xref: /dflybsd-src/sys/vm/vm_pageout.c (revision dff3cc789c2674bc2f429018d233f39a8e366dea)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1991 Regents of the University of California.
5  * All rights reserved.
6  * Copyright (c) 1994 John S. Dyson
7  * All rights reserved.
8  * Copyright (c) 1994 David Greenman
9  * All rights reserved.
10  *
11  * This code is derived from software contributed to Berkeley by
12  * The Mach Operating System project at Carnegie-Mellon University.
13  *
14  * Redistribution and use in source and binary forms, with or without
15  * modification, are permitted provided that the following conditions
16  * are met:
17  * 1. Redistributions of source code must retain the above copyright
18  *    notice, this list of conditions and the following disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
39  *
40  *
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  *
66  * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
67  */
68 
69 /*
70  *	The proverbial page-out daemon.
71  */
72 
73 #include "opt_vm.h"
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/kernel.h>
77 #include <sys/proc.h>
78 #include <sys/kthread.h>
79 #include <sys/resourcevar.h>
80 #include <sys/signalvar.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sysctl.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <sys/lock.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_extern.h>
95 
96 #include <sys/thread2.h>
97 #include <vm/vm_page2.h>
98 
99 /*
100  * System initialization
101  */
102 
103 /* the kernel process "vm_pageout"*/
104 static int vm_pageout_clean (vm_page_t);
105 static int vm_pageout_scan (int pass);
106 static int vm_pageout_free_page_calc (vm_size_t count);
107 struct thread *pagethread;
108 
109 #if !defined(NO_SWAPPING)
110 /* the kernel process "vm_daemon"*/
111 static void vm_daemon (void);
112 static struct	thread *vmthread;
113 
114 static struct kproc_desc vm_kp = {
115 	"vmdaemon",
116 	vm_daemon,
117 	&vmthread
118 };
119 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp)
120 #endif
121 
122 
123 int vm_pages_needed=0;		/* Event on which pageout daemon sleeps */
124 int vm_pageout_deficit=0;	/* Estimated number of pages deficit */
125 int vm_pageout_pages_needed=0;	/* flag saying that the pageout daemon needs pages */
126 
127 #if !defined(NO_SWAPPING)
128 static int vm_pageout_req_swapout;	/* XXX */
129 static int vm_daemon_needed;
130 #endif
131 static int vm_max_launder = 32;
132 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
133 static int vm_pageout_full_stats_interval = 0;
134 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
135 static int defer_swap_pageouts=0;
136 static int disable_swap_pageouts=0;
137 
138 #if defined(NO_SWAPPING)
139 static int vm_swap_enabled=0;
140 static int vm_swap_idle_enabled=0;
141 #else
142 static int vm_swap_enabled=1;
143 static int vm_swap_idle_enabled=0;
144 #endif
145 
146 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm,
147 	CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt");
148 
149 SYSCTL_INT(_vm, OID_AUTO, max_launder,
150 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
151 
152 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
153 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
154 
155 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
156 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
157 
158 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
159 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
160 
161 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
162 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
163 
164 #if defined(NO_SWAPPING)
165 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
166 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
167 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
168 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
169 #else
170 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
171 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
172 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
173 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
174 #endif
175 
176 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
177 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
178 
179 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
180 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
181 
182 static int pageout_lock_miss;
183 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
184 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
185 
186 int vm_load;
187 SYSCTL_INT(_vm, OID_AUTO, vm_load,
188 	CTLFLAG_RD, &vm_load, 0, "load on the VM system");
189 int vm_load_enable = 1;
190 SYSCTL_INT(_vm, OID_AUTO, vm_load_enable,
191 	CTLFLAG_RW, &vm_load_enable, 0, "enable vm_load rate limiting");
192 #ifdef INVARIANTS
193 int vm_load_debug;
194 SYSCTL_INT(_vm, OID_AUTO, vm_load_debug,
195 	CTLFLAG_RW, &vm_load_debug, 0, "debug vm_load");
196 #endif
197 
198 #define VM_PAGEOUT_PAGE_COUNT 16
199 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT;
200 
201 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
202 
203 #if !defined(NO_SWAPPING)
204 typedef void freeer_fcn_t (vm_map_t, vm_object_t, vm_pindex_t, int);
205 static void vm_pageout_map_deactivate_pages (vm_map_t, vm_pindex_t);
206 static freeer_fcn_t vm_pageout_object_deactivate_pages;
207 static void vm_req_vmdaemon (void);
208 #endif
209 static void vm_pageout_page_stats(void);
210 
211 /*
212  * Update vm_load to slow down faulting processes.
213  *
214  * SMP races ok.
215  * No requirements.
216  */
217 void
218 vm_fault_ratecheck(void)
219 {
220 	if (vm_pages_needed) {
221 		if (vm_load < 1000)
222 			++vm_load;
223 	} else {
224 		if (vm_load > 0)
225 			--vm_load;
226 	}
227 }
228 
229 /*
230  * vm_pageout_clean:
231  *
232  * Clean the page and remove it from the laundry.  The page must not be
233  * busy on-call.
234  *
235  * We set the busy bit to cause potential page faults on this page to
236  * block.  Note the careful timing, however, the busy bit isn't set till
237  * late and we cannot do anything that will mess with the page.
238  *
239  * The caller must hold vm_token.
240  */
241 static int
242 vm_pageout_clean(vm_page_t m)
243 {
244 	vm_object_t object;
245 	vm_page_t mc[2*vm_pageout_page_count];
246 	int pageout_count;
247 	int ib, is, page_base;
248 	vm_pindex_t pindex = m->pindex;
249 
250 	object = m->object;
251 
252 	/*
253 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
254 	 * with the new swapper, but we could have serious problems paging
255 	 * out other object types if there is insufficient memory.
256 	 *
257 	 * Unfortunately, checking free memory here is far too late, so the
258 	 * check has been moved up a procedural level.
259 	 */
260 
261 	/*
262 	 * Don't mess with the page if it's busy, held, or special
263 	 */
264 	if ((m->hold_count != 0) ||
265 	    ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) {
266 		return 0;
267 	}
268 
269 	mc[vm_pageout_page_count] = m;
270 	pageout_count = 1;
271 	page_base = vm_pageout_page_count;
272 	ib = 1;
273 	is = 1;
274 
275 	/*
276 	 * Scan object for clusterable pages.
277 	 *
278 	 * We can cluster ONLY if: ->> the page is NOT
279 	 * clean, wired, busy, held, or mapped into a
280 	 * buffer, and one of the following:
281 	 * 1) The page is inactive, or a seldom used
282 	 *    active page.
283 	 * -or-
284 	 * 2) we force the issue.
285 	 *
286 	 * During heavy mmap/modification loads the pageout
287 	 * daemon can really fragment the underlying file
288 	 * due to flushing pages out of order and not trying
289 	 * align the clusters (which leave sporatic out-of-order
290 	 * holes).  To solve this problem we do the reverse scan
291 	 * first and attempt to align our cluster, then do a
292 	 * forward scan if room remains.
293 	 */
294 
295 	vm_object_hold(object);
296 more:
297 	while (ib && pageout_count < vm_pageout_page_count) {
298 		vm_page_t p;
299 
300 		if (ib > pindex) {
301 			ib = 0;
302 			break;
303 		}
304 
305 		if ((p = vm_page_lookup(object, pindex - ib)) == NULL) {
306 			ib = 0;
307 			break;
308 		}
309 		if (((p->queue - p->pc) == PQ_CACHE) ||
310 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
311 			ib = 0;
312 			break;
313 		}
314 		vm_page_test_dirty(p);
315 		if ((p->dirty & p->valid) == 0 ||
316 		    p->queue != PQ_INACTIVE ||
317 		    p->wire_count != 0 ||	/* may be held by buf cache */
318 		    p->hold_count != 0) {	/* may be undergoing I/O */
319 			ib = 0;
320 			break;
321 		}
322 		mc[--page_base] = p;
323 		++pageout_count;
324 		++ib;
325 		/*
326 		 * alignment boundry, stop here and switch directions.  Do
327 		 * not clear ib.
328 		 */
329 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
330 			break;
331 	}
332 
333 	while (pageout_count < vm_pageout_page_count &&
334 	    pindex + is < object->size) {
335 		vm_page_t p;
336 
337 		if ((p = vm_page_lookup(object, pindex + is)) == NULL)
338 			break;
339 		if (((p->queue - p->pc) == PQ_CACHE) ||
340 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
341 			break;
342 		}
343 		vm_page_test_dirty(p);
344 		if ((p->dirty & p->valid) == 0 ||
345 		    p->queue != PQ_INACTIVE ||
346 		    p->wire_count != 0 ||	/* may be held by buf cache */
347 		    p->hold_count != 0) {	/* may be undergoing I/O */
348 			break;
349 		}
350 		mc[page_base + pageout_count] = p;
351 		++pageout_count;
352 		++is;
353 	}
354 
355 	/*
356 	 * If we exhausted our forward scan, continue with the reverse scan
357 	 * when possible, even past a page boundry.  This catches boundry
358 	 * conditions.
359 	 */
360 	if (ib && pageout_count < vm_pageout_page_count)
361 		goto more;
362 
363 	vm_object_drop(object);
364 
365 	/*
366 	 * we allow reads during pageouts...
367 	 */
368 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
369 }
370 
371 /*
372  * vm_pageout_flush() - launder the given pages
373  *
374  *	The given pages are laundered.  Note that we setup for the start of
375  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
376  *	reference count all in here rather then in the parent.  If we want
377  *	the parent to do more sophisticated things we may have to change
378  *	the ordering.
379  *
380  * The caller must hold vm_token.
381  */
382 int
383 vm_pageout_flush(vm_page_t *mc, int count, int flags)
384 {
385 	vm_object_t object;
386 	int pageout_status[count];
387 	int numpagedout = 0;
388 	int i;
389 
390 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
391 
392 	/*
393 	 * Initiate I/O.  Bump the vm_page_t->busy counter.
394 	 */
395 	for (i = 0; i < count; i++) {
396 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, ("vm_pageout_flush page %p index %d/%d: partially invalid page", mc[i], i, count));
397 		vm_page_io_start(mc[i]);
398 	}
399 
400 	/*
401 	 * We must make the pages read-only.  This will also force the
402 	 * modified bit in the related pmaps to be cleared.  The pager
403 	 * cannot clear the bit for us since the I/O completion code
404 	 * typically runs from an interrupt.  The act of making the page
405 	 * read-only handles the case for us.
406 	 */
407 	for (i = 0; i < count; i++) {
408 		vm_page_protect(mc[i], VM_PROT_READ);
409 	}
410 
411 	object = mc[0]->object;
412 	vm_object_pip_add(object, count);
413 
414 	vm_pager_put_pages(object, mc, count,
415 	    (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
416 	    pageout_status);
417 
418 	for (i = 0; i < count; i++) {
419 		vm_page_t mt = mc[i];
420 
421 		switch (pageout_status[i]) {
422 		case VM_PAGER_OK:
423 			numpagedout++;
424 			break;
425 		case VM_PAGER_PEND:
426 			numpagedout++;
427 			break;
428 		case VM_PAGER_BAD:
429 			/*
430 			 * Page outside of range of object. Right now we
431 			 * essentially lose the changes by pretending it
432 			 * worked.
433 			 */
434 			pmap_clear_modify(mt);
435 			vm_page_undirty(mt);
436 			break;
437 		case VM_PAGER_ERROR:
438 		case VM_PAGER_FAIL:
439 			/*
440 			 * A page typically cannot be paged out when we
441 			 * have run out of swap.  We leave the page
442 			 * marked inactive and will try to page it out
443 			 * again later.
444 			 *
445 			 * Starvation of the active page list is used to
446 			 * determine when the system is massively memory
447 			 * starved.
448 			 */
449 			break;
450 		case VM_PAGER_AGAIN:
451 			break;
452 		}
453 
454 		/*
455 		 * If the operation is still going, leave the page busy to
456 		 * block all other accesses. Also, leave the paging in
457 		 * progress indicator set so that we don't attempt an object
458 		 * collapse.
459 		 *
460 		 * For any pages which have completed synchronously,
461 		 * deactivate the page if we are under a severe deficit.
462 		 * Do not try to enter them into the cache, though, they
463 		 * might still be read-heavy.
464 		 */
465 		if (pageout_status[i] != VM_PAGER_PEND) {
466 			if (vm_page_count_severe())
467 				vm_page_deactivate(mt);
468 #if 0
469 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
470 				vm_page_protect(mt, VM_PROT_READ);
471 #endif
472 			vm_page_io_finish(mt);
473 			vm_object_pip_wakeup(object);
474 		}
475 	}
476 	return numpagedout;
477 }
478 
479 #if !defined(NO_SWAPPING)
480 /*
481  *	vm_pageout_object_deactivate_pages
482  *
483  *	deactivate enough pages to satisfy the inactive target
484  *	requirements or if vm_page_proc_limit is set, then
485  *	deactivate all of the pages in the object and its
486  *	backing_objects.
487  *
488  * The map must be locked.
489  * The caller must hold vm_token.
490  * The caller must hold the vm_object.
491  */
492 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
493 
494 static void
495 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
496 				   vm_pindex_t desired, int map_remove_only)
497 {
498 	struct rb_vm_page_scan_info info;
499 	vm_object_t tmp;
500 	int remove_mode;
501 
502 	while (object) {
503 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
504 			return;
505 
506 		vm_object_hold(object);
507 
508 		if (object->type == OBJT_DEVICE || object->type == OBJT_PHYS) {
509 			vm_object_drop(object);
510 			return;
511 		}
512 		if (object->paging_in_progress) {
513 			vm_object_drop(object);
514 			return;
515 		}
516 
517 		remove_mode = map_remove_only;
518 		if (object->shadow_count > 1)
519 			remove_mode = 1;
520 
521 		/*
522 		 * scan the objects entire memory queue.  We hold the
523 		 * object's token so the scan should not race anything.
524 		 */
525 		info.limit = remove_mode;
526 		info.map = map;
527 		info.desired = desired;
528 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
529 				vm_pageout_object_deactivate_pages_callback,
530 				&info
531 		);
532 		tmp = object->backing_object;
533 		vm_object_drop(object);
534 		object = tmp;
535 	}
536 }
537 
538 /*
539  * The caller must hold vm_token.
540  * The caller must hold the vm_object.
541  */
542 static int
543 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
544 {
545 	struct rb_vm_page_scan_info *info = data;
546 	int actcount;
547 
548 	if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
549 		return(-1);
550 	}
551 	mycpu->gd_cnt.v_pdpages++;
552 	if (p->wire_count != 0 || p->hold_count != 0 || p->busy != 0 ||
553 	    (p->flags & (PG_BUSY|PG_UNMANAGED)) ||
554 	    !pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
555 		return(0);
556 	}
557 
558 	actcount = pmap_ts_referenced(p);
559 	if (actcount) {
560 		vm_page_flag_set(p, PG_REFERENCED);
561 	} else if (p->flags & PG_REFERENCED) {
562 		actcount = 1;
563 	}
564 
565 	if ((p->queue != PQ_ACTIVE) &&
566 		(p->flags & PG_REFERENCED)) {
567 		vm_page_activate(p);
568 		p->act_count += actcount;
569 		vm_page_flag_clear(p, PG_REFERENCED);
570 	} else if (p->queue == PQ_ACTIVE) {
571 		if ((p->flags & PG_REFERENCED) == 0) {
572 			p->act_count -= min(p->act_count, ACT_DECLINE);
573 			if (!info->limit && (vm_pageout_algorithm || (p->act_count == 0))) {
574 				vm_page_busy(p);
575 				vm_page_protect(p, VM_PROT_NONE);
576 				vm_page_deactivate(p);
577 				vm_page_wakeup(p);
578 			} else {
579 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
580 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
581 			}
582 		} else {
583 			vm_page_activate(p);
584 			vm_page_flag_clear(p, PG_REFERENCED);
585 			if (p->act_count < (ACT_MAX - ACT_ADVANCE))
586 				p->act_count += ACT_ADVANCE;
587 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
588 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, p, pageq);
589 		}
590 	} else if (p->queue == PQ_INACTIVE) {
591 		vm_page_busy(p);
592 		vm_page_protect(p, VM_PROT_NONE);
593 		vm_page_wakeup(p);
594 	}
595 	return(0);
596 }
597 
598 /*
599  * Deactivate some number of pages in a map, try to do it fairly, but
600  * that is really hard to do.
601  *
602  * The caller must hold vm_token.
603  */
604 static void
605 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
606 {
607 	vm_map_entry_t tmpe;
608 	vm_object_t obj, bigobj;
609 	int nothingwired;
610 
611 	if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
612 		return;
613 	}
614 
615 	bigobj = NULL;
616 	nothingwired = TRUE;
617 
618 	/*
619 	 * first, search out the biggest object, and try to free pages from
620 	 * that.
621 	 */
622 	tmpe = map->header.next;
623 	while (tmpe != &map->header) {
624 		switch(tmpe->maptype) {
625 		case VM_MAPTYPE_NORMAL:
626 		case VM_MAPTYPE_VPAGETABLE:
627 			obj = tmpe->object.vm_object;
628 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
629 				((bigobj == NULL) ||
630 				 (bigobj->resident_page_count < obj->resident_page_count))) {
631 				bigobj = obj;
632 			}
633 			break;
634 		default:
635 			break;
636 		}
637 		if (tmpe->wired_count > 0)
638 			nothingwired = FALSE;
639 		tmpe = tmpe->next;
640 	}
641 
642 	if (bigobj)
643 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
644 
645 	/*
646 	 * Next, hunt around for other pages to deactivate.  We actually
647 	 * do this search sort of wrong -- .text first is not the best idea.
648 	 */
649 	tmpe = map->header.next;
650 	while (tmpe != &map->header) {
651 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
652 			break;
653 		switch(tmpe->maptype) {
654 		case VM_MAPTYPE_NORMAL:
655 		case VM_MAPTYPE_VPAGETABLE:
656 			obj = tmpe->object.vm_object;
657 			if (obj)
658 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
659 			break;
660 		default:
661 			break;
662 		}
663 		tmpe = tmpe->next;
664 	};
665 
666 	/*
667 	 * Remove all mappings if a process is swapped out, this will free page
668 	 * table pages.
669 	 */
670 	if (desired == 0 && nothingwired)
671 		pmap_remove(vm_map_pmap(map),
672 			    VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
673 	vm_map_unlock(map);
674 }
675 #endif
676 
677 /*
678  * Called when the pageout scan wants to free a page.  We no longer
679  * try to cycle the vm_object here with a reference & dealloc, which can
680  * cause a non-trivial object collapse in a critical path.
681  *
682  * It is unclear why we cycled the ref_count in the past, perhaps to try
683  * to optimize shadow chain collapses but I don't quite see why it would
684  * be necessary.  An OBJ_DEAD object should terminate any and all vm_pages
685  * synchronously and not have to be kicked-start.
686  *
687  * The caller must hold vm_token.
688  */
689 static void
690 vm_pageout_page_free(vm_page_t m)
691 {
692 	vm_page_busy(m);
693 	vm_page_protect(m, VM_PROT_NONE);
694 	vm_page_free(m);
695 }
696 
697 /*
698  * vm_pageout_scan does the dirty work for the pageout daemon.
699  */
700 struct vm_pageout_scan_info {
701 	struct proc *bigproc;
702 	vm_offset_t bigsize;
703 };
704 
705 static int vm_pageout_scan_callback(struct proc *p, void *data);
706 
707 /*
708  * The caller must hold vm_token.
709  */
710 static int
711 vm_pageout_scan(int pass)
712 {
713 	struct vm_pageout_scan_info info;
714 	vm_page_t m, next;
715 	struct vm_page marker;
716 	struct vnode *vpfailed;		/* warning, allowed to be stale */
717 	int maxscan, pcount;
718 	int recycle_count;
719 	int inactive_shortage, active_shortage;
720 	int inactive_original_shortage;
721 	vm_object_t object;
722 	int actcount;
723 	int vnodes_skipped = 0;
724 	int maxlaunder;
725 
726 	/*
727 	 * Do whatever cleanup that the pmap code can.
728 	 */
729 	pmap_collect();
730 
731 	/*
732 	 * Calculate our target for the number of free+cache pages we
733 	 * want to get to.  This is higher then the number that causes
734 	 * allocations to stall (severe) in order to provide hysteresis,
735 	 * and if we don't make it all the way but get to the minimum
736 	 * we're happy.
737 	 */
738 	inactive_shortage = vm_paging_target() + vm_pageout_deficit;
739 	inactive_original_shortage = inactive_shortage;
740 	vm_pageout_deficit = 0;
741 
742 	/*
743 	 * Initialize our marker
744 	 */
745 	bzero(&marker, sizeof(marker));
746 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
747 	marker.queue = PQ_INACTIVE;
748 	marker.wire_count = 1;
749 
750 	/*
751 	 * Start scanning the inactive queue for pages we can move to the
752 	 * cache or free.  The scan will stop when the target is reached or
753 	 * we have scanned the entire inactive queue.  Note that m->act_count
754 	 * is not used to form decisions for the inactive queue, only for the
755 	 * active queue.
756 	 *
757 	 * maxlaunder limits the number of dirty pages we flush per scan.
758 	 * For most systems a smaller value (16 or 32) is more robust under
759 	 * extreme memory and disk pressure because any unnecessary writes
760 	 * to disk can result in extreme performance degredation.  However,
761 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
762 	 * used) will die horribly with limited laundering.  If the pageout
763 	 * daemon cannot clean enough pages in the first pass, we let it go
764 	 * all out in succeeding passes.
765 	 */
766 	if ((maxlaunder = vm_max_launder) <= 1)
767 		maxlaunder = 1;
768 	if (pass)
769 		maxlaunder = 10000;
770 
771 	/*
772 	 * We will generally be in a critical section throughout the
773 	 * scan, but we can release it temporarily when we are sitting on a
774 	 * non-busy page without fear.  this is required to prevent an
775 	 * interrupt from unbusying or freeing a page prior to our busy
776 	 * check, leaving us on the wrong queue or checking the wrong
777 	 * page.
778 	 */
779 rescan0:
780 	vpfailed = NULL;
781 	maxscan = vmstats.v_inactive_count;
782 	for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl);
783 	     m != NULL && maxscan-- > 0 && inactive_shortage > 0;
784 	     m = next
785 	 ) {
786 		mycpu->gd_cnt.v_pdpages++;
787 
788 		/*
789 		 * It's easier for some of the conditions below to just loop
790 		 * and catch queue changes here rather then check everywhere
791 		 * else.
792 		 */
793 		if (m->queue != PQ_INACTIVE)
794 			goto rescan0;
795 		next = TAILQ_NEXT(m, pageq);
796 
797 		/*
798 		 * skip marker pages
799 		 */
800 		if (m->flags & PG_MARKER)
801 			continue;
802 
803 		/*
804 		 * A held page may be undergoing I/O, so skip it.
805 		 */
806 		if (m->hold_count) {
807 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
808 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
809 			++vm_swapcache_inactive_heuristic;
810 			continue;
811 		}
812 
813 		/*
814 		 * Dont mess with busy pages, keep in the front of the
815 		 * queue, most likely are being paged out.
816 		 */
817 		if (m->busy || (m->flags & PG_BUSY)) {
818 			continue;
819 		}
820 
821 		if (m->object->ref_count == 0) {
822 			/*
823 			 * If the object is not being used, we ignore previous
824 			 * references.
825 			 */
826 			vm_page_flag_clear(m, PG_REFERENCED);
827 			pmap_clear_reference(m);
828 
829 		} else if (((m->flags & PG_REFERENCED) == 0) &&
830 			    (actcount = pmap_ts_referenced(m))) {
831 			/*
832 			 * Otherwise, if the page has been referenced while
833 			 * in the inactive queue, we bump the "activation
834 			 * count" upwards, making it less likely that the
835 			 * page will be added back to the inactive queue
836 			 * prematurely again.  Here we check the page tables
837 			 * (or emulated bits, if any), given the upper level
838 			 * VM system not knowing anything about existing
839 			 * references.
840 			 */
841 			vm_page_activate(m);
842 			m->act_count += (actcount + ACT_ADVANCE);
843 			continue;
844 		}
845 
846 		/*
847 		 * If the upper level VM system knows about any page
848 		 * references, we activate the page.  We also set the
849 		 * "activation count" higher than normal so that we will less
850 		 * likely place pages back onto the inactive queue again.
851 		 */
852 		if ((m->flags & PG_REFERENCED) != 0) {
853 			vm_page_flag_clear(m, PG_REFERENCED);
854 			actcount = pmap_ts_referenced(m);
855 			vm_page_activate(m);
856 			m->act_count += (actcount + ACT_ADVANCE + 1);
857 			continue;
858 		}
859 
860 		/*
861 		 * If the upper level VM system doesn't know anything about
862 		 * the page being dirty, we have to check for it again.  As
863 		 * far as the VM code knows, any partially dirty pages are
864 		 * fully dirty.
865 		 *
866 		 * Pages marked PG_WRITEABLE may be mapped into the user
867 		 * address space of a process running on another cpu.  A
868 		 * user process (without holding the MP lock) running on
869 		 * another cpu may be able to touch the page while we are
870 		 * trying to remove it.  vm_page_cache() will handle this
871 		 * case for us.
872 		 */
873 		if (m->dirty == 0) {
874 			vm_page_test_dirty(m);
875 		} else {
876 			vm_page_dirty(m);
877 		}
878 
879 		if (m->valid == 0) {
880 			/*
881 			 * Invalid pages can be easily freed
882 			 */
883 			vm_pageout_page_free(m);
884 			mycpu->gd_cnt.v_dfree++;
885 			--inactive_shortage;
886 		} else if (m->dirty == 0) {
887 			/*
888 			 * Clean pages can be placed onto the cache queue.
889 			 * This effectively frees them.
890 			 */
891 			vm_page_busy(m);
892 			vm_page_cache(m);
893 			--inactive_shortage;
894 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
895 			/*
896 			 * Dirty pages need to be paged out, but flushing
897 			 * a page is extremely expensive verses freeing
898 			 * a clean page.  Rather then artificially limiting
899 			 * the number of pages we can flush, we instead give
900 			 * dirty pages extra priority on the inactive queue
901 			 * by forcing them to be cycled through the queue
902 			 * twice before being flushed, after which the
903 			 * (now clean) page will cycle through once more
904 			 * before being freed.  This significantly extends
905 			 * the thrash point for a heavily loaded machine.
906 			 */
907 			vm_page_flag_set(m, PG_WINATCFLS);
908 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
909 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
910 			++vm_swapcache_inactive_heuristic;
911 		} else if (maxlaunder > 0) {
912 			/*
913 			 * We always want to try to flush some dirty pages if
914 			 * we encounter them, to keep the system stable.
915 			 * Normally this number is small, but under extreme
916 			 * pressure where there are insufficient clean pages
917 			 * on the inactive queue, we may have to go all out.
918 			 */
919 			int swap_pageouts_ok;
920 			struct vnode *vp = NULL;
921 
922 			object = m->object;
923 
924 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
925 				swap_pageouts_ok = 1;
926 			} else {
927 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
928 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
929 				vm_page_count_min(0));
930 
931 			}
932 
933 			/*
934 			 * We don't bother paging objects that are "dead".
935 			 * Those objects are in a "rundown" state.
936 			 */
937 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
938 				TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
939 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
940 				++vm_swapcache_inactive_heuristic;
941 				continue;
942 			}
943 
944 			/*
945 			 * The object is already known NOT to be dead.   It
946 			 * is possible for the vget() to block the whole
947 			 * pageout daemon, but the new low-memory handling
948 			 * code should prevent it.
949 			 *
950 			 * The previous code skipped locked vnodes and, worse,
951 			 * reordered pages in the queue.  This results in
952 			 * completely non-deterministic operation because,
953 			 * quite often, a vm_fault has initiated an I/O and
954 			 * is holding a locked vnode at just the point where
955 			 * the pageout daemon is woken up.
956 			 *
957 			 * We can't wait forever for the vnode lock, we might
958 			 * deadlock due to a vn_read() getting stuck in
959 			 * vm_wait while holding this vnode.  We skip the
960 			 * vnode if we can't get it in a reasonable amount
961 			 * of time.
962 			 *
963 			 * vpfailed is used to (try to) avoid the case where
964 			 * a large number of pages are associated with a
965 			 * locked vnode, which could cause the pageout daemon
966 			 * to stall for an excessive amount of time.
967 			 */
968 			if (object->type == OBJT_VNODE) {
969 				int flags;
970 
971 				vp = object->handle;
972 				flags = LK_EXCLUSIVE | LK_NOOBJ;
973 				if (vp == vpfailed)
974 					flags |= LK_NOWAIT;
975 				else
976 					flags |= LK_TIMELOCK;
977 				if (vget(vp, flags) != 0) {
978 					vpfailed = vp;
979 					++pageout_lock_miss;
980 					if (object->flags & OBJ_MIGHTBEDIRTY)
981 						    vnodes_skipped++;
982 					continue;
983 				}
984 
985 				/*
986 				 * The page might have been moved to another
987 				 * queue during potential blocking in vget()
988 				 * above.  The page might have been freed and
989 				 * reused for another vnode.  The object might
990 				 * have been reused for another vnode.
991 				 */
992 				if (m->queue != PQ_INACTIVE ||
993 				    m->object != object ||
994 				    object->handle != vp) {
995 					if (object->flags & OBJ_MIGHTBEDIRTY)
996 						vnodes_skipped++;
997 					vput(vp);
998 					continue;
999 				}
1000 
1001 				/*
1002 				 * The page may have been busied during the
1003 				 * blocking in vput();  We don't move the
1004 				 * page back onto the end of the queue so that
1005 				 * statistics are more correct if we don't.
1006 				 */
1007 				if (m->busy || (m->flags & PG_BUSY)) {
1008 					vput(vp);
1009 					continue;
1010 				}
1011 
1012 				/*
1013 				 * If the page has become held it might
1014 				 * be undergoing I/O, so skip it
1015 				 */
1016 				if (m->hold_count) {
1017 					TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1018 					TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE].pl, m, pageq);
1019 					++vm_swapcache_inactive_heuristic;
1020 					if (object->flags & OBJ_MIGHTBEDIRTY)
1021 						vnodes_skipped++;
1022 					vput(vp);
1023 					continue;
1024 				}
1025 			}
1026 
1027 			/*
1028 			 * If a page is dirty, then it is either being washed
1029 			 * (but not yet cleaned) or it is still in the
1030 			 * laundry.  If it is still in the laundry, then we
1031 			 * start the cleaning operation.
1032 			 *
1033 			 * This operation may cluster, invalidating the 'next'
1034 			 * pointer.  To prevent an inordinate number of
1035 			 * restarts we use our marker to remember our place.
1036 			 *
1037 			 * decrement inactive_shortage on success to account
1038 			 * for the (future) cleaned page.  Otherwise we
1039 			 * could wind up laundering or cleaning too many
1040 			 * pages.
1041 			 */
1042 			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, m, &marker, pageq);
1043 			if (vm_pageout_clean(m) != 0) {
1044 				--inactive_shortage;
1045 				--maxlaunder;
1046 			}
1047 			next = TAILQ_NEXT(&marker, pageq);
1048 			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, &marker, pageq);
1049 			if (vp != NULL)
1050 				vput(vp);
1051 		}
1052 	}
1053 
1054 	/*
1055 	 * We want to move pages from the active queue to the inactive
1056 	 * queue to get the inactive queue to the inactive target.  If
1057 	 * we still have a page shortage from above we try to directly free
1058 	 * clean pages instead of moving them.
1059 	 *
1060 	 * If we do still have a shortage we keep track of the number of
1061 	 * pages we free or cache (recycle_count) as a measure of thrashing
1062 	 * between the active and inactive queues.
1063 	 *
1064 	 * If we were able to completely satisfy the free+cache targets
1065 	 * from the inactive pool we limit the number of pages we move
1066 	 * from the active pool to the inactive pool to 2x the pages we
1067 	 * had removed from the inactive pool (with a minimum of 1/5 the
1068 	 * inactive target).  If we were not able to completely satisfy
1069 	 * the free+cache targets we go for the whole target aggressively.
1070 	 *
1071 	 * NOTE: Both variables can end up negative.
1072 	 * NOTE: We are still in a critical section.
1073 	 */
1074 	active_shortage = vmstats.v_inactive_target - vmstats.v_inactive_count;
1075 	if (inactive_original_shortage < vmstats.v_inactive_target / 10)
1076 		inactive_original_shortage = vmstats.v_inactive_target / 10;
1077 	if (inactive_shortage <= 0 &&
1078 	    active_shortage > inactive_original_shortage * 2) {
1079 		active_shortage = inactive_original_shortage * 2;
1080 	}
1081 
1082 	pcount = vmstats.v_active_count;
1083 	recycle_count = 0;
1084 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1085 
1086 	while ((m != NULL) && (pcount-- > 0) &&
1087 	       (inactive_shortage > 0 || active_shortage > 0)
1088 	) {
1089 		/*
1090 		 * If the page was ripped out from under us, just stop.
1091 		 */
1092 		if (m->queue != PQ_ACTIVE)
1093 			break;
1094 		next = TAILQ_NEXT(m, pageq);
1095 
1096 		/*
1097 		 * Don't deactivate pages that are busy.
1098 		 */
1099 		if ((m->busy != 0) ||
1100 		    (m->flags & PG_BUSY) ||
1101 		    (m->hold_count != 0)) {
1102 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1103 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1104 			m = next;
1105 			continue;
1106 		}
1107 
1108 		/*
1109 		 * The count for pagedaemon pages is done after checking the
1110 		 * page for eligibility...
1111 		 */
1112 		mycpu->gd_cnt.v_pdpages++;
1113 
1114 		/*
1115 		 * Check to see "how much" the page has been used and clear
1116 		 * the tracking access bits.  If the object has no references
1117 		 * don't bother paying the expense.
1118 		 */
1119 		actcount = 0;
1120 		if (m->object->ref_count != 0) {
1121 			if (m->flags & PG_REFERENCED)
1122 				++actcount;
1123 			actcount += pmap_ts_referenced(m);
1124 			if (actcount) {
1125 				m->act_count += ACT_ADVANCE + actcount;
1126 				if (m->act_count > ACT_MAX)
1127 					m->act_count = ACT_MAX;
1128 			}
1129 		}
1130 		vm_page_flag_clear(m, PG_REFERENCED);
1131 
1132 		/*
1133 		 * actcount is only valid if the object ref_count is non-zero.
1134 		 */
1135 		if (actcount && m->object->ref_count != 0) {
1136 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1137 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1138 		} else {
1139 			m->act_count -= min(m->act_count, ACT_DECLINE);
1140 			if (vm_pageout_algorithm ||
1141 			    m->object->ref_count == 0 ||
1142 			    m->act_count < pass + 1
1143 			) {
1144 				/*
1145 				 * Deactivate the page.  If we had a
1146 				 * shortage from our inactive scan try to
1147 				 * free (cache) the page instead.
1148 				 *
1149 				 * Don't just blindly cache the page if
1150 				 * we do not have a shortage from the
1151 				 * inactive scan, that could lead to
1152 				 * gigabytes being moved.
1153 				 */
1154 				--active_shortage;
1155 				if (inactive_shortage > 0 ||
1156 				    m->object->ref_count == 0) {
1157 					if (inactive_shortage > 0)
1158 						++recycle_count;
1159 					vm_page_busy(m);
1160 					vm_page_protect(m, VM_PROT_NONE);
1161 					if (m->dirty == 0 &&
1162 					    inactive_shortage > 0) {
1163 						--inactive_shortage;
1164 						vm_page_cache(m);
1165 					} else {
1166 						vm_page_deactivate(m);
1167 						vm_page_wakeup(m);
1168 					}
1169 				} else {
1170 					vm_page_deactivate(m);
1171 				}
1172 			} else {
1173 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1174 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1175 			}
1176 		}
1177 		m = next;
1178 	}
1179 
1180 	/*
1181 	 * The number of actually free pages can drop down to v_free_reserved,
1182 	 * we try to build the free count back above v_free_min.  Note that
1183 	 * vm_paging_needed() also returns TRUE if v_free_count is not at
1184 	 * least v_free_min so that is the minimum we must build the free
1185 	 * count to.
1186 	 *
1187 	 * We use a slightly higher target to improve hysteresis,
1188 	 * ((v_free_target + v_free_min) / 2).  Since v_free_target
1189 	 * is usually the same as v_cache_min this maintains about
1190 	 * half the pages in the free queue as are in the cache queue,
1191 	 * providing pretty good pipelining for pageout operation.
1192 	 *
1193 	 * The system operator can manipulate vm.v_cache_min and
1194 	 * vm.v_free_target to tune the pageout demon.  Be sure
1195 	 * to keep vm.v_free_min < vm.v_free_target.
1196 	 *
1197 	 * Note that the original paging target is to get at least
1198 	 * (free_min + cache_min) into (free + cache).  The slightly
1199 	 * higher target will shift additional pages from cache to free
1200 	 * without effecting the original paging target in order to
1201 	 * maintain better hysteresis and not have the free count always
1202 	 * be dead-on v_free_min.
1203 	 *
1204 	 * NOTE: we are still in a critical section.
1205 	 *
1206 	 * Pages moved from PQ_CACHE to totally free are not counted in the
1207 	 * pages_freed counter.
1208 	 */
1209 	while (vmstats.v_free_count <
1210 	       (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1211 		/*
1212 		 *
1213 		 */
1214 		static int cache_rover = 0;
1215 		m = vm_page_list_find(PQ_CACHE, cache_rover, FALSE);
1216 		if (m == NULL)
1217 			break;
1218 		if ((m->flags & (PG_BUSY|PG_UNMANAGED)) ||
1219 		    m->busy ||
1220 		    m->hold_count ||
1221 		    m->wire_count) {
1222 #ifdef INVARIANTS
1223 			kprintf("Warning: busy page %p found in cache\n", m);
1224 #endif
1225 			vm_page_deactivate(m);
1226 			continue;
1227 		}
1228 		KKASSERT((m->flags & PG_MAPPED) == 0);
1229 		KKASSERT(m->dirty == 0);
1230 		cache_rover = (cache_rover + PQ_PRIME2) & PQ_L2_MASK;
1231 		vm_pageout_page_free(m);
1232 		mycpu->gd_cnt.v_dfree++;
1233 	}
1234 
1235 #if !defined(NO_SWAPPING)
1236 	/*
1237 	 * Idle process swapout -- run once per second.
1238 	 */
1239 	if (vm_swap_idle_enabled) {
1240 		static long lsec;
1241 		if (time_second != lsec) {
1242 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1243 			vm_req_vmdaemon();
1244 			lsec = time_second;
1245 		}
1246 	}
1247 #endif
1248 
1249 	/*
1250 	 * If we didn't get enough free pages, and we have skipped a vnode
1251 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1252 	 * if we did not get enough free pages.
1253 	 */
1254 	if (vm_paging_target() > 0) {
1255 		if (vnodes_skipped && vm_page_count_min(0))
1256 			speedup_syncer();
1257 #if !defined(NO_SWAPPING)
1258 		if (vm_swap_enabled && vm_page_count_target()) {
1259 			vm_req_vmdaemon();
1260 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1261 		}
1262 #endif
1263 	}
1264 
1265 	/*
1266 	 * Handle catastrophic conditions.  Under good conditions we should
1267 	 * be at the target, well beyond our minimum.  If we could not even
1268 	 * reach our minimum the system is under heavy stress.
1269 	 *
1270 	 * Determine whether we have run out of memory.  This occurs when
1271 	 * swap_pager_full is TRUE and the only pages left in the page
1272 	 * queues are dirty.  We will still likely have page shortages.
1273 	 *
1274 	 * - swap_pager_full is set if insufficient swap was
1275 	 *   available to satisfy a requested pageout.
1276 	 *
1277 	 * - the inactive queue is bloated (4 x size of active queue),
1278 	 *   meaning it is unable to get rid of dirty pages and.
1279 	 *
1280 	 * - vm_page_count_min() without counting pages recycled from the
1281 	 *   active queue (recycle_count) means we could not recover
1282 	 *   enough pages to meet bare minimum needs.  This test only
1283 	 *   works if the inactive queue is bloated.
1284 	 *
1285 	 * - due to a positive inactive_shortage we shifted the remaining
1286 	 *   dirty pages from the active queue to the inactive queue
1287 	 *   trying to find clean ones to free.
1288 	 */
1289 	if (swap_pager_full && vm_page_count_min(recycle_count))
1290 		kprintf("Warning: system low on memory+swap!\n");
1291 	if (swap_pager_full && vm_page_count_min(recycle_count) &&
1292 	    vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1293 	    inactive_shortage > 0) {
1294 		/*
1295 		 * Kill something.
1296 		 */
1297 		info.bigproc = NULL;
1298 		info.bigsize = 0;
1299 		allproc_scan(vm_pageout_scan_callback, &info);
1300 		if (info.bigproc != NULL) {
1301 			killproc(info.bigproc, "out of swap space");
1302 			info.bigproc->p_nice = PRIO_MIN;
1303 			info.bigproc->p_usched->resetpriority(
1304 				FIRST_LWP_IN_PROC(info.bigproc));
1305 			wakeup(&vmstats.v_free_count);
1306 			PRELE(info.bigproc);
1307 		}
1308 	}
1309 	return(inactive_shortage);
1310 }
1311 
1312 /*
1313  * The caller must hold vm_token and proc_token.
1314  */
1315 static int
1316 vm_pageout_scan_callback(struct proc *p, void *data)
1317 {
1318 	struct vm_pageout_scan_info *info = data;
1319 	vm_offset_t size;
1320 
1321 	/*
1322 	 * Never kill system processes or init.  If we have configured swap
1323 	 * then try to avoid killing low-numbered pids.
1324 	 */
1325 	if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1326 	    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1327 		return (0);
1328 	}
1329 
1330 	/*
1331 	 * if the process is in a non-running type state,
1332 	 * don't touch it.
1333 	 */
1334 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1335 		return (0);
1336 
1337 	/*
1338 	 * Get the approximate process size.  Note that anonymous pages
1339 	 * with backing swap will be counted twice, but there should not
1340 	 * be too many such pages due to the stress the VM system is
1341 	 * under at this point.
1342 	 */
1343 	size = vmspace_anonymous_count(p->p_vmspace) +
1344 		vmspace_swap_count(p->p_vmspace);
1345 
1346 	/*
1347 	 * If the this process is bigger than the biggest one
1348 	 * remember it.
1349 	 */
1350 	if (info->bigsize < size) {
1351 		if (info->bigproc)
1352 			PRELE(info->bigproc);
1353 		PHOLD(p);
1354 		info->bigproc = p;
1355 		info->bigsize = size;
1356 	}
1357 	return(0);
1358 }
1359 
1360 /*
1361  * This routine tries to maintain the pseudo LRU active queue,
1362  * so that during long periods of time where there is no paging,
1363  * that some statistic accumulation still occurs.  This code
1364  * helps the situation where paging just starts to occur.
1365  *
1366  * The caller must hold vm_token.
1367  */
1368 static void
1369 vm_pageout_page_stats(void)
1370 {
1371 	vm_page_t m,next;
1372 	int pcount,tpcount;		/* Number of pages to check */
1373 	static int fullintervalcount = 0;
1374 	int page_shortage;
1375 
1376 	page_shortage =
1377 	    (vmstats.v_inactive_target + vmstats.v_cache_max + vmstats.v_free_min) -
1378 	    (vmstats.v_free_count + vmstats.v_inactive_count + vmstats.v_cache_count);
1379 
1380 	if (page_shortage <= 0)
1381 		return;
1382 
1383 	pcount = vmstats.v_active_count;
1384 	fullintervalcount += vm_pageout_stats_interval;
1385 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1386 		tpcount = (vm_pageout_stats_max * vmstats.v_active_count) / vmstats.v_page_count;
1387 		if (pcount > tpcount)
1388 			pcount = tpcount;
1389 	} else {
1390 		fullintervalcount = 0;
1391 	}
1392 
1393 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
1394 	while ((m != NULL) && (pcount-- > 0)) {
1395 		int actcount;
1396 
1397 		if (m->queue != PQ_ACTIVE) {
1398 			break;
1399 		}
1400 
1401 		next = TAILQ_NEXT(m, pageq);
1402 		/*
1403 		 * Don't deactivate pages that are busy.
1404 		 */
1405 		if ((m->busy != 0) ||
1406 		    (m->flags & PG_BUSY) ||
1407 		    (m->hold_count != 0)) {
1408 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1409 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1410 			m = next;
1411 			continue;
1412 		}
1413 
1414 		actcount = 0;
1415 		if (m->flags & PG_REFERENCED) {
1416 			vm_page_flag_clear(m, PG_REFERENCED);
1417 			actcount += 1;
1418 		}
1419 
1420 		actcount += pmap_ts_referenced(m);
1421 		if (actcount) {
1422 			m->act_count += ACT_ADVANCE + actcount;
1423 			if (m->act_count > ACT_MAX)
1424 				m->act_count = ACT_MAX;
1425 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1426 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1427 		} else {
1428 			if (m->act_count == 0) {
1429 				/*
1430 				 * We turn off page access, so that we have
1431 				 * more accurate RSS stats.  We don't do this
1432 				 * in the normal page deactivation when the
1433 				 * system is loaded VM wise, because the
1434 				 * cost of the large number of page protect
1435 				 * operations would be higher than the value
1436 				 * of doing the operation.
1437 				 */
1438 				vm_page_busy(m);
1439 				vm_page_protect(m, VM_PROT_NONE);
1440 				vm_page_deactivate(m);
1441 				vm_page_wakeup(m);
1442 			} else {
1443 				m->act_count -= min(m->act_count, ACT_DECLINE);
1444 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1445 				TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE].pl, m, pageq);
1446 			}
1447 		}
1448 
1449 		m = next;
1450 	}
1451 }
1452 
1453 /*
1454  * The caller must hold vm_token.
1455  */
1456 static int
1457 vm_pageout_free_page_calc(vm_size_t count)
1458 {
1459 	if (count < vmstats.v_page_count)
1460 		 return 0;
1461 	/*
1462 	 * free_reserved needs to include enough for the largest swap pager
1463 	 * structures plus enough for any pv_entry structs when paging.
1464 	 */
1465 	if (vmstats.v_page_count > 1024)
1466 		vmstats.v_free_min = 4 + (vmstats.v_page_count - 1024) / 200;
1467 	else
1468 		vmstats.v_free_min = 4;
1469 	vmstats.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE +
1470 		vmstats.v_interrupt_free_min;
1471 	vmstats.v_free_reserved = vm_pageout_page_count +
1472 		vmstats.v_pageout_free_min + (count / 768) + PQ_L2_SIZE;
1473 	vmstats.v_free_severe = vmstats.v_free_min / 2;
1474 	vmstats.v_free_min += vmstats.v_free_reserved;
1475 	vmstats.v_free_severe += vmstats.v_free_reserved;
1476 	return 1;
1477 }
1478 
1479 
1480 /*
1481  * vm_pageout is the high level pageout daemon.
1482  *
1483  * No requirements.
1484  */
1485 static void
1486 vm_pageout_thread(void)
1487 {
1488 	int pass;
1489 	int inactive_shortage;
1490 
1491 	/*
1492 	 * Permanently hold vm_token.
1493 	 */
1494 	lwkt_gettoken(&vm_token);
1495 
1496 	/*
1497 	 * Initialize some paging parameters.
1498 	 */
1499 	curthread->td_flags |= TDF_SYSTHREAD;
1500 
1501 	vmstats.v_interrupt_free_min = 2;
1502 	if (vmstats.v_page_count < 2000)
1503 		vm_pageout_page_count = 8;
1504 
1505 	vm_pageout_free_page_calc(vmstats.v_page_count);
1506 
1507 	/*
1508 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1509 	 * that these are more a measure of the VM cache queue hysteresis
1510 	 * then the VM free queue.  Specifically, v_free_target is the
1511 	 * high water mark (free+cache pages).
1512 	 *
1513 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1514 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1515 	 * be big enough to handle memory needs while the pageout daemon
1516 	 * is signalled and run to free more pages.
1517 	 */
1518 	if (vmstats.v_free_count > 6144)
1519 		vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1520 	else
1521 		vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1522 
1523 	/*
1524 	 * NOTE: With the new buffer cache b_act_count we want the default
1525 	 *	 inactive target to be a percentage of available memory.
1526 	 *
1527 	 *	 The inactive target essentially determines the minimum
1528 	 *	 number of 'temporary' pages capable of caching one-time-use
1529 	 *	 files when the VM system is otherwise full of pages
1530 	 *	 belonging to multi-time-use files or active program data.
1531 	 *
1532 	 * NOTE: The inactive target is aggressively persued only if the
1533 	 *	 inactive queue becomes too small.  If the inactive queue
1534 	 *	 is large enough to satisfy page movement to free+cache
1535 	 *	 then it is repopulated more slowly from the active queue.
1536 	 *	 This allows a general inactive_target default to be set.
1537 	 *
1538 	 *	 There is an issue here for processes which sit mostly idle
1539 	 *	 'overnight', such as sshd, tcsh, and X.  Any movement from
1540 	 *	 the active queue will eventually cause such pages to
1541 	 *	 recycle eventually causing a lot of paging in the morning.
1542 	 *	 To reduce the incidence of this pages cycled out of the
1543 	 *	 buffer cache are moved directly to the inactive queue if
1544 	 *	 they were only used once or twice.
1545 	 *
1546 	 *	 The vfs.vm_cycle_point sysctl can be used to adjust this.
1547 	 *	 Increasing the value (up to 64) increases the number of
1548 	 *	 buffer recyclements which go directly to the inactive queue.
1549 	 */
1550 	if (vmstats.v_free_count > 2048) {
1551 		vmstats.v_cache_min = vmstats.v_free_target;
1552 		vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1553 	} else {
1554 		vmstats.v_cache_min = 0;
1555 		vmstats.v_cache_max = 0;
1556 	}
1557 	vmstats.v_inactive_target = vmstats.v_free_count / 4;
1558 
1559 	/* XXX does not really belong here */
1560 	if (vm_page_max_wired == 0)
1561 		vm_page_max_wired = vmstats.v_free_count / 3;
1562 
1563 	if (vm_pageout_stats_max == 0)
1564 		vm_pageout_stats_max = vmstats.v_free_target;
1565 
1566 	/*
1567 	 * Set interval in seconds for stats scan.
1568 	 */
1569 	if (vm_pageout_stats_interval == 0)
1570 		vm_pageout_stats_interval = 5;
1571 	if (vm_pageout_full_stats_interval == 0)
1572 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1573 
1574 
1575 	/*
1576 	 * Set maximum free per pass
1577 	 */
1578 	if (vm_pageout_stats_free_max == 0)
1579 		vm_pageout_stats_free_max = 5;
1580 
1581 	swap_pager_swap_init();
1582 	pass = 0;
1583 
1584 	/*
1585 	 * The pageout daemon is never done, so loop forever.
1586 	 */
1587 	while (TRUE) {
1588 		int error;
1589 
1590 		/*
1591 		 * Wait for an action request.  If we timeout check to
1592 		 * see if paging is needed (in case the normal wakeup
1593 		 * code raced us).
1594 		 */
1595 		if (vm_pages_needed == 0) {
1596 			error = tsleep(&vm_pages_needed,
1597 				       0, "psleep",
1598 				       vm_pageout_stats_interval * hz);
1599 			if (error &&
1600 			    vm_paging_needed() == 0 &&
1601 			    vm_pages_needed == 0) {
1602 				vm_pageout_page_stats();
1603 				continue;
1604 			}
1605 			vm_pages_needed = 1;
1606 		}
1607 
1608 		mycpu->gd_cnt.v_pdwakeups++;
1609 
1610 		/*
1611 		 * Scan for pageout.  Try to avoid thrashing the system
1612 		 * with activity.
1613 		 */
1614 		inactive_shortage = vm_pageout_scan(pass);
1615 		if (inactive_shortage > 0) {
1616 			++pass;
1617 			if (swap_pager_full) {
1618 				/*
1619 				 * Running out of memory, catastrophic back-off
1620 				 * to one-second intervals.
1621 				 */
1622 				tsleep(&vm_pages_needed, 0, "pdelay", hz);
1623 			} else if (pass < 10 && vm_pages_needed > 1) {
1624 				/*
1625 				 * Normal operation, additional processes
1626 				 * have already kicked us.  Retry immediately.
1627 				 */
1628 			} else if (pass < 10) {
1629 				/*
1630 				 * Normal operation, fewer processes.  Delay
1631 				 * a bit but allow wakeups.
1632 				 */
1633 				vm_pages_needed = 0;
1634 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1635 				vm_pages_needed = 1;
1636 			} else {
1637 				/*
1638 				 * We've taken too many passes, forced delay.
1639 				 */
1640 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1641 			}
1642 		} else {
1643 			/*
1644 			 * Interlocked wakeup of waiters (non-optional)
1645 			 */
1646 			pass = 0;
1647 			if (vm_pages_needed && !vm_page_count_min(0)) {
1648 				wakeup(&vmstats.v_free_count);
1649 				vm_pages_needed = 0;
1650 			}
1651 		}
1652 	}
1653 }
1654 
1655 static struct kproc_desc page_kp = {
1656 	"pagedaemon",
1657 	vm_pageout_thread,
1658 	&pagethread
1659 };
1660 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
1661 
1662 
1663 /*
1664  * Called after allocating a page out of the cache or free queue
1665  * to possibly wake the pagedaemon up to replentish our supply.
1666  *
1667  * We try to generate some hysteresis by waking the pagedaemon up
1668  * when our free+cache pages go below the free_min+cache_min level.
1669  * The pagedaemon tries to get the count back up to at least the
1670  * minimum, and through to the target level if possible.
1671  *
1672  * If the pagedaemon is already active bump vm_pages_needed as a hint
1673  * that there are even more requests pending.
1674  *
1675  * SMP races ok?
1676  * No requirements.
1677  */
1678 void
1679 pagedaemon_wakeup(void)
1680 {
1681 	if (vm_paging_needed() && curthread != pagethread) {
1682 		if (vm_pages_needed == 0) {
1683 			vm_pages_needed = 1;	/* SMP race ok */
1684 			wakeup(&vm_pages_needed);
1685 		} else if (vm_page_count_min(0)) {
1686 			++vm_pages_needed;	/* SMP race ok */
1687 		}
1688 	}
1689 }
1690 
1691 #if !defined(NO_SWAPPING)
1692 
1693 /*
1694  * SMP races ok?
1695  * No requirements.
1696  */
1697 static void
1698 vm_req_vmdaemon(void)
1699 {
1700 	static int lastrun = 0;
1701 
1702 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
1703 		wakeup(&vm_daemon_needed);
1704 		lastrun = ticks;
1705 	}
1706 }
1707 
1708 static int vm_daemon_callback(struct proc *p, void *data __unused);
1709 
1710 /*
1711  * No requirements.
1712  */
1713 static void
1714 vm_daemon(void)
1715 {
1716 	/*
1717 	 * Permanently hold vm_token.
1718 	 */
1719 	lwkt_gettoken(&vm_token);
1720 
1721 	while (TRUE) {
1722 		tsleep(&vm_daemon_needed, 0, "psleep", 0);
1723 		if (vm_pageout_req_swapout) {
1724 			swapout_procs(vm_pageout_req_swapout);
1725 			vm_pageout_req_swapout = 0;
1726 		}
1727 		/*
1728 		 * scan the processes for exceeding their rlimits or if
1729 		 * process is swapped out -- deactivate pages
1730 		 */
1731 		allproc_scan(vm_daemon_callback, NULL);
1732 	}
1733 }
1734 
1735 /*
1736  * Caller must hold vm_token and proc_token.
1737  */
1738 static int
1739 vm_daemon_callback(struct proc *p, void *data __unused)
1740 {
1741 	vm_pindex_t limit, size;
1742 
1743 	/*
1744 	 * if this is a system process or if we have already
1745 	 * looked at this process, skip it.
1746 	 */
1747 	if (p->p_flag & (P_SYSTEM | P_WEXIT))
1748 		return (0);
1749 
1750 	/*
1751 	 * if the process is in a non-running type state,
1752 	 * don't touch it.
1753 	 */
1754 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1755 		return (0);
1756 
1757 	/*
1758 	 * get a limit
1759 	 */
1760 	limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
1761 			        p->p_rlimit[RLIMIT_RSS].rlim_max));
1762 
1763 	/*
1764 	 * let processes that are swapped out really be
1765 	 * swapped out.  Set the limit to nothing to get as
1766 	 * many pages out to swap as possible.
1767 	 */
1768 	if (p->p_flag & P_SWAPPEDOUT)
1769 		limit = 0;
1770 
1771 	size = vmspace_resident_count(p->p_vmspace);
1772 	if (limit >= 0 && size >= limit) {
1773 		vm_pageout_map_deactivate_pages(
1774 		    &p->p_vmspace->vm_map, limit);
1775 	}
1776 	return (0);
1777 }
1778 
1779 #endif
1780