xref: /dflybsd-src/sys/vm/vm_pageout.c (revision e90a7c45c3303ed54c0fde732b2ba32dc80ffd9b)
1 /*
2  * (MPSAFE)
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  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 3. All advertising materials mentioning features or use of this software
23  *    must display the following acknowledgement:
24  *	This product includes software developed by the University of
25  *	California, Berkeley and its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49  *
50  * Permission to use, copy, modify and distribute this software and
51  * its documentation is hereby granted, provided that both the copyright
52  * notice and this permission notice appear in all copies of the
53  * software, derivative works or modified versions, and any portions
54  * thereof, and that both notices appear in supporting documentation.
55  *
56  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
57  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
58  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59  *
60  * Carnegie Mellon requests users of this software to return to
61  *
62  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
63  *  School of Computer Science
64  *  Carnegie Mellon University
65  *  Pittsburgh PA 15213-3890
66  *
67  * any improvements or extensions that they make and grant Carnegie the
68  * rights to redistribute these changes.
69  *
70  * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
71  * $DragonFly: src/sys/vm/vm_pageout.c,v 1.36 2008/07/01 02:02:56 dillon Exp $
72  */
73 
74 /*
75  *	The proverbial page-out daemon.
76  */
77 
78 #include "opt_vm.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/proc.h>
83 #include <sys/kthread.h>
84 #include <sys/resourcevar.h>
85 #include <sys/signalvar.h>
86 #include <sys/vnode.h>
87 #include <sys/vmmeter.h>
88 #include <sys/sysctl.h>
89 
90 #include <vm/vm.h>
91 #include <vm/vm_param.h>
92 #include <sys/lock.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/swap_pager.h>
99 #include <vm/vm_extern.h>
100 
101 #include <sys/thread2.h>
102 #include <vm/vm_page2.h>
103 
104 /*
105  * System initialization
106  */
107 
108 /* the kernel process "vm_pageout"*/
109 static int vm_pageout_clean (vm_page_t);
110 static int vm_pageout_scan (int pass);
111 static int vm_pageout_free_page_calc (vm_size_t count);
112 struct thread *pagethread;
113 
114 #if !defined(NO_SWAPPING)
115 /* the kernel process "vm_daemon"*/
116 static void vm_daemon (void);
117 static struct	thread *vmthread;
118 
119 static struct kproc_desc vm_kp = {
120 	"vmdaemon",
121 	vm_daemon,
122 	&vmthread
123 };
124 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
125 #endif
126 
127 
128 int vm_pages_needed=0;		/* Event on which pageout daemon sleeps */
129 int vm_pageout_deficit=0;	/* Estimated number of pages deficit */
130 int vm_pageout_pages_needed=0;	/* flag saying that the pageout daemon needs pages */
131 
132 #if !defined(NO_SWAPPING)
133 static int vm_pageout_req_swapout;	/* XXX */
134 static int vm_daemon_needed;
135 #endif
136 static int vm_max_launder = 32;
137 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
138 static int vm_pageout_full_stats_interval = 0;
139 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
140 static int defer_swap_pageouts=0;
141 static int disable_swap_pageouts=0;
142 
143 #if defined(NO_SWAPPING)
144 static int vm_swap_enabled=0;
145 static int vm_swap_idle_enabled=0;
146 #else
147 static int vm_swap_enabled=1;
148 static int vm_swap_idle_enabled=0;
149 #endif
150 
151 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
152 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
153 
154 SYSCTL_INT(_vm, OID_AUTO, max_launder,
155 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
156 
157 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
158 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
159 
160 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
161 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
162 
163 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
164 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
165 
166 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
167 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
168 
169 #if defined(NO_SWAPPING)
170 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
171 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
172 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
173 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
174 #else
175 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
176 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
177 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
178 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
179 #endif
180 
181 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
182 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
183 
184 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
185 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
186 
187 static int pageout_lock_miss;
188 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
189 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
190 
191 int vm_load;
192 SYSCTL_INT(_vm, OID_AUTO, vm_load,
193 	CTLFLAG_RD, &vm_load, 0, "load on the VM system");
194 int vm_load_enable = 1;
195 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
196 	CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
197 #ifdef INVARIANTS
198 int vm_load_debug;
199 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
200 	CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
201 #endif
202 
203 #define VM_PAGEOUT_PAGE_COUNT 16
204 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
205 
206 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
207 
208 #if !defined(NO_SWAPPING)
209 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
210 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
211 static freeer_fcn_t vm_pageout_object_deactivate_pages;
212 static void vm_req_vmdaemon (void);
213 #endif
214 static void vm_pageout_page_stats(void);
215 
216 /*
217  * Update vm_load to slow down faulting processes.
218  *
219  * SMP races ok.
220  * No requirements.
221  */
222 void
223 vm_fault_ratecheck(void)
224 {
225 	if (vm_pages_needed) {
226 		if (vm_load < 1000)
227 			++vm_load;
228 	} else {
229 		if (vm_load > 0)
230 			--vm_load;
231 	}
232 }
233 
234 /*
235  * vm_pageout_clean:
236  *
237  * Clean the page and remove it from the laundry.  The page must not be
238  * busy on-call.
239  *
240  * We set the busy bit to cause potential page faults on this page to
241  * block.  Note the careful timing, however, the busy bit isn't set till
242  * late and we cannot do anything that will mess with the page.
243  *
244  * The caller must hold vm_token.
245  */
246 static int
247 vm_pageout_clean(vm_page_t m)
248 {
249 	vm_object_t object;
250 	vm_page_t mc[2*vm_pageout_page_count];
251 	int pageout_count;
252 	int ib, is, page_base;
253 	vm_pindex_t pindex = m->pindex;
254 
255 	object = m->object;
256 
257 	/*
258 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
259 	 * with the new swapper, but we could have serious problems paging
260 	 * out other object types if there is insufficient memory.
261 	 *
262 	 * Unfortunately, checking free memory here is far too late, so the
263 	 * check has been moved up a procedural level.
264 	 */
265 
266 	/*
267 	 * Don't mess with the page if it's busy, held, or special
268 	 */
269 	if ((m->hold_count != 0) ||
270 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
271 		return 0;
272 	}
273 
274 	mc[vm_pageout_page_count] = m;
275 	pageout_count = 1;
276 	page_base = vm_pageout_page_count;
277 	ib = 1;
278 	is = 1;
279 
280 	/*
281 	 * Scan object for clusterable pages.
282 	 *
283 	 * We can cluster ONLY if: ->> the page is NOT
284 	 * clean, wired, busy, held, or mapped into a
285 	 * buffer, and one of the following:
286 	 * 1) The page is inactive, or a seldom used
287 	 *    active page.
288 	 * -or-
289 	 * 2) we force the issue.
290 	 *
291 	 * During heavy mmap/modification loads the pageout
292 	 * daemon can really fragment the underlying file
293 	 * due to flushing pages out of order and not trying
294 	 * align the clusters (which leave sporatic out-of-order
295 	 * holes).  To solve this problem we do the reverse scan
296 	 * first and attempt to align our cluster, then do a
297 	 * forward scan if room remains.
298 	 */
299 
300 more:
301 	while (ib && pageout_count < vm_pageout_page_count) {
302 		vm_page_t p;
303 
304 		if (ib > pindex) {
305 			ib = 0;
306 			break;
307 		}
308 
309 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
310 			ib = 0;
311 			break;
312 		}
313 		if (((p->queue - p->pc) == PQ_CACHE) ||
314 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
315 			ib = 0;
316 			break;
317 		}
318 		vm_page_test_dirty(p);
319 		if ((p->dirty & p->valid) == 0 ||
320 		    p->queue != PQ_INACTIVE ||
321 		    p->wire_count != 0 ||	/* may be held by buf cache */
322 		    p->hold_count != 0) {	/* may be undergoing I/O */
323 			ib = 0;
324 			break;
325 		}
326 		mc[--page_base] = p;
327 		++pageout_count;
328 		++ib;
329 		/*
330 		 * alignment boundry, stop here and switch directions.  Do
331 		 * not clear ib.
332 		 */
333 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
334 			break;
335 	}
336 
337 	while (pageout_count < vm_pageout_page_count &&
338 	    pindex + is < object->size) {
339 		vm_page_t p;
340 
341 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
342 			break;
343 		if (((p->queue - p->pc) == PQ_CACHE) ||
344 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
345 			break;
346 		}
347 		vm_page_test_dirty(p);
348 		if ((p->dirty & p->valid) == 0 ||
349 		    p->queue != PQ_INACTIVE ||
350 		    p->wire_count != 0 ||	/* may be held by buf cache */
351 		    p->hold_count != 0) {	/* may be undergoing I/O */
352 			break;
353 		}
354 		mc[page_base + pageout_count] = p;
355 		++pageout_count;
356 		++is;
357 	}
358 
359 	/*
360 	 * If we exhausted our forward scan, continue with the reverse scan
361 	 * when possible, even past a page boundry.  This catches boundry
362 	 * conditions.
363 	 */
364 	if (ib && pageout_count < vm_pageout_page_count)
365 		goto more;
366 
367 	/*
368 	 * we allow reads during pageouts...
369 	 */
370 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
371 }
372 
373 /*
374  * vm_pageout_flush() - launder the given pages
375  *
376  *	The given pages are laundered.  Note that we setup for the start of
377  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
378  *	reference count all in here rather then in the parent.  If we want
379  *	the parent to do more sophisticated things we may have to change
380  *	the ordering.
381  *
382  * The caller must hold vm_token.
383  */
384 int
385 vm_pageout_flush(vm_page_t *mc, int count, int flags)
386 {
387 	vm_object_t object;
388 	int pageout_status[count];
389 	int numpagedout = 0;
390 	int i;
391 
392 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
393 
394 	/*
395 	 * Initiate I/O.  Bump the vm_page_t->busy counter.
396 	 */
397 	for (i = 0; i < count; i++) {
398 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
399 		vm_page_io_start(mc[i]);
400 	}
401 
402 	/*
403 	 * We must make the pages read-only.  This will also force the
404 	 * modified bit in the related pmaps to be cleared.  The pager
405 	 * cannot clear the bit for us since the I/O completion code
406 	 * typically runs from an interrupt.  The act of making the page
407 	 * read-only handles the case for us.
408 	 */
409 	for (i = 0; i < count; i++) {
410 		vm_page_protect(mc[i], VM_PROT_READ);
411 	}
412 
413 	object = mc[0]->object;
414 	vm_object_pip_add(object, count);
415 
416 	vm_pager_put_pages(object, mc, count,
417 	    (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
418 	    pageout_status);
419 
420 	for (i = 0; i < count; i++) {
421 		vm_page_t mt = mc[i];
422 
423 		switch (pageout_status[i]) {
424 		case VM_PAGER_OK:
425 			numpagedout++;
426 			break;
427 		case VM_PAGER_PEND:
428 			numpagedout++;
429 			break;
430 		case VM_PAGER_BAD:
431 			/*
432 			 * Page outside of range of object. Right now we
433 			 * essentially lose the changes by pretending it
434 			 * worked.
435 			 */
436 			pmap_clear_modify(mt);
437 			vm_page_undirty(mt);
438 			break;
439 		case VM_PAGER_ERROR:
440 		case VM_PAGER_FAIL:
441 			/*
442 			 * A page typically cannot be paged out when we
443 			 * have run out of swap.  We leave the page
444 			 * marked inactive and will try to page it out
445 			 * again later.
446 			 *
447 			 * Starvation of the active page list is used to
448 			 * determine when the system is massively memory
449 			 * starved.
450 			 */
451 			break;
452 		case VM_PAGER_AGAIN:
453 			break;
454 		}
455 
456 		/*
457 		 * If the operation is still going, leave the page busy to
458 		 * block all other accesses. Also, leave the paging in
459 		 * progress indicator set so that we don't attempt an object
460 		 * collapse.
461 		 *
462 		 * For any pages which have completed synchronously,
463 		 * deactivate the page if we are under a severe deficit.
464 		 * Do not try to enter them into the cache, though, they
465 		 * might still be read-heavy.
466 		 */
467 		if (pageout_status[i] != VM_PAGER_PEND) {
468 			if (vm_page_count_severe())
469 				vm_page_deactivate(mt);
470 #if 0
471 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
472 				vm_page_protect(mt, VM_PROT_READ);
473 #endif
474 			vm_page_io_finish(mt);
475 			vm_object_pip_wakeup(object);
476 		}
477 	}
478 	return numpagedout;
479 }
480 
481 #if !defined(NO_SWAPPING)
482 /*
483  *	vm_pageout_object_deactivate_pages
484  *
485  *	deactivate enough pages to satisfy the inactive target
486  *	requirements or if vm_page_proc_limit is set, then
487  *	deactivate all of the pages in the object and its
488  *	backing_objects.
489  *
490  * The map must be locked.
491  * The caller must hold vm_token.
492  */
493 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
494 
495 static void
496 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
497 				   vm_pindex_t desired, int map_remove_only)
498 {
499 	struct rb_vm_page_scan_info info;
500 	int remove_mode;
501 
502 	if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
503 		return;
504 
505 	while (object) {
506 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
507 			return;
508 		if (object->paging_in_progress)
509 			return;
510 
511 		remove_mode = map_remove_only;
512 		if (object->shadow_count > 1)
513 			remove_mode = 1;
514 
515 		/*
516 		 * scan the objects entire memory queue.  spl protection is
517 		 * required to avoid an interrupt unbusy/free race against
518 		 * our busy check.
519 		 */
520 		crit_enter();
521 		info.limit = remove_mode;
522 		info.map = map;
523 		info.desired = desired;
524 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
525 				vm_pageout_object_deactivate_pages_callback,
526 				&info
527 		);
528 		crit_exit();
529 		object = object->backing_object;
530 	}
531 }
532 
533 /*
534  * The caller must hold vm_token.
535  */
536 static int
537 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
538 {
539 	struct rb_vm_page_scan_info *info = data;
540 	int actcount;
541 
542 	if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
543 		return(-1);
544 	}
545 	mycpu->gd_cnt.v_pdpages++;
546 	if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 ||
547 	    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
548 	    !pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
549 		return(0);
550 	}
551 
552 	actcount = pmap_ts_referenced(p);
553 	if (actcount) {
554 		vm_page_flag_set(p, PG_REFERENCED);
555 	} else if (p->flags & PG_REFERENCED) {
556 		actcount = 1;
557 	}
558 
559 	if ((p->queue != PQ_ACTIVE) &&
560 		(p->flags & PG_REFERENCED)) {
561 		vm_page_activate(p);
562 		p->act_count += actcount;
563 		vm_page_flag_clear(p, PG_REFERENCED);
564 	} else if (p->queue == PQ_ACTIVE) {
565 		if ((p->flags & PG_REFERENCED) == 0) {
566 			p->act_count -= min(p->act_count, ACT_DECLINE);
567 			if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) {
568 				vm_page_busy(p);
569 				vm_page_protect(p, VM_PROT_NONE);
570 				vm_page_deactivate(p);
571 				vm_page_wakeup(p);
572 			} else {
573 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
574 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
575 			}
576 		} else {
577 			vm_page_activate(p);
578 			vm_page_flag_clear(p, PG_REFERENCED);
579 			if (p->act_count < (ACT_MAX - ACT_ADVANCE))
580 				p->act_count += ACT_ADVANCE;
581 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
582 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
583 		}
584 	} else if (p->queue == PQ_INACTIVE) {
585 		vm_page_busy(p);
586 		vm_page_protect(p, VM_PROT_NONE);
587 		vm_page_wakeup(p);
588 	}
589 	return(0);
590 }
591 
592 /*
593  * Deactivate some number of pages in a map, try to do it fairly, but
594  * that is really hard to do.
595  *
596  * The caller must hold vm_token.
597  */
598 static void
599 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
600 {
601 	vm_map_entry_t tmpe;
602 	vm_object_t obj, bigobj;
603 	int nothingwired;
604 
605 	if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
606 		return;
607 	}
608 
609 	bigobj = NULL;
610 	nothingwired = TRUE;
611 
612 	/*
613 	 * first, search out the biggest object, and try to free pages from
614 	 * that.
615 	 */
616 	tmpe = map->header.next;
617 	while (tmpe != &map->header) {
618 		switch(tmpe->maptype) {
619 		case VM_MAPTYPE_NORMAL:
620 		case VM_MAPTYPE_VPAGETABLE:
621 			obj = tmpe->object.vm_object;
622 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
623 				((bigobj == NULL) ||
624 				 (bigobj->resident_page_count < obj->resident_page_count))) {
625 				bigobj = obj;
626 			}
627 			break;
628 		default:
629 			break;
630 		}
631 		if (tmpe->wired_count > 0)
632 			nothingwired = FALSE;
633 		tmpe = tmpe->next;
634 	}
635 
636 	if (bigobj)
637 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
638 
639 	/*
640 	 * Next, hunt around for other pages to deactivate.  We actually
641 	 * do this search sort of wrong -- .text first is not the best idea.
642 	 */
643 	tmpe = map->header.next;
644 	while (tmpe != &map->header) {
645 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
646 			break;
647 		switch(tmpe->maptype) {
648 		case VM_MAPTYPE_NORMAL:
649 		case VM_MAPTYPE_VPAGETABLE:
650 			obj = tmpe->object.vm_object;
651 			if (obj)
652 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
653 			break;
654 		default:
655 			break;
656 		}
657 		tmpe = tmpe->next;
658 	};
659 
660 	/*
661 	 * Remove all mappings if a process is swapped out, this will free page
662 	 * table pages.
663 	 */
664 	if (desired == 0 && nothingwired)
665 		pmap_remove(vm_map_pmap(map),
666 			    VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
667 	vm_map_unlock(map);
668 }
669 #endif
670 
671 /*
672  * Don't try to be fancy - being fancy can lead to vnode deadlocks.   We
673  * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
674  * be trivially freed.
675  *
676  * The caller must hold vm_token.
677  *
678  * WARNING: vm_object_reference() can block.
679  */
680 static void
681 vm_pageout_page_free(vm_page_t m)
682 {
683 	vm_object_t object = m->object;
684 	int type = object->type;
685 
686 	vm_page_busy(m);
687 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
688 		vm_object_reference(object);
689 	vm_page_protect(m, VM_PROT_NONE);
690 	vm_page_free(m);
691 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
692 		vm_object_deallocate(object);
693 }
694 
695 /*
696  * vm_pageout_scan does the dirty work for the pageout daemon.
697  */
698 struct vm_pageout_scan_info {
699 	struct proc *bigproc;
700 	vm_offset_t bigsize;
701 };
702 
703 static int vm_pageout_scan_callback(struct proc *p, void *data);
704 
705 /*
706  * The caller must hold vm_token.
707  */
708 static int
709 vm_pageout_scan(int pass)
710 {
711 	struct vm_pageout_scan_info info;
712 	vm_page_t m, next;
713 	struct vm_page marker;
714 	struct vnode *vpfailed;		/* warning, allowed to be stale */
715 	int maxscan, pcount;
716 	int recycle_count;
717 	int inactive_shortage, active_shortage;
718 	int inactive_original_shortage;
719 	vm_object_t object;
720 	int actcount;
721 	int vnodes_skipped = 0;
722 	int maxlaunder;
723 
724 	/*
725 	 * Do whatever cleanup that the pmap code can.
726 	 */
727 	pmap_collect();
728 
729 	/*
730 	 * Calculate our target for the number of free+cache pages we
731 	 * want to get to.  This is higher then the number that causes
732 	 * allocations to stall (severe) in order to provide hysteresis,
733 	 * and if we don't make it all the way but get to the minimum
734 	 * we're happy.
735 	 */
736 	inactive_shortage = vm_paging_target() + vm_pageout_deficit;
737 	inactive_original_shortage = inactive_shortage;
738 	vm_pageout_deficit = 0;
739 
740 	/*
741 	 * Initialize our marker
742 	 */
743 	bzero(&marker, sizeof(marker));
744 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
745 	marker.queue = PQ_INACTIVE;
746 	marker.wire_count = 1;
747 
748 	/*
749 	 * Start scanning the inactive queue for pages we can move to the
750 	 * cache or free.  The scan will stop when the target is reached or
751 	 * we have scanned the entire inactive queue.  Note that m->act_count
752 	 * is not used to form decisions for the inactive queue, only for the
753 	 * active queue.
754 	 *
755 	 * maxlaunder limits the number of dirty pages we flush per scan.
756 	 * For most systems a smaller value (16 or 32) is more robust under
757 	 * extreme memory and disk pressure because any unnecessary writes
758 	 * to disk can result in extreme performance degredation.  However,
759 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
760 	 * used) will die horribly with limited laundering.  If the pageout
761 	 * daemon cannot clean enough pages in the first pass, we let it go
762 	 * all out in succeeding passes.
763 	 */
764 	if ((maxlaunder = vm_max_launder) <= 1)
765 		maxlaunder = 1;
766 	if (pass)
767 		maxlaunder = 10000;
768 
769 	/*
770 	 * We will generally be in a critical section throughout the
771 	 * scan, but we can release it temporarily when we are sitting on a
772 	 * non-busy page without fear.  this is required to prevent an
773 	 * interrupt from unbusying or freeing a page prior to our busy
774 	 * check, leaving us on the wrong queue or checking the wrong
775 	 * page.
776 	 */
777 	crit_enter();
778 rescan0:
779 	vpfailed = NULL;
780 	maxscan = vmstats.v_inactive_count;
781 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
782 	     m != NULL && maxscan-- > 0 && inactive_shortage > 0;
783 	     m = next
784 	 ) {
785 		mycpu->gd_cnt.v_pdpages++;
786 
787 		/*
788 		 * Give interrupts a chance
789 		 */
790 		crit_exit();
791 		crit_enter();
792 
793 		/*
794 		 * It's easier for some of the conditions below to just loop
795 		 * and catch queue changes here rather then check everywhere
796 		 * else.
797 		 */
798 		if (m->queue != PQ_INACTIVE)
799 			goto rescan0;
800 		next = TAILQ_NEXT(m, pageq);
801 
802 		/*
803 		 * skip marker pages
804 		 */
805 		if (m->flags & PG_MARKER)
806 			continue;
807 
808 		/*
809 		 * A held page may be undergoing I/O, so skip it.
810 		 */
811 		if (m->hold_count) {
812 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
813 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
814 			++vm_swapcache_inactive_heuristic;
815 			continue;
816 		}
817 
818 		/*
819 		 * Dont mess with busy pages, keep in the front of the
820 		 * queue, most likely are being paged out.
821 		 */
822 		if (m->busy || (m->flags & PG_BUSY)) {
823 			continue;
824 		}
825 
826 		if (m->object->ref_count == 0) {
827 			/*
828 			 * If the object is not being used, we ignore previous
829 			 * references.
830 			 */
831 			vm_page_flag_clear(m, PG_REFERENCED);
832 			pmap_clear_reference(m);
833 
834 		} else if (((m->flags & PG_REFERENCED) == 0) &&
835 			    (actcount = pmap_ts_referenced(m))) {
836 			/*
837 			 * Otherwise, if the page has been referenced while
838 			 * in the inactive queue, we bump the "activation
839 			 * count" upwards, making it less likely that the
840 			 * page will be added back to the inactive queue
841 			 * prematurely again.  Here we check the page tables
842 			 * (or emulated bits, if any), given the upper level
843 			 * VM system not knowing anything about existing
844 			 * references.
845 			 */
846 			vm_page_activate(m);
847 			m->act_count += (actcount + ACT_ADVANCE);
848 			continue;
849 		}
850 
851 		/*
852 		 * If the upper level VM system knows about any page
853 		 * references, we activate the page.  We also set the
854 		 * "activation count" higher than normal so that we will less
855 		 * likely place pages back onto the inactive queue again.
856 		 */
857 		if ((m->flags & PG_REFERENCED) != 0) {
858 			vm_page_flag_clear(m, PG_REFERENCED);
859 			actcount = pmap_ts_referenced(m);
860 			vm_page_activate(m);
861 			m->act_count += (actcount + ACT_ADVANCE + 1);
862 			continue;
863 		}
864 
865 		/*
866 		 * If the upper level VM system doesn't know anything about
867 		 * the page being dirty, we have to check for it again.  As
868 		 * far as the VM code knows, any partially dirty pages are
869 		 * fully dirty.
870 		 *
871 		 * Pages marked PG_WRITEABLE may be mapped into the user
872 		 * address space of a process running on another cpu.  A
873 		 * user process (without holding the MP lock) running on
874 		 * another cpu may be able to touch the page while we are
875 		 * trying to remove it.  vm_page_cache() will handle this
876 		 * case for us.
877 		 */
878 		if (m->dirty == 0) {
879 			vm_page_test_dirty(m);
880 		} else {
881 			vm_page_dirty(m);
882 		}
883 
884 		if (m->valid == 0) {
885 			/*
886 			 * Invalid pages can be easily freed
887 			 */
888 			vm_pageout_page_free(m);
889 			mycpu->gd_cnt.v_dfree++;
890 			--inactive_shortage;
891 		} else if (m->dirty == 0) {
892 			/*
893 			 * Clean pages can be placed onto the cache queue.
894 			 * This effectively frees them.
895 			 */
896 			vm_page_busy(m);
897 			vm_page_cache(m);
898 			--inactive_shortage;
899 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
900 			/*
901 			 * Dirty pages need to be paged out, but flushing
902 			 * a page is extremely expensive verses freeing
903 			 * a clean page.  Rather then artificially limiting
904 			 * the number of pages we can flush, we instead give
905 			 * dirty pages extra priority on the inactive queue
906 			 * by forcing them to be cycled through the queue
907 			 * twice before being flushed, after which the
908 			 * (now clean) page will cycle through once more
909 			 * before being freed.  This significantly extends
910 			 * the thrash point for a heavily loaded machine.
911 			 */
912 			vm_page_flag_set(m, PG_WINATCFLS);
913 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
914 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
915 			++vm_swapcache_inactive_heuristic;
916 		} else if (maxlaunder > 0) {
917 			/*
918 			 * We always want to try to flush some dirty pages if
919 			 * we encounter them, to keep the system stable.
920 			 * Normally this number is small, but under extreme
921 			 * pressure where there are insufficient clean pages
922 			 * on the inactive queue, we may have to go all out.
923 			 */
924 			int swap_pageouts_ok;
925 			struct vnode *vp = NULL;
926 
927 			object = m->object;
928 
929 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
930 				swap_pageouts_ok = 1;
931 			} else {
932 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
933 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
934 				vm_page_count_min(0));
935 
936 			}
937 
938 			/*
939 			 * We don't bother paging objects that are "dead".
940 			 * Those objects are in a "rundown" state.
941 			 */
942 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
943 				TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
944 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
945 				++vm_swapcache_inactive_heuristic;
946 				continue;
947 			}
948 
949 			/*
950 			 * The object is already known NOT to be dead.   It
951 			 * is possible for the vget() to block the whole
952 			 * pageout daemon, but the new low-memory handling
953 			 * code should prevent it.
954 			 *
955 			 * The previous code skipped locked vnodes and, worse,
956 			 * reordered pages in the queue.  This results in
957 			 * completely non-deterministic operation because,
958 			 * quite often, a vm_fault has initiated an I/O and
959 			 * is holding a locked vnode at just the point where
960 			 * the pageout daemon is woken up.
961 			 *
962 			 * We can't wait forever for the vnode lock, we might
963 			 * deadlock due to a vn_read() getting stuck in
964 			 * vm_wait while holding this vnode.  We skip the
965 			 * vnode if we can't get it in a reasonable amount
966 			 * of time.
967 			 *
968 			 * vpfailed is used to (try to) avoid the case where
969 			 * a large number of pages are associated with a
970 			 * locked vnode, which could cause the pageout daemon
971 			 * to stall for an excessive amount of time.
972 			 */
973 			if (object->type == OBJT_VNODE) {
974 				int flags;
975 
976 				vp = object->handle;
977 				flags = LK_EXCLUSIVE | LK_NOOBJ;
978 				if (vp == vpfailed)
979 					flags |= LK_NOWAIT;
980 				else
981 					flags |= LK_TIMELOCK;
982 				if (vget(vp, flags) != 0) {
983 					vpfailed = vp;
984 					++pageout_lock_miss;
985 					if (object->flags & OBJ_MIGHTBEDIRTY)
986 						    vnodes_skipped++;
987 					continue;
988 				}
989 
990 				/*
991 				 * The page might have been moved to another
992 				 * queue during potential blocking in vget()
993 				 * above.  The page might have been freed and
994 				 * reused for another vnode.  The object might
995 				 * have been reused for another vnode.
996 				 */
997 				if (m->queue != PQ_INACTIVE ||
998 				    m->object != object ||
999 				    object->handle != vp) {
1000 					if (object->flags & OBJ_MIGHTBEDIRTY)
1001 						vnodes_skipped++;
1002 					vput(vp);
1003 					continue;
1004 				}
1005 
1006 				/*
1007 				 * The page may have been busied during the
1008 				 * blocking in vput();  We don't move the
1009 				 * page back onto the end of the queue so that
1010 				 * statistics are more correct if we don't.
1011 				 */
1012 				if (m->busy || (m->flags & PG_BUSY)) {
1013 					vput(vp);
1014 					continue;
1015 				}
1016 
1017 				/*
1018 				 * If the page has become held it might
1019 				 * be undergoing I/O, so skip it
1020 				 */
1021 				if (m->hold_count) {
1022 					TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1023 					TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1024 					++vm_swapcache_inactive_heuristic;
1025 					if (object->flags & OBJ_MIGHTBEDIRTY)
1026 						vnodes_skipped++;
1027 					vput(vp);
1028 					continue;
1029 				}
1030 			}
1031 
1032 			/*
1033 			 * If a page is dirty, then it is either being washed
1034 			 * (but not yet cleaned) or it is still in the
1035 			 * laundry.  If it is still in the laundry, then we
1036 			 * start the cleaning operation.
1037 			 *
1038 			 * This operation may cluster, invalidating the 'next'
1039 			 * pointer.  To prevent an inordinate number of
1040 			 * restarts we use our marker to remember our place.
1041 			 *
1042 			 * decrement inactive_shortage on success to account
1043 			 * for the (future) cleaned page.  Otherwise we
1044 			 * could wind up laundering or cleaning too many
1045 			 * pages.
1046 			 */
1047 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1048 			if (vm_pageout_clean(m) != 0) {
1049 				--inactive_shortage;
1050 				--maxlaunder;
1051 			}
1052 			next = TAILQ_NEXT(&marker, pageq);
1053 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1054 			if (vp != NULL)
1055 				vput(vp);
1056 		}
1057 	}
1058 
1059 	/*
1060 	 * We want to move pages from the active queue to the inactive
1061 	 * queue to get the inactive queue to the inactive target.  If
1062 	 * we still have a page shortage from above we try to directly free
1063 	 * clean pages instead of moving them.
1064 	 *
1065 	 * If we do still have a shortage we keep track of the number of
1066 	 * pages we free or cache (recycle_count) as a measure of thrashing
1067 	 * between the active and inactive queues.
1068 	 *
1069 	 * If we were able to completely satisfy the free+cache targets
1070 	 * from the inactive pool we limit the number of pages we move
1071 	 * from the active pool to the inactive pool to 2x the pages we
1072 	 * had removed from the inactive pool (with a minimum of 1/5 the
1073 	 * inactive target).  If we were not able to completely satisfy
1074 	 * the free+cache targets we go for the whole target aggressively.
1075 	 *
1076 	 * NOTE: Both variables can end up negative.
1077 	 * NOTE: We are still in a critical section.
1078 	 */
1079 	active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count;
1080 	if (inactive_original_shortage < vmstats.v_inactive_target / 10)
1081 		inactive_original_shortage = vmstats.v_inactive_target / 10;
1082 	if (inactive_shortage <= 0 &&
1083 	    active_shortage > inactive_original_shortage * 2) {
1084 		active_shortage = inactive_original_shortage * 2;
1085 	}
1086 
1087 	pcount = vmstats.v_active_count;
1088 	recycle_count = 0;
1089 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1090 
1091 	while ((m != NULL) && (pcount-- > 0) &&
1092 	       (inactive_shortage > 0 || active_shortage > 0)
1093 	) {
1094 		/*
1095 		 * Give interrupts a chance.
1096 		 */
1097 		crit_exit();
1098 		crit_enter();
1099 
1100 		/*
1101 		 * If the page was ripped out from under us, just stop.
1102 		 */
1103 		if (m->queue != PQ_ACTIVE)
1104 			break;
1105 		next = TAILQ_NEXT(m, pageq);
1106 
1107 		/*
1108 		 * Don't deactivate pages that are busy.
1109 		 */
1110 		if ((m->busy != 0) ||
1111 		    (m->flags & PG_BUSY) ||
1112 		    (m->hold_count != 0)) {
1113 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1114 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1115 			m = next;
1116 			continue;
1117 		}
1118 
1119 		/*
1120 		 * The count for pagedaemon pages is done after checking the
1121 		 * page for eligibility...
1122 		 */
1123 		mycpu->gd_cnt.v_pdpages++;
1124 
1125 		/*
1126 		 * Check to see "how much" the page has been used and clear
1127 		 * the tracking access bits.  If the object has no references
1128 		 * don't bother paying the expense.
1129 		 */
1130 		actcount = 0;
1131 		if (m->object->ref_count != 0) {
1132 			if (m->flags & PG_REFERENCED)
1133 				++actcount;
1134 			actcount += pmap_ts_referenced(m);
1135 			if (actcount) {
1136 				m->act_count += ACT_ADVANCE + actcount;
1137 				if (m->act_count > ACT_MAX)
1138 					m->act_count = ACT_MAX;
1139 			}
1140 		}
1141 		vm_page_flag_clear(m, PG_REFERENCED);
1142 
1143 		/*
1144 		 * actcount is only valid if the object ref_count is non-zero.
1145 		 */
1146 		if (actcount && m->object->ref_count != 0) {
1147 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1148 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1149 		} else {
1150 			m->act_count -= min(m->act_count, ACT_DECLINE);
1151 			if (vm_pageout_algorithm ||
1152 			    m->object->ref_count == 0 ||
1153 			    m->act_count < pass + 1
1154 			) {
1155 				/*
1156 				 * Deactivate the page.  If we had a
1157 				 * shortage from our inactive scan try to
1158 				 * free (cache) the page instead.
1159 				 *
1160 				 * Don't just blindly cache the page if
1161 				 * we do not have a shortage from the
1162 				 * inactive scan, that could lead to
1163 				 * gigabytes being moved.
1164 				 */
1165 				--active_shortage;
1166 				if (inactive_shortage > 0 ||
1167 				    m->object->ref_count == 0) {
1168 					if (inactive_shortage > 0)
1169 						++recycle_count;
1170 					vm_page_busy(m);
1171 					vm_page_protect(m, VM_PROT_NONE);
1172 					if (m->dirty == 0 &&
1173 					    inactive_shortage > 0) {
1174 						--inactive_shortage;
1175 						vm_page_cache(m);
1176 					} else {
1177 						vm_page_deactivate(m);
1178 						vm_page_wakeup(m);
1179 					}
1180 				} else {
1181 					vm_page_deactivate(m);
1182 				}
1183 			} else {
1184 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1185 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1186 			}
1187 		}
1188 		m = next;
1189 	}
1190 
1191 	/*
1192 	 * The number of actually free pages can drop down to v_free_reserved,
1193 	 * we try to build the free count back above v_free_min.  Note that
1194 	 * vm_paging_needed() also returns TRUE if v_free_count is not at
1195 	 * least v_free_min so that is the minimum we must build the free
1196 	 * count to.
1197 	 *
1198 	 * We use a slightly higher target to improve hysteresis,
1199 	 * ((v_free_target + v_free_min) / 2).  Since v_free_target
1200 	 * is usually the same as v_cache_min this maintains about
1201 	 * half the pages in the free queue as are in the cache queue,
1202 	 * providing pretty good pipelining for pageout operation.
1203 	 *
1204 	 * The system operator can manipulate vm.v_cache_min and
1205 	 * vm.v_free_target to tune the pageout demon.  Be sure
1206 	 * to keep vm.v_free_min < vm.v_free_target.
1207 	 *
1208 	 * Note that the original paging target is to get at least
1209 	 * (free_min + cache_min) into (free + cache).  The slightly
1210 	 * higher target will shift additional pages from cache to free
1211 	 * without effecting the original paging target in order to
1212 	 * maintain better hysteresis and not have the free count always
1213 	 * be dead-on v_free_min.
1214 	 *
1215 	 * NOTE: we are still in a critical section.
1216 	 *
1217 	 * Pages moved from PQ_CACHE to totally free are not counted in the
1218 	 * pages_freed counter.
1219 	 */
1220 	while (vmstats.v_free_count <
1221 	       (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1222 		/*
1223 		 *
1224 		 */
1225 		static int cache_rover = 0;
1226 		m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1227 		if (m == NULL)
1228 			break;
1229 		if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1230 		    m->busy ||
1231 		    m->hold_count ||
1232 		    m->wire_count) {
1233 #ifdef INVARIANTS
1234 			kprintf("Warning: busy page %p found in cache\n", m);
1235 #endif
1236 			vm_page_deactivate(m);
1237 			continue;
1238 		}
1239 		KKASSERT((m->flags & PG_MAPPED) == 0);
1240 		KKASSERT(m->dirty == 0);
1241 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1242 		vm_pageout_page_free(m);
1243 		mycpu->gd_cnt.v_dfree++;
1244 	}
1245 
1246 	crit_exit();
1247 
1248 #if !defined(NO_SWAPPING)
1249 	/*
1250 	 * Idle process swapout -- run once per second.
1251 	 */
1252 	if (vm_swap_idle_enabled) {
1253 		static long lsec;
1254 		if (time_second != lsec) {
1255 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1256 			vm_req_vmdaemon();
1257 			lsec = time_second;
1258 		}
1259 	}
1260 #endif
1261 
1262 	/*
1263 	 * If we didn't get enough free pages, and we have skipped a vnode
1264 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1265 	 * if we did not get enough free pages.
1266 	 */
1267 	if (vm_paging_target() > 0) {
1268 		if (vnodes_skipped && vm_page_count_min(0))
1269 			speedup_syncer();
1270 #if !defined(NO_SWAPPING)
1271 		if (vm_swap_enabled && vm_page_count_target()) {
1272 			vm_req_vmdaemon();
1273 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1274 		}
1275 #endif
1276 	}
1277 
1278 	/*
1279 	 * Handle catastrophic conditions.  Under good conditions we should
1280 	 * be at the target, well beyond our minimum.  If we could not even
1281 	 * reach our minimum the system is under heavy stress.
1282 	 *
1283 	 * Determine whether we have run out of memory.  This occurs when
1284 	 * swap_pager_full is TRUE and the only pages left in the page
1285 	 * queues are dirty.  We will still likely have page shortages.
1286 	 *
1287 	 * - swap_pager_full is set if insufficient swap was
1288 	 *   available to satisfy a requested pageout.
1289 	 *
1290 	 * - the inactive queue is bloated (4 x size of active queue),
1291 	 *   meaning it is unable to get rid of dirty pages and.
1292 	 *
1293 	 * - vm_page_count_min() without counting pages recycled from the
1294 	 *   active queue (recycle_count) means we could not recover
1295 	 *   enough pages to meet bare minimum needs.  This test only
1296 	 *   works if the inactive queue is bloated.
1297 	 *
1298 	 * - due to a positive inactive_shortage we shifted the remaining
1299 	 *   dirty pages from the active queue to the inactive queue
1300 	 *   trying to find clean ones to free.
1301 	 */
1302 	if (swap_pager_full && vm_page_count_min(recycle_count))
1303 		kprintf("Warning: system low on memory+swap!\n");
1304 	if (swap_pager_full && vm_page_count_min(recycle_count) &&
1305 	    vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1306 	    inactive_shortage > 0) {
1307 		/*
1308 		 * Kill something.
1309 		 */
1310 		info.bigproc = NULL;
1311 		info.bigsize = 0;
1312 		allproc_scan(vm_pageout_scan_callback, &info);
1313 		if (info.bigproc != NULL) {
1314 			killproc(info.bigproc, "out of swap space");
1315 			info.bigproc->p_nice = PRIO_MIN;
1316 			info.bigproc->p_usched->resetpriority(
1317 				FIRST_LWP_IN_PROC(info.bigproc));
1318 			wakeup(&vmstats.v_free_count);
1319 			PRELE(info.bigproc);
1320 		}
1321 	}
1322 	return(inactive_shortage);
1323 }
1324 
1325 /*
1326  * The caller must hold vm_token and proc_token.
1327  */
1328 static int
1329 vm_pageout_scan_callback(struct proc *p, void *data)
1330 {
1331 	struct vm_pageout_scan_info *info = data;
1332 	vm_offset_t size;
1333 
1334 	/*
1335 	 * Never kill system processes or init.  If we have configured swap
1336 	 * then try to avoid killing low-numbered pids.
1337 	 */
1338 	if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1339 	    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1340 		return (0);
1341 	}
1342 
1343 	/*
1344 	 * if the process is in a non-running type state,
1345 	 * don't touch it.
1346 	 */
1347 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1348 		return (0);
1349 
1350 	/*
1351 	 * Get the approximate process size.  Note that anonymous pages
1352 	 * with backing swap will be counted twice, but there should not
1353 	 * be too many such pages due to the stress the VM system is
1354 	 * under at this point.
1355 	 */
1356 	size = vmspace_anonymous_count(p->p_vmspace) +
1357 		vmspace_swap_count(p->p_vmspace);
1358 
1359 	/*
1360 	 * If the this process is bigger than the biggest one
1361 	 * remember it.
1362 	 */
1363 	if (info->bigsize < size) {
1364 		if (info->bigproc)
1365 			PRELE(info->bigproc);
1366 		PHOLD(p);
1367 		info->bigproc = p;
1368 		info->bigsize = size;
1369 	}
1370 	return(0);
1371 }
1372 
1373 /*
1374  * This routine tries to maintain the pseudo LRU active queue,
1375  * so that during long periods of time where there is no paging,
1376  * that some statistic accumulation still occurs.  This code
1377  * helps the situation where paging just starts to occur.
1378  *
1379  * The caller must hold vm_token.
1380  */
1381 static void
1382 vm_pageout_page_stats(void)
1383 {
1384 	vm_page_t m,next;
1385 	int pcount,tpcount;		/* Number of pages to check */
1386 	static int fullintervalcount = 0;
1387 	int page_shortage;
1388 
1389 	page_shortage =
1390 	    (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1391 	    (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1392 
1393 	if (page_shortage <= 0)
1394 		return;
1395 
1396 	crit_enter();
1397 
1398 	pcount = vmstats.v_active_count;
1399 	fullintervalcount += vm_pageout_stats_interval;
1400 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1401 		tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1402 		if (pcount > tpcount)
1403 			pcount = tpcount;
1404 	} else {
1405 		fullintervalcount = 0;
1406 	}
1407 
1408 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1409 	while ((m != NULL) && (pcount-- > 0)) {
1410 		int actcount;
1411 
1412 		if (m->queue != PQ_ACTIVE) {
1413 			break;
1414 		}
1415 
1416 		next = TAILQ_NEXT(m, pageq);
1417 		/*
1418 		 * Don't deactivate pages that are busy.
1419 		 */
1420 		if ((m->busy != 0) ||
1421 		    (m->flags & PG_BUSY) ||
1422 		    (m->hold_count != 0)) {
1423 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1424 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1425 			m = next;
1426 			continue;
1427 		}
1428 
1429 		actcount = 0;
1430 		if (m->flags & PG_REFERENCED) {
1431 			vm_page_flag_clear(m, PG_REFERENCED);
1432 			actcount += 1;
1433 		}
1434 
1435 		actcount += pmap_ts_referenced(m);
1436 		if (actcount) {
1437 			m->act_count += ACT_ADVANCE + actcount;
1438 			if (m->act_count > ACT_MAX)
1439 				m->act_count = ACT_MAX;
1440 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1441 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1442 		} else {
1443 			if (m->act_count == 0) {
1444 				/*
1445 				 * We turn off page access, so that we have
1446 				 * more accurate RSS stats.  We don't do this
1447 				 * in the normal page deactivation when the
1448 				 * system is loaded VM wise, because the
1449 				 * cost of the large number of page protect
1450 				 * operations would be higher than the value
1451 				 * of doing the operation.
1452 				 */
1453 				vm_page_busy(m);
1454 				vm_page_protect(m, VM_PROT_NONE);
1455 				vm_page_deactivate(m);
1456 				vm_page_wakeup(m);
1457 			} else {
1458 				m->act_count -= min(m->act_count, ACT_DECLINE);
1459 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1460 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1461 			}
1462 		}
1463 
1464 		m = next;
1465 	}
1466 	crit_exit();
1467 }
1468 
1469 /*
1470  * The caller must hold vm_token.
1471  */
1472 static int
1473 vm_pageout_free_page_calc(vm_size_t count)
1474 {
1475 	if (count < vmstats.v_page_count)
1476 		 return 0;
1477 	/*
1478 	 * free_reserved needs to include enough for the largest swap pager
1479 	 * structures plus enough for any pv_entry structs when paging.
1480 	 */
1481 	if (vmstats.v_page_count > 1024)
1482 		vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1483 	else
1484 		vmstats.v_free_min = 4;
1485 	vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1486 		vmstats.v_interrupt_free_min;
1487 	vmstats.v_free_reserved = vm_pageout_page_count +
1488 		vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1489 	vmstats.v_free_severe = vmstats.v_free_min / 2;
1490 	vmstats.v_free_min += vmstats.v_free_reserved;
1491 	vmstats.v_free_severe += vmstats.v_free_reserved;
1492 	return 1;
1493 }
1494 
1495 
1496 /*
1497  * vm_pageout is the high level pageout daemon.
1498  *
1499  * No requirements.
1500  */
1501 static void
1502 vm_pageout_thread(void)
1503 {
1504 	int pass;
1505 	int inactive_shortage;
1506 
1507 	/*
1508 	 * Permanently hold vm_token.
1509 	 */
1510 	lwkt_gettoken(&vm_token);
1511 
1512 	/*
1513 	 * Initialize some paging parameters.
1514 	 */
1515 	curthread->td_flags |= TDF_SYSTHREAD;
1516 
1517 	vmstats.v_interrupt_free_min = 2;
1518 	if (vmstats.v_page_count < 2000)
1519 		vm_pageout_page_count = 8;
1520 
1521 	vm_pageout_free_page_calc(vmstats.v_page_count);
1522 
1523 	/*
1524 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1525 	 * that these are more a measure of the VM cache queue hysteresis
1526 	 * then the VM free queue.  Specifically, v_free_target is the
1527 	 * high water mark (free+cache pages).
1528 	 *
1529 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1530 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1531 	 * be big enough to handle memory needs while the pageout daemon
1532 	 * is signalled and run to free more pages.
1533 	 */
1534 	if (vmstats.v_free_count > 6144)
1535 		vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1536 	else
1537 		vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1538 
1539 	/*
1540 	 * NOTE: With the new buffer cache b_act_count we want the default
1541 	 *	 inactive target to be a percentage of available memory.
1542 	 *
1543 	 *	 The inactive target essentially determines the minimum
1544 	 *	 number of 'temporary' pages capable of caching one-time-use
1545 	 *	 files when the VM system is otherwise full of pages
1546 	 *	 belonging to multi-time-use files or active program data.
1547 	 *
1548 	 * NOTE: The inactive target is aggressively persued only if the
1549 	 *	 inactive queue becomes too small.  If the inactive queue
1550 	 *	 is large enough to satisfy page movement to free+cache
1551 	 *	 then it is repopulated more slowly from the active queue.
1552 	 *	 This allows a general inactive_target default to be set.
1553 	 *
1554 	 *	 There is an issue here for processes which sit mostly idle
1555 	 *	 'overnight', such as sshd, tcsh, and X.  Any movement from
1556 	 *	 the active queue will eventually cause such pages to
1557 	 *	 recycle eventually causing a lot of paging in the morning.
1558 	 *	 To reduce the incidence of this pages cycled out of the
1559 	 *	 buffer cache are moved directly to the inactive queue if
1560 	 *	 they were only used once or twice.
1561 	 *
1562 	 *	 The vfs.vm_cycle_point sysctl can be used to adjust this.
1563 	 *	 Increasing the value (up to 64) increases the number of
1564 	 *	 buffer recyclements which go directly to the inactive queue.
1565 	 */
1566 	if (vmstats.v_free_count > 2048) {
1567 		vmstats.v_cache_min = vmstats.v_free_target;
1568 		vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1569 	} else {
1570 		vmstats.v_cache_min = 0;
1571 		vmstats.v_cache_max = 0;
1572 	}
1573 	vmstats.v_inactive_target = vmstats.v_free_count / 4;
1574 
1575 	/* XXX does not really belong here */
1576 	if (vm_page_max_wired == 0)
1577 		vm_page_max_wired = vmstats.v_free_count / 3;
1578 
1579 	if (vm_pageout_stats_max == 0)
1580 		vm_pageout_stats_max = vmstats.v_free_target;
1581 
1582 	/*
1583 	 * Set interval in seconds for stats scan.
1584 	 */
1585 	if (vm_pageout_stats_interval == 0)
1586 		vm_pageout_stats_interval = 5;
1587 	if (vm_pageout_full_stats_interval == 0)
1588 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1589 
1590 
1591 	/*
1592 	 * Set maximum free per pass
1593 	 */
1594 	if (vm_pageout_stats_free_max == 0)
1595 		vm_pageout_stats_free_max = 5;
1596 
1597 	swap_pager_swap_init();
1598 	pass = 0;
1599 
1600 	/*
1601 	 * The pageout daemon is never done, so loop forever.
1602 	 */
1603 	while (TRUE) {
1604 		int error;
1605 
1606 		/*
1607 		 * Wait for an action request.  If we timeout check to
1608 		 * see if paging is needed (in case the normal wakeup
1609 		 * code raced us).
1610 		 */
1611 		if (vm_pages_needed == 0) {
1612 			error = tsleep(&vm_pages_needed,
1613 				       0, "psleep",
1614 				       vm_pageout_stats_interval * hz);
1615 			if (error &&
1616 			    vm_paging_needed() == 0 &&
1617 			    vm_pages_needed == 0) {
1618 				vm_pageout_page_stats();
1619 				continue;
1620 			}
1621 			vm_pages_needed = 1;
1622 		}
1623 
1624 		mycpu->gd_cnt.v_pdwakeups++;
1625 
1626 		/*
1627 		 * Scan for pageout.  Try to avoid thrashing the system
1628 		 * with activity.
1629 		 */
1630 		inactive_shortage = vm_pageout_scan(pass);
1631 		if (inactive_shortage > 0) {
1632 			++pass;
1633 			if (swap_pager_full) {
1634 				/*
1635 				 * Running out of memory, catastrophic back-off
1636 				 * to one-second intervals.
1637 				 */
1638 				tsleep(&vm_pages_needed, 0, "pdelay", hz);
1639 			} else if (pass < 10 && vm_pages_needed > 1) {
1640 				/*
1641 				 * Normal operation, additional processes
1642 				 * have already kicked us.  Retry immediately.
1643 				 */
1644 			} else if (pass < 10) {
1645 				/*
1646 				 * Normal operation, fewer processes.  Delay
1647 				 * a bit but allow wakeups.
1648 				 */
1649 				vm_pages_needed = 0;
1650 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1651 				vm_pages_needed = 1;
1652 			} else {
1653 				/*
1654 				 * We've taken too many passes, forced delay.
1655 				 */
1656 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1657 			}
1658 		} else {
1659 			/*
1660 			 * Interlocked wakeup of waiters (non-optional)
1661 			 */
1662 			pass = 0;
1663 			if (vm_pages_needed && !vm_page_count_min(0)) {
1664 				wakeup(&vmstats.v_free_count);
1665 				vm_pages_needed = 0;
1666 			}
1667 		}
1668 	}
1669 }
1670 
1671 static struct kproc_desc page_kp = {
1672 	"pagedaemon",
1673 	vm_pageout_thread,
1674 	&pagethread
1675 };
1676 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
1677 
1678 
1679 /*
1680  * Called after allocating a page out of the cache or free queue
1681  * to possibly wake the pagedaemon up to replentish our supply.
1682  *
1683  * We try to generate some hysteresis by waking the pagedaemon up
1684  * when our free+cache pages go below the free_min+cache_min level.
1685  * The pagedaemon tries to get the count back up to at least the
1686  * minimum, and through to the target level if possible.
1687  *
1688  * If the pagedaemon is already active bump vm_pages_needed as a hint
1689  * that there are even more requests pending.
1690  *
1691  * SMP races ok?
1692  * No requirements.
1693  */
1694 void
1695 pagedaemon_wakeup(void)
1696 {
1697 	if (vm_paging_needed() && curthread != pagethread) {
1698 		if (vm_pages_needed == 0) {
1699 			vm_pages_needed = 1;	/* SMP race ok */
1700 			wakeup(&vm_pages_needed);
1701 		} else if (vm_page_count_min(0)) {
1702 			++vm_pages_needed;	/* SMP race ok */
1703 		}
1704 	}
1705 }
1706 
1707 #if !defined(NO_SWAPPING)
1708 
1709 /*
1710  * SMP races ok?
1711  * No requirements.
1712  */
1713 static void
1714 vm_req_vmdaemon(void)
1715 {
1716 	static int lastrun = 0;
1717 
1718 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1719 		wakeup(&vm_daemon_needed);
1720 		lastrun = ticks;
1721 	}
1722 }
1723 
1724 static int vm_daemon_callback(struct proc *p, void *data __unused);
1725 
1726 /*
1727  * No requirements.
1728  */
1729 static void
1730 vm_daemon(void)
1731 {
1732 	/*
1733 	 * Permanently hold vm_token.
1734 	 */
1735 	lwkt_gettoken(&vm_token);
1736 
1737 	while (TRUE) {
1738 		tsleep(&vm_daemon_needed, 0, "psleep", 0);
1739 		if (vm_pageout_req_swapout) {
1740 			swapout_procs(vm_pageout_req_swapout);
1741 			vm_pageout_req_swapout = 0;
1742 		}
1743 		/*
1744 		 * scan the processes for exceeding their rlimits or if
1745 		 * process is swapped out -- deactivate pages
1746 		 */
1747 		allproc_scan(vm_daemon_callback, NULL);
1748 	}
1749 }
1750 
1751 /*
1752  * Caller must hold vm_token and proc_token.
1753  */
1754 static int
1755 vm_daemon_callback(struct proc *p, void *data __unused)
1756 {
1757 	vm_pindex_t limit, size;
1758 
1759 	/*
1760 	 * if this is a system process or if we have already
1761 	 * looked at this process, skip it.
1762 	 */
1763 	if (p->p_flag & (P_SYSTEM | P_WEXIT))
1764 		return (0);
1765 
1766 	/*
1767 	 * if the process is in a non-running type state,
1768 	 * don't touch it.
1769 	 */
1770 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1771 		return (0);
1772 
1773 	/*
1774 	 * get a limit
1775 	 */
1776 	limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1777 			        p->p_rlimit[RLIMIT_RSS].rlim_max));
1778 
1779 	/*
1780 	 * let processes that are swapped out really be
1781 	 * swapped out.  Set the limit to nothing to get as
1782 	 * many pages out to swap as possible.
1783 	 */
1784 	if (p->p_flag & P_SWAPPEDOUT)
1785 		limit = 0;
1786 
1787 	size = vmspace_resident_count(p->p_vmspace);
1788 	if (limit >= 0 && size >= limit) {
1789 		vm_pageout_map_deactivate_pages(
1790 		    &p->p_vmspace->vm_map, limit);
1791 	}
1792 	return (0);
1793 }
1794 
1795 #endif
1796