xref: /dflybsd-src/sys/vm/vm_pageout.c (revision ae788f37fe53d5d1ca1e12a184a662192caad3c5)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by the University of
23  *	California, Berkeley and its contributors.
24  * 4. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
41  *
42  *
43  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
44  * All rights reserved.
45  *
46  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  *
68  * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
69  * $DragonFly: src/sys/vm/vm_pageout.c,v 1.36 2008/07/01 02:02:56 dillon Exp $
70  */
71 
72 /*
73  *	The proverbial page-out daemon.
74  */
75 
76 #include "opt_vm.h"
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/proc.h>
81 #include <sys/kthread.h>
82 #include <sys/resourcevar.h>
83 #include <sys/signalvar.h>
84 #include <sys/vnode.h>
85 #include <sys/vmmeter.h>
86 #include <sys/sysctl.h>
87 
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <sys/lock.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pager.h>
96 #include <vm/swap_pager.h>
97 #include <vm/vm_extern.h>
98 
99 #include <sys/thread2.h>
100 #include <vm/vm_page2.h>
101 
102 /*
103  * System initialization
104  */
105 
106 /* the kernel process "vm_pageout"*/
107 static void vm_pageout (void);
108 static int vm_pageout_clean (vm_page_t);
109 static int vm_pageout_scan (int pass);
110 static int vm_pageout_free_page_calc (vm_size_t count);
111 struct thread *pagethread;
112 
113 static struct kproc_desc page_kp = {
114 	"pagedaemon",
115 	vm_pageout,
116 	&pagethread
117 };
118 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
119 
120 #if !defined(NO_SWAPPING)
121 /* the kernel process "vm_daemon"*/
122 static void vm_daemon (void);
123 static struct	thread *vmthread;
124 
125 static struct kproc_desc vm_kp = {
126 	"vmdaemon",
127 	vm_daemon,
128 	&vmthread
129 };
130 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
131 #endif
132 
133 
134 int vm_pages_needed=0;		/* Event on which pageout daemon sleeps */
135 int vm_pageout_deficit=0;	/* Estimated number of pages deficit */
136 int vm_pageout_pages_needed=0;	/* flag saying that the pageout daemon needs pages */
137 
138 #if !defined(NO_SWAPPING)
139 static int vm_pageout_req_swapout;	/* XXX */
140 static int vm_daemon_needed;
141 #endif
142 extern int vm_swap_size;
143 static int vm_max_launder = 32;
144 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
145 static int vm_pageout_full_stats_interval = 0;
146 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
147 static int defer_swap_pageouts=0;
148 static int disable_swap_pageouts=0;
149 
150 #if defined(NO_SWAPPING)
151 static int vm_swap_enabled=0;
152 static int vm_swap_idle_enabled=0;
153 #else
154 static int vm_swap_enabled=1;
155 static int vm_swap_idle_enabled=0;
156 #endif
157 
158 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
159 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
160 
161 SYSCTL_INT(_vm, OID_AUTO, max_launder,
162 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
163 
164 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
165 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
166 
167 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
168 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
169 
170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
171 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
172 
173 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
174 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
175 
176 #if defined(NO_SWAPPING)
177 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
178 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
179 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
180 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
181 #else
182 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
183 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
184 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
185 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
186 #endif
187 
188 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
189 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
190 
191 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
192 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
193 
194 static int pageout_lock_miss;
195 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
196 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
197 
198 int vm_load;
199 SYSCTL_INT(_vm, OID_AUTO, vm_load,
200 	CTLFLAG_RD, &vm_load, 0, "load on the VM system");
201 int vm_load_enable = 1;
202 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
203 	CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
204 #ifdef INVARIANTS
205 int vm_load_debug;
206 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
207 	CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
208 #endif
209 
210 #define VM_PAGEOUT_PAGE_COUNT 16
211 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
212 
213 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
214 
215 #if !defined(NO_SWAPPING)
216 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
217 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
218 static freeer_fcn_t vm_pageout_object_deactivate_pages;
219 static void vm_req_vmdaemon (void);
220 #endif
221 static void vm_pageout_page_stats(void);
222 
223 /*
224  * Update vm_load to slow down faulting processes.
225  */
226 void
227 vm_fault_ratecheck(void)
228 {
229 	if (vm_pages_needed) {
230 		if (vm_load < 1000)
231 			++vm_load;
232 	} else {
233 		if (vm_load > 0)
234 			--vm_load;
235 	}
236 }
237 
238 /*
239  * vm_pageout_clean:
240  *
241  * Clean the page and remove it from the laundry.  The page must not be
242  * busy on-call.
243  *
244  * We set the busy bit to cause potential page faults on this page to
245  * block.  Note the careful timing, however, the busy bit isn't set till
246  * late and we cannot do anything that will mess with the page.
247  */
248 
249 static int
250 vm_pageout_clean(vm_page_t m)
251 {
252 	vm_object_t object;
253 	vm_page_t mc[2*vm_pageout_page_count];
254 	int pageout_count;
255 	int ib, is, page_base;
256 	vm_pindex_t pindex = m->pindex;
257 
258 	object = m->object;
259 
260 	/*
261 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
262 	 * with the new swapper, but we could have serious problems paging
263 	 * out other object types if there is insufficient memory.
264 	 *
265 	 * Unfortunately, checking free memory here is far too late, so the
266 	 * check has been moved up a procedural level.
267 	 */
268 
269 	/*
270 	 * Don't mess with the page if it's busy, held, or special
271 	 */
272 	if ((m->hold_count != 0) ||
273 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
274 		return 0;
275 	}
276 
277 	mc[vm_pageout_page_count] = m;
278 	pageout_count = 1;
279 	page_base = vm_pageout_page_count;
280 	ib = 1;
281 	is = 1;
282 
283 	/*
284 	 * Scan object for clusterable pages.
285 	 *
286 	 * We can cluster ONLY if: ->> the page is NOT
287 	 * clean, wired, busy, held, or mapped into a
288 	 * buffer, and one of the following:
289 	 * 1) The page is inactive, or a seldom used
290 	 *    active page.
291 	 * -or-
292 	 * 2) we force the issue.
293 	 *
294 	 * During heavy mmap/modification loads the pageout
295 	 * daemon can really fragment the underlying file
296 	 * due to flushing pages out of order and not trying
297 	 * align the clusters (which leave sporatic out-of-order
298 	 * holes).  To solve this problem we do the reverse scan
299 	 * first and attempt to align our cluster, then do a
300 	 * forward scan if room remains.
301 	 */
302 
303 more:
304 	while (ib && pageout_count < vm_pageout_page_count) {
305 		vm_page_t p;
306 
307 		if (ib > pindex) {
308 			ib = 0;
309 			break;
310 		}
311 
312 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
313 			ib = 0;
314 			break;
315 		}
316 		if (((p->queue - p->pc) == PQ_CACHE) ||
317 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
318 			ib = 0;
319 			break;
320 		}
321 		vm_page_test_dirty(p);
322 		if ((p->dirty & p->valid) == 0 ||
323 		    p->queue != PQ_INACTIVE ||
324 		    p->wire_count != 0 ||	/* may be held by buf cache */
325 		    p->hold_count != 0) {	/* may be undergoing I/O */
326 			ib = 0;
327 			break;
328 		}
329 		mc[--page_base] = p;
330 		++pageout_count;
331 		++ib;
332 		/*
333 		 * alignment boundry, stop here and switch directions.  Do
334 		 * not clear ib.
335 		 */
336 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
337 			break;
338 	}
339 
340 	while (pageout_count < vm_pageout_page_count &&
341 	    pindex + is < object->size) {
342 		vm_page_t p;
343 
344 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
345 			break;
346 		if (((p->queue - p->pc) == PQ_CACHE) ||
347 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
348 			break;
349 		}
350 		vm_page_test_dirty(p);
351 		if ((p->dirty & p->valid) == 0 ||
352 		    p->queue != PQ_INACTIVE ||
353 		    p->wire_count != 0 ||	/* may be held by buf cache */
354 		    p->hold_count != 0) {	/* may be undergoing I/O */
355 			break;
356 		}
357 		mc[page_base + pageout_count] = p;
358 		++pageout_count;
359 		++is;
360 	}
361 
362 	/*
363 	 * If we exhausted our forward scan, continue with the reverse scan
364 	 * when possible, even past a page boundry.  This catches boundry
365 	 * conditions.
366 	 */
367 	if (ib && pageout_count < vm_pageout_page_count)
368 		goto more;
369 
370 	/*
371 	 * we allow reads during pageouts...
372 	 */
373 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
374 }
375 
376 /*
377  * vm_pageout_flush() - launder the given pages
378  *
379  *	The given pages are laundered.  Note that we setup for the start of
380  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
381  *	reference count all in here rather then in the parent.  If we want
382  *	the parent to do more sophisticated things we may have to change
383  *	the ordering.
384  */
385 int
386 vm_pageout_flush(vm_page_t *mc, int count, int flags)
387 {
388 	vm_object_t object;
389 	int pageout_status[count];
390 	int numpagedout = 0;
391 	int i;
392 
393 	/*
394 	 * Initiate I/O.  Bump the vm_page_t->busy counter.
395 	 */
396 	for (i = 0; i < count; i++) {
397 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
398 		vm_page_io_start(mc[i]);
399 	}
400 
401 	/*
402 	 * We must make the pages read-only.  This will also force the
403 	 * modified bit in the related pmaps to be cleared.  The pager
404 	 * cannot clear the bit for us since the I/O completion code
405 	 * typically runs from an interrupt.  The act of making the page
406 	 * read-only handles the case for us.
407 	 */
408 	for (i = 0; i < count; i++) {
409 		vm_page_protect(mc[i], VM_PROT_READ);
410 	}
411 
412 	object = mc[0]->object;
413 	vm_object_pip_add(object, count);
414 
415 	vm_pager_put_pages(object, mc, count,
416 	    (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
417 	    pageout_status);
418 
419 	for (i = 0; i < count; i++) {
420 		vm_page_t mt = mc[i];
421 
422 		switch (pageout_status[i]) {
423 		case VM_PAGER_OK:
424 			numpagedout++;
425 			break;
426 		case VM_PAGER_PEND:
427 			numpagedout++;
428 			break;
429 		case VM_PAGER_BAD:
430 			/*
431 			 * Page outside of range of object. Right now we
432 			 * essentially lose the changes by pretending it
433 			 * worked.
434 			 */
435 			pmap_clear_modify(mt);
436 			vm_page_undirty(mt);
437 			break;
438 		case VM_PAGER_ERROR:
439 		case VM_PAGER_FAIL:
440 			/*
441 			 * A page typically cannot be paged out when we
442 			 * have run out of swap.  We leave the page
443 			 * marked inactive and will try to page it out
444 			 * again later.
445 			 *
446 			 * Starvation of the active page list is used to
447 			 * determine when the system is massively memory
448 			 * starved.
449 			 */
450 			break;
451 		case VM_PAGER_AGAIN:
452 			break;
453 		}
454 
455 		/*
456 		 * If the operation is still going, leave the page busy to
457 		 * block all other accesses. Also, leave the paging in
458 		 * progress indicator set so that we don't attempt an object
459 		 * collapse.
460 		 *
461 		 * For any pages which have completed synchronously,
462 		 * deactivate the page if we are under a severe deficit.
463 		 * Do not try to enter them into the cache, though, they
464 		 * might still be read-heavy.
465 		 */
466 		if (pageout_status[i] != VM_PAGER_PEND) {
467 			vm_object_pip_wakeup(object);
468 			vm_page_io_finish(mt);
469 			if (vm_page_count_severe())
470 				vm_page_deactivate(mt);
471 #if 0
472 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
473 				vm_page_protect(mt, VM_PROT_READ);
474 #endif
475 		}
476 	}
477 	return numpagedout;
478 }
479 
480 #if !defined(NO_SWAPPING)
481 /*
482  *	vm_pageout_object_deactivate_pages
483  *
484  *	deactivate enough pages to satisfy the inactive target
485  *	requirements or if vm_page_proc_limit is set, then
486  *	deactivate all of the pages in the object and its
487  *	backing_objects.
488  *
489  *	The object and map must be locked.
490  */
491 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
492 
493 static void
494 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
495 	vm_pindex_t desired, int map_remove_only)
496 {
497 	struct rb_vm_page_scan_info info;
498 	int remove_mode;
499 
500 	if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
501 		return;
502 
503 	while (object) {
504 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
505 			return;
506 		if (object->paging_in_progress)
507 			return;
508 
509 		remove_mode = map_remove_only;
510 		if (object->shadow_count > 1)
511 			remove_mode = 1;
512 
513 		/*
514 		 * scan the objects entire memory queue.  spl protection is
515 		 * required to avoid an interrupt unbusy/free race against
516 		 * our busy check.
517 		 */
518 		crit_enter();
519 		info.limit = remove_mode;
520 		info.map = map;
521 		info.desired = desired;
522 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
523 				vm_pageout_object_deactivate_pages_callback,
524 				&info
525 		);
526 		crit_exit();
527 		object = object->backing_object;
528 	}
529 }
530 
531 static int
532 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
533 {
534 	struct rb_vm_page_scan_info *info = data;
535 	int actcount;
536 
537 	if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
538 		return(-1);
539 	}
540 	mycpu->gd_cnt.v_pdpages++;
541 	if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 ||
542 	    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
543 	    !pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
544 		return(0);
545 	}
546 
547 	actcount = pmap_ts_referenced(p);
548 	if (actcount) {
549 		vm_page_flag_set(p, PG_REFERENCED);
550 	} else if (p->flags & PG_REFERENCED) {
551 		actcount = 1;
552 	}
553 
554 	if ((p->queue != PQ_ACTIVE) &&
555 		(p->flags & PG_REFERENCED)) {
556 		vm_page_activate(p);
557 		p->act_count += actcount;
558 		vm_page_flag_clear(p, PG_REFERENCED);
559 	} else if (p->queue == PQ_ACTIVE) {
560 		if ((p->flags & PG_REFERENCED) == 0) {
561 			p->act_count -= min(p->act_count, ACT_DECLINE);
562 			if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) {
563 				vm_page_busy(p);
564 				vm_page_protect(p, VM_PROT_NONE);
565 				vm_page_wakeup(p);
566 				vm_page_deactivate(p);
567 			} else {
568 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
569 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
570 			}
571 		} else {
572 			vm_page_activate(p);
573 			vm_page_flag_clear(p, PG_REFERENCED);
574 			if (p->act_count < (ACT_MAX - ACT_ADVANCE))
575 				p->act_count += ACT_ADVANCE;
576 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
577 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
578 		}
579 	} else if (p->queue == PQ_INACTIVE) {
580 		vm_page_busy(p);
581 		vm_page_protect(p, VM_PROT_NONE);
582 		vm_page_wakeup(p);
583 	}
584 	return(0);
585 }
586 
587 /*
588  * deactivate some number of pages in a map, try to do it fairly, but
589  * that is really hard to do.
590  */
591 static void
592 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
593 {
594 	vm_map_entry_t tmpe;
595 	vm_object_t obj, bigobj;
596 	int nothingwired;
597 
598 	if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
599 		return;
600 	}
601 
602 	bigobj = NULL;
603 	nothingwired = TRUE;
604 
605 	/*
606 	 * first, search out the biggest object, and try to free pages from
607 	 * that.
608 	 */
609 	tmpe = map->header.next;
610 	while (tmpe != &map->header) {
611 		switch(tmpe->maptype) {
612 		case VM_MAPTYPE_NORMAL:
613 		case VM_MAPTYPE_VPAGETABLE:
614 			obj = tmpe->object.vm_object;
615 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
616 				((bigobj == NULL) ||
617 				 (bigobj->resident_page_count < obj->resident_page_count))) {
618 				bigobj = obj;
619 			}
620 			break;
621 		default:
622 			break;
623 		}
624 		if (tmpe->wired_count > 0)
625 			nothingwired = FALSE;
626 		tmpe = tmpe->next;
627 	}
628 
629 	if (bigobj)
630 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
631 
632 	/*
633 	 * Next, hunt around for other pages to deactivate.  We actually
634 	 * do this search sort of wrong -- .text first is not the best idea.
635 	 */
636 	tmpe = map->header.next;
637 	while (tmpe != &map->header) {
638 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
639 			break;
640 		switch(tmpe->maptype) {
641 		case VM_MAPTYPE_NORMAL:
642 		case VM_MAPTYPE_VPAGETABLE:
643 			obj = tmpe->object.vm_object;
644 			if (obj)
645 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
646 			break;
647 		default:
648 			break;
649 		}
650 		tmpe = tmpe->next;
651 	};
652 
653 	/*
654 	 * Remove all mappings if a process is swapped out, this will free page
655 	 * table pages.
656 	 */
657 	if (desired == 0 && nothingwired)
658 		pmap_remove(vm_map_pmap(map),
659 			    VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
660 	vm_map_unlock(map);
661 }
662 #endif
663 
664 /*
665  * Don't try to be fancy - being fancy can lead to vnode deadlocks.   We
666  * only do it for OBJT_DEFAULT and OBJT_SWAP objects which we know can
667  * be trivially freed.
668  */
669 void
670 vm_pageout_page_free(vm_page_t m)
671 {
672 	vm_object_t object = m->object;
673 	int type = object->type;
674 
675 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
676 		vm_object_reference(object);
677 	vm_page_busy(m);
678 	vm_page_protect(m, VM_PROT_NONE);
679 	vm_page_free(m);
680 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
681 		vm_object_deallocate(object);
682 }
683 
684 /*
685  * vm_pageout_scan does the dirty work for the pageout daemon.
686  */
687 struct vm_pageout_scan_info {
688 	struct proc *bigproc;
689 	vm_offset_t bigsize;
690 };
691 
692 static int vm_pageout_scan_callback(struct proc *p, void *data);
693 
694 static int
695 vm_pageout_scan(int pass)
696 {
697 	struct vm_pageout_scan_info info;
698 	vm_page_t m, next;
699 	struct vm_page marker;
700 	int maxscan, pcount;
701 	int recycle_count;
702 	int inactive_shortage, active_shortage;
703 	int inactive_original_shortage;
704 	vm_object_t object;
705 	int actcount;
706 	int vnodes_skipped = 0;
707 	int maxlaunder;
708 
709 	/*
710 	 * Do whatever cleanup that the pmap code can.
711 	 */
712 	pmap_collect();
713 
714 	/*
715 	 * Calculate our target for the number of free+cache pages we
716 	 * want to get to.  This is higher then the number that causes
717 	 * allocations to stall (severe) in order to provide hysteresis,
718 	 * and if we don't make it all the way but get to the minimum
719 	 * we're happy.
720 	 */
721 	inactive_shortage = vm_paging_target() + vm_pageout_deficit;
722 	inactive_original_shortage = inactive_shortage;
723 	vm_pageout_deficit = 0;
724 
725 	/*
726 	 * Initialize our marker
727 	 */
728 	bzero(&marker, sizeof(marker));
729 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
730 	marker.queue = PQ_INACTIVE;
731 	marker.wire_count = 1;
732 
733 	/*
734 	 * Start scanning the inactive queue for pages we can move to the
735 	 * cache or free.  The scan will stop when the target is reached or
736 	 * we have scanned the entire inactive queue.  Note that m->act_count
737 	 * is not used to form decisions for the inactive queue, only for the
738 	 * active queue.
739 	 *
740 	 * maxlaunder limits the number of dirty pages we flush per scan.
741 	 * For most systems a smaller value (16 or 32) is more robust under
742 	 * extreme memory and disk pressure because any unnecessary writes
743 	 * to disk can result in extreme performance degredation.  However,
744 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
745 	 * used) will die horribly with limited laundering.  If the pageout
746 	 * daemon cannot clean enough pages in the first pass, we let it go
747 	 * all out in succeeding passes.
748 	 */
749 	if ((maxlaunder = vm_max_launder) <= 1)
750 		maxlaunder = 1;
751 	if (pass)
752 		maxlaunder = 10000;
753 
754 	/*
755 	 * We will generally be in a critical section throughout the
756 	 * scan, but we can release it temporarily when we are sitting on a
757 	 * non-busy page without fear.  this is required to prevent an
758 	 * interrupt from unbusying or freeing a page prior to our busy
759 	 * check, leaving us on the wrong queue or checking the wrong
760 	 * page.
761 	 */
762 	crit_enter();
763 rescan0:
764 	maxscan = vmstats.v_inactive_count;
765 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
766 	     m != NULL && maxscan-- > 0 && inactive_shortage > 0;
767 	     m = next
768 	 ) {
769 		mycpu->gd_cnt.v_pdpages++;
770 
771 		/*
772 		 * Give interrupts a chance
773 		 */
774 		crit_exit();
775 		crit_enter();
776 
777 		/*
778 		 * It's easier for some of the conditions below to just loop
779 		 * and catch queue changes here rather then check everywhere
780 		 * else.
781 		 */
782 		if (m->queue != PQ_INACTIVE)
783 			goto rescan0;
784 		next = TAILQ_NEXT(m, pageq);
785 
786 		/*
787 		 * skip marker pages
788 		 */
789 		if (m->flags & PG_MARKER)
790 			continue;
791 
792 		/*
793 		 * A held page may be undergoing I/O, so skip it.
794 		 */
795 		if (m->hold_count) {
796 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
797 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
798 			continue;
799 		}
800 
801 		/*
802 		 * Dont mess with busy pages, keep in the front of the
803 		 * queue, most likely are being paged out.
804 		 */
805 		if (m->busy || (m->flags & PG_BUSY)) {
806 			continue;
807 		}
808 
809 		if (m->object->ref_count == 0) {
810 			/*
811 			 * If the object is not being used, we ignore previous
812 			 * references.
813 			 */
814 			vm_page_flag_clear(m, PG_REFERENCED);
815 			pmap_clear_reference(m);
816 
817 		} else if (((m->flags & PG_REFERENCED) == 0) &&
818 			    (actcount = pmap_ts_referenced(m))) {
819 			/*
820 			 * Otherwise, if the page has been referenced while
821 			 * in the inactive queue, we bump the "activation
822 			 * count" upwards, making it less likely that the
823 			 * page will be added back to the inactive queue
824 			 * prematurely again.  Here we check the page tables
825 			 * (or emulated bits, if any), given the upper level
826 			 * VM system not knowing anything about existing
827 			 * references.
828 			 */
829 			vm_page_activate(m);
830 			m->act_count += (actcount + ACT_ADVANCE);
831 			continue;
832 		}
833 
834 		/*
835 		 * If the upper level VM system knows about any page
836 		 * references, we activate the page.  We also set the
837 		 * "activation count" higher than normal so that we will less
838 		 * likely place pages back onto the inactive queue again.
839 		 */
840 		if ((m->flags & PG_REFERENCED) != 0) {
841 			vm_page_flag_clear(m, PG_REFERENCED);
842 			actcount = pmap_ts_referenced(m);
843 			vm_page_activate(m);
844 			m->act_count += (actcount + ACT_ADVANCE + 1);
845 			continue;
846 		}
847 
848 		/*
849 		 * If the upper level VM system doesn't know anything about
850 		 * the page being dirty, we have to check for it again.  As
851 		 * far as the VM code knows, any partially dirty pages are
852 		 * fully dirty.
853 		 *
854 		 * Pages marked PG_WRITEABLE may be mapped into the user
855 		 * address space of a process running on another cpu.  A
856 		 * user process (without holding the MP lock) running on
857 		 * another cpu may be able to touch the page while we are
858 		 * trying to remove it.  vm_page_cache() will handle this
859 		 * case for us.
860 		 */
861 		if (m->dirty == 0) {
862 			vm_page_test_dirty(m);
863 		} else {
864 			vm_page_dirty(m);
865 		}
866 
867 		if (m->valid == 0) {
868 			/*
869 			 * Invalid pages can be easily freed
870 			 */
871 			vm_pageout_page_free(m);
872 			mycpu->gd_cnt.v_dfree++;
873 			--inactive_shortage;
874 		} else if (m->dirty == 0) {
875 			/*
876 			 * Clean pages can be placed onto the cache queue.
877 			 * This effectively frees them.
878 			 */
879 			vm_page_cache(m);
880 			--inactive_shortage;
881 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
882 			/*
883 			 * Dirty pages need to be paged out, but flushing
884 			 * a page is extremely expensive verses freeing
885 			 * a clean page.  Rather then artificially limiting
886 			 * the number of pages we can flush, we instead give
887 			 * dirty pages extra priority on the inactive queue
888 			 * by forcing them to be cycled through the queue
889 			 * twice before being flushed, after which the
890 			 * (now clean) page will cycle through once more
891 			 * before being freed.  This significantly extends
892 			 * the thrash point for a heavily loaded machine.
893 			 */
894 			vm_page_flag_set(m, PG_WINATCFLS);
895 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
896 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
897 		} else if (maxlaunder > 0) {
898 			/*
899 			 * We always want to try to flush some dirty pages if
900 			 * we encounter them, to keep the system stable.
901 			 * Normally this number is small, but under extreme
902 			 * pressure where there are insufficient clean pages
903 			 * on the inactive queue, we may have to go all out.
904 			 */
905 			int swap_pageouts_ok;
906 			struct vnode *vp = NULL;
907 
908 			object = m->object;
909 
910 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
911 				swap_pageouts_ok = 1;
912 			} else {
913 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
914 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
915 				vm_page_count_min(0));
916 
917 			}
918 
919 			/*
920 			 * We don't bother paging objects that are "dead".
921 			 * Those objects are in a "rundown" state.
922 			 */
923 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
924 				TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
925 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
926 				continue;
927 			}
928 
929 			/*
930 			 * The object is already known NOT to be dead.   It
931 			 * is possible for the vget() to block the whole
932 			 * pageout daemon, but the new low-memory handling
933 			 * code should prevent it.
934 			 *
935 			 * The previous code skipped locked vnodes and, worse,
936 			 * reordered pages in the queue.  This results in
937 			 * completely non-deterministic operation because,
938 			 * quite often, a vm_fault has initiated an I/O and
939 			 * is holding a locked vnode at just the point where
940 			 * the pageout daemon is woken up.
941 			 *
942 			 * We can't wait forever for the vnode lock, we might
943 			 * deadlock due to a vn_read() getting stuck in
944 			 * vm_wait while holding this vnode.  We skip the
945 			 * vnode if we can't get it in a reasonable amount
946 			 * of time.
947 			 */
948 
949 			if (object->type == OBJT_VNODE) {
950 				vp = object->handle;
951 
952 				if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK)) {
953 					++pageout_lock_miss;
954 					if (object->flags & OBJ_MIGHTBEDIRTY)
955 						    vnodes_skipped++;
956 					continue;
957 				}
958 
959 				/*
960 				 * The page might have been moved to another
961 				 * queue during potential blocking in vget()
962 				 * above.  The page might have been freed and
963 				 * reused for another vnode.  The object might
964 				 * have been reused for another vnode.
965 				 */
966 				if (m->queue != PQ_INACTIVE ||
967 				    m->object != object ||
968 				    object->handle != vp) {
969 					if (object->flags & OBJ_MIGHTBEDIRTY)
970 						vnodes_skipped++;
971 					vput(vp);
972 					continue;
973 				}
974 
975 				/*
976 				 * The page may have been busied during the
977 				 * blocking in vput();  We don't move the
978 				 * page back onto the end of the queue so that
979 				 * statistics are more correct if we don't.
980 				 */
981 				if (m->busy || (m->flags & PG_BUSY)) {
982 					vput(vp);
983 					continue;
984 				}
985 
986 				/*
987 				 * If the page has become held it might
988 				 * be undergoing I/O, so skip it
989 				 */
990 				if (m->hold_count) {
991 					TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
992 					TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
993 					if (object->flags & OBJ_MIGHTBEDIRTY)
994 						vnodes_skipped++;
995 					vput(vp);
996 					continue;
997 				}
998 			}
999 
1000 			/*
1001 			 * If a page is dirty, then it is either being washed
1002 			 * (but not yet cleaned) or it is still in the
1003 			 * laundry.  If it is still in the laundry, then we
1004 			 * start the cleaning operation.
1005 			 *
1006 			 * This operation may cluster, invalidating the 'next'
1007 			 * pointer.  To prevent an inordinate number of
1008 			 * restarts we use our marker to remember our place.
1009 			 *
1010 			 * decrement inactive_shortage on success to account
1011 			 * for the (future) cleaned page.  Otherwise we
1012 			 * could wind up laundering or cleaning too many
1013 			 * pages.
1014 			 */
1015 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1016 			if (vm_pageout_clean(m) != 0) {
1017 				--inactive_shortage;
1018 				--maxlaunder;
1019 			}
1020 			next = TAILQ_NEXT(&marker, pageq);
1021 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1022 			if (vp != NULL)
1023 				vput(vp);
1024 		}
1025 	}
1026 
1027 	/*
1028 	 * We want to move pages from the active queue to the inactive
1029 	 * queue to get the inactive queue to the inactive target.  If
1030 	 * we still have a page shortage from above we try to directly free
1031 	 * clean pages instead of moving them.
1032 	 *
1033 	 * If we do still have a shortage we keep track of the number of
1034 	 * pages we free or cache (recycle_count) as a measure of thrashing
1035 	 * between the active and inactive queues.
1036 	 *
1037 	 * If we were able to completely satisfy the free+cache targets
1038 	 * from the inactive pool we limit the number of pages we move
1039 	 * from the active pool to the inactive pool to 2x the pages we
1040 	 * had removed from the inactive pool.  If we were not able to
1041 	 * completel satisfy the free+cache targets we go for the whole
1042 	 * target aggressively.
1043 	 *
1044 	 * NOTE: Both variables can end up negative.
1045 	 * NOTE: We are still in a critical section.
1046 	 */
1047 	active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count;
1048 	if (inactive_shortage <= 0 &&
1049 	    active_shortage > inactive_original_shortage * 2) {
1050 		active_shortage = inactive_original_shortage * 2;
1051 	}
1052 
1053 	pcount = vmstats.v_active_count;
1054 	recycle_count = 0;
1055 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1056 
1057 	while ((m != NULL) && (pcount-- > 0) &&
1058 	       (inactive_shortage > 0 || active_shortage > 0)
1059 	) {
1060 		/*
1061 		 * Give interrupts a chance.
1062 		 */
1063 		crit_exit();
1064 		crit_enter();
1065 
1066 		/*
1067 		 * If the page was ripped out from under us, just stop.
1068 		 */
1069 		if (m->queue != PQ_ACTIVE)
1070 			break;
1071 		next = TAILQ_NEXT(m, pageq);
1072 
1073 		/*
1074 		 * Don't deactivate pages that are busy.
1075 		 */
1076 		if ((m->busy != 0) ||
1077 		    (m->flags & PG_BUSY) ||
1078 		    (m->hold_count != 0)) {
1079 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1080 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1081 			m = next;
1082 			continue;
1083 		}
1084 
1085 		/*
1086 		 * The count for pagedaemon pages is done after checking the
1087 		 * page for eligibility...
1088 		 */
1089 		mycpu->gd_cnt.v_pdpages++;
1090 
1091 		/*
1092 		 * Check to see "how much" the page has been used and clear
1093 		 * the tracking access bits.  If the object has no references
1094 		 * don't bother paying the expense.
1095 		 */
1096 		actcount = 0;
1097 		if (m->object->ref_count != 0) {
1098 			if (m->flags & PG_REFERENCED)
1099 				++actcount;
1100 			actcount += pmap_ts_referenced(m);
1101 			if (actcount) {
1102 				m->act_count += ACT_ADVANCE + actcount;
1103 				if (m->act_count > ACT_MAX)
1104 					m->act_count = ACT_MAX;
1105 			}
1106 		}
1107 		vm_page_flag_clear(m, PG_REFERENCED);
1108 
1109 		/*
1110 		 * actcount is only valid if the object ref_count is non-zero.
1111 		 */
1112 		if (actcount && m->object->ref_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 		} else {
1116 			m->act_count -= min(m->act_count, ACT_DECLINE);
1117 			if (vm_pageout_algorithm ||
1118 			    m->object->ref_count == 0 ||
1119 			    m->act_count < pass + 1
1120 			) {
1121 				/*
1122 				 * Deactivate the page.  If we had a
1123 				 * shortage from our inactive scan try to
1124 				 * free (cache) the page instead.
1125 				 */
1126 				--active_shortage;
1127 				if (inactive_shortage > 0 ||
1128 				    m->object->ref_count == 0) {
1129 					if (inactive_shortage > 0)
1130 						++recycle_count;
1131 					vm_page_busy(m);
1132 					vm_page_protect(m, VM_PROT_NONE);
1133 					vm_page_wakeup(m);
1134 					if (m->dirty == 0) {
1135 						--inactive_shortage;
1136 						vm_page_cache(m);
1137 					} else {
1138 						vm_page_deactivate(m);
1139 					}
1140 				} else {
1141 					vm_page_deactivate(m);
1142 				}
1143 			} else {
1144 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1145 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1146 			}
1147 		}
1148 		m = next;
1149 	}
1150 
1151 	/*
1152 	 * We try to maintain some *really* free pages, this allows interrupt
1153 	 * code to be guaranteed space.  Since both cache and free queues
1154 	 * are considered basically 'free', moving pages from cache to free
1155 	 * does not effect other calculations.
1156 	 *
1157 	 * NOTE: we are still in a critical section.
1158 	 *
1159 	 * Pages moved from PQ_CACHE to totally free are not counted in the
1160 	 * pages_freed counter.
1161 	 */
1162 	while (vmstats.v_free_count < vmstats.v_free_reserved) {
1163 		static int cache_rover = 0;
1164 		m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1165 		if (m == NULL)
1166 			break;
1167 		if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1168 		    m->busy ||
1169 		    m->hold_count ||
1170 		    m->wire_count) {
1171 #ifdef INVARIANTS
1172 			kprintf("Warning: busy page %p found in cache\n", m);
1173 #endif
1174 			vm_page_deactivate(m);
1175 			continue;
1176 		}
1177 		KKASSERT((m->flags & PG_MAPPED) == 0);
1178 		KKASSERT(m->dirty == 0);
1179 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1180 		vm_pageout_page_free(m);
1181 		mycpu->gd_cnt.v_dfree++;
1182 	}
1183 
1184 	crit_exit();
1185 
1186 #if !defined(NO_SWAPPING)
1187 	/*
1188 	 * Idle process swapout -- run once per second.
1189 	 */
1190 	if (vm_swap_idle_enabled) {
1191 		static long lsec;
1192 		if (time_second != lsec) {
1193 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1194 			vm_req_vmdaemon();
1195 			lsec = time_second;
1196 		}
1197 	}
1198 #endif
1199 
1200 	/*
1201 	 * If we didn't get enough free pages, and we have skipped a vnode
1202 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1203 	 * if we did not get enough free pages.
1204 	 */
1205 	if (vm_paging_target() > 0) {
1206 		if (vnodes_skipped && vm_page_count_min(0))
1207 			speedup_syncer();
1208 #if !defined(NO_SWAPPING)
1209 		if (vm_swap_enabled && vm_page_count_target()) {
1210 			vm_req_vmdaemon();
1211 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1212 		}
1213 #endif
1214 	}
1215 
1216 	/*
1217 	 * Handle catastrophic conditions.  Under good conditions we should
1218 	 * be at the target, well beyond our minimum.  If we could not even
1219 	 * reach our minimum the system is under heavy stress.
1220 	 *
1221 	 * Determine whether we have run out of memory.  This occurs when
1222 	 * swap_pager_full is TRUE and the only pages left in the page
1223 	 * queues are dirty.  We will still likely have page shortages.
1224 	 *
1225 	 * - swap_pager_full is set if insufficient swap was
1226 	 *   available to satisfy a requested pageout.
1227 	 *
1228 	 * - the inactive queue is bloated (4 x size of active queue),
1229 	 *   meaning it is unable to get rid of dirty pages and.
1230 	 *
1231 	 * - vm_page_count_min() without counting pages recycled from the
1232 	 *   active queue (recycle_count) means we could not recover
1233 	 *   enough pages to meet bare minimum needs.  This test only
1234 	 *   works if the inactive queue is bloated.
1235 	 *
1236 	 * - due to a positive inactive_shortage we shifted the remaining
1237 	 *   dirty pages from the active queue to the inactive queue
1238 	 *   trying to find clean ones to free.
1239 	 */
1240 	if (swap_pager_full && vm_page_count_min(recycle_count))
1241 		kprintf("Warning: system low on memory+swap!\n");
1242 	if (swap_pager_full && vm_page_count_min(recycle_count) &&
1243 	    vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1244 	    inactive_shortage > 0) {
1245 		/*
1246 		 * Kill something.
1247 		 */
1248 		info.bigproc = NULL;
1249 		info.bigsize = 0;
1250 		allproc_scan(vm_pageout_scan_callback, &info);
1251 		if (info.bigproc != NULL) {
1252 			killproc(info.bigproc, "out of swap space");
1253 			info.bigproc->p_nice = PRIO_MIN;
1254 			info.bigproc->p_usched->resetpriority(
1255 				FIRST_LWP_IN_PROC(info.bigproc));
1256 			wakeup(&vmstats.v_free_count);
1257 			PRELE(info.bigproc);
1258 		}
1259 	}
1260 	return(inactive_shortage);
1261 }
1262 
1263 static int
1264 vm_pageout_scan_callback(struct proc *p, void *data)
1265 {
1266 	struct vm_pageout_scan_info *info = data;
1267 	vm_offset_t size;
1268 
1269 	/*
1270 	 * Never kill system processes or init.  If we have configured swap
1271 	 * then try to avoid killing low-numbered pids.
1272 	 */
1273 	if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1274 	    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1275 		return (0);
1276 	}
1277 
1278 	/*
1279 	 * if the process is in a non-running type state,
1280 	 * don't touch it.
1281 	 */
1282 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1283 		return (0);
1284 
1285 	/*
1286 	 * Get the approximate process size.  Note that anonymous pages
1287 	 * with backing swap will be counted twice, but there should not
1288 	 * be too many such pages due to the stress the VM system is
1289 	 * under at this point.
1290 	 */
1291 	size = vmspace_anonymous_count(p->p_vmspace) +
1292 		vmspace_swap_count(p->p_vmspace);
1293 
1294 	/*
1295 	 * If the this process is bigger than the biggest one
1296 	 * remember it.
1297 	 */
1298 	if (info->bigsize < size) {
1299 		if (info->bigproc)
1300 			PRELE(info->bigproc);
1301 		PHOLD(p);
1302 		info->bigproc = p;
1303 		info->bigsize = size;
1304 	}
1305 	return(0);
1306 }
1307 
1308 /*
1309  * This routine tries to maintain the pseudo LRU active queue,
1310  * so that during long periods of time where there is no paging,
1311  * that some statistic accumulation still occurs.  This code
1312  * helps the situation where paging just starts to occur.
1313  */
1314 static void
1315 vm_pageout_page_stats(void)
1316 {
1317 	vm_page_t m,next;
1318 	int pcount,tpcount;		/* Number of pages to check */
1319 	static int fullintervalcount = 0;
1320 	int page_shortage;
1321 
1322 	page_shortage =
1323 	    (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1324 	    (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1325 
1326 	if (page_shortage <= 0)
1327 		return;
1328 
1329 	crit_enter();
1330 
1331 	pcount = vmstats.v_active_count;
1332 	fullintervalcount += vm_pageout_stats_interval;
1333 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1334 		tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1335 		if (pcount > tpcount)
1336 			pcount = tpcount;
1337 	} else {
1338 		fullintervalcount = 0;
1339 	}
1340 
1341 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1342 	while ((m != NULL) && (pcount-- > 0)) {
1343 		int actcount;
1344 
1345 		if (m->queue != PQ_ACTIVE) {
1346 			break;
1347 		}
1348 
1349 		next = TAILQ_NEXT(m, pageq);
1350 		/*
1351 		 * Don't deactivate pages that are busy.
1352 		 */
1353 		if ((m->busy != 0) ||
1354 		    (m->flags & PG_BUSY) ||
1355 		    (m->hold_count != 0)) {
1356 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1357 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1358 			m = next;
1359 			continue;
1360 		}
1361 
1362 		actcount = 0;
1363 		if (m->flags & PG_REFERENCED) {
1364 			vm_page_flag_clear(m, PG_REFERENCED);
1365 			actcount += 1;
1366 		}
1367 
1368 		actcount += pmap_ts_referenced(m);
1369 		if (actcount) {
1370 			m->act_count += ACT_ADVANCE + actcount;
1371 			if (m->act_count > ACT_MAX)
1372 				m->act_count = ACT_MAX;
1373 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1374 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1375 		} else {
1376 			if (m->act_count == 0) {
1377 				/*
1378 				 * We turn off page access, so that we have
1379 				 * more accurate RSS stats.  We don't do this
1380 				 * in the normal page deactivation when the
1381 				 * system is loaded VM wise, because the
1382 				 * cost of the large number of page protect
1383 				 * operations would be higher than the value
1384 				 * of doing the operation.
1385 				 */
1386 				vm_page_busy(m);
1387 				vm_page_protect(m, VM_PROT_NONE);
1388 				vm_page_wakeup(m);
1389 				vm_page_deactivate(m);
1390 			} else {
1391 				m->act_count -= min(m->act_count, ACT_DECLINE);
1392 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1393 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1394 			}
1395 		}
1396 
1397 		m = next;
1398 	}
1399 	crit_exit();
1400 }
1401 
1402 static int
1403 vm_pageout_free_page_calc(vm_size_t count)
1404 {
1405 	if (count < vmstats.v_page_count)
1406 		 return 0;
1407 	/*
1408 	 * free_reserved needs to include enough for the largest swap pager
1409 	 * structures plus enough for any pv_entry structs when paging.
1410 	 */
1411 	if (vmstats.v_page_count > 1024)
1412 		vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1413 	else
1414 		vmstats.v_free_min = 4;
1415 	vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1416 		vmstats.v_interrupt_free_min;
1417 	vmstats.v_free_reserved = vm_pageout_page_count +
1418 		vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1419 	vmstats.v_free_severe = vmstats.v_free_min / 2;
1420 	vmstats.v_free_min += vmstats.v_free_reserved;
1421 	vmstats.v_free_severe += vmstats.v_free_reserved;
1422 	return 1;
1423 }
1424 
1425 
1426 /*
1427  * vm_pageout is the high level pageout daemon.
1428  */
1429 static void
1430 vm_pageout(void)
1431 {
1432 	int pass;
1433 	int inactive_shortage;
1434 
1435 	/*
1436 	 * Initialize some paging parameters.
1437 	 */
1438 	curthread->td_flags |= TDF_SYSTHREAD;
1439 
1440 	vmstats.v_interrupt_free_min = 2;
1441 	if (vmstats.v_page_count < 2000)
1442 		vm_pageout_page_count = 8;
1443 
1444 	vm_pageout_free_page_calc(vmstats.v_page_count);
1445 
1446 	/*
1447 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1448 	 * that these are more a measure of the VM cache queue hysteresis
1449 	 * then the VM free queue.  Specifically, v_free_target is the
1450 	 * high water mark (free+cache pages).
1451 	 *
1452 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1453 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1454 	 * be big enough to handle memory needs while the pageout daemon
1455 	 * is signalled and run to free more pages.
1456 	 */
1457 	if (vmstats.v_free_count > 6144)
1458 		vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1459 	else
1460 		vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1461 
1462 	/*
1463 	 * NOTE: With the new buffer cache b_act_count we want the default
1464 	 *	 inactive target to be a percentage of available memory.
1465 	 *
1466 	 *	 The inactive target essentially determines the minimum
1467 	 *	 number of 'temporary' pages capable of caching one-time-use
1468 	 *	 files when the VM system is otherwise full of pages
1469 	 *	 belonging to multi-time-use files or active program data.
1470 	 *
1471 	 * NOTE: The inactive target is aggressively persued only if the
1472 	 *	 inactive queue becomes too small.  If the inactive queue
1473 	 *	 is large enough to satisfy page movement to free+cache
1474 	 *	 then it is repopulated more slowly from the active queue.
1475 	 *	 This allows a generate inactive_target default to be set.
1476 	 *
1477 	 *	 There is an issue here for processes which sit mostly idle
1478 	 *	 'overnight', such as sshd, tcsh, and X.  Any movement from
1479 	 *	 the active queue will eventually cause such pages to
1480 	 *	 recycle eventually causing a lot of paging in the morning.
1481 	 *	 To reduce the incidence of this pages cycled out of the
1482 	 *	 buffer cache are moved directly to the inactive queue if
1483 	 *	 they were only used once or twice.  The vfs.vm_cycle_point
1484 	 *	 sysctl can be used to adjust this.
1485 	 */
1486 	if (vmstats.v_free_count > 2048) {
1487 		vmstats.v_cache_min = vmstats.v_free_target;
1488 		vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1489 	} else {
1490 		vmstats.v_cache_min = 0;
1491 		vmstats.v_cache_max = 0;
1492 	}
1493 	vmstats.v_inactive_target = vmstats.v_free_count / 2;
1494 
1495 	/* XXX does not really belong here */
1496 	if (vm_page_max_wired == 0)
1497 		vm_page_max_wired = vmstats.v_free_count / 3;
1498 
1499 	if (vm_pageout_stats_max == 0)
1500 		vm_pageout_stats_max = vmstats.v_free_target;
1501 
1502 	/*
1503 	 * Set interval in seconds for stats scan.
1504 	 */
1505 	if (vm_pageout_stats_interval == 0)
1506 		vm_pageout_stats_interval = 5;
1507 	if (vm_pageout_full_stats_interval == 0)
1508 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1509 
1510 
1511 	/*
1512 	 * Set maximum free per pass
1513 	 */
1514 	if (vm_pageout_stats_free_max == 0)
1515 		vm_pageout_stats_free_max = 5;
1516 
1517 	swap_pager_swap_init();
1518 	pass = 0;
1519 
1520 	/*
1521 	 * The pageout daemon is never done, so loop forever.
1522 	 */
1523 	while (TRUE) {
1524 		int error;
1525 
1526 		if (vm_pages_needed == 0) {
1527 			/*
1528 			 * Wait for an action request
1529 			 */
1530 			error = tsleep(&vm_pages_needed,
1531 				       0, "psleep",
1532 				       vm_pageout_stats_interval * hz);
1533 			if (error && vm_pages_needed == 0) {
1534 				vm_pageout_page_stats();
1535 				continue;
1536 			}
1537 			vm_pages_needed = 1;
1538 		}
1539 
1540 		/*
1541 		 * If we have enough free memory, wakeup waiters.
1542 		 */
1543 		crit_enter();
1544 		if (!vm_page_count_min(0))
1545 			wakeup(&vmstats.v_free_count);
1546 		mycpu->gd_cnt.v_pdwakeups++;
1547 		crit_exit();
1548 		inactive_shortage = vm_pageout_scan(pass);
1549 
1550 		/*
1551 		 * Try to avoid thrashing the system with activity.
1552 		 */
1553 		if (inactive_shortage > 0) {
1554 			++pass;
1555 			if (swap_pager_full) {
1556 				/*
1557 				 * Running out of memory, catastrophic back-off
1558 				 * to one-second intervals.
1559 				 */
1560 				tsleep(&vm_pages_needed, 0, "pdelay", hz);
1561 			} else if (pass < 10 && vm_pages_needed > 1) {
1562 				/*
1563 				 * Normal operation, additional processes
1564 				 * have already kicked us.  Retry immediately.
1565 				 */
1566 			} else if (pass < 10) {
1567 				/*
1568 				 * Normal operation, fewer processes.  Delay
1569 				 * a bit but allow wakeups.
1570 				 */
1571 				vm_pages_needed = 0;
1572 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1573 				vm_pages_needed = 1;
1574 			} else {
1575 				/*
1576 				 * We've taken too many passes, forced delay.
1577 				 */
1578 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1579 			}
1580 		} else {
1581 			pass = 0;
1582 			vm_pages_needed = 0;
1583 		}
1584 	}
1585 }
1586 
1587 /*
1588  * Called after allocating a page out of the cache or free queue
1589  * to possibly wake the pagedaemon up to replentish our supply.
1590  *
1591  * We try to generate some hysteresis by waking the pagedaemon up
1592  * when our free+cache pages go below the severe level.  The pagedaemon
1593  * tries to get the count back up to at least the minimum, and through
1594  * to the target level if possible.
1595  *
1596  * If the pagedaemon is already active bump vm_pages_needed as a hint
1597  * that there are even more requests pending.
1598  */
1599 void
1600 pagedaemon_wakeup(void)
1601 {
1602 	if (vm_page_count_severe() && curthread != pagethread) {
1603 		if (vm_pages_needed == 0) {
1604 			vm_pages_needed = 1;
1605 			wakeup(&vm_pages_needed);
1606 		} else if (vm_page_count_min(0)) {
1607 			++vm_pages_needed;
1608 		}
1609 	}
1610 }
1611 
1612 #if !defined(NO_SWAPPING)
1613 static void
1614 vm_req_vmdaemon(void)
1615 {
1616 	static int lastrun = 0;
1617 
1618 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1619 		wakeup(&vm_daemon_needed);
1620 		lastrun = ticks;
1621 	}
1622 }
1623 
1624 static int vm_daemon_callback(struct proc *p, void *data __unused);
1625 
1626 static void
1627 vm_daemon(void)
1628 {
1629 	while (TRUE) {
1630 		tsleep(&vm_daemon_needed, 0, "psleep", 0);
1631 		if (vm_pageout_req_swapout) {
1632 			swapout_procs(vm_pageout_req_swapout);
1633 			vm_pageout_req_swapout = 0;
1634 		}
1635 		/*
1636 		 * scan the processes for exceeding their rlimits or if
1637 		 * process is swapped out -- deactivate pages
1638 		 */
1639 		allproc_scan(vm_daemon_callback, NULL);
1640 	}
1641 }
1642 
1643 static int
1644 vm_daemon_callback(struct proc *p, void *data __unused)
1645 {
1646 	vm_pindex_t limit, size;
1647 
1648 	/*
1649 	 * if this is a system process or if we have already
1650 	 * looked at this process, skip it.
1651 	 */
1652 	if (p->p_flag & (P_SYSTEM | P_WEXIT))
1653 		return (0);
1654 
1655 	/*
1656 	 * if the process is in a non-running type state,
1657 	 * don't touch it.
1658 	 */
1659 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1660 		return (0);
1661 
1662 	/*
1663 	 * get a limit
1664 	 */
1665 	limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1666 			        p->p_rlimit[RLIMIT_RSS].rlim_max));
1667 
1668 	/*
1669 	 * let processes that are swapped out really be
1670 	 * swapped out.  Set the limit to nothing to get as
1671 	 * many pages out to swap as possible.
1672 	 */
1673 	if (p->p_flag & P_SWAPPEDOUT)
1674 		limit = 0;
1675 
1676 	size = vmspace_resident_count(p->p_vmspace);
1677 	if (limit >= 0 && size >= limit) {
1678 		vm_pageout_map_deactivate_pages(
1679 		    &p->p_vmspace->vm_map, limit);
1680 	}
1681 	return (0);
1682 }
1683 
1684 #endif
1685