xref: /freebsd/sys/vm/vm_pageout.c (revision b52b9d56d4e96089873a75f9e29062eec19fabba)
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  * $FreeBSD$
69  */
70 
71 /*
72  *	The proverbial page-out daemon.
73  */
74 
75 #include "opt_vm.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/lock.h>
80 #include <sys/mutex.h>
81 #include <sys/proc.h>
82 #include <sys/kthread.h>
83 #include <sys/ktr.h>
84 #include <sys/resourcevar.h>
85 #include <sys/signalvar.h>
86 #include <sys/vnode.h>
87 #include <sys/vmmeter.h>
88 #include <sys/sx.h>
89 #include <sys/sysctl.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_param.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 #include <vm/uma.h>
101 
102 #include <machine/mutex.h>
103 
104 /*
105  * System initialization
106  */
107 
108 /* the kernel process "vm_pageout"*/
109 static void vm_pageout(void);
110 static int vm_pageout_clean(vm_page_t);
111 static void vm_pageout_scan(int pass);
112 static int vm_pageout_free_page_calc(vm_size_t count);
113 struct proc *pageproc;
114 
115 static struct kproc_desc page_kp = {
116 	"pagedaemon",
117 	vm_pageout,
118 	&pageproc
119 };
120 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
121 
122 #if !defined(NO_SWAPPING)
123 /* the kernel process "vm_daemon"*/
124 static void vm_daemon(void);
125 static struct	proc *vmproc;
126 
127 static struct kproc_desc vm_kp = {
128 	"vmdaemon",
129 	vm_daemon,
130 	&vmproc
131 };
132 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
133 #endif
134 
135 
136 int vm_pages_needed=0;		/* Event on which pageout daemon sleeps */
137 int vm_pageout_deficit=0;	/* Estimated number of pages deficit */
138 int vm_pageout_pages_needed=0;	/* flag saying that the pageout daemon needs pages */
139 
140 #if !defined(NO_SWAPPING)
141 static int vm_pageout_req_swapout;	/* XXX */
142 static int vm_daemon_needed;
143 #endif
144 extern int vm_swap_size;
145 static int vm_max_launder = 32;
146 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
147 static int vm_pageout_full_stats_interval = 0;
148 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
149 static int defer_swap_pageouts=0;
150 static int disable_swap_pageouts=0;
151 
152 #if defined(NO_SWAPPING)
153 static int vm_swap_enabled=0;
154 static int vm_swap_idle_enabled=0;
155 #else
156 static int vm_swap_enabled=1;
157 static int vm_swap_idle_enabled=0;
158 #endif
159 
160 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
161 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
162 
163 SYSCTL_INT(_vm, OID_AUTO, max_launder,
164 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
165 
166 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
167 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
168 
169 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
170 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
171 
172 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
173 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
174 
175 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
176 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
177 
178 #if defined(NO_SWAPPING)
179 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
180 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
181 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
182 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
183 #else
184 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
185 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
186 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
187 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
188 #endif
189 
190 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
191 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
192 
193 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
194 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
195 
196 static int pageout_lock_miss;
197 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
198 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
199 
200 #define VM_PAGEOUT_PAGE_COUNT 16
201 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
202 
203 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
204 
205 #if !defined(NO_SWAPPING)
206 typedef void freeer_fcn_t(vm_map_t, vm_object_t, vm_pindex_t, int);
207 static void vm_pageout_map_deactivate_pages(vm_map_t, vm_pindex_t);
208 static freeer_fcn_t vm_pageout_object_deactivate_pages;
209 static void vm_req_vmdaemon(void);
210 #endif
211 static void vm_pageout_page_stats(void);
212 
213 /*
214  * vm_pageout_clean:
215  *
216  * Clean the page and remove it from the laundry.
217  *
218  * We set the busy bit to cause potential page faults on this page to
219  * block.  Note the careful timing, however, the busy bit isn't set till
220  * late and we cannot do anything that will mess with the page.
221  */
222 static int
223 vm_pageout_clean(m)
224 	vm_page_t m;
225 {
226 	vm_object_t object;
227 	vm_page_t mc[2*vm_pageout_page_count];
228 	int pageout_count;
229 	int ib, is, page_base;
230 	vm_pindex_t pindex = m->pindex;
231 
232 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
233 
234 	object = m->object;
235 
236 	/*
237 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
238 	 * with the new swapper, but we could have serious problems paging
239 	 * out other object types if there is insufficient memory.
240 	 *
241 	 * Unfortunately, checking free memory here is far too late, so the
242 	 * check has been moved up a procedural level.
243 	 */
244 
245 	/*
246 	 * Don't mess with the page if it's busy, held, or special
247 	 */
248 	if ((m->hold_count != 0) ||
249 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
250 		return 0;
251 	}
252 
253 	mc[vm_pageout_page_count] = m;
254 	pageout_count = 1;
255 	page_base = vm_pageout_page_count;
256 	ib = 1;
257 	is = 1;
258 
259 	/*
260 	 * Scan object for clusterable pages.
261 	 *
262 	 * We can cluster ONLY if: ->> the page is NOT
263 	 * clean, wired, busy, held, or mapped into a
264 	 * buffer, and one of the following:
265 	 * 1) The page is inactive, or a seldom used
266 	 *    active page.
267 	 * -or-
268 	 * 2) we force the issue.
269 	 *
270 	 * During heavy mmap/modification loads the pageout
271 	 * daemon can really fragment the underlying file
272 	 * due to flushing pages out of order and not trying
273 	 * align the clusters (which leave sporatic out-of-order
274 	 * holes).  To solve this problem we do the reverse scan
275 	 * first and attempt to align our cluster, then do a
276 	 * forward scan if room remains.
277 	 */
278 more:
279 	while (ib && pageout_count < vm_pageout_page_count) {
280 		vm_page_t p;
281 
282 		if (ib > pindex) {
283 			ib = 0;
284 			break;
285 		}
286 
287 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
288 			ib = 0;
289 			break;
290 		}
291 		if (((p->queue - p->pc) == PQ_CACHE) ||
292 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
293 			ib = 0;
294 			break;
295 		}
296 		vm_page_test_dirty(p);
297 		if ((p->dirty & p->valid) == 0 ||
298 		    p->queue != PQ_INACTIVE ||
299 		    p->wire_count != 0 ||	/* may be held by buf cache */
300 		    p->hold_count != 0) {	/* may be undergoing I/O */
301 			ib = 0;
302 			break;
303 		}
304 		mc[--page_base] = p;
305 		++pageout_count;
306 		++ib;
307 		/*
308 		 * alignment boundry, stop here and switch directions.  Do
309 		 * not clear ib.
310 		 */
311 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
312 			break;
313 	}
314 
315 	while (pageout_count < vm_pageout_page_count &&
316 	    pindex + is < object->size) {
317 		vm_page_t p;
318 
319 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
320 			break;
321 		if (((p->queue - p->pc) == PQ_CACHE) ||
322 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
323 			break;
324 		}
325 		vm_page_test_dirty(p);
326 		if ((p->dirty & p->valid) == 0 ||
327 		    p->queue != PQ_INACTIVE ||
328 		    p->wire_count != 0 ||	/* may be held by buf cache */
329 		    p->hold_count != 0) {	/* may be undergoing I/O */
330 			break;
331 		}
332 		mc[page_base + pageout_count] = p;
333 		++pageout_count;
334 		++is;
335 	}
336 
337 	/*
338 	 * If we exhausted our forward scan, continue with the reverse scan
339 	 * when possible, even past a page boundry.  This catches boundry
340 	 * conditions.
341 	 */
342 	if (ib && pageout_count < vm_pageout_page_count)
343 		goto more;
344 
345 	/*
346 	 * we allow reads during pageouts...
347 	 */
348 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
349 }
350 
351 /*
352  * vm_pageout_flush() - launder the given pages
353  *
354  *	The given pages are laundered.  Note that we setup for the start of
355  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
356  *	reference count all in here rather then in the parent.  If we want
357  *	the parent to do more sophisticated things we may have to change
358  *	the ordering.
359  */
360 int
361 vm_pageout_flush(mc, count, flags)
362 	vm_page_t *mc;
363 	int count;
364 	int flags;
365 {
366 	vm_object_t object;
367 	int pageout_status[count];
368 	int numpagedout = 0;
369 	int i;
370 
371 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
372 	/*
373 	 * Initiate I/O.  Bump the vm_page_t->busy counter and
374 	 * mark the pages read-only.
375 	 *
376 	 * We do not have to fixup the clean/dirty bits here... we can
377 	 * allow the pager to do it after the I/O completes.
378 	 *
379 	 * NOTE! mc[i]->dirty may be partial or fragmented due to an
380 	 * edge case with file fragments.
381 	 */
382 	for (i = 0; i < count; i++) {
383 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
384 		vm_page_io_start(mc[i]);
385 		vm_page_protect(mc[i], VM_PROT_READ);
386 	}
387 	object = mc[0]->object;
388 	vm_page_unlock_queues();
389 	vm_object_pip_add(object, count);
390 
391 	vm_pager_put_pages(object, mc, count,
392 	    (flags | ((object == kernel_object) ? OBJPC_SYNC : 0)),
393 	    pageout_status);
394 
395 	vm_page_lock_queues();
396 	for (i = 0; i < count; i++) {
397 		vm_page_t mt = mc[i];
398 
399 		switch (pageout_status[i]) {
400 		case VM_PAGER_OK:
401 			numpagedout++;
402 			break;
403 		case VM_PAGER_PEND:
404 			numpagedout++;
405 			break;
406 		case VM_PAGER_BAD:
407 			/*
408 			 * Page outside of range of object. Right now we
409 			 * essentially lose the changes by pretending it
410 			 * worked.
411 			 */
412 			pmap_clear_modify(mt);
413 			vm_page_undirty(mt);
414 			break;
415 		case VM_PAGER_ERROR:
416 		case VM_PAGER_FAIL:
417 			/*
418 			 * If page couldn't be paged out, then reactivate the
419 			 * page so it doesn't clog the inactive list.  (We
420 			 * will try paging out it again later).
421 			 */
422 			vm_page_activate(mt);
423 			break;
424 		case VM_PAGER_AGAIN:
425 			break;
426 		}
427 
428 		/*
429 		 * If the operation is still going, leave the page busy to
430 		 * block all other accesses. Also, leave the paging in
431 		 * progress indicator set so that we don't attempt an object
432 		 * collapse.
433 		 */
434 		if (pageout_status[i] != VM_PAGER_PEND) {
435 			vm_object_pip_wakeup(object);
436 			vm_page_io_finish(mt);
437 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
438 				vm_page_protect(mt, VM_PROT_READ);
439 		}
440 	}
441 	return numpagedout;
442 }
443 
444 #if !defined(NO_SWAPPING)
445 /*
446  *	vm_pageout_object_deactivate_pages
447  *
448  *	deactivate enough pages to satisfy the inactive target
449  *	requirements or if vm_page_proc_limit is set, then
450  *	deactivate all of the pages in the object and its
451  *	backing_objects.
452  *
453  *	The object and map must be locked.
454  */
455 static void
456 vm_pageout_object_deactivate_pages(map, object, desired, map_remove_only)
457 	vm_map_t map;
458 	vm_object_t object;
459 	vm_pindex_t desired;
460 	int map_remove_only;
461 {
462 	vm_page_t p, next;
463 	int actcount, rcount, remove_mode;
464 
465 	GIANT_REQUIRED;
466 	if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS)
467 		return;
468 
469 	while (object) {
470 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
471 			return;
472 		if (object->paging_in_progress)
473 			return;
474 
475 		remove_mode = map_remove_only;
476 		if (object->shadow_count > 1)
477 			remove_mode = 1;
478 		/*
479 		 * scan the objects entire memory queue
480 		 */
481 		rcount = object->resident_page_count;
482 		p = TAILQ_FIRST(&object->memq);
483 		vm_page_lock_queues();
484 		while (p && (rcount-- > 0)) {
485 			if (pmap_resident_count(map->pmap) <= desired) {
486 				vm_page_unlock_queues();
487 				return;
488 			}
489 			next = TAILQ_NEXT(p, listq);
490 			cnt.v_pdpages++;
491 			if (p->wire_count != 0 ||
492 			    p->hold_count != 0 ||
493 			    p->busy != 0 ||
494 			    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
495 			    !pmap_page_exists_quick(vm_map_pmap(map), p)) {
496 				p = next;
497 				continue;
498 			}
499 			actcount = pmap_ts_referenced(p);
500 			if (actcount) {
501 				vm_page_flag_set(p, PG_REFERENCED);
502 			} else if (p->flags & PG_REFERENCED) {
503 				actcount = 1;
504 			}
505 			if ((p->queue != PQ_ACTIVE) &&
506 				(p->flags & PG_REFERENCED)) {
507 				vm_page_activate(p);
508 				p->act_count += actcount;
509 				vm_page_flag_clear(p, PG_REFERENCED);
510 			} else if (p->queue == PQ_ACTIVE) {
511 				if ((p->flags & PG_REFERENCED) == 0) {
512 					p->act_count -= min(p->act_count, ACT_DECLINE);
513 					if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) {
514 						vm_page_protect(p, VM_PROT_NONE);
515 						vm_page_deactivate(p);
516 					} else {
517 						vm_pageq_requeue(p);
518 					}
519 				} else {
520 					vm_page_activate(p);
521 					vm_page_flag_clear(p, PG_REFERENCED);
522 					if (p->act_count < (ACT_MAX - ACT_ADVANCE))
523 						p->act_count += ACT_ADVANCE;
524 					vm_pageq_requeue(p);
525 				}
526 			} else if (p->queue == PQ_INACTIVE) {
527 				vm_page_protect(p, VM_PROT_NONE);
528 			}
529 			p = next;
530 		}
531 		vm_page_unlock_queues();
532 		object = object->backing_object;
533 	}
534 }
535 
536 /*
537  * deactivate some number of pages in a map, try to do it fairly, but
538  * that is really hard to do.
539  */
540 static void
541 vm_pageout_map_deactivate_pages(map, desired)
542 	vm_map_t map;
543 	vm_pindex_t desired;
544 {
545 	vm_map_entry_t tmpe;
546 	vm_object_t obj, bigobj;
547 	int nothingwired;
548 
549 	GIANT_REQUIRED;
550 	if (!vm_map_trylock(map))
551 		return;
552 
553 	bigobj = NULL;
554 	nothingwired = TRUE;
555 
556 	/*
557 	 * first, search out the biggest object, and try to free pages from
558 	 * that.
559 	 */
560 	tmpe = map->header.next;
561 	while (tmpe != &map->header) {
562 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
563 			obj = tmpe->object.vm_object;
564 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
565 				((bigobj == NULL) ||
566 				 (bigobj->resident_page_count < obj->resident_page_count))) {
567 				bigobj = obj;
568 			}
569 		}
570 		if (tmpe->wired_count > 0)
571 			nothingwired = FALSE;
572 		tmpe = tmpe->next;
573 	}
574 
575 	if (bigobj)
576 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
577 
578 	/*
579 	 * Next, hunt around for other pages to deactivate.  We actually
580 	 * do this search sort of wrong -- .text first is not the best idea.
581 	 */
582 	tmpe = map->header.next;
583 	while (tmpe != &map->header) {
584 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
585 			break;
586 		if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) {
587 			obj = tmpe->object.vm_object;
588 			if (obj)
589 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
590 		}
591 		tmpe = tmpe->next;
592 	};
593 
594 	/*
595 	 * Remove all mappings if a process is swapped out, this will free page
596 	 * table pages.
597 	 */
598 	if (desired == 0 && nothingwired)
599 		pmap_remove(vm_map_pmap(map),
600 			VM_MIN_ADDRESS, VM_MAXUSER_ADDRESS);
601 	vm_map_unlock(map);
602 	return;
603 }
604 #endif		/* !defined(NO_SWAPPING) */
605 
606 /*
607  * Don't try to be fancy - being fancy can lead to VOP_LOCK's and therefore
608  * to vnode deadlocks.  We only do it for OBJT_DEFAULT and OBJT_SWAP objects
609  * which we know can be trivially freed.
610  */
611 void
612 vm_pageout_page_free(vm_page_t m) {
613 	vm_object_t object = m->object;
614 	int type = object->type;
615 
616 	GIANT_REQUIRED;
617 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
618 		vm_object_reference(object);
619 	vm_page_busy(m);
620 	vm_page_protect(m, VM_PROT_NONE);
621 	vm_page_free(m);
622 	cnt.v_dfree++;
623 	if (type == OBJT_SWAP || type == OBJT_DEFAULT)
624 		vm_object_deallocate(object);
625 }
626 
627 /*
628  *	vm_pageout_scan does the dirty work for the pageout daemon.
629  */
630 static void
631 vm_pageout_scan(int pass)
632 {
633 	vm_page_t m, next;
634 	struct vm_page marker;
635 	int save_page_shortage;
636 	int save_inactive_count;
637 	int page_shortage, maxscan, pcount;
638 	int addl_page_shortage, addl_page_shortage_init;
639 	struct proc *p, *bigproc;
640 	vm_offset_t size, bigsize;
641 	vm_object_t object;
642 	int actcount;
643 	int vnodes_skipped = 0;
644 	int maxlaunder;
645 	int s;
646 	struct thread *td;
647 
648 	GIANT_REQUIRED;
649 	/*
650 	 * Do whatever cleanup that the pmap code can.
651 	 */
652 	pmap_collect();
653 	uma_reclaim();
654 
655 	addl_page_shortage_init = vm_pageout_deficit;
656 	vm_pageout_deficit = 0;
657 
658 	/*
659 	 * Calculate the number of pages we want to either free or move
660 	 * to the cache.
661 	 */
662 	page_shortage = vm_paging_target() + addl_page_shortage_init;
663 	save_page_shortage = page_shortage;
664 	save_inactive_count = cnt.v_inactive_count;
665 
666 	/*
667 	 * Initialize our marker
668 	 */
669 	bzero(&marker, sizeof(marker));
670 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
671 	marker.queue = PQ_INACTIVE;
672 	marker.wire_count = 1;
673 
674 	/*
675 	 * Start scanning the inactive queue for pages we can move to the
676 	 * cache or free.  The scan will stop when the target is reached or
677 	 * we have scanned the entire inactive queue.  Note that m->act_count
678 	 * is not used to form decisions for the inactive queue, only for the
679 	 * active queue.
680 	 *
681 	 * maxlaunder limits the number of dirty pages we flush per scan.
682 	 * For most systems a smaller value (16 or 32) is more robust under
683 	 * extreme memory and disk pressure because any unnecessary writes
684 	 * to disk can result in extreme performance degredation.  However,
685 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
686 	 * used) will die horribly with limited laundering.  If the pageout
687 	 * daemon cannot clean enough pages in the first pass, we let it go
688 	 * all out in succeeding passes.
689 	 */
690 	if ((maxlaunder = vm_max_launder) <= 1)
691 		maxlaunder = 1;
692 	if (pass)
693 		maxlaunder = 10000;
694 rescan0:
695 	addl_page_shortage = addl_page_shortage_init;
696 	maxscan = cnt.v_inactive_count;
697 
698 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
699 	     m != NULL && maxscan-- > 0 && page_shortage > 0;
700 	     m = next) {
701 
702 		cnt.v_pdpages++;
703 
704 		if (m->queue != PQ_INACTIVE) {
705 			goto rescan0;
706 		}
707 
708 		next = TAILQ_NEXT(m, pageq);
709 
710 		/*
711 		 * skip marker pages
712 		 */
713 		if (m->flags & PG_MARKER)
714 			continue;
715 
716 		/*
717 		 * A held page may be undergoing I/O, so skip it.
718 		 */
719 		if (m->hold_count) {
720 			vm_pageq_requeue(m);
721 			addl_page_shortage++;
722 			continue;
723 		}
724 		/*
725 		 * Don't mess with busy pages, keep in the front of the
726 		 * queue, most likely are being paged out.
727 		 */
728 		if (m->busy || (m->flags & PG_BUSY)) {
729 			addl_page_shortage++;
730 			continue;
731 		}
732 
733 		/*
734 		 * If the object is not being used, we ignore previous
735 		 * references.
736 		 */
737 		if (m->object->ref_count == 0) {
738 			vm_page_flag_clear(m, PG_REFERENCED);
739 			pmap_clear_reference(m);
740 
741 		/*
742 		 * Otherwise, if the page has been referenced while in the
743 		 * inactive queue, we bump the "activation count" upwards,
744 		 * making it less likely that the page will be added back to
745 		 * the inactive queue prematurely again.  Here we check the
746 		 * page tables (or emulated bits, if any), given the upper
747 		 * level VM system not knowing anything about existing
748 		 * references.
749 		 */
750 		} else if (((m->flags & PG_REFERENCED) == 0) &&
751 			(actcount = pmap_ts_referenced(m))) {
752 			vm_page_activate(m);
753 			m->act_count += (actcount + ACT_ADVANCE);
754 			continue;
755 		}
756 
757 		/*
758 		 * If the upper level VM system knows about any page
759 		 * references, we activate the page.  We also set the
760 		 * "activation count" higher than normal so that we will less
761 		 * likely place pages back onto the inactive queue again.
762 		 */
763 		if ((m->flags & PG_REFERENCED) != 0) {
764 			vm_page_flag_clear(m, PG_REFERENCED);
765 			actcount = pmap_ts_referenced(m);
766 			vm_page_activate(m);
767 			m->act_count += (actcount + ACT_ADVANCE + 1);
768 			continue;
769 		}
770 
771 		/*
772 		 * If the upper level VM system doesn't know anything about
773 		 * the page being dirty, we have to check for it again.  As
774 		 * far as the VM code knows, any partially dirty pages are
775 		 * fully dirty.
776 		 */
777 		if (m->dirty == 0) {
778 			vm_page_test_dirty(m);
779 		} else {
780 			vm_page_dirty(m);
781 		}
782 
783 		/*
784 		 * Invalid pages can be easily freed
785 		 */
786 		if (m->valid == 0) {
787 			vm_page_lock_queues();
788 			vm_pageout_page_free(m);
789 			vm_page_unlock_queues();
790 			--page_shortage;
791 
792 		/*
793 		 * Clean pages can be placed onto the cache queue.  This
794 		 * effectively frees them.
795 		 */
796 		} else if (m->dirty == 0) {
797 			vm_page_lock_queues();
798 			vm_page_cache(m);
799 			vm_page_unlock_queues();
800 			--page_shortage;
801 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
802 			/*
803 			 * Dirty pages need to be paged out, but flushing
804 			 * a page is extremely expensive verses freeing
805 			 * a clean page.  Rather then artificially limiting
806 			 * the number of pages we can flush, we instead give
807 			 * dirty pages extra priority on the inactive queue
808 			 * by forcing them to be cycled through the queue
809 			 * twice before being flushed, after which the
810 			 * (now clean) page will cycle through once more
811 			 * before being freed.  This significantly extends
812 			 * the thrash point for a heavily loaded machine.
813 			 */
814 			vm_page_flag_set(m, PG_WINATCFLS);
815 			vm_pageq_requeue(m);
816 		} else if (maxlaunder > 0) {
817 			/*
818 			 * We always want to try to flush some dirty pages if
819 			 * we encounter them, to keep the system stable.
820 			 * Normally this number is small, but under extreme
821 			 * pressure where there are insufficient clean pages
822 			 * on the inactive queue, we may have to go all out.
823 			 */
824 			int swap_pageouts_ok;
825 			struct vnode *vp = NULL;
826 			struct mount *mp;
827 
828 			object = m->object;
829 
830 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
831 				swap_pageouts_ok = 1;
832 			} else {
833 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
834 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
835 				vm_page_count_min());
836 
837 			}
838 
839 			/*
840 			 * We don't bother paging objects that are "dead".
841 			 * Those objects are in a "rundown" state.
842 			 */
843 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
844 				vm_pageq_requeue(m);
845 				continue;
846 			}
847 
848 			/*
849 			 * The object is already known NOT to be dead.   It
850 			 * is possible for the vget() to block the whole
851 			 * pageout daemon, but the new low-memory handling
852 			 * code should prevent it.
853 			 *
854 			 * The previous code skipped locked vnodes and, worse,
855 			 * reordered pages in the queue.  This results in
856 			 * completely non-deterministic operation and, on a
857 			 * busy system, can lead to extremely non-optimal
858 			 * pageouts.  For example, it can cause clean pages
859 			 * to be freed and dirty pages to be moved to the end
860 			 * of the queue.  Since dirty pages are also moved to
861 			 * the end of the queue once-cleaned, this gives
862 			 * way too large a weighting to defering the freeing
863 			 * of dirty pages.
864 			 *
865 			 * We can't wait forever for the vnode lock, we might
866 			 * deadlock due to a vn_read() getting stuck in
867 			 * vm_wait while holding this vnode.  We skip the
868 			 * vnode if we can't get it in a reasonable amount
869 			 * of time.
870 			 */
871 			if (object->type == OBJT_VNODE) {
872 				vp = object->handle;
873 
874 				mp = NULL;
875 				if (vp->v_type == VREG)
876 					vn_start_write(vp, &mp, V_NOWAIT);
877 				if (vget(vp, LK_EXCLUSIVE|LK_NOOBJ|LK_TIMELOCK, curthread)) {
878 					++pageout_lock_miss;
879 					vn_finished_write(mp);
880 					if (object->flags & OBJ_MIGHTBEDIRTY)
881 						vnodes_skipped++;
882 					continue;
883 				}
884 
885 				/*
886 				 * The page might have been moved to another
887 				 * queue during potential blocking in vget()
888 				 * above.  The page might have been freed and
889 				 * reused for another vnode.  The object might
890 				 * have been reused for another vnode.
891 				 */
892 				if (m->queue != PQ_INACTIVE ||
893 				    m->object != object ||
894 				    object->handle != vp) {
895 					if (object->flags & OBJ_MIGHTBEDIRTY)
896 						vnodes_skipped++;
897 					vput(vp);
898 					vn_finished_write(mp);
899 					continue;
900 				}
901 
902 				/*
903 				 * The page may have been busied during the
904 				 * blocking in vput();  We don't move the
905 				 * page back onto the end of the queue so that
906 				 * statistics are more correct if we don't.
907 				 */
908 				if (m->busy || (m->flags & PG_BUSY)) {
909 					vput(vp);
910 					vn_finished_write(mp);
911 					continue;
912 				}
913 
914 				/*
915 				 * If the page has become held it might
916 				 * be undergoing I/O, so skip it
917 				 */
918 				if (m->hold_count) {
919 					vm_pageq_requeue(m);
920 					if (object->flags & OBJ_MIGHTBEDIRTY)
921 						vnodes_skipped++;
922 					vput(vp);
923 					vn_finished_write(mp);
924 					continue;
925 				}
926 			}
927 
928 			/*
929 			 * If a page is dirty, then it is either being washed
930 			 * (but not yet cleaned) or it is still in the
931 			 * laundry.  If it is still in the laundry, then we
932 			 * start the cleaning operation.
933 			 *
934 			 * This operation may cluster, invalidating the 'next'
935 			 * pointer.  To prevent an inordinate number of
936 			 * restarts we use our marker to remember our place.
937 			 *
938 			 * decrement page_shortage on success to account for
939 			 * the (future) cleaned page.  Otherwise we could wind
940 			 * up laundering or cleaning too many pages.
941 			 */
942 			vm_page_lock_queues();
943 			s = splvm();
944 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
945 			splx(s);
946 			if (vm_pageout_clean(m) != 0) {
947 				--page_shortage;
948 				--maxlaunder;
949 			}
950 			s = splvm();
951 			next = TAILQ_NEXT(&marker, pageq);
952 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
953 			splx(s);
954 			vm_page_unlock_queues();
955 			if (vp) {
956 				vput(vp);
957 				vn_finished_write(mp);
958 			}
959 		}
960 	}
961 
962 	/*
963 	 * Compute the number of pages we want to try to move from the
964 	 * active queue to the inactive queue.
965 	 */
966 	page_shortage = vm_paging_target() +
967 		cnt.v_inactive_target - cnt.v_inactive_count;
968 	page_shortage += addl_page_shortage;
969 
970 	vm_page_lock_queues();
971 	/*
972 	 * Scan the active queue for things we can deactivate. We nominally
973 	 * track the per-page activity counter and use it to locate
974 	 * deactivation candidates.
975 	 */
976 	pcount = cnt.v_active_count;
977 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
978 
979 	while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) {
980 
981 		/*
982 		 * This is a consistency check, and should likely be a panic
983 		 * or warning.
984 		 */
985 		if (m->queue != PQ_ACTIVE) {
986 			break;
987 		}
988 
989 		next = TAILQ_NEXT(m, pageq);
990 		/*
991 		 * Don't deactivate pages that are busy.
992 		 */
993 		if ((m->busy != 0) ||
994 		    (m->flags & PG_BUSY) ||
995 		    (m->hold_count != 0)) {
996 			vm_pageq_requeue(m);
997 			m = next;
998 			continue;
999 		}
1000 
1001 		/*
1002 		 * The count for pagedaemon pages is done after checking the
1003 		 * page for eligibility...
1004 		 */
1005 		cnt.v_pdpages++;
1006 
1007 		/*
1008 		 * Check to see "how much" the page has been used.
1009 		 */
1010 		actcount = 0;
1011 		if (m->object->ref_count != 0) {
1012 			if (m->flags & PG_REFERENCED) {
1013 				actcount += 1;
1014 			}
1015 			actcount += pmap_ts_referenced(m);
1016 			if (actcount) {
1017 				m->act_count += ACT_ADVANCE + actcount;
1018 				if (m->act_count > ACT_MAX)
1019 					m->act_count = ACT_MAX;
1020 			}
1021 		}
1022 
1023 		/*
1024 		 * Since we have "tested" this bit, we need to clear it now.
1025 		 */
1026 		vm_page_flag_clear(m, PG_REFERENCED);
1027 
1028 		/*
1029 		 * Only if an object is currently being used, do we use the
1030 		 * page activation count stats.
1031 		 */
1032 		if (actcount && (m->object->ref_count != 0)) {
1033 			vm_pageq_requeue(m);
1034 		} else {
1035 			m->act_count -= min(m->act_count, ACT_DECLINE);
1036 			if (vm_pageout_algorithm ||
1037 			    m->object->ref_count == 0 ||
1038 			    m->act_count == 0) {
1039 				page_shortage--;
1040 				if (m->object->ref_count == 0) {
1041 					vm_page_protect(m, VM_PROT_NONE);
1042 					if (m->dirty == 0)
1043 						vm_page_cache(m);
1044 					else
1045 						vm_page_deactivate(m);
1046 				} else {
1047 					vm_page_deactivate(m);
1048 				}
1049 			} else {
1050 				vm_pageq_requeue(m);
1051 			}
1052 		}
1053 		m = next;
1054 	}
1055 	s = splvm();
1056 
1057 	/*
1058 	 * We try to maintain some *really* free pages, this allows interrupt
1059 	 * code to be guaranteed space.  Since both cache and free queues
1060 	 * are considered basically 'free', moving pages from cache to free
1061 	 * does not effect other calculations.
1062 	 */
1063 	while (cnt.v_free_count < cnt.v_free_reserved) {
1064 		static int cache_rover = 0;
1065 		m = vm_pageq_find(PQ_CACHE, cache_rover, FALSE);
1066 		if (!m)
1067 			break;
1068 		if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1069 		    m->busy ||
1070 		    m->hold_count ||
1071 		    m->wire_count) {
1072 #ifdef INVARIANTS
1073 			printf("Warning: busy page %p found in cache\n", m);
1074 #endif
1075 			vm_page_deactivate(m);
1076 			continue;
1077 		}
1078 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1079 		vm_pageout_page_free(m);
1080 	}
1081 	splx(s);
1082 	vm_page_unlock_queues();
1083 #if !defined(NO_SWAPPING)
1084 	/*
1085 	 * Idle process swapout -- run once per second.
1086 	 */
1087 	if (vm_swap_idle_enabled) {
1088 		static long lsec;
1089 		if (time_second != lsec) {
1090 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1091 			vm_req_vmdaemon();
1092 			lsec = time_second;
1093 		}
1094 	}
1095 #endif
1096 
1097 	/*
1098 	 * If we didn't get enough free pages, and we have skipped a vnode
1099 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1100 	 * if we did not get enough free pages.
1101 	 */
1102 	if (vm_paging_target() > 0) {
1103 		if (vnodes_skipped && vm_page_count_min())
1104 			(void) speedup_syncer();
1105 #if !defined(NO_SWAPPING)
1106 		if (vm_swap_enabled && vm_page_count_target()) {
1107 			vm_req_vmdaemon();
1108 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1109 		}
1110 #endif
1111 	}
1112 
1113 	/*
1114 	 * If we are out of swap and were not able to reach our paging
1115 	 * target, kill the largest process.
1116 	 *
1117 	 * We keep the process bigproc locked once we find it to keep anyone
1118 	 * from messing with it; however, there is a possibility of
1119 	 * deadlock if process B is bigproc and one of it's child processes
1120 	 * attempts to propagate a signal to B while we are waiting for A's
1121 	 * lock while walking this list.  To avoid this, we don't block on
1122 	 * the process lock but just skip a process if it is already locked.
1123 	 */
1124 	if ((vm_swap_size < 64 && vm_page_count_min()) ||
1125 	    (swap_pager_full && vm_paging_target() > 0)) {
1126 #if 0
1127 	if ((vm_swap_size < 64 || swap_pager_full) && vm_page_count_min()) {
1128 #endif
1129 		bigproc = NULL;
1130 		bigsize = 0;
1131 		sx_slock(&allproc_lock);
1132 		FOREACH_PROC_IN_SYSTEM(p) {
1133 			int breakout;
1134 			/*
1135 			 * If this process is already locked, skip it.
1136 			 */
1137 			if (PROC_TRYLOCK(p) == 0)
1138 				continue;
1139 			/*
1140 			 * if this is a system process, skip it
1141 			 */
1142 			if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1143 			    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1144 				PROC_UNLOCK(p);
1145 				continue;
1146 			}
1147 			/*
1148 			 * if the process is in a non-running type state,
1149 			 * don't touch it. Check all the threads individually.
1150 			 */
1151 			mtx_lock_spin(&sched_lock);
1152 			breakout = 0;
1153 			FOREACH_THREAD_IN_PROC(p, td) {
1154 				if (td->td_state != TDS_RUNQ &&
1155 				    td->td_state != TDS_RUNNING &&
1156 				    td->td_state != TDS_SLP) {
1157 					breakout = 1;
1158 					break;
1159 				}
1160 			}
1161 			if (breakout) {
1162 				mtx_unlock_spin(&sched_lock);
1163 				PROC_UNLOCK(p);
1164 				continue;
1165 			}
1166 			mtx_unlock_spin(&sched_lock);
1167 			/*
1168 			 * get the process size
1169 			 */
1170 			size = vmspace_resident_count(p->p_vmspace) +
1171 				vmspace_swap_count(p->p_vmspace);
1172 			/*
1173 			 * if the this process is bigger than the biggest one
1174 			 * remember it.
1175 			 */
1176 			if (size > bigsize) {
1177 				if (bigproc != NULL)
1178 					PROC_UNLOCK(bigproc);
1179 				bigproc = p;
1180 				bigsize = size;
1181 			} else
1182 				PROC_UNLOCK(p);
1183 		}
1184 		sx_sunlock(&allproc_lock);
1185 		if (bigproc != NULL) {
1186 			struct ksegrp *kg;
1187 			killproc(bigproc, "out of swap space");
1188 			mtx_lock_spin(&sched_lock);
1189 			FOREACH_KSEGRP_IN_PROC(bigproc, kg) {
1190 				kg->kg_estcpu = 0;
1191 				kg->kg_nice = PRIO_MIN; /* XXXKSE ??? */
1192 				resetpriority(kg);
1193 			}
1194 			mtx_unlock_spin(&sched_lock);
1195 			PROC_UNLOCK(bigproc);
1196 			wakeup(&cnt.v_free_count);
1197 		}
1198 	}
1199 }
1200 
1201 /*
1202  * This routine tries to maintain the pseudo LRU active queue,
1203  * so that during long periods of time where there is no paging,
1204  * that some statistic accumulation still occurs.  This code
1205  * helps the situation where paging just starts to occur.
1206  */
1207 static void
1208 vm_pageout_page_stats()
1209 {
1210 	vm_page_t m,next;
1211 	int pcount,tpcount;		/* Number of pages to check */
1212 	static int fullintervalcount = 0;
1213 	int page_shortage;
1214 	int s0;
1215 
1216 	page_shortage =
1217 	    (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) -
1218 	    (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count);
1219 
1220 	if (page_shortage <= 0)
1221 		return;
1222 
1223 	s0 = splvm();
1224 	vm_page_lock_queues();
1225 	pcount = cnt.v_active_count;
1226 	fullintervalcount += vm_pageout_stats_interval;
1227 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1228 		tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count;
1229 		if (pcount > tpcount)
1230 			pcount = tpcount;
1231 	} else {
1232 		fullintervalcount = 0;
1233 	}
1234 
1235 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1236 	while ((m != NULL) && (pcount-- > 0)) {
1237 		int actcount;
1238 
1239 		if (m->queue != PQ_ACTIVE) {
1240 			break;
1241 		}
1242 
1243 		next = TAILQ_NEXT(m, pageq);
1244 		/*
1245 		 * Don't deactivate pages that are busy.
1246 		 */
1247 		if ((m->busy != 0) ||
1248 		    (m->flags & PG_BUSY) ||
1249 		    (m->hold_count != 0)) {
1250 			vm_pageq_requeue(m);
1251 			m = next;
1252 			continue;
1253 		}
1254 
1255 		actcount = 0;
1256 		if (m->flags & PG_REFERENCED) {
1257 			vm_page_flag_clear(m, PG_REFERENCED);
1258 			actcount += 1;
1259 		}
1260 
1261 		actcount += pmap_ts_referenced(m);
1262 		if (actcount) {
1263 			m->act_count += ACT_ADVANCE + actcount;
1264 			if (m->act_count > ACT_MAX)
1265 				m->act_count = ACT_MAX;
1266 			vm_pageq_requeue(m);
1267 		} else {
1268 			if (m->act_count == 0) {
1269 				/*
1270 				 * We turn off page access, so that we have
1271 				 * more accurate RSS stats.  We don't do this
1272 				 * in the normal page deactivation when the
1273 				 * system is loaded VM wise, because the
1274 				 * cost of the large number of page protect
1275 				 * operations would be higher than the value
1276 				 * of doing the operation.
1277 				 */
1278 				vm_page_protect(m, VM_PROT_NONE);
1279 				vm_page_deactivate(m);
1280 			} else {
1281 				m->act_count -= min(m->act_count, ACT_DECLINE);
1282 				vm_pageq_requeue(m);
1283 			}
1284 		}
1285 
1286 		m = next;
1287 	}
1288 	vm_page_unlock_queues();
1289 	splx(s0);
1290 }
1291 
1292 static int
1293 vm_pageout_free_page_calc(count)
1294 vm_size_t count;
1295 {
1296 	if (count < cnt.v_page_count)
1297 		 return 0;
1298 	/*
1299 	 * free_reserved needs to include enough for the largest swap pager
1300 	 * structures plus enough for any pv_entry structs when paging.
1301 	 */
1302 	if (cnt.v_page_count > 1024)
1303 		cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200;
1304 	else
1305 		cnt.v_free_min = 4;
1306 	cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1307 		cnt.v_interrupt_free_min;
1308 	cnt.v_free_reserved = vm_pageout_page_count +
1309 		cnt.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1310 	cnt.v_free_severe = cnt.v_free_min / 2;
1311 	cnt.v_free_min += cnt.v_free_reserved;
1312 	cnt.v_free_severe += cnt.v_free_reserved;
1313 	return 1;
1314 }
1315 
1316 /*
1317  *	vm_pageout is the high level pageout daemon.
1318  */
1319 static void
1320 vm_pageout()
1321 {
1322 	int pass;
1323 
1324 	mtx_lock(&Giant);
1325 
1326 	/*
1327 	 * Initialize some paging parameters.
1328 	 */
1329 	cnt.v_interrupt_free_min = 2;
1330 	if (cnt.v_page_count < 2000)
1331 		vm_pageout_page_count = 8;
1332 
1333 	vm_pageout_free_page_calc(cnt.v_page_count);
1334 	/*
1335 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1336 	 * that these are more a measure of the VM cache queue hysteresis
1337 	 * then the VM free queue.  Specifically, v_free_target is the
1338 	 * high water mark (free+cache pages).
1339 	 *
1340 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1341 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1342 	 * be big enough to handle memory needs while the pageout daemon
1343 	 * is signalled and run to free more pages.
1344 	 */
1345 	if (cnt.v_free_count > 6144)
1346 		cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved;
1347 	else
1348 		cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved;
1349 
1350 	if (cnt.v_free_count > 2048) {
1351 		cnt.v_cache_min = cnt.v_free_target;
1352 		cnt.v_cache_max = 2 * cnt.v_cache_min;
1353 		cnt.v_inactive_target = (3 * cnt.v_free_target) / 2;
1354 	} else {
1355 		cnt.v_cache_min = 0;
1356 		cnt.v_cache_max = 0;
1357 		cnt.v_inactive_target = cnt.v_free_count / 4;
1358 	}
1359 	if (cnt.v_inactive_target > cnt.v_free_count / 3)
1360 		cnt.v_inactive_target = cnt.v_free_count / 3;
1361 
1362 	/* XXX does not really belong here */
1363 	if (vm_page_max_wired == 0)
1364 		vm_page_max_wired = cnt.v_free_count / 3;
1365 
1366 	if (vm_pageout_stats_max == 0)
1367 		vm_pageout_stats_max = cnt.v_free_target;
1368 
1369 	/*
1370 	 * Set interval in seconds for stats scan.
1371 	 */
1372 	if (vm_pageout_stats_interval == 0)
1373 		vm_pageout_stats_interval = 5;
1374 	if (vm_pageout_full_stats_interval == 0)
1375 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1376 
1377 	/*
1378 	 * Set maximum free per pass
1379 	 */
1380 	if (vm_pageout_stats_free_max == 0)
1381 		vm_pageout_stats_free_max = 5;
1382 
1383 	swap_pager_swap_init();
1384 	pass = 0;
1385 	/*
1386 	 * The pageout daemon is never done, so loop forever.
1387 	 */
1388 	while (TRUE) {
1389 		int error;
1390 		int s = splvm();
1391 
1392 		/*
1393 		 * If we have enough free memory, wakeup waiters.  Do
1394 		 * not clear vm_pages_needed until we reach our target,
1395 		 * otherwise we may be woken up over and over again and
1396 		 * waste a lot of cpu.
1397 		 */
1398 		if (vm_pages_needed && !vm_page_count_min()) {
1399 			if (vm_paging_needed() <= 0)
1400 				vm_pages_needed = 0;
1401 			wakeup(&cnt.v_free_count);
1402 		}
1403 		if (vm_pages_needed) {
1404 			/*
1405 			 * Still not done, take a second pass without waiting
1406 			 * (unlimited dirty cleaning), otherwise sleep a bit
1407 			 * and try again.
1408 			 */
1409 			++pass;
1410 			if (pass > 1)
1411 				tsleep(&vm_pages_needed, PVM,
1412 				       "psleep", hz/2);
1413 		} else {
1414 			/*
1415 			 * Good enough, sleep & handle stats.  Prime the pass
1416 			 * for the next run.
1417 			 */
1418 			if (pass > 1)
1419 				pass = 1;
1420 			else
1421 				pass = 0;
1422 			error = tsleep(&vm_pages_needed, PVM,
1423 				    "psleep", vm_pageout_stats_interval * hz);
1424 			if (error && !vm_pages_needed) {
1425 				splx(s);
1426 				pass = 0;
1427 				vm_pageout_page_stats();
1428 				continue;
1429 			}
1430 		}
1431 
1432 		if (vm_pages_needed)
1433 			cnt.v_pdwakeups++;
1434 		splx(s);
1435 		vm_pageout_scan(pass);
1436 		vm_pageout_deficit = 0;
1437 	}
1438 }
1439 
1440 void
1441 pagedaemon_wakeup()
1442 {
1443 	if (!vm_pages_needed && curthread->td_proc != pageproc) {
1444 		vm_pages_needed++;
1445 		wakeup(&vm_pages_needed);
1446 	}
1447 }
1448 
1449 #if !defined(NO_SWAPPING)
1450 static void
1451 vm_req_vmdaemon()
1452 {
1453 	static int lastrun = 0;
1454 
1455 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1456 		wakeup(&vm_daemon_needed);
1457 		lastrun = ticks;
1458 	}
1459 }
1460 
1461 static void
1462 vm_daemon()
1463 {
1464 	struct proc *p;
1465 	int breakout;
1466 	struct thread *td;
1467 
1468 	mtx_lock(&Giant);
1469 	while (TRUE) {
1470 		tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0);
1471 		if (vm_pageout_req_swapout) {
1472 			swapout_procs(vm_pageout_req_swapout);
1473 			vm_pageout_req_swapout = 0;
1474 		}
1475 		/*
1476 		 * scan the processes for exceeding their rlimits or if
1477 		 * process is swapped out -- deactivate pages
1478 		 */
1479 		sx_slock(&allproc_lock);
1480 		LIST_FOREACH(p, &allproc, p_list) {
1481 			vm_pindex_t limit, size;
1482 
1483 			/*
1484 			 * if this is a system process or if we have already
1485 			 * looked at this process, skip it.
1486 			 */
1487 			if (p->p_flag & (P_SYSTEM | P_WEXIT)) {
1488 				continue;
1489 			}
1490 			/*
1491 			 * if the process is in a non-running type state,
1492 			 * don't touch it.
1493 			 */
1494 			mtx_lock_spin(&sched_lock);
1495 			breakout = 0;
1496 			FOREACH_THREAD_IN_PROC(p, td) {
1497 				if (td->td_state != TDS_RUNQ &&
1498 				    td->td_state != TDS_RUNNING &&
1499 				    td->td_state != TDS_SLP) {
1500 					breakout = 1;
1501 					break;
1502 				}
1503 			}
1504 			if (breakout) {
1505 				mtx_unlock_spin(&sched_lock);
1506 				continue;
1507 			}
1508 			/*
1509 			 * get a limit
1510 			 */
1511 			limit = OFF_TO_IDX(
1512 			    qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1513 				p->p_rlimit[RLIMIT_RSS].rlim_max));
1514 
1515 			/*
1516 			 * let processes that are swapped out really be
1517 			 * swapped out set the limit to nothing (will force a
1518 			 * swap-out.)
1519 			 */
1520 			if ((p->p_sflag & PS_INMEM) == 0)
1521 				limit = 0;	/* XXX */
1522 			mtx_unlock_spin(&sched_lock);
1523 
1524 			size = vmspace_resident_count(p->p_vmspace);
1525 			if (limit >= 0 && size >= limit) {
1526 				vm_pageout_map_deactivate_pages(
1527 				    &p->p_vmspace->vm_map, limit);
1528 			}
1529 		}
1530 		sx_sunlock(&allproc_lock);
1531 	}
1532 }
1533 #endif			/* !defined(NO_SWAPPING) */
1534