xref: /freebsd/sys/vm/vm_pageout.c (revision 3eeb75112272ef9ef244a00bc73beef3a146856a)
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 bool vm_pageout_scan(struct vm_domain *vmd, int pass, int shortage);
128 static void vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
129     int starting_page_shortage);
130 
131 SYSINIT(pagedaemon_init, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, vm_pageout_init,
132     NULL);
133 
134 struct proc *pageproc;
135 
136 static struct kproc_desc page_kp = {
137 	"pagedaemon",
138 	vm_pageout,
139 	&pageproc
140 };
141 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_SECOND, kproc_start,
142     &page_kp);
143 
144 SDT_PROVIDER_DEFINE(vm);
145 SDT_PROBE_DEFINE(vm, , , vm__lowmem_scan);
146 
147 /* Pagedaemon activity rates, in subdivisions of one second. */
148 #define	VM_LAUNDER_RATE		10
149 #define	VM_INACT_SCAN_RATE	10
150 
151 static int vm_pageout_oom_seq = 12;
152 
153 static int vm_pageout_update_period;
154 static int disable_swap_pageouts;
155 static int lowmem_period = 10;
156 static time_t lowmem_uptime;
157 static int swapdev_enabled;
158 
159 static int vm_panic_on_oom = 0;
160 
161 SYSCTL_INT(_vm, OID_AUTO, panic_on_oom,
162 	CTLFLAG_RWTUN, &vm_panic_on_oom, 0,
163 	"panic on out of memory instead of killing the largest process");
164 
165 SYSCTL_INT(_vm, OID_AUTO, pageout_update_period,
166 	CTLFLAG_RWTUN, &vm_pageout_update_period, 0,
167 	"Maximum active LRU update period");
168 
169 SYSCTL_INT(_vm, OID_AUTO, lowmem_period, CTLFLAG_RWTUN, &lowmem_period, 0,
170 	"Low memory callback period");
171 
172 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
173 	CTLFLAG_RWTUN, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
174 
175 static int pageout_lock_miss;
176 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
177 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
178 
179 SYSCTL_INT(_vm, OID_AUTO, pageout_oom_seq,
180 	CTLFLAG_RWTUN, &vm_pageout_oom_seq, 0,
181 	"back-to-back calls to oom detector to start OOM");
182 
183 static int act_scan_laundry_weight = 3;
184 SYSCTL_INT(_vm, OID_AUTO, act_scan_laundry_weight, CTLFLAG_RWTUN,
185     &act_scan_laundry_weight, 0,
186     "weight given to clean vs. dirty pages in active queue scans");
187 
188 static u_int vm_background_launder_rate = 4096;
189 SYSCTL_UINT(_vm, OID_AUTO, background_launder_rate, CTLFLAG_RWTUN,
190     &vm_background_launder_rate, 0,
191     "background laundering rate, in kilobytes per second");
192 
193 static u_int vm_background_launder_max = 20 * 1024;
194 SYSCTL_UINT(_vm, OID_AUTO, background_launder_max, CTLFLAG_RWTUN,
195     &vm_background_launder_max, 0, "background laundering cap, in kilobytes");
196 
197 int vm_pageout_page_count = 32;
198 
199 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
200 SYSCTL_INT(_vm, OID_AUTO, max_wired,
201 	CTLFLAG_RW, &vm_page_max_wired, 0, "System-wide limit to wired page count");
202 
203 static u_int isqrt(u_int num);
204 static boolean_t vm_pageout_fallback_object_lock(vm_page_t, vm_page_t *);
205 static int vm_pageout_launder(struct vm_domain *vmd, int launder,
206     bool in_shortfall);
207 static void vm_pageout_laundry_worker(void *arg);
208 static boolean_t vm_pageout_page_lock(vm_page_t, vm_page_t *);
209 
210 /*
211  * Initialize a dummy page for marking the caller's place in the specified
212  * paging queue.  In principle, this function only needs to set the flag
213  * PG_MARKER.  Nonetheless, it write busies and initializes the hold count
214  * to one as safety precautions.
215  */
216 static void
217 vm_pageout_init_marker(vm_page_t marker, u_short queue)
218 {
219 
220 	bzero(marker, sizeof(*marker));
221 	marker->flags = PG_MARKER;
222 	marker->busy_lock = VPB_SINGLE_EXCLUSIVER;
223 	marker->queue = queue;
224 	marker->hold_count = 1;
225 }
226 
227 /*
228  * vm_pageout_fallback_object_lock:
229  *
230  * Lock vm object currently associated with `m'. VM_OBJECT_TRYWLOCK is
231  * known to have failed and page queue must be either PQ_ACTIVE or
232  * PQ_INACTIVE.  To avoid lock order violation, unlock the page queue
233  * while locking the vm object.  Use marker page to detect page queue
234  * changes and maintain notion of next page on page queue.  Return
235  * TRUE if no changes were detected, FALSE otherwise.  vm object is
236  * locked on return.
237  *
238  * This function depends on both the lock portion of struct vm_object
239  * and normal struct vm_page being type stable.
240  */
241 static boolean_t
242 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next)
243 {
244 	struct vm_page marker;
245 	struct vm_pagequeue *pq;
246 	boolean_t unchanged;
247 	u_short queue;
248 	vm_object_t object;
249 
250 	queue = m->queue;
251 	vm_pageout_init_marker(&marker, queue);
252 	pq = vm_page_pagequeue(m);
253 	object = m->object;
254 
255 	TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
256 	vm_pagequeue_unlock(pq);
257 	vm_page_unlock(m);
258 	VM_OBJECT_WLOCK(object);
259 	vm_page_lock(m);
260 	vm_pagequeue_lock(pq);
261 
262 	/*
263 	 * The page's object might have changed, and/or the page might
264 	 * have moved from its original position in the queue.  If the
265 	 * page's object has changed, then the caller should abandon
266 	 * processing the page because the wrong object lock was
267 	 * acquired.  Use the marker's plinks.q, not the page's, to
268 	 * determine if the page has been moved.  The state of the
269 	 * page's plinks.q can be indeterminate; whereas, the marker's
270 	 * plinks.q must be valid.
271 	 */
272 	*next = TAILQ_NEXT(&marker, plinks.q);
273 	unchanged = m->object == object &&
274 	    m == TAILQ_PREV(&marker, pglist, plinks.q);
275 	KASSERT(!unchanged || m->queue == queue,
276 	    ("page %p queue %d %d", m, queue, m->queue));
277 	TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
278 	return (unchanged);
279 }
280 
281 /*
282  * Lock the page while holding the page queue lock.  Use marker page
283  * to detect page queue changes and maintain notion of next page on
284  * page queue.  Return TRUE if no changes were detected, FALSE
285  * otherwise.  The page is locked on return. The page queue lock might
286  * be dropped and reacquired.
287  *
288  * This function depends on normal struct vm_page being type stable.
289  */
290 static boolean_t
291 vm_pageout_page_lock(vm_page_t m, vm_page_t *next)
292 {
293 	struct vm_page marker;
294 	struct vm_pagequeue *pq;
295 	boolean_t unchanged;
296 	u_short queue;
297 
298 	vm_page_lock_assert(m, MA_NOTOWNED);
299 	if (vm_page_trylock(m))
300 		return (TRUE);
301 
302 	queue = m->queue;
303 	vm_pageout_init_marker(&marker, queue);
304 	pq = vm_page_pagequeue(m);
305 
306 	TAILQ_INSERT_AFTER(&pq->pq_pl, m, &marker, plinks.q);
307 	vm_pagequeue_unlock(pq);
308 	vm_page_lock(m);
309 	vm_pagequeue_lock(pq);
310 
311 	/* Page queue might have changed. */
312 	*next = TAILQ_NEXT(&marker, plinks.q);
313 	unchanged = m == TAILQ_PREV(&marker, pglist, plinks.q);
314 	KASSERT(!unchanged || m->queue == queue,
315 	    ("page %p queue %d %d", m, queue, m->queue));
316 	TAILQ_REMOVE(&pq->pq_pl, &marker, plinks.q);
317 	return (unchanged);
318 }
319 
320 /*
321  * Scan for pages at adjacent offsets within the given page's object that are
322  * eligible for laundering, form a cluster of these pages and the given page,
323  * and launder that cluster.
324  */
325 static int
326 vm_pageout_cluster(vm_page_t m)
327 {
328 	vm_object_t object;
329 	vm_page_t mc[2 * vm_pageout_page_count], p, pb, ps;
330 	vm_pindex_t pindex;
331 	int ib, is, page_base, pageout_count;
332 
333 	vm_page_assert_locked(m);
334 	object = m->object;
335 	VM_OBJECT_ASSERT_WLOCKED(object);
336 	pindex = m->pindex;
337 
338 	vm_page_assert_unbusied(m);
339 	KASSERT(!vm_page_held(m), ("page %p is held", m));
340 
341 	pmap_remove_write(m);
342 	vm_page_unlock(m);
343 
344 	mc[vm_pageout_page_count] = pb = ps = m;
345 	pageout_count = 1;
346 	page_base = vm_pageout_page_count;
347 	ib = 1;
348 	is = 1;
349 
350 	/*
351 	 * We can cluster only if the page is not clean, busy, or held, and
352 	 * the page is in the laundry queue.
353 	 *
354 	 * During heavy mmap/modification loads the pageout
355 	 * daemon can really fragment the underlying file
356 	 * due to flushing pages out of order and not trying to
357 	 * align the clusters (which leaves sporadic out-of-order
358 	 * holes).  To solve this problem we do the reverse scan
359 	 * first and attempt to align our cluster, then do a
360 	 * forward scan if room remains.
361 	 */
362 more:
363 	while (ib != 0 && pageout_count < vm_pageout_page_count) {
364 		if (ib > pindex) {
365 			ib = 0;
366 			break;
367 		}
368 		if ((p = vm_page_prev(pb)) == NULL || vm_page_busied(p)) {
369 			ib = 0;
370 			break;
371 		}
372 		vm_page_test_dirty(p);
373 		if (p->dirty == 0) {
374 			ib = 0;
375 			break;
376 		}
377 		vm_page_lock(p);
378 		if (!vm_page_in_laundry(p) || vm_page_held(p)) {
379 			vm_page_unlock(p);
380 			ib = 0;
381 			break;
382 		}
383 		pmap_remove_write(p);
384 		vm_page_unlock(p);
385 		mc[--page_base] = pb = p;
386 		++pageout_count;
387 		++ib;
388 
389 		/*
390 		 * We are at an alignment boundary.  Stop here, and switch
391 		 * directions.  Do not clear ib.
392 		 */
393 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
394 			break;
395 	}
396 	while (pageout_count < vm_pageout_page_count &&
397 	    pindex + is < object->size) {
398 		if ((p = vm_page_next(ps)) == NULL || vm_page_busied(p))
399 			break;
400 		vm_page_test_dirty(p);
401 		if (p->dirty == 0)
402 			break;
403 		vm_page_lock(p);
404 		if (!vm_page_in_laundry(p) || vm_page_held(p)) {
405 			vm_page_unlock(p);
406 			break;
407 		}
408 		pmap_remove_write(p);
409 		vm_page_unlock(p);
410 		mc[page_base + pageout_count] = ps = p;
411 		++pageout_count;
412 		++is;
413 	}
414 
415 	/*
416 	 * If we exhausted our forward scan, continue with the reverse scan
417 	 * when possible, even past an alignment boundary.  This catches
418 	 * boundary conditions.
419 	 */
420 	if (ib != 0 && pageout_count < vm_pageout_page_count)
421 		goto more;
422 
423 	return (vm_pageout_flush(&mc[page_base], pageout_count,
424 	    VM_PAGER_PUT_NOREUSE, 0, NULL, NULL));
425 }
426 
427 /*
428  * vm_pageout_flush() - launder the given pages
429  *
430  *	The given pages are laundered.  Note that we setup for the start of
431  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
432  *	reference count all in here rather then in the parent.  If we want
433  *	the parent to do more sophisticated things we may have to change
434  *	the ordering.
435  *
436  *	Returned runlen is the count of pages between mreq and first
437  *	page after mreq with status VM_PAGER_AGAIN.
438  *	*eio is set to TRUE if pager returned VM_PAGER_ERROR or VM_PAGER_FAIL
439  *	for any page in runlen set.
440  */
441 int
442 vm_pageout_flush(vm_page_t *mc, int count, int flags, int mreq, int *prunlen,
443     boolean_t *eio)
444 {
445 	vm_object_t object = mc[0]->object;
446 	int pageout_status[count];
447 	int numpagedout = 0;
448 	int i, runlen;
449 
450 	VM_OBJECT_ASSERT_WLOCKED(object);
451 
452 	/*
453 	 * Initiate I/O.  Mark the pages busy and verify that they're valid
454 	 * and read-only.
455 	 *
456 	 * We do not have to fixup the clean/dirty bits here... we can
457 	 * allow the pager to do it after the I/O completes.
458 	 *
459 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
460 	 * edge case with file fragments.
461 	 */
462 	for (i = 0; i < count; i++) {
463 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
464 		    ("vm_pageout_flush: partially invalid page %p index %d/%d",
465 			mc[i], i, count));
466 		KASSERT((mc[i]->aflags & PGA_WRITEABLE) == 0,
467 		    ("vm_pageout_flush: writeable page %p", mc[i]));
468 		vm_page_sbusy(mc[i]);
469 	}
470 	vm_object_pip_add(object, count);
471 
472 	vm_pager_put_pages(object, mc, count, flags, pageout_status);
473 
474 	runlen = count - mreq;
475 	if (eio != NULL)
476 		*eio = FALSE;
477 	for (i = 0; i < count; i++) {
478 		vm_page_t mt = mc[i];
479 
480 		KASSERT(pageout_status[i] == VM_PAGER_PEND ||
481 		    !pmap_page_is_write_mapped(mt),
482 		    ("vm_pageout_flush: page %p is not write protected", mt));
483 		switch (pageout_status[i]) {
484 		case VM_PAGER_OK:
485 			vm_page_lock(mt);
486 			if (vm_page_in_laundry(mt))
487 				vm_page_deactivate_noreuse(mt);
488 			vm_page_unlock(mt);
489 			/* FALLTHROUGH */
490 		case VM_PAGER_PEND:
491 			numpagedout++;
492 			break;
493 		case VM_PAGER_BAD:
494 			/*
495 			 * The page is outside the object's range.  We pretend
496 			 * that the page out worked and clean the page, so the
497 			 * changes will be lost if the page is reclaimed by
498 			 * the page daemon.
499 			 */
500 			vm_page_undirty(mt);
501 			vm_page_lock(mt);
502 			if (vm_page_in_laundry(mt))
503 				vm_page_deactivate_noreuse(mt);
504 			vm_page_unlock(mt);
505 			break;
506 		case VM_PAGER_ERROR:
507 		case VM_PAGER_FAIL:
508 			/*
509 			 * If the page couldn't be paged out to swap because the
510 			 * pager wasn't able to find space, place the page in
511 			 * the PQ_UNSWAPPABLE holding queue.  This is an
512 			 * optimization that prevents the page daemon from
513 			 * wasting CPU cycles on pages that cannot be reclaimed
514 			 * becase no swap device is configured.
515 			 *
516 			 * Otherwise, reactivate the page so that it doesn't
517 			 * clog the laundry and inactive queues.  (We will try
518 			 * paging it out again later.)
519 			 */
520 			vm_page_lock(mt);
521 			if (object->type == OBJT_SWAP &&
522 			    pageout_status[i] == VM_PAGER_FAIL) {
523 				vm_page_unswappable(mt);
524 				numpagedout++;
525 			} else
526 				vm_page_activate(mt);
527 			vm_page_unlock(mt);
528 			if (eio != NULL && i >= mreq && i - mreq < runlen)
529 				*eio = TRUE;
530 			break;
531 		case VM_PAGER_AGAIN:
532 			if (i >= mreq && i - mreq < runlen)
533 				runlen = i - mreq;
534 			break;
535 		}
536 
537 		/*
538 		 * If the operation is still going, leave the page busy to
539 		 * block all other accesses. Also, leave the paging in
540 		 * progress indicator set so that we don't attempt an object
541 		 * collapse.
542 		 */
543 		if (pageout_status[i] != VM_PAGER_PEND) {
544 			vm_object_pip_wakeup(object);
545 			vm_page_sunbusy(mt);
546 		}
547 	}
548 	if (prunlen != NULL)
549 		*prunlen = runlen;
550 	return (numpagedout);
551 }
552 
553 static void
554 vm_pageout_swapon(void *arg __unused, struct swdevt *sp __unused)
555 {
556 
557 	atomic_store_rel_int(&swapdev_enabled, 1);
558 }
559 
560 static void
561 vm_pageout_swapoff(void *arg __unused, struct swdevt *sp __unused)
562 {
563 
564 	if (swap_pager_nswapdev() == 1)
565 		atomic_store_rel_int(&swapdev_enabled, 0);
566 }
567 
568 /*
569  * Attempt to acquire all of the necessary locks to launder a page and
570  * then call through the clustering layer to PUTPAGES.  Wait a short
571  * time for a vnode lock.
572  *
573  * Requires the page and object lock on entry, releases both before return.
574  * Returns 0 on success and an errno otherwise.
575  */
576 static int
577 vm_pageout_clean(vm_page_t m, int *numpagedout)
578 {
579 	struct vnode *vp;
580 	struct mount *mp;
581 	vm_object_t object;
582 	vm_pindex_t pindex;
583 	int error, lockmode;
584 
585 	vm_page_assert_locked(m);
586 	object = m->object;
587 	VM_OBJECT_ASSERT_WLOCKED(object);
588 	error = 0;
589 	vp = NULL;
590 	mp = NULL;
591 
592 	/*
593 	 * The object is already known NOT to be dead.   It
594 	 * is possible for the vget() to block the whole
595 	 * pageout daemon, but the new low-memory handling
596 	 * code should prevent it.
597 	 *
598 	 * We can't wait forever for the vnode lock, we might
599 	 * deadlock due to a vn_read() getting stuck in
600 	 * vm_wait while holding this vnode.  We skip the
601 	 * vnode if we can't get it in a reasonable amount
602 	 * of time.
603 	 */
604 	if (object->type == OBJT_VNODE) {
605 		vm_page_unlock(m);
606 		vp = object->handle;
607 		if (vp->v_type == VREG &&
608 		    vn_start_write(vp, &mp, V_NOWAIT) != 0) {
609 			mp = NULL;
610 			error = EDEADLK;
611 			goto unlock_all;
612 		}
613 		KASSERT(mp != NULL,
614 		    ("vp %p with NULL v_mount", vp));
615 		vm_object_reference_locked(object);
616 		pindex = m->pindex;
617 		VM_OBJECT_WUNLOCK(object);
618 		lockmode = MNT_SHARED_WRITES(vp->v_mount) ?
619 		    LK_SHARED : LK_EXCLUSIVE;
620 		if (vget(vp, lockmode | LK_TIMELOCK, curthread)) {
621 			vp = NULL;
622 			error = EDEADLK;
623 			goto unlock_mp;
624 		}
625 		VM_OBJECT_WLOCK(object);
626 
627 		/*
628 		 * Ensure that the object and vnode were not disassociated
629 		 * while locks were dropped.
630 		 */
631 		if (vp->v_object != object) {
632 			error = ENOENT;
633 			goto unlock_all;
634 		}
635 		vm_page_lock(m);
636 
637 		/*
638 		 * While the object and page were unlocked, the page
639 		 * may have been:
640 		 * (1) moved to a different queue,
641 		 * (2) reallocated to a different object,
642 		 * (3) reallocated to a different offset, or
643 		 * (4) cleaned.
644 		 */
645 		if (!vm_page_in_laundry(m) || m->object != object ||
646 		    m->pindex != pindex || m->dirty == 0) {
647 			vm_page_unlock(m);
648 			error = ENXIO;
649 			goto unlock_all;
650 		}
651 
652 		/*
653 		 * The page may have been busied or referenced while the object
654 		 * and page locks were released.
655 		 */
656 		if (vm_page_busied(m) || vm_page_held(m)) {
657 			vm_page_unlock(m);
658 			error = EBUSY;
659 			goto unlock_all;
660 		}
661 	}
662 
663 	/*
664 	 * If a page is dirty, then it is either being washed
665 	 * (but not yet cleaned) or it is still in the
666 	 * laundry.  If it is still in the laundry, then we
667 	 * start the cleaning operation.
668 	 */
669 	if ((*numpagedout = vm_pageout_cluster(m)) == 0)
670 		error = EIO;
671 
672 unlock_all:
673 	VM_OBJECT_WUNLOCK(object);
674 
675 unlock_mp:
676 	vm_page_lock_assert(m, MA_NOTOWNED);
677 	if (mp != NULL) {
678 		if (vp != NULL)
679 			vput(vp);
680 		vm_object_deallocate(object);
681 		vn_finished_write(mp);
682 	}
683 
684 	return (error);
685 }
686 
687 /*
688  * Attempt to launder the specified number of pages.
689  *
690  * Returns the number of pages successfully laundered.
691  */
692 static int
693 vm_pageout_launder(struct vm_domain *vmd, int launder, bool in_shortfall)
694 {
695 	struct vm_pagequeue *pq;
696 	vm_object_t object;
697 	vm_page_t m, next;
698 	int act_delta, error, maxscan, numpagedout, starting_target;
699 	int vnodes_skipped;
700 	bool pageout_ok, queue_locked;
701 
702 	starting_target = launder;
703 	vnodes_skipped = 0;
704 
705 	/*
706 	 * Scan the laundry queues for pages eligible to be laundered.  We stop
707 	 * once the target number of dirty pages have been laundered, or once
708 	 * we've reached the end of the queue.  A single iteration of this loop
709 	 * may cause more than one page to be laundered because of clustering.
710 	 *
711 	 * maxscan ensures that we don't re-examine requeued pages.  Any
712 	 * additional pages written as part of a cluster are subtracted from
713 	 * maxscan since they must be taken from the laundry queue.
714 	 *
715 	 * As an optimization, we avoid laundering from PQ_UNSWAPPABLE when no
716 	 * swap devices are configured.
717 	 */
718 	if (atomic_load_acq_int(&swapdev_enabled))
719 		pq = &vmd->vmd_pagequeues[PQ_UNSWAPPABLE];
720 	else
721 		pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
722 
723 scan:
724 	vm_pagequeue_lock(pq);
725 	maxscan = pq->pq_cnt;
726 	queue_locked = true;
727 	for (m = TAILQ_FIRST(&pq->pq_pl);
728 	    m != NULL && maxscan-- > 0 && launder > 0;
729 	    m = next) {
730 		vm_pagequeue_assert_locked(pq);
731 		KASSERT(queue_locked, ("unlocked laundry queue"));
732 		KASSERT(vm_page_in_laundry(m),
733 		    ("page %p has an inconsistent queue", m));
734 		next = TAILQ_NEXT(m, plinks.q);
735 		if ((m->flags & PG_MARKER) != 0)
736 			continue;
737 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
738 		    ("PG_FICTITIOUS page %p cannot be in laundry queue", m));
739 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
740 		    ("VPO_UNMANAGED page %p cannot be in laundry queue", m));
741 		if (!vm_pageout_page_lock(m, &next) || m->hold_count != 0) {
742 			vm_page_unlock(m);
743 			continue;
744 		}
745 		if (m->wire_count != 0) {
746 			vm_page_dequeue_locked(m);
747 			vm_page_unlock(m);
748 			continue;
749 		}
750 		object = m->object;
751 		if ((!VM_OBJECT_TRYWLOCK(object) &&
752 		    (!vm_pageout_fallback_object_lock(m, &next) ||
753 		    vm_page_held(m))) || vm_page_busied(m)) {
754 			VM_OBJECT_WUNLOCK(object);
755 			if (m->wire_count != 0 && vm_page_pagequeue(m) == pq)
756 				vm_page_dequeue_locked(m);
757 			vm_page_unlock(m);
758 			continue;
759 		}
760 
761 		/*
762 		 * Unlock the laundry queue, invalidating the 'next' pointer.
763 		 * Use a marker to remember our place in the laundry queue.
764 		 */
765 		TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_laundry_marker,
766 		    plinks.q);
767 		vm_pagequeue_unlock(pq);
768 		queue_locked = false;
769 
770 		/*
771 		 * Invalid pages can be easily freed.  They cannot be
772 		 * mapped; vm_page_free() asserts this.
773 		 */
774 		if (m->valid == 0)
775 			goto free_page;
776 
777 		/*
778 		 * If the page has been referenced and the object is not dead,
779 		 * reactivate or requeue the page depending on whether the
780 		 * object is mapped.
781 		 */
782 		if ((m->aflags & PGA_REFERENCED) != 0) {
783 			vm_page_aflag_clear(m, PGA_REFERENCED);
784 			act_delta = 1;
785 		} else
786 			act_delta = 0;
787 		if (object->ref_count != 0)
788 			act_delta += pmap_ts_referenced(m);
789 		else {
790 			KASSERT(!pmap_page_is_mapped(m),
791 			    ("page %p is mapped", m));
792 		}
793 		if (act_delta != 0) {
794 			if (object->ref_count != 0) {
795 				VM_CNT_INC(v_reactivated);
796 				vm_page_activate(m);
797 
798 				/*
799 				 * Increase the activation count if the page
800 				 * was referenced while in the laundry queue.
801 				 * This makes it less likely that the page will
802 				 * be returned prematurely to the inactive
803 				 * queue.
804  				 */
805 				m->act_count += act_delta + ACT_ADVANCE;
806 
807 				/*
808 				 * If this was a background laundering, count
809 				 * activated pages towards our target.  The
810 				 * purpose of background laundering is to ensure
811 				 * that pages are eventually cycled through the
812 				 * laundry queue, and an activation is a valid
813 				 * way out.
814 				 */
815 				if (!in_shortfall)
816 					launder--;
817 				goto drop_page;
818 			} else if ((object->flags & OBJ_DEAD) == 0)
819 				goto requeue_page;
820 		}
821 
822 		/*
823 		 * If the page appears to be clean at the machine-independent
824 		 * layer, then remove all of its mappings from the pmap in
825 		 * anticipation of freeing it.  If, however, any of the page's
826 		 * mappings allow write access, then the page may still be
827 		 * modified until the last of those mappings are removed.
828 		 */
829 		if (object->ref_count != 0) {
830 			vm_page_test_dirty(m);
831 			if (m->dirty == 0)
832 				pmap_remove_all(m);
833 		}
834 
835 		/*
836 		 * Clean pages are freed, and dirty pages are paged out unless
837 		 * they belong to a dead object.  Requeueing dirty pages from
838 		 * dead objects is pointless, as they are being paged out and
839 		 * freed by the thread that destroyed the object.
840 		 */
841 		if (m->dirty == 0) {
842 free_page:
843 			vm_page_free(m);
844 			VM_CNT_INC(v_dfree);
845 		} else if ((object->flags & OBJ_DEAD) == 0) {
846 			if (object->type != OBJT_SWAP &&
847 			    object->type != OBJT_DEFAULT)
848 				pageout_ok = true;
849 			else if (disable_swap_pageouts)
850 				pageout_ok = false;
851 			else
852 				pageout_ok = true;
853 			if (!pageout_ok) {
854 requeue_page:
855 				vm_pagequeue_lock(pq);
856 				queue_locked = true;
857 				vm_page_requeue_locked(m);
858 				goto drop_page;
859 			}
860 
861 			/*
862 			 * Form a cluster with adjacent, dirty pages from the
863 			 * same object, and page out that entire cluster.
864 			 *
865 			 * The adjacent, dirty pages must also be in the
866 			 * laundry.  However, their mappings are not checked
867 			 * for new references.  Consequently, a recently
868 			 * referenced page may be paged out.  However, that
869 			 * page will not be prematurely reclaimed.  After page
870 			 * out, the page will be placed in the inactive queue,
871 			 * where any new references will be detected and the
872 			 * page reactivated.
873 			 */
874 			error = vm_pageout_clean(m, &numpagedout);
875 			if (error == 0) {
876 				launder -= numpagedout;
877 				maxscan -= numpagedout - 1;
878 			} else if (error == EDEADLK) {
879 				pageout_lock_miss++;
880 				vnodes_skipped++;
881 			}
882 			goto relock_queue;
883 		}
884 drop_page:
885 		vm_page_unlock(m);
886 		VM_OBJECT_WUNLOCK(object);
887 relock_queue:
888 		if (!queue_locked) {
889 			vm_pagequeue_lock(pq);
890 			queue_locked = true;
891 		}
892 		next = TAILQ_NEXT(&vmd->vmd_laundry_marker, plinks.q);
893 		TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_laundry_marker, plinks.q);
894 	}
895 	vm_pagequeue_unlock(pq);
896 
897 	if (launder > 0 && pq == &vmd->vmd_pagequeues[PQ_UNSWAPPABLE]) {
898 		pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
899 		goto scan;
900 	}
901 
902 	/*
903 	 * Wakeup the sync daemon if we skipped a vnode in a writeable object
904 	 * and we didn't launder enough pages.
905 	 */
906 	if (vnodes_skipped > 0 && launder > 0)
907 		(void)speedup_syncer();
908 
909 	return (starting_target - launder);
910 }
911 
912 /*
913  * Compute the integer square root.
914  */
915 static u_int
916 isqrt(u_int num)
917 {
918 	u_int bit, root, tmp;
919 
920 	bit = 1u << ((NBBY * sizeof(u_int)) - 2);
921 	while (bit > num)
922 		bit >>= 2;
923 	root = 0;
924 	while (bit != 0) {
925 		tmp = root + bit;
926 		root >>= 1;
927 		if (num >= tmp) {
928 			num -= tmp;
929 			root += bit;
930 		}
931 		bit >>= 2;
932 	}
933 	return (root);
934 }
935 
936 /*
937  * Perform the work of the laundry thread: periodically wake up and determine
938  * whether any pages need to be laundered.  If so, determine the number of pages
939  * that need to be laundered, and launder them.
940  */
941 static void
942 vm_pageout_laundry_worker(void *arg)
943 {
944 	struct vm_domain *vmd;
945 	struct vm_pagequeue *pq;
946 	uint64_t nclean, ndirty, nfreed;
947 	int domain, last_target, launder, shortfall, shortfall_cycle, target;
948 	bool in_shortfall;
949 
950 	domain = (uintptr_t)arg;
951 	vmd = VM_DOMAIN(domain);
952 	pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
953 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
954 	vm_pageout_init_marker(&vmd->vmd_laundry_marker, PQ_LAUNDRY);
955 
956 	shortfall = 0;
957 	in_shortfall = false;
958 	shortfall_cycle = 0;
959 	target = 0;
960 	nfreed = 0;
961 
962 	/*
963 	 * Calls to these handlers are serialized by the swap syscall lock.
964 	 */
965 	(void)EVENTHANDLER_REGISTER(swapon, vm_pageout_swapon, vmd,
966 	    EVENTHANDLER_PRI_ANY);
967 	(void)EVENTHANDLER_REGISTER(swapoff, vm_pageout_swapoff, vmd,
968 	    EVENTHANDLER_PRI_ANY);
969 
970 	/*
971 	 * The pageout laundry worker is never done, so loop forever.
972 	 */
973 	for (;;) {
974 		KASSERT(target >= 0, ("negative target %d", target));
975 		KASSERT(shortfall_cycle >= 0,
976 		    ("negative cycle %d", shortfall_cycle));
977 		launder = 0;
978 
979 		/*
980 		 * First determine whether we need to launder pages to meet a
981 		 * shortage of free pages.
982 		 */
983 		if (shortfall > 0) {
984 			in_shortfall = true;
985 			shortfall_cycle = VM_LAUNDER_RATE / VM_INACT_SCAN_RATE;
986 			target = shortfall;
987 		} else if (!in_shortfall)
988 			goto trybackground;
989 		else if (shortfall_cycle == 0 || vm_laundry_target(vmd) <= 0) {
990 			/*
991 			 * We recently entered shortfall and began laundering
992 			 * pages.  If we have completed that laundering run
993 			 * (and we are no longer in shortfall) or we have met
994 			 * our laundry target through other activity, then we
995 			 * can stop laundering pages.
996 			 */
997 			in_shortfall = false;
998 			target = 0;
999 			goto trybackground;
1000 		}
1001 		launder = target / shortfall_cycle--;
1002 		goto dolaundry;
1003 
1004 		/*
1005 		 * There's no immediate need to launder any pages; see if we
1006 		 * meet the conditions to perform background laundering:
1007 		 *
1008 		 * 1. The ratio of dirty to clean inactive pages exceeds the
1009 		 *    background laundering threshold, or
1010 		 * 2. we haven't yet reached the target of the current
1011 		 *    background laundering run.
1012 		 *
1013 		 * The background laundering threshold is not a constant.
1014 		 * Instead, it is a slowly growing function of the number of
1015 		 * clean pages freed by the page daemon since the last
1016 		 * background laundering.  Thus, as the ratio of dirty to
1017 		 * clean inactive pages grows, the amount of memory pressure
1018 		 * required to trigger laundering decreases.  We ensure
1019 		 * that the threshold is non-zero after an inactive queue
1020 		 * scan, even if that scan failed to free a single clean page.
1021 		 */
1022 trybackground:
1023 		nclean = vmd->vmd_free_count +
1024 		    vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt;
1025 		ndirty = vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt;
1026 		if (target == 0 && ndirty * isqrt(howmany(nfreed + 1,
1027 		    vmd->vmd_free_target - vmd->vmd_free_min)) >= nclean) {
1028 			target = vmd->vmd_background_launder_target;
1029 		}
1030 
1031 		/*
1032 		 * We have a non-zero background laundering target.  If we've
1033 		 * laundered up to our maximum without observing a page daemon
1034 		 * request, just stop.  This is a safety belt that ensures we
1035 		 * don't launder an excessive amount if memory pressure is low
1036 		 * and the ratio of dirty to clean pages is large.  Otherwise,
1037 		 * proceed at the background laundering rate.
1038 		 */
1039 		if (target > 0) {
1040 			if (nfreed > 0) {
1041 				nfreed = 0;
1042 				last_target = target;
1043 			} else if (last_target - target >=
1044 			    vm_background_launder_max * PAGE_SIZE / 1024) {
1045 				target = 0;
1046 			}
1047 			launder = vm_background_launder_rate * PAGE_SIZE / 1024;
1048 			launder /= VM_LAUNDER_RATE;
1049 			if (launder > target)
1050 				launder = target;
1051 		}
1052 
1053 dolaundry:
1054 		if (launder > 0) {
1055 			/*
1056 			 * Because of I/O clustering, the number of laundered
1057 			 * pages could exceed "target" by the maximum size of
1058 			 * a cluster minus one.
1059 			 */
1060 			target -= min(vm_pageout_launder(vmd, launder,
1061 			    in_shortfall), target);
1062 			pause("laundp", hz / VM_LAUNDER_RATE);
1063 		}
1064 
1065 		/*
1066 		 * If we're not currently laundering pages and the page daemon
1067 		 * hasn't posted a new request, sleep until the page daemon
1068 		 * kicks us.
1069 		 */
1070 		vm_pagequeue_lock(pq);
1071 		if (target == 0 && vmd->vmd_laundry_request == VM_LAUNDRY_IDLE)
1072 			(void)mtx_sleep(&vmd->vmd_laundry_request,
1073 			    vm_pagequeue_lockptr(pq), PVM, "launds", 0);
1074 
1075 		/*
1076 		 * If the pagedaemon has indicated that it's in shortfall, start
1077 		 * a shortfall laundering unless we're already in the middle of
1078 		 * one.  This may preempt a background laundering.
1079 		 */
1080 		if (vmd->vmd_laundry_request == VM_LAUNDRY_SHORTFALL &&
1081 		    (!in_shortfall || shortfall_cycle == 0)) {
1082 			shortfall = vm_laundry_target(vmd) +
1083 			    vmd->vmd_pageout_deficit;
1084 			target = 0;
1085 		} else
1086 			shortfall = 0;
1087 
1088 		if (target == 0)
1089 			vmd->vmd_laundry_request = VM_LAUNDRY_IDLE;
1090 		nfreed += vmd->vmd_clean_pages_freed;
1091 		vmd->vmd_clean_pages_freed = 0;
1092 		vm_pagequeue_unlock(pq);
1093 	}
1094 }
1095 
1096 /*
1097  *	vm_pageout_scan does the dirty work for the pageout daemon.
1098  *
1099  *	pass == 0: Update active LRU/deactivate pages
1100  *	pass >= 1: Free inactive pages
1101  *
1102  * Returns true if pass was zero or enough pages were freed by the inactive
1103  * queue scan to meet the target.
1104  */
1105 static bool
1106 vm_pageout_scan(struct vm_domain *vmd, int pass, int shortage)
1107 {
1108 	vm_page_t m, next;
1109 	struct vm_pagequeue *pq;
1110 	vm_object_t object;
1111 	long min_scan;
1112 	int act_delta, addl_page_shortage, deficit, inactq_shortage, maxscan;
1113 	int page_shortage, scan_tick, scanned, starting_page_shortage;
1114 	boolean_t queue_locked;
1115 
1116 	/*
1117 	 * If we need to reclaim memory ask kernel caches to return
1118 	 * some.  We rate limit to avoid thrashing.
1119 	 */
1120 	if (vmd == VM_DOMAIN(0) && pass > 0 &&
1121 	    (time_uptime - lowmem_uptime) >= lowmem_period) {
1122 		/*
1123 		 * Decrease registered cache sizes.
1124 		 */
1125 		SDT_PROBE0(vm, , , vm__lowmem_scan);
1126 		EVENTHANDLER_INVOKE(vm_lowmem, VM_LOW_PAGES);
1127 		/*
1128 		 * We do this explicitly after the caches have been
1129 		 * drained above.
1130 		 */
1131 		uma_reclaim();
1132 		lowmem_uptime = time_uptime;
1133 	}
1134 
1135 	/*
1136 	 * The addl_page_shortage is the number of temporarily
1137 	 * stuck pages in the inactive queue.  In other words, the
1138 	 * number of pages from the inactive count that should be
1139 	 * discounted in setting the target for the active queue scan.
1140 	 */
1141 	addl_page_shortage = 0;
1142 
1143 	/*
1144 	 * Calculate the number of pages that we want to free.  This number
1145 	 * can be negative if many pages are freed between the wakeup call to
1146 	 * the page daemon and this calculation.
1147 	 */
1148 	if (pass > 0) {
1149 		deficit = atomic_readandclear_int(&vmd->vmd_pageout_deficit);
1150 		page_shortage = shortage + deficit;
1151 	} else
1152 		page_shortage = deficit = 0;
1153 	starting_page_shortage = page_shortage;
1154 
1155 	/*
1156 	 * Start scanning the inactive queue for pages that we can free.  The
1157 	 * scan will stop when we reach the target or we have scanned the
1158 	 * entire queue.  (Note that m->act_count is not used to make
1159 	 * decisions for the inactive queue, only for the active queue.)
1160 	 */
1161 	pq = &vmd->vmd_pagequeues[PQ_INACTIVE];
1162 	maxscan = pq->pq_cnt;
1163 	vm_pagequeue_lock(pq);
1164 	queue_locked = TRUE;
1165 	for (m = TAILQ_FIRST(&pq->pq_pl);
1166 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
1167 	     m = next) {
1168 		vm_pagequeue_assert_locked(pq);
1169 		KASSERT(queue_locked, ("unlocked inactive queue"));
1170 		KASSERT(vm_page_inactive(m), ("Inactive queue %p", m));
1171 
1172 		VM_CNT_INC(v_pdpages);
1173 		next = TAILQ_NEXT(m, plinks.q);
1174 
1175 		/*
1176 		 * skip marker pages
1177 		 */
1178 		if (m->flags & PG_MARKER)
1179 			continue;
1180 
1181 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
1182 		    ("Fictitious page %p cannot be in inactive queue", m));
1183 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1184 		    ("Unmanaged page %p cannot be in inactive queue", m));
1185 
1186 		/*
1187 		 * The page or object lock acquisitions fail if the
1188 		 * page was removed from the queue or moved to a
1189 		 * different position within the queue.  In either
1190 		 * case, addl_page_shortage should not be incremented.
1191 		 */
1192 		if (!vm_pageout_page_lock(m, &next))
1193 			goto unlock_page;
1194 		else if (m->wire_count != 0) {
1195 			/*
1196 			 * Wired pages may not be freed, and unwiring a queued
1197 			 * page will cause it to be requeued.  Thus, remove them
1198 			 * from the queue now to avoid unnecessary revisits.
1199 			 */
1200 			vm_page_dequeue_locked(m);
1201 			addl_page_shortage++;
1202 			goto unlock_page;
1203 		} else if (m->hold_count != 0) {
1204 			/*
1205 			 * Held pages are essentially stuck in the
1206 			 * queue.  So, they ought to be discounted
1207 			 * from the inactive count.  See the
1208 			 * calculation of inactq_shortage before the
1209 			 * loop over the active queue below.
1210 			 */
1211 			addl_page_shortage++;
1212 			goto unlock_page;
1213 		}
1214 		object = m->object;
1215 		if (!VM_OBJECT_TRYWLOCK(object)) {
1216 			if (!vm_pageout_fallback_object_lock(m, &next))
1217 				goto unlock_object;
1218 			else if (m->wire_count != 0) {
1219 				vm_page_dequeue_locked(m);
1220 				addl_page_shortage++;
1221 				goto unlock_object;
1222 			} else if (m->hold_count != 0) {
1223 				addl_page_shortage++;
1224 				goto unlock_object;
1225 			}
1226 		}
1227 		if (vm_page_busied(m)) {
1228 			/*
1229 			 * Don't mess with busy pages.  Leave them at
1230 			 * the front of the queue.  Most likely, they
1231 			 * are being paged out and will leave the
1232 			 * queue shortly after the scan finishes.  So,
1233 			 * they ought to be discounted from the
1234 			 * inactive count.
1235 			 */
1236 			addl_page_shortage++;
1237 unlock_object:
1238 			VM_OBJECT_WUNLOCK(object);
1239 unlock_page:
1240 			vm_page_unlock(m);
1241 			continue;
1242 		}
1243 		KASSERT(!vm_page_held(m), ("Held page %p", m));
1244 
1245 		/*
1246 		 * Dequeue the inactive page and unlock the inactive page
1247 		 * queue, invalidating the 'next' pointer.  Dequeueing the
1248 		 * page here avoids a later reacquisition (and release) of
1249 		 * the inactive page queue lock when vm_page_activate(),
1250 		 * vm_page_free(), or vm_page_launder() is called.  Use a
1251 		 * marker to remember our place in the inactive queue.
1252 		 */
1253 		TAILQ_INSERT_AFTER(&pq->pq_pl, m, &vmd->vmd_marker, plinks.q);
1254 		vm_page_dequeue_locked(m);
1255 		vm_pagequeue_unlock(pq);
1256 		queue_locked = FALSE;
1257 
1258 		/*
1259 		 * Invalid pages can be easily freed. They cannot be
1260 		 * mapped, vm_page_free() asserts this.
1261 		 */
1262 		if (m->valid == 0)
1263 			goto free_page;
1264 
1265 		/*
1266 		 * If the page has been referenced and the object is not dead,
1267 		 * reactivate or requeue the page depending on whether the
1268 		 * object is mapped.
1269 		 */
1270 		if ((m->aflags & PGA_REFERENCED) != 0) {
1271 			vm_page_aflag_clear(m, PGA_REFERENCED);
1272 			act_delta = 1;
1273 		} else
1274 			act_delta = 0;
1275 		if (object->ref_count != 0) {
1276 			act_delta += pmap_ts_referenced(m);
1277 		} else {
1278 			KASSERT(!pmap_page_is_mapped(m),
1279 			    ("vm_pageout_scan: page %p is mapped", m));
1280 		}
1281 		if (act_delta != 0) {
1282 			if (object->ref_count != 0) {
1283 				VM_CNT_INC(v_reactivated);
1284 				vm_page_activate(m);
1285 
1286 				/*
1287 				 * Increase the activation count if the page
1288 				 * was referenced while in the inactive queue.
1289 				 * This makes it less likely that the page will
1290 				 * be returned prematurely to the inactive
1291 				 * queue.
1292  				 */
1293 				m->act_count += act_delta + ACT_ADVANCE;
1294 				goto drop_page;
1295 			} else if ((object->flags & OBJ_DEAD) == 0) {
1296 				vm_pagequeue_lock(pq);
1297 				queue_locked = TRUE;
1298 				m->queue = PQ_INACTIVE;
1299 				TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
1300 				vm_pagequeue_cnt_inc(pq);
1301 				goto drop_page;
1302 			}
1303 		}
1304 
1305 		/*
1306 		 * If the page appears to be clean at the machine-independent
1307 		 * layer, then remove all of its mappings from the pmap in
1308 		 * anticipation of freeing it.  If, however, any of the page's
1309 		 * mappings allow write access, then the page may still be
1310 		 * modified until the last of those mappings are removed.
1311 		 */
1312 		if (object->ref_count != 0) {
1313 			vm_page_test_dirty(m);
1314 			if (m->dirty == 0)
1315 				pmap_remove_all(m);
1316 		}
1317 
1318 		/*
1319 		 * Clean pages can be freed, but dirty pages must be sent back
1320 		 * to the laundry, unless they belong to a dead object.
1321 		 * Requeueing dirty pages from dead objects is pointless, as
1322 		 * they are being paged out and freed by the thread that
1323 		 * destroyed the object.
1324 		 */
1325 		if (m->dirty == 0) {
1326 free_page:
1327 			vm_page_free(m);
1328 			VM_CNT_INC(v_dfree);
1329 			--page_shortage;
1330 		} else if ((object->flags & OBJ_DEAD) == 0)
1331 			vm_page_launder(m);
1332 drop_page:
1333 		vm_page_unlock(m);
1334 		VM_OBJECT_WUNLOCK(object);
1335 		if (!queue_locked) {
1336 			vm_pagequeue_lock(pq);
1337 			queue_locked = TRUE;
1338 		}
1339 		next = TAILQ_NEXT(&vmd->vmd_marker, plinks.q);
1340 		TAILQ_REMOVE(&pq->pq_pl, &vmd->vmd_marker, plinks.q);
1341 	}
1342 	vm_pagequeue_unlock(pq);
1343 
1344 	/*
1345 	 * Wake up the laundry thread so that it can perform any needed
1346 	 * laundering.  If we didn't meet our target, we're in shortfall and
1347 	 * need to launder more aggressively.  If PQ_LAUNDRY is empty and no
1348 	 * swap devices are configured, the laundry thread has no work to do, so
1349 	 * don't bother waking it up.
1350 	 *
1351 	 * The laundry thread uses the number of inactive queue scans elapsed
1352 	 * since the last laundering to determine whether to launder again, so
1353 	 * keep count.
1354 	 */
1355 	if (starting_page_shortage > 0) {
1356 		pq = &vmd->vmd_pagequeues[PQ_LAUNDRY];
1357 		vm_pagequeue_lock(pq);
1358 		if (vmd->vmd_laundry_request == VM_LAUNDRY_IDLE &&
1359 		    (pq->pq_cnt > 0 || atomic_load_acq_int(&swapdev_enabled))) {
1360 			if (page_shortage > 0) {
1361 				vmd->vmd_laundry_request = VM_LAUNDRY_SHORTFALL;
1362 				VM_CNT_INC(v_pdshortfalls);
1363 			} else if (vmd->vmd_laundry_request !=
1364 			    VM_LAUNDRY_SHORTFALL)
1365 				vmd->vmd_laundry_request =
1366 				    VM_LAUNDRY_BACKGROUND;
1367 			wakeup(&vmd->vmd_laundry_request);
1368 		}
1369 		vmd->vmd_clean_pages_freed +=
1370 		    starting_page_shortage - page_shortage;
1371 		vm_pagequeue_unlock(pq);
1372 	}
1373 
1374 	/*
1375 	 * Wakeup the swapout daemon if we didn't free the targeted number of
1376 	 * pages.
1377 	 */
1378 	if (page_shortage > 0)
1379 		vm_swapout_run();
1380 
1381 	/*
1382 	 * If the inactive queue scan fails repeatedly to meet its
1383 	 * target, kill the largest process.
1384 	 */
1385 	vm_pageout_mightbe_oom(vmd, page_shortage, starting_page_shortage);
1386 
1387 	/*
1388 	 * Compute the number of pages we want to try to move from the
1389 	 * active queue to either the inactive or laundry queue.
1390 	 *
1391 	 * When scanning active pages, we make clean pages count more heavily
1392 	 * towards the page shortage than dirty pages.  This is because dirty
1393 	 * pages must be laundered before they can be reused and thus have less
1394 	 * utility when attempting to quickly alleviate a shortage.  However,
1395 	 * this weighting also causes the scan to deactivate dirty pages more
1396 	 * more aggressively, improving the effectiveness of clustering and
1397 	 * ensuring that they can eventually be reused.
1398 	 */
1399 	inactq_shortage = vmd->vmd_inactive_target - (pq->pq_cnt +
1400 	    vmd->vmd_pagequeues[PQ_LAUNDRY].pq_cnt / act_scan_laundry_weight) +
1401 	    vm_paging_target(vmd) + deficit + addl_page_shortage;
1402 	inactq_shortage *= act_scan_laundry_weight;
1403 
1404 	pq = &vmd->vmd_pagequeues[PQ_ACTIVE];
1405 	vm_pagequeue_lock(pq);
1406 	maxscan = pq->pq_cnt;
1407 
1408 	/*
1409 	 * If we're just idle polling attempt to visit every
1410 	 * active page within 'update_period' seconds.
1411 	 */
1412 	scan_tick = ticks;
1413 	if (vm_pageout_update_period != 0) {
1414 		min_scan = pq->pq_cnt;
1415 		min_scan *= scan_tick - vmd->vmd_last_active_scan;
1416 		min_scan /= hz * vm_pageout_update_period;
1417 	} else
1418 		min_scan = 0;
1419 	if (min_scan > 0 || (inactq_shortage > 0 && maxscan > 0))
1420 		vmd->vmd_last_active_scan = scan_tick;
1421 
1422 	/*
1423 	 * Scan the active queue for pages that can be deactivated.  Update
1424 	 * the per-page activity counter and use it to identify deactivation
1425 	 * candidates.  Held pages may be deactivated.
1426 	 */
1427 	for (m = TAILQ_FIRST(&pq->pq_pl), scanned = 0; m != NULL && (scanned <
1428 	    min_scan || (inactq_shortage > 0 && scanned < maxscan)); m = next,
1429 	    scanned++) {
1430 		KASSERT(m->queue == PQ_ACTIVE,
1431 		    ("vm_pageout_scan: page %p isn't active", m));
1432 		next = TAILQ_NEXT(m, plinks.q);
1433 		if ((m->flags & PG_MARKER) != 0)
1434 			continue;
1435 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
1436 		    ("Fictitious page %p cannot be in active queue", m));
1437 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1438 		    ("Unmanaged page %p cannot be in active queue", m));
1439 		if (!vm_pageout_page_lock(m, &next)) {
1440 			vm_page_unlock(m);
1441 			continue;
1442 		}
1443 
1444 		/*
1445 		 * The count for page daemon pages is updated after checking
1446 		 * the page for eligibility.
1447 		 */
1448 		VM_CNT_INC(v_pdpages);
1449 
1450 		/*
1451 		 * Wired pages are dequeued lazily.
1452 		 */
1453 		if (m->wire_count != 0) {
1454 			vm_page_dequeue_locked(m);
1455 			vm_page_unlock(m);
1456 			continue;
1457 		}
1458 
1459 		/*
1460 		 * Check to see "how much" the page has been used.
1461 		 */
1462 		if ((m->aflags & PGA_REFERENCED) != 0) {
1463 			vm_page_aflag_clear(m, PGA_REFERENCED);
1464 			act_delta = 1;
1465 		} else
1466 			act_delta = 0;
1467 
1468 		/*
1469 		 * Perform an unsynchronized object ref count check.  While
1470 		 * the page lock ensures that the page is not reallocated to
1471 		 * another object, in particular, one with unmanaged mappings
1472 		 * that cannot support pmap_ts_referenced(), two races are,
1473 		 * nonetheless, possible:
1474 		 * 1) The count was transitioning to zero, but we saw a non-
1475 		 *    zero value.  pmap_ts_referenced() will return zero
1476 		 *    because the page is not mapped.
1477 		 * 2) The count was transitioning to one, but we saw zero.
1478 		 *    This race delays the detection of a new reference.  At
1479 		 *    worst, we will deactivate and reactivate the page.
1480 		 */
1481 		if (m->object->ref_count != 0)
1482 			act_delta += pmap_ts_referenced(m);
1483 
1484 		/*
1485 		 * Advance or decay the act_count based on recent usage.
1486 		 */
1487 		if (act_delta != 0) {
1488 			m->act_count += ACT_ADVANCE + act_delta;
1489 			if (m->act_count > ACT_MAX)
1490 				m->act_count = ACT_MAX;
1491 		} else
1492 			m->act_count -= min(m->act_count, ACT_DECLINE);
1493 
1494 		/*
1495 		 * Move this page to the tail of the active, inactive or laundry
1496 		 * queue depending on usage.
1497 		 */
1498 		if (m->act_count == 0) {
1499 			/* Dequeue to avoid later lock recursion. */
1500 			vm_page_dequeue_locked(m);
1501 
1502 			/*
1503 			 * When not short for inactive pages, let dirty pages go
1504 			 * through the inactive queue before moving to the
1505 			 * laundry queues.  This gives them some extra time to
1506 			 * be reactivated, potentially avoiding an expensive
1507 			 * pageout.  During a page shortage, the inactive queue
1508 			 * is necessarily small, so we may move dirty pages
1509 			 * directly to the laundry queue.
1510 			 */
1511 			if (inactq_shortage <= 0)
1512 				vm_page_deactivate(m);
1513 			else {
1514 				/*
1515 				 * Calling vm_page_test_dirty() here would
1516 				 * require acquisition of the object's write
1517 				 * lock.  However, during a page shortage,
1518 				 * directing dirty pages into the laundry
1519 				 * queue is only an optimization and not a
1520 				 * requirement.  Therefore, we simply rely on
1521 				 * the opportunistic updates to the page's
1522 				 * dirty field by the pmap.
1523 				 */
1524 				if (m->dirty == 0) {
1525 					vm_page_deactivate(m);
1526 					inactq_shortage -=
1527 					    act_scan_laundry_weight;
1528 				} else {
1529 					vm_page_launder(m);
1530 					inactq_shortage--;
1531 				}
1532 			}
1533 		} else
1534 			vm_page_requeue_locked(m);
1535 		vm_page_unlock(m);
1536 	}
1537 	vm_pagequeue_unlock(pq);
1538 	if (pass > 0)
1539 		vm_swapout_run_idle();
1540 	return (page_shortage <= 0);
1541 }
1542 
1543 static int vm_pageout_oom_vote;
1544 
1545 /*
1546  * The pagedaemon threads randlomly select one to perform the
1547  * OOM.  Trying to kill processes before all pagedaemons
1548  * failed to reach free target is premature.
1549  */
1550 static void
1551 vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
1552     int starting_page_shortage)
1553 {
1554 	int old_vote;
1555 
1556 	if (starting_page_shortage <= 0 || starting_page_shortage !=
1557 	    page_shortage)
1558 		vmd->vmd_oom_seq = 0;
1559 	else
1560 		vmd->vmd_oom_seq++;
1561 	if (vmd->vmd_oom_seq < vm_pageout_oom_seq) {
1562 		if (vmd->vmd_oom) {
1563 			vmd->vmd_oom = FALSE;
1564 			atomic_subtract_int(&vm_pageout_oom_vote, 1);
1565 		}
1566 		return;
1567 	}
1568 
1569 	/*
1570 	 * Do not follow the call sequence until OOM condition is
1571 	 * cleared.
1572 	 */
1573 	vmd->vmd_oom_seq = 0;
1574 
1575 	if (vmd->vmd_oom)
1576 		return;
1577 
1578 	vmd->vmd_oom = TRUE;
1579 	old_vote = atomic_fetchadd_int(&vm_pageout_oom_vote, 1);
1580 	if (old_vote != vm_ndomains - 1)
1581 		return;
1582 
1583 	/*
1584 	 * The current pagedaemon thread is the last in the quorum to
1585 	 * start OOM.  Initiate the selection and signaling of the
1586 	 * victim.
1587 	 */
1588 	vm_pageout_oom(VM_OOM_MEM);
1589 
1590 	/*
1591 	 * After one round of OOM terror, recall our vote.  On the
1592 	 * next pass, current pagedaemon would vote again if the low
1593 	 * memory condition is still there, due to vmd_oom being
1594 	 * false.
1595 	 */
1596 	vmd->vmd_oom = FALSE;
1597 	atomic_subtract_int(&vm_pageout_oom_vote, 1);
1598 }
1599 
1600 /*
1601  * The OOM killer is the page daemon's action of last resort when
1602  * memory allocation requests have been stalled for a prolonged period
1603  * of time because it cannot reclaim memory.  This function computes
1604  * the approximate number of physical pages that could be reclaimed if
1605  * the specified address space is destroyed.
1606  *
1607  * Private, anonymous memory owned by the address space is the
1608  * principal resource that we expect to recover after an OOM kill.
1609  * Since the physical pages mapped by the address space's COW entries
1610  * are typically shared pages, they are unlikely to be released and so
1611  * they are not counted.
1612  *
1613  * To get to the point where the page daemon runs the OOM killer, its
1614  * efforts to write-back vnode-backed pages may have stalled.  This
1615  * could be caused by a memory allocation deadlock in the write path
1616  * that might be resolved by an OOM kill.  Therefore, physical pages
1617  * belonging to vnode-backed objects are counted, because they might
1618  * be freed without being written out first if the address space holds
1619  * the last reference to an unlinked vnode.
1620  *
1621  * Similarly, physical pages belonging to OBJT_PHYS objects are
1622  * counted because the address space might hold the last reference to
1623  * the object.
1624  */
1625 static long
1626 vm_pageout_oom_pagecount(struct vmspace *vmspace)
1627 {
1628 	vm_map_t map;
1629 	vm_map_entry_t entry;
1630 	vm_object_t obj;
1631 	long res;
1632 
1633 	map = &vmspace->vm_map;
1634 	KASSERT(!map->system_map, ("system map"));
1635 	sx_assert(&map->lock, SA_LOCKED);
1636 	res = 0;
1637 	for (entry = map->header.next; entry != &map->header;
1638 	    entry = entry->next) {
1639 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0)
1640 			continue;
1641 		obj = entry->object.vm_object;
1642 		if (obj == NULL)
1643 			continue;
1644 		if ((entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0 &&
1645 		    obj->ref_count != 1)
1646 			continue;
1647 		switch (obj->type) {
1648 		case OBJT_DEFAULT:
1649 		case OBJT_SWAP:
1650 		case OBJT_PHYS:
1651 		case OBJT_VNODE:
1652 			res += obj->resident_page_count;
1653 			break;
1654 		}
1655 	}
1656 	return (res);
1657 }
1658 
1659 void
1660 vm_pageout_oom(int shortage)
1661 {
1662 	struct proc *p, *bigproc;
1663 	vm_offset_t size, bigsize;
1664 	struct thread *td;
1665 	struct vmspace *vm;
1666 	bool breakout;
1667 
1668 	/*
1669 	 * We keep the process bigproc locked once we find it to keep anyone
1670 	 * from messing with it; however, there is a possibility of
1671 	 * deadlock if process B is bigproc and one of its child processes
1672 	 * attempts to propagate a signal to B while we are waiting for A's
1673 	 * lock while walking this list.  To avoid this, we don't block on
1674 	 * the process lock but just skip a process if it is already locked.
1675 	 */
1676 	bigproc = NULL;
1677 	bigsize = 0;
1678 	sx_slock(&allproc_lock);
1679 	FOREACH_PROC_IN_SYSTEM(p) {
1680 		PROC_LOCK(p);
1681 
1682 		/*
1683 		 * If this is a system, protected or killed process, skip it.
1684 		 */
1685 		if (p->p_state != PRS_NORMAL || (p->p_flag & (P_INEXEC |
1686 		    P_PROTECTED | P_SYSTEM | P_WEXIT)) != 0 ||
1687 		    p->p_pid == 1 || P_KILLED(p) ||
1688 		    (p->p_pid < 48 && swap_pager_avail != 0)) {
1689 			PROC_UNLOCK(p);
1690 			continue;
1691 		}
1692 		/*
1693 		 * If the process is in a non-running type state,
1694 		 * don't touch it.  Check all the threads individually.
1695 		 */
1696 		breakout = false;
1697 		FOREACH_THREAD_IN_PROC(p, td) {
1698 			thread_lock(td);
1699 			if (!TD_ON_RUNQ(td) &&
1700 			    !TD_IS_RUNNING(td) &&
1701 			    !TD_IS_SLEEPING(td) &&
1702 			    !TD_IS_SUSPENDED(td) &&
1703 			    !TD_IS_SWAPPED(td)) {
1704 				thread_unlock(td);
1705 				breakout = true;
1706 				break;
1707 			}
1708 			thread_unlock(td);
1709 		}
1710 		if (breakout) {
1711 			PROC_UNLOCK(p);
1712 			continue;
1713 		}
1714 		/*
1715 		 * get the process size
1716 		 */
1717 		vm = vmspace_acquire_ref(p);
1718 		if (vm == NULL) {
1719 			PROC_UNLOCK(p);
1720 			continue;
1721 		}
1722 		_PHOLD_LITE(p);
1723 		PROC_UNLOCK(p);
1724 		sx_sunlock(&allproc_lock);
1725 		if (!vm_map_trylock_read(&vm->vm_map)) {
1726 			vmspace_free(vm);
1727 			sx_slock(&allproc_lock);
1728 			PRELE(p);
1729 			continue;
1730 		}
1731 		size = vmspace_swap_count(vm);
1732 		if (shortage == VM_OOM_MEM)
1733 			size += vm_pageout_oom_pagecount(vm);
1734 		vm_map_unlock_read(&vm->vm_map);
1735 		vmspace_free(vm);
1736 		sx_slock(&allproc_lock);
1737 
1738 		/*
1739 		 * If this process is bigger than the biggest one,
1740 		 * remember it.
1741 		 */
1742 		if (size > bigsize) {
1743 			if (bigproc != NULL)
1744 				PRELE(bigproc);
1745 			bigproc = p;
1746 			bigsize = size;
1747 		} else {
1748 			PRELE(p);
1749 		}
1750 	}
1751 	sx_sunlock(&allproc_lock);
1752 	if (bigproc != NULL) {
1753 		if (vm_panic_on_oom != 0)
1754 			panic("out of swap space");
1755 		PROC_LOCK(bigproc);
1756 		killproc(bigproc, "out of swap space");
1757 		sched_nice(bigproc, PRIO_MIN);
1758 		_PRELE(bigproc);
1759 		PROC_UNLOCK(bigproc);
1760 	}
1761 }
1762 
1763 static void
1764 vm_pageout_worker(void *arg)
1765 {
1766 	struct vm_domain *vmd;
1767 	int domain, pass, shortage;
1768 	bool target_met;
1769 
1770 	domain = (uintptr_t)arg;
1771 	vmd = VM_DOMAIN(domain);
1772 	pass = 0;
1773 	shortage = 0;
1774 	target_met = true;
1775 
1776 	/*
1777 	 * XXXKIB It could be useful to bind pageout daemon threads to
1778 	 * the cores belonging to the domain, from which vm_page_array
1779 	 * is allocated.
1780 	 */
1781 
1782 	KASSERT(vmd->vmd_segs != 0, ("domain without segments"));
1783 	vmd->vmd_last_active_scan = ticks;
1784 	vm_pageout_init_marker(&vmd->vmd_marker, PQ_INACTIVE);
1785 	vm_pageout_init_marker(&vmd->vmd_inacthead, PQ_INACTIVE);
1786 	TAILQ_INSERT_HEAD(&vmd->vmd_pagequeues[PQ_INACTIVE].pq_pl,
1787 	    &vmd->vmd_inacthead, plinks.q);
1788 
1789 	/*
1790 	 * The pageout daemon worker is never done, so loop forever.
1791 	 */
1792 	while (TRUE) {
1793 		vm_domain_pageout_lock(vmd);
1794 		/*
1795 		 * We need to clear wanted before we check the limits.  This
1796 		 * prevents races with wakers who will check wanted after they
1797 		 * reach the limit.
1798 		 */
1799 		atomic_store_int(&vmd->vmd_pageout_wanted, 0);
1800 
1801 		/*
1802 		 * Might the page daemon need to run again?
1803 		 */
1804 		if (vm_paging_needed(vmd, vmd->vmd_free_count)) {
1805 			/*
1806 			 * Yes, the scan failed to free enough pages.  If
1807 			 * we have performed a level >= 1 (page reclamation)
1808 			 * scan, then sleep a bit and try again.
1809 			 */
1810 			vm_domain_pageout_unlock(vmd);
1811 			if (pass > 1)
1812 				pause("pwait", hz / VM_INACT_SCAN_RATE);
1813 		} else {
1814 			/*
1815 			 * No, sleep until the next wakeup or until pages
1816 			 * need to have their reference stats updated.
1817 			 */
1818 			if (mtx_sleep(&vmd->vmd_pageout_wanted,
1819 			    vm_domain_pageout_lockptr(vmd), PDROP | PVM,
1820 			    "psleep", hz / VM_INACT_SCAN_RATE) == 0)
1821 				VM_CNT_INC(v_pdwakeups);
1822 		}
1823 		/* Prevent spurious wakeups by ensuring that wanted is set. */
1824 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1825 
1826 		/*
1827 		 * Use the controller to calculate how many pages to free in
1828 		 * this interval.
1829 		 */
1830 		shortage = pidctrl_daemon(&vmd->vmd_pid, vmd->vmd_free_count);
1831 		if (shortage && pass == 0)
1832 			pass = 1;
1833 
1834 		target_met = vm_pageout_scan(vmd, pass, shortage);
1835 		/*
1836 		 * If the target was not met we must increase the pass to
1837 		 * more aggressively reclaim.
1838 		 */
1839 		if (!target_met)
1840 			pass++;
1841 	}
1842 }
1843 
1844 /*
1845  *	vm_pageout_init initialises basic pageout daemon settings.
1846  */
1847 static void
1848 vm_pageout_init_domain(int domain)
1849 {
1850 	struct vm_domain *vmd;
1851 	struct sysctl_oid *oid;
1852 
1853 	vmd = VM_DOMAIN(domain);
1854 	vmd->vmd_interrupt_free_min = 2;
1855 
1856 	/*
1857 	 * v_free_reserved needs to include enough for the largest
1858 	 * swap pager structures plus enough for any pv_entry structs
1859 	 * when paging.
1860 	 */
1861 	if (vmd->vmd_page_count > 1024)
1862 		vmd->vmd_free_min = 4 + (vmd->vmd_page_count - 1024) / 200;
1863 	else
1864 		vmd->vmd_free_min = 4;
1865 	vmd->vmd_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1866 	    vmd->vmd_interrupt_free_min;
1867 	vmd->vmd_free_reserved = vm_pageout_page_count +
1868 	    vmd->vmd_pageout_free_min + (vmd->vmd_page_count / 768);
1869 	vmd->vmd_free_severe = vmd->vmd_free_min / 2;
1870 	vmd->vmd_free_target = 4 * vmd->vmd_free_min + vmd->vmd_free_reserved;
1871 	vmd->vmd_free_min += vmd->vmd_free_reserved;
1872 	vmd->vmd_free_severe += vmd->vmd_free_reserved;
1873 	vmd->vmd_inactive_target = (3 * vmd->vmd_free_target) / 2;
1874 	if (vmd->vmd_inactive_target > vmd->vmd_free_count / 3)
1875 		vmd->vmd_inactive_target = vmd->vmd_free_count / 3;
1876 
1877 	/*
1878 	 * Set the default wakeup threshold to be 10% below the paging
1879 	 * target.  This keeps the steady state out of shortfall.
1880 	 */
1881 	vmd->vmd_pageout_wakeup_thresh = (vmd->vmd_free_target / 10) * 9;
1882 
1883 	/*
1884 	 * Target amount of memory to move out of the laundry queue during a
1885 	 * background laundering.  This is proportional to the amount of system
1886 	 * memory.
1887 	 */
1888 	vmd->vmd_background_launder_target = (vmd->vmd_free_target -
1889 	    vmd->vmd_free_min) / 10;
1890 
1891 	/* Initialize the pageout daemon pid controller. */
1892 	pidctrl_init(&vmd->vmd_pid, hz / VM_INACT_SCAN_RATE,
1893 	    vmd->vmd_free_target, PIDCTRL_BOUND,
1894 	    PIDCTRL_KPD, PIDCTRL_KID, PIDCTRL_KDD);
1895 	oid = SYSCTL_ADD_NODE(NULL, SYSCTL_CHILDREN(vmd->vmd_oid), OID_AUTO,
1896 	    "pidctrl", CTLFLAG_RD, NULL, "");
1897 	pidctrl_init_sysctl(&vmd->vmd_pid, SYSCTL_CHILDREN(oid));
1898 }
1899 
1900 static void
1901 vm_pageout_init(void)
1902 {
1903 	u_int freecount;
1904 	int i;
1905 
1906 	/*
1907 	 * Initialize some paging parameters.
1908 	 */
1909 	if (vm_cnt.v_page_count < 2000)
1910 		vm_pageout_page_count = 8;
1911 
1912 	freecount = 0;
1913 	for (i = 0; i < vm_ndomains; i++) {
1914 		struct vm_domain *vmd;
1915 
1916 		vm_pageout_init_domain(i);
1917 		vmd = VM_DOMAIN(i);
1918 		vm_cnt.v_free_reserved += vmd->vmd_free_reserved;
1919 		vm_cnt.v_free_target += vmd->vmd_free_target;
1920 		vm_cnt.v_free_min += vmd->vmd_free_min;
1921 		vm_cnt.v_inactive_target += vmd->vmd_inactive_target;
1922 		vm_cnt.v_pageout_free_min += vmd->vmd_pageout_free_min;
1923 		vm_cnt.v_interrupt_free_min += vmd->vmd_interrupt_free_min;
1924 		vm_cnt.v_free_severe += vmd->vmd_free_severe;
1925 		freecount += vmd->vmd_free_count;
1926 	}
1927 
1928 	/*
1929 	 * Set interval in seconds for active scan.  We want to visit each
1930 	 * page at least once every ten minutes.  This is to prevent worst
1931 	 * case paging behaviors with stale active LRU.
1932 	 */
1933 	if (vm_pageout_update_period == 0)
1934 		vm_pageout_update_period = 600;
1935 
1936 	if (vm_page_max_wired == 0)
1937 		vm_page_max_wired = freecount / 3;
1938 }
1939 
1940 /*
1941  *     vm_pageout is the high level pageout daemon.
1942  */
1943 static void
1944 vm_pageout(void)
1945 {
1946 	int error;
1947 	int i;
1948 
1949 	swap_pager_swap_init();
1950 	snprintf(curthread->td_name, sizeof(curthread->td_name), "dom0");
1951 	error = kthread_add(vm_pageout_laundry_worker, NULL, curproc, NULL,
1952 	    0, 0, "laundry: dom0");
1953 	if (error != 0)
1954 		panic("starting laundry for domain 0, error %d", error);
1955 	for (i = 1; i < vm_ndomains; i++) {
1956 		error = kthread_add(vm_pageout_worker, (void *)(uintptr_t)i,
1957 		    curproc, NULL, 0, 0, "dom%d", i);
1958 		if (error != 0) {
1959 			panic("starting pageout for domain %d, error %d\n",
1960 			    i, error);
1961 		}
1962 		error = kthread_add(vm_pageout_laundry_worker,
1963 		    (void *)(uintptr_t)i, curproc, NULL, 0, 0,
1964 		    "laundry: dom%d", i);
1965 		if (error != 0)
1966 			panic("starting laundry for domain %d, error %d",
1967 			    i, error);
1968 	}
1969 	error = kthread_add(uma_reclaim_worker, NULL, curproc, NULL,
1970 	    0, 0, "uma");
1971 	if (error != 0)
1972 		panic("starting uma_reclaim helper, error %d\n", error);
1973 	vm_pageout_worker((void *)(uintptr_t)0);
1974 }
1975 
1976 /*
1977  * Perform an advisory wakeup of the page daemon.
1978  */
1979 void
1980 pagedaemon_wakeup(int domain)
1981 {
1982 	struct vm_domain *vmd;
1983 
1984 	vmd = VM_DOMAIN(domain);
1985 	vm_domain_pageout_assert_unlocked(vmd);
1986 	if (curproc == pageproc)
1987 		return;
1988 
1989 	if (atomic_fetchadd_int(&vmd->vmd_pageout_wanted, 1) == 0) {
1990 		vm_domain_pageout_lock(vmd);
1991 		atomic_store_int(&vmd->vmd_pageout_wanted, 1);
1992 		wakeup(&vmd->vmd_pageout_wanted);
1993 		vm_domain_pageout_unlock(vmd);
1994 	}
1995 }
1996