_delta = 0;
1503 		}
1504 		if ((m->aflags & PGA_REFERENCED) != 0) {
1505 			vm_page_aflag_clear(m, PGA_REFERENCED);
1506 			act_delta++;
1507 		}
1508 		if (act_delta != 0) {
1509 			if (object->ref_count != 0) {
1510 				VM_CNT_INC(v_reactivated);
1511 				vm_page_activate(m);
1512 
1513 				/*
1514 				 * Increase the activation count if the page
1515 				 * was referenced while in the inactive queue.
1516 				 * This makes it less likely that the page will
1517 				 * be returned prematurely to the inactive
1518 				 * queue.
1519  				 */
1520 				m->act_count += act_delta + ACT_ADVANCE;
1521 				continue;
1522 			} else if ((object->flags & OBJ_DEAD) == 0) {
1523 				vm_page_aflag_set(m, PGA_REQUEUE);
1524 				goto reinsert;
1525 			}
1526 		}
1527 
1528 		/*
1529 		 * If the page appears to be clean at the machine-independent
1530 		 * layer, then remove all of its mappings from the pmap in
1531 		 * anticipation of freeing it.  If, however, any of the page's
1532 		 * mappings allow write access, then the page may still be
1533 		 * modified until the last of those mappings are removed.
1534 		 */
1535 		if (object->ref_count != 0) {
1536 			vm_page_test_dirty(m);
1537 			if (m->dirty == 0)
1538 				pmap_remove_all(m);
1539 		}
1540 
1541 		/*
1542 		 * Clean pages can be freed, but dirty pages must be sent back
1543 		 * to the laundry, unless they belong to a dead object.
1544 		 * Requeueing dirty pages from dead objects is pointless, as
1545 		 * they are being paged out and freed by the thread that
1546 		 * destroyed the object.
1547 		 */
1548 		if (m->dirty == 0) {
1549 free_page:
1550 			/*
1551 			 * Because we dequeued the page and have already
1552 			 * checked for concurrent dequeue and enqueue
1553 			 * requests, we can safely disassociate the page
1554 			 * from the inactive queue.
1555 			 */
1556 			KASSERT((m->aflags & PGA_QUEUE_STATE_MASK) == 0,
1557 			    ("page %p has queue state", m));
1558 			m->queue = PQ_NONE;
1559 			vm_page_free(m);
1560 			page_shortage--;
1561 		} else if ((object->flags & OBJ_DEAD) == 0)
1562 			vm_page_launder(m);
1563 		continue;
1564 reinsert:
1565 		vm_pageout_reinsert_inactive(&ss, &rq, m);
1566 	}
1567 	if (mtx != NULL) {
1568 		mtx_unlock(mtx);
1569 		mtx = NULL;
1570 	}
1571 	if (obj_locked) {
1572 		VM_OBJECT_WUNLOCK(object);
1573 		obj_locked = false;
1574 	}
1575 	vm_pageout_reinsert_inactive(&ss, &rq, NULL);
1576 	vm_pageout_reinsert_inactive(&ss, &ss.bq, NULL);
1577 	vm_pagequeue_lock(pq);
1578 	vm_pageout_end_scan(&ss);
1579 	vm_pagequeue_unlock(pq);
1580 
1581 	VM_CNT_ADD(v_dfree, starting_page_shortage - page_shortage);
1582 
1583 	/*
1584 	 * Wake up the laundry thread so that it can perform any needed
1585 	 * laundering.  If we didn't meet our target, we're in shortfall and
1586 	 * need to launder more aggressively.  If PQ_LAUNDRY is empty and no
1587 	 * swap devices are configured, the laundry thread has no work to do, so
1588 	 * don't bother waking it up.
1589 	 *
1590 	 * The laundry thread uses the number of inactive queue scans elapsed
1591 	 * since the last laundering to determine whether to launder again, so
1592 	 * keep count.
1593 	 */
1594 	if (starting_page_shortage > 0) {
1595 		pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1596 		vm_pagequeue_lock(pq);
1597 		if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1598 		    (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1599 			if (page_shortage > 0) {
1600 				vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1601 				VM_CNT_INC(v_pdshortfalls);
1602 			} else if (vmd->vmd_laundry_request !=
1603 			    VM_LAUNDRY_SHORTFALL)
1604 				vmd->vmd_laundry_request =
1605 				    VM_LAUNDRY_BACKGROUND;
1606 			wakeup(&vmd->vmd_laundry_request);
1607 		}
1608 		vmd->vmd_clean_pages_freed +=
1609 		    starting_page_shortage - page_shortage;
1610 		vm_pagequeue_unlock(pq);
1611 	}
1612 
1613 	/*
1614 	 * Wakeup the swapout daemon if we didn't free the targeted number of
1615 	 * pages.
1616 	 */
1617 	if (page_shortage > 0)
1618 		vm_swapout_run();
1619 
1620 	/*
1621 	 * If the inactive queue scan fails repeatedly to meet its
1622 	 * target, kill the largest process.
1623 	 */
1624 	vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1625 
1626 	/*
1627 	 * Reclaim pages by swapping out idle processes, if configured to do so.
1628 	 */
1629 	vm_swapout_run_idle();
1630 
1631 	/*
1632 	 * See the description of addl_page_shortage above.
1633 	 */
1634 	*addl_shortage = addl_page_shortage + deficit;
1635 
1636 	return (page_shortage <= 0);
1637 }
1638 
1639 static int vm_pageout_oom_vote;
1640 
1641 /*
1642  * The pagedaemon threads randlomly select one to perform the
1643  * OOM.  Trying to kill processes before all pagedaemons
1644  * failed to reach free target is premature.
1645  */
1646 static void
1647 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1648     int starting_page_shortage)
1649 {
1650 	int old_vote;
1651 
1652 	if (starting_page_shortage <= 0 || starting_page_shortage !=
1653 	    page_shortage)
1654 		vmd->vmd_oom_seq = 0;
1655 	else
1656 		vmd->vmd_oom_seq++;
1657 	if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1658 		if (vmd->vmd_oom) {
1659 			vmd->vmd_oom = FALSE;
1660 			atomic_subtract_int(&vm_pageout_oom_vote, 1);
1661 		}
1662 		return;
1663 	}
1664 
1665 	/*
1666 	 * Do not follow the call sequence until OOM condition is
1667 	 * cleared.
1668 	 */
1669 	vmd->vmd_oom_seq = 0;
1670 
1671 	if (vmd->vmd_oom)
1672 		return;
1673 
1674 	vmd->vmd_oom = TRUE;
1675 	old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1676 	if (old_vote != vm_ndomains - 1)
1677 		return;
1678 
1679 	/*
1680 	 * The current pagedaemon thread is the last in the quorum to
1681 	 * start OOM.  Initiate the selection and signaling of the
1682 	 * victim.
1683 	 */
1684 	vm_pageout_oom(VM_OOM_MEM);
1685 
1686 	/*
1687 	 * After one round of OOM terror, recall our vote.  On the
1688 	 * next pass, current pagedaemon would vote again if the low
1689 	 * memory condition is still there, due to vmd_oom being
1690 	 * false.
1691 	 */
1692 	vmd->vmd_oom = FALSE;
1693 	atomic_subtract_int(&vm_pageout_oom_vote, 1);
1694 }
1695 
1696 /*
1697  * The OOM killer is the page daemon's action of last resort when
1698  * memory allocation requests have been stalled for a prolonged period
1699  * of time because it cannot reclaim memory.  This function computes
1700  * the approximate number of physical pages that could be reclaimed if
1701  * the specified address space is destroyed.
1702  *
1703  * Private, anonymous memory owned by the address space is the
1704  * principal resource that we expect to recover after an OOM kill.
1705  * Since the physical pages mapped by the address space's COW entries
1706  * are typically shared pages, they are unlikely to be released and so
1707  * they are not counted.
1708  *
1709  * To get to the point where the page daemon runs the OOM killer, its
1710  * efforts to write-back vnode-backed pages may have stalled.  This
1711  * could be caused by a memory allocation deadlock in the write path
1712  * that might be resolved by an OOM kill.  Therefore, physical pages
1713  * belonging to vnode-backed objects are counted, because they might
1714  * be freed without being written out first if the address space holds
1715  * the last reference to an unlinked vnode.
1716  *
1717  * Similarly, physical pages belonging to OBJT_PHYS objects are
1718  * counted because the address space might hold the last reference to
1719  * the object.
1720  */
1721 static long
1722 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1723 {
1724 	vm_map_t map;
1725 	vm_map_entry_t entry;
1726 	vm_object_t obj;
1727 	long res;
1728 
1729 	map = &vmspace->vm_map;
1730 	KASSERT(!map->system_map, ("system map"));
1731 	sx_assert(&map->lock, SA_LOCKED);
1732 	res = 0;
1733 	for (entry = map->header.next; entry != &map->header;
1734 	    entry = entry->next) {
1735 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1736 			continue;
1737 		obj = entry->object.vm_object;
1738 		if (obj == NULL)
1739 			continue;
1740 		if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1741 		    obj->ref_count != 1)
1742 			continue;
1743 		switch (obj->type) {
1744 		case OBJT_DEFAULT:
1745 		case OBJT_SWAP:
1746 		case OBJT_PHYS:
1747 		case OBJT_VNODE:
1748 			res += obj->resident_page_count;
1749 			break;
1750 		}
1751 	}
1752 	return (res);
1753 }
1754 
1755 void
1756 vm_pageout_oom(int shortage)
1757 {
1758 	struct proc *p, *bigproc;
1759 	vm_offset_t size, bigsize;
1760 	struct thread *td;
1761 	struct vmspace *vm;
1762 	bool breakout;
1763 
1764 	/*
1765 	 * We keep the process bigproc locked once we find it to keep anyone
1766 	 * from messing with it; however, there is a possibility of
1767 	 * deadlock if process B is bigproc and one of its child processes
1768 	 * attempts to propagate a signal to B while we are waiting for A's
1769 	 * lock while walking this list.  To avoid this, we don't block on
1770 	 * the process lock but just skip a process if it is already locked.
1771 	 */
1772 	bigproc = NULL;
1773 	bigsize = 0;
1774 	sx_slock(&allproc_lock);
1775 	FOREACH_PROC_IN_SYSTEM(p) {
1776 		PROC_LOCK(p);
1777 
1778 		/*
1779 		 * If this is a system, protected or killed process, skip it.
1780 		 */
1781 		if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1782 		    P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1783 		    p->p_pid == 1 || P_KILLED(p) ||
1784 		    (p->p_pid < 48 && swap_pager_avail != 0)) {
1785 			PROC_UNLOCK(p);
1786 			continue;
1787 		}
1788 		/*
1789 		 * If the process is in a non-running type state,
1790 		 * don't touch it.  Check all the threads individually.
1791 		 */
1792 		breakout = false;
1793 		FOREACH_THREAD_IN_PROC(p, td) {
1794 			thread_lock(td);
1795 			if (!TD_ON_RUNQ(td) &&
1796 			    !TD_IS_RUNNING(td) &&
1797 			    !TD_IS_SLEEPING(td) &&
1798 			    !TD_IS_SUSPENDED(td) &&
1799 			    !TD_IS_SWAPPED(td)) {
1800 				thread_unlock(td);
1801 				breakout = true;
1802 				break;
1803 			}
1804 			thread_unlock(td);
1805 		}
1806 		if (breakout) {
1807 			PROC_UNLOCK(p);
1808 			continue;
1809 		}
1810 		/*
1811 		 * get the process size
1812 		 */
1813 		vm = vmspace_acquire_ref(p);
1814 		if (vm == NULL) {
1815 			PROC_UNLOCK(p);
1816 			continue;
1817 		}
1818 		_PHOLD_LITE(p);
1819 		PROC_UNLOCK(p);
1820 		sx_sunlock(&allproc_lock);
1821 		if (!vm_map_trylock_read(&vm->vm_map)) {
1822 			vmspace_free(vm);
1823 			sx_slock(&allproc_lock);
1824 			PRELE(p);
1825 			continue;
1826 		}
1827 		size = vmspace_swap_count(vm);
1828 		if (shortage == VM_OOM_MEM)
1829 			size += vm_pageout_oom_pagecount(vm);
1830 		vm_map_unlock_read(&vm->vm_map);
1831 		vmspace_free(vm);
1832 		sx_slock(&allproc_lock);
1833 
1834 		/*
1835 		 * If this process is bigger than the biggest one,
1836 		 * remember it.
1837 		 */
1838 		if (size > bigsize) {
1839 			if (bigproc != NULL)
1840 				PRELE(bigproc);
1841 			bigproc = p;
1842 			bigsize = size;
1843 		} else {
1844 			PRELE(p);
1845 		}
1846 	}
1847 	sx_sunlock(&allproc_lock);
1848 	if (bigproc != NULL) {
1849 		if (vm_panic_on_oom != 0)
1850 			panic("out of swap space");
1851 		PROC_LOCK(bigproc);
1852 		killproc(bigproc, "out of swap space");
1853 		sched_nice(bigproc, PRIO_MIN);
1854 		_PRELE(bigproc);
1855 		PROC_UNLOCK(bigproc);
1856 	}
1857 }
1858 
1859 static void
1860 vm_pageout_lowmem(struct vm_domain *vmd)
1861 {
1862 
1863 	if (vmd == VM_DOMAIN(0) &&
1864 	    time_uptime - lowmem_uptime >= lowmem_period) {
1865 		/*
1866 		 * Decrease registered cache sizes.
1867 		 */
1868 		SDT_PROBE0(vm, , , vm__lowmem_scan);
1869 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1870 
1871 		/*
1872 		 * We do this explicitly after the caches have been
1873 		 * drained above.
1874 		 */
1875 		uma_reclaim();
1876 		lowmem_uptime = time_uptime;
1877 	}
1878 }
1879 
1880 static void
1881 vm_pageout_worker(void *arg)
1882 {
1883 	struct vm_domain *vmd;
1884 	int addl_shortage, domain, shortage;
1885 	bool target_met;
1886 
1887 	domain = (uintptr_t)arg;
1888 	vmd = VM_DOMAIN(domain);
1889 	shortage = 0;
1890 	target_met = true;
1891 
1892 	/*
1893 	 * XXXKIB It could be useful to bind pageout daemon threads to
1894 	 * the cores belonging to the domain, from which vm_page_array
1895 	 * is allocated.
1896 	 */
1897 
1898 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1899 	vmd->vmd_last_active_scan = ticks;
1900 
1901 	/*
1902 	 * The pageout daemon worker is never done, so loop forever.
1903 	 */
1904 	while (TRUE) {
1905 		vm_domain_pageout_lock(vmd);
1906 
1907 		/*
1908 		 * We need to clear wanted before we check the limits.  This
1909 		 * prevents races with wakers who will check wanted after they
1910 		 * reach the limit.
1911 		 */
1912 		atomic_store_int(&vmd->vmd_pageout_wanted, 0);
1913 
1914 		/*
1915 		 * Might the page daemon need to run again?
1916 		 */
1917 		if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
1918 			/*
1919 			 * Yes.  If the scan failed to produce enough free
1920 			 * pages, sleep uninterruptibly for some time in the
1921 			 * hope that the laundry thread will clean some pages.
1922 			 */
1923 			vm_domain_pageout_unlock(vmd);
1924 			if (!target_met)
1925 				pause("pwait", hz / VM_INACT_SCAN_RATE);
1926 		} else {
1927 			/*
1928 			 * No, sleep until the next wakeup or until pages
1929 			 * need to have their reference stats updated.
1930 			 */
1931 			if (mtx_sleep(&vmd->vmd_pageout_wanted,
1932 			    vm_domain_pageout_lockptr(vmd), PDROP | PVM,
1933 			    "psleep", hz / VM_INACT_SCAN_RATE) == 0)
1934 				VM_CNT_INC(v_pdwakeups);
1935 		}
1936 
1937 		/* Prevent spurious wakeups by ensuring that wanted is set. */
1938 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1939 
1940 		/*
1941 		 * Use the controller to calculate how many pages to free in
1942 		 * this interval, and scan the inactive queue.
1943 		 */
1944 		shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
1945 		if (shortage > 0) {
1946 			vm_pageout_lowmem(vmd);
1947 			target_met = vm_pageout_scan_inactive(vmd, shortage,
1948 			    &addl_shortage);
1949 		} else
1950 			addl_shortage = 0;
1951 
1952 		/*
1953 		 * Scan the active queue.  A positive value for shortage
1954 		 * indicates that we must aggressively deactivate pages to avoid
1955 		 * a shortfall.
1956 		 */
1957 		shortage = vm_pageout_active_target(vmd) + addl_shortage;
1958 		vm_pageout_scan_active(vmd, shortage);
1959 	}
1960 }
1961 
1962 /*
1963  *	vm_pageout_init initialises basic pageout daemon settings.
1964  */
1965 static void
1966 vm_pageout_init_domain(int domain)
1967 {
1968 	struct vm_domain *vmd;
1969 	struct sysctl_oid *oid;
1970 
1971 	vmd = VM_DOMAIN(domain);
1972 	vmd->vmd_interrupt_free_min = 2;
1973 
1974 	/*
1975 	 * v_free_reserved needs to include enough for the largest
1976 	 * swap pager structures plus enough for any pv_entry structs
1977 	 * when paging.
1978 	 */
1979 	if (vmd->vmd_page_count > 1024)
1980 		vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200;
1981 	else
1982 		vmd->vmd_free_min = 4;
1983 	vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1984 	    vmd->vmd_interrupt_free_min;
1985 	vmd->vmd_free_reserved = vm_pageout_page_count +
1986 	    vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768);
1987 	vmd->vmd_free_severe = vmd->vmd_free_min / 2;
1988 	vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
1989 	vmd->vmd_free_min += vmd->vmd_free_reserved;
1990 	vmd->vmd_free_severe += vmd->vmd_free_reserved;
1991 	vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
1992 	if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
1993 		vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
1994 
1995 	/*
1996 	 * Set the default wakeup threshold to be 10% below the paging
1997 	 * target.  This keeps the steady state out of shortfall.
1998 	 */
1999 	vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
2000 
2001 	/*
2002 	 * Target amount of memory to move out of the laundry queue during a
2003 	 * background laundering.  This is proportional to the amount of system
2004 	 * memory.
2005 	 */
2006 	vmd->vmd_background_launder_target = (vmd->vmd_free_target -
2007 	    vmd->vmd_free_min) / 10;
2008 
2009 	/* Initialize the pageout daemon pid controller. */
2010 	pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
2011 	    vmd->vmd_free_target, PIDCTRL_BOUND,
2012 	    PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
2013 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
2014 	    "pidctrl", CTLFLAG_RD, NULL, "");
2015 	pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
2016 }
2017 
2018 static void
2019 vm_pageout_init(void)
2020 {
2021 	u_int freecount;
2022 	int i;
2023 
2024 	/*
2025 	 * Initialize some paging parameters.
2026 	 */
2027 	if (vm_cnt.v_page_count < 2000)
2028 		vm_pageout_page_count = 8;
2029 
2030 	freecount = 0;
2031 	for (i = 0; i < vm_ndomains; i++) {
2032 		struct vm_domain *vmd;
2033 
2034 		vm_pageout_init_domain(i);
2035 		vmd = VM_DOMAIN(i);
2036 		vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
2037 		vm_cnt.v_free_target += vmd->vmd_free_target;
2038 		vm_cnt.v_free_min += vmd->vmd_free_min;
2039 		vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
2040 		vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
2041 		vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
2042 		vm_cnt.v_free_severe += vmd->vmd_free_severe;
2043 		freecount += vmd->vmd_free_count;
2044 	}
2045 
2046 	/*
2047 	 * Set interval in seconds for active scan.  We want to visit each
2048 	 * page at least once every ten minutes.  This is to prevent worst
2049 	 * case paging behaviors with stale active LRU.
2050 	 */
2051 	if (vm_pageout_update_period == 0)
2052 		vm_pageout_update_period = 600;
2053 
2054 	if (vm_page_max_wired == 0)
2055 		vm_page_max_wired = freecount / 3;
2056 }
2057 
2058 /*
2059  *     vm_pageout is the high level pageout daemon.
2060  */
2061 static void
2062 vm_pageout(void)
2063 {
2064 	int error;
2065 	int i;
2066 
2067 	swap_pager_swap_init();
2068 	snprintf(curthread->td_name, sizeof(curthread->td_name), "dom0");
2069 	error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
2070 	    0, 0, "laundry: dom0");
2071 	if (error != 0)
2072 		panic("starting laundry for domain 0, error %d", error);
2073 	for (i = 1; i < vm_ndomains; i++) {
2074 		error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
2075 		    curproc, NULL, 0, 0, "dom%d", i);
2076 		if (error != 0) {
2077 			panic("starting pageout for domain %d, error %d\n",
2078 			    i, error);
2079 		}
2080 		error = kthread_add(vm_pageout_laundry_worker,
2081 		    (void *)(uintptr_t)i, curproc, NULL, 0, 0,
2082 		    "laundry: dom%d", i);
2083 		if (error != 0)
2084 			panic("starting laundry for domain %d, error %d",
2085 			    i, error);
2086 	}
2087 	error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
2088 	    0, 0, "uma");
2089 	if (error != 0)
2090 		panic("starting uma_reclaim helper, error %d\n", error);
2091 	vm_pageout_worker((void *)(uintptr_t)0);
2092 }
2093 
2094 /*
2095  * Perform an advisory wakeup of the page daemon.
2096  */
2097 void
2098 pagedaemon_wakeup(int domain)
2099 {
2100 	struct vm_domain *vmd;
2101 
2102 	vmd = VM_DOMAIN(domain);
2103 	vm_domain_pageout_assert_unlocked(vmd);
2104 	if (curproc == pageproc)
2105 		return;
2106 
2107 	if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
2108 		vm_domain_pageout_lock(vmd);
2109 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
2110 		wakeup(&vmd->vmd_pageout_wanted);
2111 		vm_domain_pageout_unlock(vmd);
2112 	}
2113 }
2114