xref: /dflybsd-src/sys/vm/vm_pageout.c (revision dae741e33c840b92a8a53bf9f01157ede145e256)
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 <sys/spinlock2.h>
98 #include <vm/vm_page2.h>
99 
100 /*
101  * System initialization
102  */
103 
104 /* the kernel process "vm_pageout"*/
105 static int vm_pageout_clean (vm_page_t);
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(int q);
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 static int
240 vm_pageout_clean(vm_page_t m)
241 {
242 	vm_object_t object;
243 	vm_page_t mc[2*vm_pageout_page_count];
244 	int pageout_count;
245 	int error;
246 	int ib, is, page_base;
247 	vm_pindex_t pindex = m->pindex;
248 
249 	object = m->object;
250 
251 	/*
252 	 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP
253 	 * with the new swapper, but we could have serious problems paging
254 	 * out other object types if there is insufficient memory.
255 	 *
256 	 * Unfortunately, checking free memory here is far too late, so the
257 	 * check has been moved up a procedural level.
258 	 */
259 
260 	/*
261 	 * Don't mess with the page if it's busy, held, or special
262 	 *
263 	 * XXX do we really need to check hold_count here?  hold_count
264 	 * isn't supposed to mess with vm_page ops except prevent the
265 	 * page from being reused.
266 	 */
267 	if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
268 		vm_page_wakeup(m);
269 		return 0;
270 	}
271 
272 	mc[vm_pageout_page_count] = m;
273 	pageout_count = 1;
274 	page_base = vm_pageout_page_count;
275 	ib = 1;
276 	is = 1;
277 
278 	/*
279 	 * Scan object for clusterable pages.
280 	 *
281 	 * We can cluster ONLY if: ->> the page is NOT
282 	 * clean, wired, busy, held, or mapped into a
283 	 * buffer, and one of the following:
284 	 * 1) The page is inactive, or a seldom used
285 	 *    active page.
286 	 * -or-
287 	 * 2) we force the issue.
288 	 *
289 	 * During heavy mmap/modification loads the pageout
290 	 * daemon can really fragment the underlying file
291 	 * due to flushing pages out of order and not trying
292 	 * align the clusters (which leave sporatic out-of-order
293 	 * holes).  To solve this problem we do the reverse scan
294 	 * first and attempt to align our cluster, then do a
295 	 * forward scan if room remains.
296 	 */
297 
298 	vm_object_hold(object);
299 more:
300 	while (ib && pageout_count < vm_pageout_page_count) {
301 		vm_page_t p;
302 
303 		if (ib > pindex) {
304 			ib = 0;
305 			break;
306 		}
307 
308 		p = vm_page_lookup_busy_try(object, pindex - ib, TRUE, &error);
309 		if (error || p == NULL) {
310 			ib = 0;
311 			break;
312 		}
313 		if ((p->queue - p->pc) == PQ_CACHE ||
314 		    (p->flags & PG_UNMANAGED)) {
315 			vm_page_wakeup(p);
316 			ib = 0;
317 			break;
318 		}
319 		vm_page_test_dirty(p);
320 		if ((p->dirty & p->valid) == 0 ||
321 		    p->queue - p->pc != PQ_INACTIVE ||
322 		    p->wire_count != 0 ||	/* may be held by buf cache */
323 		    p->hold_count != 0) {	/* may be undergoing I/O */
324 			vm_page_wakeup(p);
325 			ib = 0;
326 			break;
327 		}
328 		mc[--page_base] = p;
329 		++pageout_count;
330 		++ib;
331 		/*
332 		 * alignment boundry, stop here and switch directions.  Do
333 		 * not clear ib.
334 		 */
335 		if ((pindex - (ib - 1)) % vm_pageout_page_count == 0)
336 			break;
337 	}
338 
339 	while (pageout_count < vm_pageout_page_count &&
340 	    pindex + is < object->size) {
341 		vm_page_t p;
342 
343 		p = vm_page_lookup_busy_try(object, pindex + is, TRUE, &error);
344 		if (error || p == NULL)
345 			break;
346 		if (((p->queue - p->pc) == PQ_CACHE) ||
347 		    (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) {
348 			vm_page_wakeup(p);
349 			break;
350 		}
351 		vm_page_test_dirty(p);
352 		if ((p->dirty & p->valid) == 0 ||
353 		    p->queue - p->pc != PQ_INACTIVE ||
354 		    p->wire_count != 0 ||	/* may be held by buf cache */
355 		    p->hold_count != 0) {	/* may be undergoing I/O */
356 			vm_page_wakeup(p);
357 			break;
358 		}
359 		mc[page_base + pageout_count] = p;
360 		++pageout_count;
361 		++is;
362 	}
363 
364 	/*
365 	 * If we exhausted our forward scan, continue with the reverse scan
366 	 * when possible, even past a page boundry.  This catches boundry
367 	 * conditions.
368 	 */
369 	if (ib && pageout_count < vm_pageout_page_count)
370 		goto more;
371 
372 	vm_object_drop(object);
373 
374 	/*
375 	 * we allow reads during pageouts...
376 	 */
377 	return vm_pageout_flush(&mc[page_base], pageout_count, 0);
378 }
379 
380 /*
381  * vm_pageout_flush() - launder the given pages
382  *
383  *	The given pages are laundered.  Note that we setup for the start of
384  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
385  *	reference count all in here rather then in the parent.  If we want
386  *	the parent to do more sophisticated things we may have to change
387  *	the ordering.
388  *
389  *	The pages in the array must be busied by the caller and will be
390  *	unbusied by this function.
391  */
392 int
393 vm_pageout_flush(vm_page_t *mc, int count, int flags)
394 {
395 	vm_object_t object;
396 	int pageout_status[count];
397 	int numpagedout = 0;
398 	int i;
399 
400 	/*
401 	 * Initiate I/O.  Bump the vm_page_t->busy counter.
402 	 */
403 	for (i = 0; i < count; i++) {
404 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
405 			("vm_pageout_flush page %p index %d/%d: partially "
406 			 "invalid page", mc[i], i, count));
407 		vm_page_io_start(mc[i]);
408 	}
409 
410 	/*
411 	 * We must make the pages read-only.  This will also force the
412 	 * modified bit in the related pmaps to be cleared.  The pager
413 	 * cannot clear the bit for us since the I/O completion code
414 	 * typically runs from an interrupt.  The act of making the page
415 	 * read-only handles the case for us.
416 	 *
417 	 * Then we can unbusy the pages, we still hold a reference by virtue
418 	 * of our soft-busy.
419 	 */
420 	for (i = 0; i < count; i++) {
421 		vm_page_protect(mc[i], VM_PROT_READ);
422 		vm_page_wakeup(mc[i]);
423 	}
424 
425 	object = mc[0]->object;
426 	vm_object_pip_add(object, count);
427 
428 	vm_pager_put_pages(object, mc, count,
429 	    (flags | ((object == &kernel_object) ? VM_PAGER_PUT_SYNC : 0)),
430 	    pageout_status);
431 
432 	for (i = 0; i < count; i++) {
433 		vm_page_t mt = mc[i];
434 
435 		switch (pageout_status[i]) {
436 		case VM_PAGER_OK:
437 			numpagedout++;
438 			break;
439 		case VM_PAGER_PEND:
440 			numpagedout++;
441 			break;
442 		case VM_PAGER_BAD:
443 			/*
444 			 * Page outside of range of object. Right now we
445 			 * essentially lose the changes by pretending it
446 			 * worked.
447 			 */
448 			vm_page_busy_wait(mt, FALSE, "pgbad");
449 			pmap_clear_modify(mt);
450 			vm_page_undirty(mt);
451 			vm_page_wakeup(mt);
452 			break;
453 		case VM_PAGER_ERROR:
454 		case VM_PAGER_FAIL:
455 			/*
456 			 * A page typically cannot be paged out when we
457 			 * have run out of swap.  We leave the page
458 			 * marked inactive and will try to page it out
459 			 * again later.
460 			 *
461 			 * Starvation of the active page list is used to
462 			 * determine when the system is massively memory
463 			 * starved.
464 			 */
465 			break;
466 		case VM_PAGER_AGAIN:
467 			break;
468 		}
469 
470 		/*
471 		 * If the operation is still going, leave the page busy to
472 		 * block all other accesses. Also, leave the paging in
473 		 * progress indicator set so that we don't attempt an object
474 		 * collapse.
475 		 *
476 		 * For any pages which have completed synchronously,
477 		 * deactivate the page if we are under a severe deficit.
478 		 * Do not try to enter them into the cache, though, they
479 		 * might still be read-heavy.
480 		 */
481 		if (pageout_status[i] != VM_PAGER_PEND) {
482 			vm_page_busy_wait(mt, FALSE, "pgouw");
483 			if (vm_page_count_severe())
484 				vm_page_deactivate(mt);
485 #if 0
486 			if (!vm_page_count_severe() || !vm_page_try_to_cache(mt))
487 				vm_page_protect(mt, VM_PROT_READ);
488 #endif
489 			vm_page_io_finish(mt);
490 			vm_page_wakeup(mt);
491 			vm_object_pip_wakeup(object);
492 		}
493 	}
494 	return numpagedout;
495 }
496 
497 #if !defined(NO_SWAPPING)
498 /*
499  * deactivate enough pages to satisfy the inactive target
500  * requirements or if vm_page_proc_limit is set, then
501  * deactivate all of the pages in the object and its
502  * backing_objects.
503  *
504  * The map must be locked.
505  * The caller must hold the vm_object.
506  */
507 static int vm_pageout_object_deactivate_pages_callback(vm_page_t, void *);
508 
509 static void
510 vm_pageout_object_deactivate_pages(vm_map_t map, vm_object_t object,
511 				   vm_pindex_t desired, int map_remove_only)
512 {
513 	struct rb_vm_page_scan_info info;
514 	vm_object_t lobject;
515 	vm_object_t tobject;
516 	int remove_mode;
517 
518 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
519 	lobject = object;
520 
521 	while (lobject) {
522 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
523 			break;
524 		if (lobject->type == OBJT_DEVICE || lobject->type == OBJT_PHYS)
525 			break;
526 		if (lobject->paging_in_progress)
527 			break;
528 
529 		remove_mode = map_remove_only;
530 		if (lobject->shadow_count > 1)
531 			remove_mode = 1;
532 
533 		/*
534 		 * scan the objects entire memory queue.  We hold the
535 		 * object's token so the scan should not race anything.
536 		 */
537 		info.limit = remove_mode;
538 		info.map = map;
539 		info.desired = desired;
540 		vm_page_rb_tree_RB_SCAN(&lobject->rb_memq, NULL,
541 				vm_pageout_object_deactivate_pages_callback,
542 				&info
543 		);
544 		while ((tobject = lobject->backing_object) != NULL) {
545 			KKASSERT(tobject != object);
546 			vm_object_hold(tobject);
547 			if (tobject == lobject->backing_object)
548 				break;
549 			vm_object_drop(tobject);
550 		}
551 		if (lobject != object) {
552 			vm_object_lock_swap();
553 			vm_object_drop(lobject);
554 		}
555 		lobject = tobject;
556 	}
557 	if (lobject != object)
558 		vm_object_drop(lobject);
559 }
560 
561 /*
562  * The caller must hold the vm_object.
563  */
564 static int
565 vm_pageout_object_deactivate_pages_callback(vm_page_t p, void *data)
566 {
567 	struct rb_vm_page_scan_info *info = data;
568 	int actcount;
569 
570 	if (pmap_resident_count(vm_map_pmap(info->map)) <= info->desired) {
571 		return(-1);
572 	}
573 	mycpu->gd_cnt.v_pdpages++;
574 
575 	if (vm_page_busy_try(p, TRUE))
576 		return(0);
577 	if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
578 		vm_page_wakeup(p);
579 		return(0);
580 	}
581 	if (!pmap_page_exists_quick(vm_map_pmap(info->map), p)) {
582 		vm_page_wakeup(p);
583 		return(0);
584 	}
585 
586 	actcount = pmap_ts_referenced(p);
587 	if (actcount) {
588 		vm_page_flag_set(p, PG_REFERENCED);
589 	} else if (p->flags & PG_REFERENCED) {
590 		actcount = 1;
591 	}
592 
593 	vm_page_and_queue_spin_lock(p);
594 	if (p->queue - p->pc != PQ_ACTIVE && (p->flags & PG_REFERENCED)) {
595 		vm_page_and_queue_spin_unlock(p);
596 		vm_page_activate(p);
597 		p->act_count += actcount;
598 		vm_page_flag_clear(p, PG_REFERENCED);
599 	} else if (p->queue - p->pc == PQ_ACTIVE) {
600 		if ((p->flags & PG_REFERENCED) == 0) {
601 			p->act_count -= min(p->act_count, ACT_DECLINE);
602 			if (!info->limit &&
603 			    (vm_pageout_algorithm || (p->act_count == 0))) {
604 				vm_page_and_queue_spin_unlock(p);
605 				vm_page_protect(p, VM_PROT_NONE);
606 				vm_page_deactivate(p);
607 			} else {
608 				TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
609 					     p, pageq);
610 				TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
611 						  p, pageq);
612 				vm_page_and_queue_spin_unlock(p);
613 			}
614 		} else {
615 			vm_page_and_queue_spin_unlock(p);
616 			vm_page_activate(p);
617 			vm_page_flag_clear(p, PG_REFERENCED);
618 
619 			vm_page_and_queue_spin_lock(p);
620 			if (p->queue - p->pc == PQ_ACTIVE) {
621 				if (p->act_count < (ACT_MAX - ACT_ADVANCE))
622 					p->act_count += ACT_ADVANCE;
623 				TAILQ_REMOVE(&vm_page_queues[p->queue].pl,
624 					     p, pageq);
625 				TAILQ_INSERT_TAIL(&vm_page_queues[p->queue].pl,
626 						  p, pageq);
627 			}
628 			vm_page_and_queue_spin_unlock(p);
629 		}
630 	} else if (p->queue - p->pc == PQ_INACTIVE) {
631 		vm_page_and_queue_spin_unlock(p);
632 		vm_page_protect(p, VM_PROT_NONE);
633 	} else {
634 		vm_page_and_queue_spin_unlock(p);
635 	}
636 	vm_page_wakeup(p);
637 	return(0);
638 }
639 
640 /*
641  * Deactivate some number of pages in a map, try to do it fairly, but
642  * that is really hard to do.
643  */
644 static void
645 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t desired)
646 {
647 	vm_map_entry_t tmpe;
648 	vm_object_t obj, bigobj;
649 	int nothingwired;
650 
651 	if (lockmgr(&map->lock, LK_EXCLUSIVE | LK_NOWAIT)) {
652 		return;
653 	}
654 
655 	bigobj = NULL;
656 	nothingwired = TRUE;
657 
658 	/*
659 	 * first, search out the biggest object, and try to free pages from
660 	 * that.
661 	 */
662 	tmpe = map->header.next;
663 	while (tmpe != &map->header) {
664 		switch(tmpe->maptype) {
665 		case VM_MAPTYPE_NORMAL:
666 		case VM_MAPTYPE_VPAGETABLE:
667 			obj = tmpe->object.vm_object;
668 			if ((obj != NULL) && (obj->shadow_count <= 1) &&
669 				((bigobj == NULL) ||
670 				 (bigobj->resident_page_count < obj->resident_page_count))) {
671 				bigobj = obj;
672 			}
673 			break;
674 		default:
675 			break;
676 		}
677 		if (tmpe->wired_count > 0)
678 			nothingwired = FALSE;
679 		tmpe = tmpe->next;
680 	}
681 
682 	if (bigobj)  {
683 		vm_object_hold(bigobj);
684 		vm_pageout_object_deactivate_pages(map, bigobj, desired, 0);
685 		vm_object_drop(bigobj);
686 	}
687 
688 	/*
689 	 * Next, hunt around for other pages to deactivate.  We actually
690 	 * do this search sort of wrong -- .text first is not the best idea.
691 	 */
692 	tmpe = map->header.next;
693 	while (tmpe != &map->header) {
694 		if (pmap_resident_count(vm_map_pmap(map)) <= desired)
695 			break;
696 		switch(tmpe->maptype) {
697 		case VM_MAPTYPE_NORMAL:
698 		case VM_MAPTYPE_VPAGETABLE:
699 			obj = tmpe->object.vm_object;
700 			if (obj) {
701 				vm_object_hold(obj);
702 				vm_pageout_object_deactivate_pages(map, obj, desired, 0);
703 				vm_object_drop(obj);
704 			}
705 			break;
706 		default:
707 			break;
708 		}
709 		tmpe = tmpe->next;
710 	};
711 
712 	/*
713 	 * Remove all mappings if a process is swapped out, this will free page
714 	 * table pages.
715 	 */
716 	if (desired == 0 && nothingwired)
717 		pmap_remove(vm_map_pmap(map),
718 			    VM_MIN_USER_ADDRESS, VM_MAX_USER_ADDRESS);
719 	vm_map_unlock(map);
720 }
721 #endif
722 
723 /*
724  * Called when the pageout scan wants to free a page.  We no longer
725  * try to cycle the vm_object here with a reference & dealloc, which can
726  * cause a non-trivial object collapse in a critical path.
727  *
728  * It is unclear why we cycled the ref_count in the past, perhaps to try
729  * to optimize shadow chain collapses but I don't quite see why it would
730  * be necessary.  An OBJ_DEAD object should terminate any and all vm_pages
731  * synchronously and not have to be kicked-start.
732  */
733 static void
734 vm_pageout_page_free(vm_page_t m)
735 {
736 	vm_page_protect(m, VM_PROT_NONE);
737 	vm_page_free(m);
738 }
739 
740 /*
741  * vm_pageout_scan does the dirty work for the pageout daemon.
742  */
743 struct vm_pageout_scan_info {
744 	struct proc *bigproc;
745 	vm_offset_t bigsize;
746 };
747 
748 static int vm_pageout_scan_callback(struct proc *p, void *data);
749 
750 static int
751 vm_pageout_scan_inactive(int pass, int q, int inactive_shortage,
752 			 int *vnodes_skippedp)
753 {
754 	vm_page_t m;
755 	struct vm_page marker;
756 	struct vnode *vpfailed;		/* warning, allowed to be stale */
757 	int maxscan;
758 	int delta = 0;
759 	vm_object_t object;
760 	int actcount;
761 	int maxlaunder;
762 
763 	/*
764 	 * Start scanning the inactive queue for pages we can move to the
765 	 * cache or free.  The scan will stop when the target is reached or
766 	 * we have scanned the entire inactive queue.  Note that m->act_count
767 	 * is not used to form decisions for the inactive queue, only for the
768 	 * active queue.
769 	 *
770 	 * maxlaunder limits the number of dirty pages we flush per scan.
771 	 * For most systems a smaller value (16 or 32) is more robust under
772 	 * extreme memory and disk pressure because any unnecessary writes
773 	 * to disk can result in extreme performance degredation.  However,
774 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
775 	 * used) will die horribly with limited laundering.  If the pageout
776 	 * daemon cannot clean enough pages in the first pass, we let it go
777 	 * all out in succeeding passes.
778 	 */
779 	if ((maxlaunder = vm_max_launder) <= 1)
780 		maxlaunder = 1;
781 	if (pass)
782 		maxlaunder = 10000;
783 
784 	/*
785 	 * Initialize our marker
786 	 */
787 	bzero(&marker, sizeof(marker));
788 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
789 	marker.queue = PQ_INACTIVE + q;
790 	marker.pc = q;
791 	marker.wire_count = 1;
792 
793 	/*
794 	 * Inactive queue scan.
795 	 *
796 	 * NOTE: The vm_page must be spinlocked before the queue to avoid
797 	 *	 deadlocks, so it is easiest to simply iterate the loop
798 	 *	 with the queue unlocked at the top.
799 	 */
800 	vpfailed = NULL;
801 
802 	vm_page_queues_spin_lock(PQ_INACTIVE + q);
803 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
804 	maxscan = vmstats.v_inactive_count;
805 	vm_page_queues_spin_unlock(PQ_INACTIVE + q);
806 
807 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
808 	       maxscan-- > 0 && inactive_shortage - delta > 0)
809 	{
810 		vm_page_and_queue_spin_lock(m);
811 		if (m != TAILQ_NEXT(&marker, pageq)) {
812 			vm_page_and_queue_spin_unlock(m);
813 			++maxscan;
814 			continue;
815 		}
816 		KKASSERT(m->queue - m->pc == PQ_INACTIVE);
817 		TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
818 			     &marker, pageq);
819 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
820 				   &marker, pageq);
821 		mycpu->gd_cnt.v_pdpages++;
822 
823 		/*
824 		 * Skip marker pages
825 		 */
826 		if (m->flags & PG_MARKER) {
827 			vm_page_and_queue_spin_unlock(m);
828 			continue;
829 		}
830 
831 		/*
832 		 * Try to busy the page.  Don't mess with pages which are
833 		 * already busy or reorder them in the queue.
834 		 */
835 		if (vm_page_busy_try(m, TRUE)) {
836 			vm_page_and_queue_spin_unlock(m);
837 			continue;
838 		}
839 		vm_page_and_queue_spin_unlock(m);
840 		KKASSERT(m->queue - m->pc == PQ_INACTIVE);
841 
842 		lwkt_yield();
843 
844 		/*
845 		 * The page has been successfully busied and is now no
846 		 * longer spinlocked.  The queue is no longer spinlocked
847 		 * either.
848 		 */
849 
850 		/*
851 		 * A held page may be undergoing I/O, so skip it.
852 		 */
853 		if (m->hold_count) {
854 			vm_page_and_queue_spin_lock(m);
855 			if (m->queue - m->pc == PQ_INACTIVE) {
856 				TAILQ_REMOVE(
857 					&vm_page_queues[PQ_INACTIVE + q].pl,
858 					m, pageq);
859 				TAILQ_INSERT_TAIL(
860 					&vm_page_queues[PQ_INACTIVE + q].pl,
861 					m, pageq);
862 			}
863 			vm_page_and_queue_spin_unlock(m);
864 			++vm_swapcache_inactive_heuristic;
865 			vm_page_wakeup(m);
866 			continue;
867 		}
868 
869 		if (m->object->ref_count == 0) {
870 			/*
871 			 * If the object is not being used, we ignore previous
872 			 * references.
873 			 */
874 			vm_page_flag_clear(m, PG_REFERENCED);
875 			pmap_clear_reference(m);
876 			/* fall through to end */
877 		} else if (((m->flags & PG_REFERENCED) == 0) &&
878 			    (actcount = pmap_ts_referenced(m))) {
879 			/*
880 			 * Otherwise, if the page has been referenced while
881 			 * in the inactive queue, we bump the "activation
882 			 * count" upwards, making it less likely that the
883 			 * page will be added back to the inactive queue
884 			 * prematurely again.  Here we check the page tables
885 			 * (or emulated bits, if any), given the upper level
886 			 * VM system not knowing anything about existing
887 			 * references.
888 			 */
889 			vm_page_activate(m);
890 			m->act_count += (actcount + ACT_ADVANCE);
891 			vm_page_wakeup(m);
892 			continue;
893 		}
894 
895 		/*
896 		 * (m) is still busied.
897 		 *
898 		 * If the upper level VM system knows about any page
899 		 * references, we activate the page.  We also set the
900 		 * "activation count" higher than normal so that we will less
901 		 * likely place pages back onto the inactive queue again.
902 		 */
903 		if ((m->flags & PG_REFERENCED) != 0) {
904 			vm_page_flag_clear(m, PG_REFERENCED);
905 			actcount = pmap_ts_referenced(m);
906 			vm_page_activate(m);
907 			m->act_count += (actcount + ACT_ADVANCE + 1);
908 			vm_page_wakeup(m);
909 			continue;
910 		}
911 
912 		/*
913 		 * If the upper level VM system doesn't know anything about
914 		 * the page being dirty, we have to check for it again.  As
915 		 * far as the VM code knows, any partially dirty pages are
916 		 * fully dirty.
917 		 *
918 		 * Pages marked PG_WRITEABLE may be mapped into the user
919 		 * address space of a process running on another cpu.  A
920 		 * user process (without holding the MP lock) running on
921 		 * another cpu may be able to touch the page while we are
922 		 * trying to remove it.  vm_page_cache() will handle this
923 		 * case for us.
924 		 */
925 		if (m->dirty == 0) {
926 			vm_page_test_dirty(m);
927 		} else {
928 			vm_page_dirty(m);
929 		}
930 
931 		if (m->valid == 0) {
932 			/*
933 			 * Invalid pages can be easily freed
934 			 */
935 			vm_pageout_page_free(m);
936 			mycpu->gd_cnt.v_dfree++;
937 			++delta;
938 		} else if (m->dirty == 0) {
939 			/*
940 			 * Clean pages can be placed onto the cache queue.
941 			 * This effectively frees them.
942 			 */
943 			vm_page_cache(m);
944 			++delta;
945 		} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
946 			/*
947 			 * Dirty pages need to be paged out, but flushing
948 			 * a page is extremely expensive verses freeing
949 			 * a clean page.  Rather then artificially limiting
950 			 * the number of pages we can flush, we instead give
951 			 * dirty pages extra priority on the inactive queue
952 			 * by forcing them to be cycled through the queue
953 			 * twice before being flushed, after which the
954 			 * (now clean) page will cycle through once more
955 			 * before being freed.  This significantly extends
956 			 * the thrash point for a heavily loaded machine.
957 			 */
958 			vm_page_flag_set(m, PG_WINATCFLS);
959 			vm_page_and_queue_spin_lock(m);
960 			if (m->queue - m->pc == PQ_INACTIVE) {
961 				TAILQ_REMOVE(
962 					&vm_page_queues[PQ_INACTIVE + q].pl,
963 					m, pageq);
964 				TAILQ_INSERT_TAIL(
965 					&vm_page_queues[PQ_INACTIVE + q].pl,
966 					m, pageq);
967 			}
968 			vm_page_and_queue_spin_unlock(m);
969 			++vm_swapcache_inactive_heuristic;
970 			vm_page_wakeup(m);
971 		} else if (maxlaunder > 0) {
972 			/*
973 			 * We always want to try to flush some dirty pages if
974 			 * we encounter them, to keep the system stable.
975 			 * Normally this number is small, but under extreme
976 			 * pressure where there are insufficient clean pages
977 			 * on the inactive queue, we may have to go all out.
978 			 */
979 			int swap_pageouts_ok;
980 			struct vnode *vp = NULL;
981 
982 			object = m->object;
983 
984 			if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) {
985 				swap_pageouts_ok = 1;
986 			} else {
987 				swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts);
988 				swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts &&
989 				vm_page_count_min(0));
990 
991 			}
992 
993 			/*
994 			 * We don't bother paging objects that are "dead".
995 			 * Those objects are in a "rundown" state.
996 			 */
997 			if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) {
998 				vm_page_and_queue_spin_lock(m);
999 				if (m->queue - m->pc == PQ_INACTIVE) {
1000 					TAILQ_REMOVE(
1001 					    &vm_page_queues[PQ_INACTIVE + q].pl,
1002 					    m, pageq);
1003 					TAILQ_INSERT_TAIL(
1004 					    &vm_page_queues[PQ_INACTIVE + q].pl,
1005 					    m, pageq);
1006 				}
1007 				vm_page_and_queue_spin_unlock(m);
1008 				++vm_swapcache_inactive_heuristic;
1009 				vm_page_wakeup(m);
1010 				continue;
1011 			}
1012 
1013 			/*
1014 			 * (m) is still busied.
1015 			 *
1016 			 * The object is already known NOT to be dead.   It
1017 			 * is possible for the vget() to block the whole
1018 			 * pageout daemon, but the new low-memory handling
1019 			 * code should prevent it.
1020 			 *
1021 			 * The previous code skipped locked vnodes and, worse,
1022 			 * reordered pages in the queue.  This results in
1023 			 * completely non-deterministic operation because,
1024 			 * quite often, a vm_fault has initiated an I/O and
1025 			 * is holding a locked vnode at just the point where
1026 			 * the pageout daemon is woken up.
1027 			 *
1028 			 * We can't wait forever for the vnode lock, we might
1029 			 * deadlock due to a vn_read() getting stuck in
1030 			 * vm_wait while holding this vnode.  We skip the
1031 			 * vnode if we can't get it in a reasonable amount
1032 			 * of time.
1033 			 *
1034 			 * vpfailed is used to (try to) avoid the case where
1035 			 * a large number of pages are associated with a
1036 			 * locked vnode, which could cause the pageout daemon
1037 			 * to stall for an excessive amount of time.
1038 			 */
1039 			if (object->type == OBJT_VNODE) {
1040 				int flags;
1041 
1042 				vp = object->handle;
1043 				flags = LK_EXCLUSIVE | LK_NOOBJ;
1044 				if (vp == vpfailed)
1045 					flags |= LK_NOWAIT;
1046 				else
1047 					flags |= LK_TIMELOCK;
1048 				vm_page_hold(m);
1049 				vm_page_wakeup(m);
1050 
1051 				/*
1052 				 * We have unbusied (m) temporarily so we can
1053 				 * acquire the vp lock without deadlocking.
1054 				 * (m) is held to prevent destruction.
1055 				 */
1056 				if (vget(vp, flags) != 0) {
1057 					vpfailed = vp;
1058 					++pageout_lock_miss;
1059 					if (object->flags & OBJ_MIGHTBEDIRTY)
1060 						    ++*vnodes_skippedp;
1061 					vm_page_unhold(m);
1062 					continue;
1063 				}
1064 
1065 				/*
1066 				 * The page might have been moved to another
1067 				 * queue during potential blocking in vget()
1068 				 * above.  The page might have been freed and
1069 				 * reused for another vnode.  The object might
1070 				 * have been reused for another vnode.
1071 				 */
1072 				if (m->queue - m->pc != PQ_INACTIVE ||
1073 				    m->object != object ||
1074 				    object->handle != vp) {
1075 					if (object->flags & OBJ_MIGHTBEDIRTY)
1076 						++*vnodes_skippedp;
1077 					vput(vp);
1078 					vm_page_unhold(m);
1079 					continue;
1080 				}
1081 
1082 				/*
1083 				 * The page may have been busied during the
1084 				 * blocking in vput();  We don't move the
1085 				 * page back onto the end of the queue so that
1086 				 * statistics are more correct if we don't.
1087 				 */
1088 				if (vm_page_busy_try(m, TRUE)) {
1089 					vput(vp);
1090 					vm_page_unhold(m);
1091 					continue;
1092 				}
1093 				vm_page_unhold(m);
1094 
1095 				/*
1096 				 * (m) is busied again
1097 				 *
1098 				 * We own the busy bit and remove our hold
1099 				 * bit.  If the page is still held it
1100 				 * might be undergoing I/O, so skip it.
1101 				 */
1102 				if (m->hold_count) {
1103 					vm_page_and_queue_spin_lock(m);
1104 					if (m->queue - m->pc == PQ_INACTIVE) {
1105 						TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1106 						TAILQ_INSERT_TAIL(&vm_page_queues[PQ_INACTIVE + q].pl, m, pageq);
1107 					}
1108 					vm_page_and_queue_spin_unlock(m);
1109 					++vm_swapcache_inactive_heuristic;
1110 					if (object->flags & OBJ_MIGHTBEDIRTY)
1111 						++*vnodes_skippedp;
1112 					vm_page_wakeup(m);
1113 					vput(vp);
1114 					continue;
1115 				}
1116 				/* (m) is left busied as we fall through */
1117 			}
1118 
1119 			/*
1120 			 * page is busy and not held here.
1121 			 *
1122 			 * If a page is dirty, then it is either being washed
1123 			 * (but not yet cleaned) or it is still in the
1124 			 * laundry.  If it is still in the laundry, then we
1125 			 * start the cleaning operation.
1126 			 *
1127 			 * decrement inactive_shortage on success to account
1128 			 * for the (future) cleaned page.  Otherwise we
1129 			 * could wind up laundering or cleaning too many
1130 			 * pages.
1131 			 */
1132 			if (vm_pageout_clean(m) != 0) {
1133 				++delta;
1134 				--maxlaunder;
1135 			}
1136 			/* clean ate busy, page no longer accessible */
1137 			if (vp != NULL)
1138 				vput(vp);
1139 		} else {
1140 			vm_page_wakeup(m);
1141 		}
1142 	}
1143 	vm_page_queues_spin_lock(PQ_INACTIVE + q);
1144 	TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
1145 	vm_page_queues_spin_unlock(PQ_INACTIVE + q);
1146 
1147 	return (delta);
1148 }
1149 
1150 static int
1151 vm_pageout_scan_active(int pass, int q,
1152 		       int inactive_shortage, int active_shortage,
1153 		       int *recycle_countp)
1154 {
1155 	struct vm_page marker;
1156 	vm_page_t m;
1157 	int actcount;
1158 	int delta = 0;
1159 	int pcount;
1160 
1161 	/*
1162 	 * We want to move pages from the active queue to the inactive
1163 	 * queue to get the inactive queue to the inactive target.  If
1164 	 * we still have a page shortage from above we try to directly free
1165 	 * clean pages instead of moving them.
1166 	 *
1167 	 * If we do still have a shortage we keep track of the number of
1168 	 * pages we free or cache (recycle_count) as a measure of thrashing
1169 	 * between the active and inactive queues.
1170 	 *
1171 	 * If we were able to completely satisfy the free+cache targets
1172 	 * from the inactive pool we limit the number of pages we move
1173 	 * from the active pool to the inactive pool to 2x the pages we
1174 	 * had removed from the inactive pool (with a minimum of 1/5 the
1175 	 * inactive target).  If we were not able to completely satisfy
1176 	 * the free+cache targets we go for the whole target aggressively.
1177 	 *
1178 	 * NOTE: Both variables can end up negative.
1179 	 * NOTE: We are still in a critical section.
1180 	 */
1181 
1182 	bzero(&marker, sizeof(marker));
1183 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1184 	marker.queue = PQ_ACTIVE + q;
1185 	marker.pc = q;
1186 	marker.wire_count = 1;
1187 
1188 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1189 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1190 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1191 	pcount = vmstats.v_active_count;
1192 
1193 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1194 	       pcount-- > 0 && (inactive_shortage - delta > 0 ||
1195 				active_shortage > 0))
1196 	{
1197 		vm_page_and_queue_spin_lock(m);
1198 		if (m != TAILQ_NEXT(&marker, pageq)) {
1199 			vm_page_and_queue_spin_unlock(m);
1200 			++pcount;
1201 			continue;
1202 		}
1203 		KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1204 		TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1205 			     &marker, pageq);
1206 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1207 				   &marker, pageq);
1208 
1209 		/*
1210 		 * Skip marker pages
1211 		 */
1212 		if (m->flags & PG_MARKER) {
1213 			vm_page_and_queue_spin_unlock(m);
1214 			continue;
1215 		}
1216 
1217 		/*
1218 		 * Try to busy the page.  Don't mess with pages which are
1219 		 * already busy or reorder them in the queue.
1220 		 */
1221 		if (vm_page_busy_try(m, TRUE)) {
1222 			vm_page_and_queue_spin_unlock(m);
1223 			continue;
1224 		}
1225 
1226 		/*
1227 		 * Don't deactivate pages that are held, even if we can
1228 		 * busy them.  (XXX why not?)
1229 		 */
1230 		if (m->hold_count != 0) {
1231 			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1232 				     m, pageq);
1233 			TAILQ_INSERT_TAIL(&vm_page_queues[PQ_ACTIVE + q].pl,
1234 					  m, pageq);
1235 			vm_page_and_queue_spin_unlock(m);
1236 			vm_page_wakeup(m);
1237 			continue;
1238 		}
1239 		vm_page_and_queue_spin_unlock(m);
1240 		lwkt_yield();
1241 
1242 		/*
1243 		 * The page has been successfully busied and the page and
1244 		 * queue are no longer locked.
1245 		 */
1246 
1247 		/*
1248 		 * The count for pagedaemon pages is done after checking the
1249 		 * page for eligibility...
1250 		 */
1251 		mycpu->gd_cnt.v_pdpages++;
1252 
1253 		/*
1254 		 * Check to see "how much" the page has been used and clear
1255 		 * the tracking access bits.  If the object has no references
1256 		 * don't bother paying the expense.
1257 		 */
1258 		actcount = 0;
1259 		if (m->object->ref_count != 0) {
1260 			if (m->flags & PG_REFERENCED)
1261 				++actcount;
1262 			actcount += pmap_ts_referenced(m);
1263 			if (actcount) {
1264 				m->act_count += ACT_ADVANCE + actcount;
1265 				if (m->act_count > ACT_MAX)
1266 					m->act_count = ACT_MAX;
1267 			}
1268 		}
1269 		vm_page_flag_clear(m, PG_REFERENCED);
1270 
1271 		/*
1272 		 * actcount is only valid if the object ref_count is non-zero.
1273 		 */
1274 		if (actcount && m->object->ref_count != 0) {
1275 			vm_page_and_queue_spin_lock(m);
1276 			if (m->queue - m->pc == PQ_ACTIVE) {
1277 				TAILQ_REMOVE(
1278 					&vm_page_queues[PQ_ACTIVE + q].pl,
1279 					m, pageq);
1280 				TAILQ_INSERT_TAIL(
1281 					&vm_page_queues[PQ_ACTIVE + q].pl,
1282 					m, pageq);
1283 			}
1284 			vm_page_and_queue_spin_unlock(m);
1285 			vm_page_wakeup(m);
1286 		} else {
1287 			m->act_count -= min(m->act_count, ACT_DECLINE);
1288 			if (vm_pageout_algorithm ||
1289 			    m->object->ref_count == 0 ||
1290 			    m->act_count < pass + 1
1291 			) {
1292 				/*
1293 				 * Deactivate the page.  If we had a
1294 				 * shortage from our inactive scan try to
1295 				 * free (cache) the page instead.
1296 				 *
1297 				 * Don't just blindly cache the page if
1298 				 * we do not have a shortage from the
1299 				 * inactive scan, that could lead to
1300 				 * gigabytes being moved.
1301 				 */
1302 				--active_shortage;
1303 				if (inactive_shortage - delta > 0 ||
1304 				    m->object->ref_count == 0) {
1305 					if (inactive_shortage - delta > 0)
1306 						++*recycle_countp;
1307 					vm_page_protect(m, VM_PROT_NONE);
1308 					if (m->dirty == 0 &&
1309 					    inactive_shortage - delta > 0) {
1310 						++delta;
1311 						vm_page_cache(m);
1312 					} else {
1313 						vm_page_deactivate(m);
1314 						vm_page_wakeup(m);
1315 					}
1316 				} else {
1317 					vm_page_deactivate(m);
1318 					vm_page_wakeup(m);
1319 				}
1320 			} else {
1321 				vm_page_and_queue_spin_lock(m);
1322 				if (m->queue - m->pc == PQ_ACTIVE) {
1323 					TAILQ_REMOVE(
1324 					    &vm_page_queues[PQ_ACTIVE + q].pl,
1325 					    m, pageq);
1326 					TAILQ_INSERT_TAIL(
1327 					    &vm_page_queues[PQ_ACTIVE + q].pl,
1328 					    m, pageq);
1329 				}
1330 				vm_page_and_queue_spin_unlock(m);
1331 				vm_page_wakeup(m);
1332 			}
1333 		}
1334 	}
1335 
1336 	/*
1337 	 * Clean out our local marker.
1338 	 */
1339 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1340 	TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1341 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1342 
1343 	return (delta);
1344 }
1345 
1346 /*
1347  * The number of actually free pages can drop down to v_free_reserved,
1348  * we try to build the free count back above v_free_min.  Note that
1349  * vm_paging_needed() also returns TRUE if v_free_count is not at
1350  * least v_free_min so that is the minimum we must build the free
1351  * count to.
1352  *
1353  * We use a slightly higher target to improve hysteresis,
1354  * ((v_free_target + v_free_min) / 2).  Since v_free_target
1355  * is usually the same as v_cache_min this maintains about
1356  * half the pages in the free queue as are in the cache queue,
1357  * providing pretty good pipelining for pageout operation.
1358  *
1359  * The system operator can manipulate vm.v_cache_min and
1360  * vm.v_free_target to tune the pageout demon.  Be sure
1361  * to keep vm.v_free_min < vm.v_free_target.
1362  *
1363  * Note that the original paging target is to get at least
1364  * (free_min + cache_min) into (free + cache).  The slightly
1365  * higher target will shift additional pages from cache to free
1366  * without effecting the original paging target in order to
1367  * maintain better hysteresis and not have the free count always
1368  * be dead-on v_free_min.
1369  *
1370  * NOTE: we are still in a critical section.
1371  *
1372  * Pages moved from PQ_CACHE to totally free are not counted in the
1373  * pages_freed counter.
1374  */
1375 static void
1376 vm_pageout_scan_cache(int inactive_shortage,
1377 		      int vnodes_skipped, int recycle_count)
1378 {
1379 	struct vm_pageout_scan_info info;
1380 	vm_page_t m;
1381 
1382 	while (vmstats.v_free_count <
1383 	       (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1384 		/*
1385 		 * This steals some code from vm/vm_page.c
1386 		 */
1387 		static int cache_rover = 0;
1388 
1389 		m = vm_page_list_find(PQ_CACHE, cache_rover & PQ_L2_MASK, FALSE);
1390 		if (m == NULL)
1391 			break;
1392 		/* page is returned removed from its queue and spinlocked */
1393 		if (vm_page_busy_try(m, TRUE)) {
1394 			vm_page_deactivate_locked(m);
1395 			vm_page_spin_unlock(m);
1396 #ifdef INVARIANTS
1397 			kprintf("Warning: busy page %p found in cache\n", m);
1398 #endif
1399 			continue;
1400 		}
1401 		vm_page_spin_unlock(m);
1402 		pagedaemon_wakeup();
1403 		lwkt_yield();
1404 
1405 		/*
1406 		 * Page has been successfully busied and it and its queue
1407 		 * is no longer spinlocked.
1408 		 */
1409 		if ((m->flags & PG_UNMANAGED) ||
1410 		    m->hold_count ||
1411 		    m->wire_count) {
1412 			vm_page_deactivate(m);
1413 			vm_page_wakeup(m);
1414 			continue;
1415 		}
1416 		KKASSERT((m->flags & PG_MAPPED) == 0);
1417 		KKASSERT(m->dirty == 0);
1418 		cache_rover += PQ_PRIME2;
1419 		vm_pageout_page_free(m);
1420 		mycpu->gd_cnt.v_dfree++;
1421 	}
1422 
1423 #if !defined(NO_SWAPPING)
1424 	/*
1425 	 * Idle process swapout -- run once per second.
1426 	 */
1427 	if (vm_swap_idle_enabled) {
1428 		static long lsec;
1429 		if (time_second != lsec) {
1430 			vm_pageout_req_swapout |= VM_SWAP_IDLE;
1431 			vm_req_vmdaemon();
1432 			lsec = time_second;
1433 		}
1434 	}
1435 #endif
1436 
1437 	/*
1438 	 * If we didn't get enough free pages, and we have skipped a vnode
1439 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1440 	 * if we did not get enough free pages.
1441 	 */
1442 	if (vm_paging_target() > 0) {
1443 		if (vnodes_skipped && vm_page_count_min(0))
1444 			speedup_syncer();
1445 #if !defined(NO_SWAPPING)
1446 		if (vm_swap_enabled && vm_page_count_target()) {
1447 			vm_req_vmdaemon();
1448 			vm_pageout_req_swapout |= VM_SWAP_NORMAL;
1449 		}
1450 #endif
1451 	}
1452 
1453 	/*
1454 	 * Handle catastrophic conditions.  Under good conditions we should
1455 	 * be at the target, well beyond our minimum.  If we could not even
1456 	 * reach our minimum the system is under heavy stress.
1457 	 *
1458 	 * Determine whether we have run out of memory.  This occurs when
1459 	 * swap_pager_full is TRUE and the only pages left in the page
1460 	 * queues are dirty.  We will still likely have page shortages.
1461 	 *
1462 	 * - swap_pager_full is set if insufficient swap was
1463 	 *   available to satisfy a requested pageout.
1464 	 *
1465 	 * - the inactive queue is bloated (4 x size of active queue),
1466 	 *   meaning it is unable to get rid of dirty pages and.
1467 	 *
1468 	 * - vm_page_count_min() without counting pages recycled from the
1469 	 *   active queue (recycle_count) means we could not recover
1470 	 *   enough pages to meet bare minimum needs.  This test only
1471 	 *   works if the inactive queue is bloated.
1472 	 *
1473 	 * - due to a positive inactive_shortage we shifted the remaining
1474 	 *   dirty pages from the active queue to the inactive queue
1475 	 *   trying to find clean ones to free.
1476 	 */
1477 	if (swap_pager_full && vm_page_count_min(recycle_count))
1478 		kprintf("Warning: system low on memory+swap!\n");
1479 	if (swap_pager_full && vm_page_count_min(recycle_count) &&
1480 	    vmstats.v_inactive_count > vmstats.v_active_count * 4 &&
1481 	    inactive_shortage > 0) {
1482 		/*
1483 		 * Kill something.
1484 		 */
1485 		info.bigproc = NULL;
1486 		info.bigsize = 0;
1487 		allproc_scan(vm_pageout_scan_callback, &info);
1488 		if (info.bigproc != NULL) {
1489 			killproc(info.bigproc, "out of swap space");
1490 			info.bigproc->p_nice = PRIO_MIN;
1491 			info.bigproc->p_usched->resetpriority(
1492 				FIRST_LWP_IN_PROC(info.bigproc));
1493 			wakeup(&vmstats.v_free_count);
1494 			PRELE(info.bigproc);
1495 		}
1496 	}
1497 }
1498 
1499 /*
1500  * The caller must hold proc_token.
1501  */
1502 static int
1503 vm_pageout_scan_callback(struct proc *p, void *data)
1504 {
1505 	struct vm_pageout_scan_info *info = data;
1506 	vm_offset_t size;
1507 
1508 	/*
1509 	 * Never kill system processes or init.  If we have configured swap
1510 	 * then try to avoid killing low-numbered pids.
1511 	 */
1512 	if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) ||
1513 	    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1514 		return (0);
1515 	}
1516 
1517 	/*
1518 	 * if the process is in a non-running type state,
1519 	 * don't touch it.
1520 	 */
1521 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
1522 		return (0);
1523 
1524 	/*
1525 	 * Get the approximate process size.  Note that anonymous pages
1526 	 * with backing swap will be counted twice, but there should not
1527 	 * be too many such pages due to the stress the VM system is
1528 	 * under at this point.
1529 	 */
1530 	size = vmspace_anonymous_count(p->p_vmspace) +
1531 		vmspace_swap_count(p->p_vmspace);
1532 
1533 	/*
1534 	 * If the this process is bigger than the biggest one
1535 	 * remember it.
1536 	 */
1537 	if (info->bigsize < size) {
1538 		if (info->bigproc)
1539 			PRELE(info->bigproc);
1540 		PHOLD(p);
1541 		info->bigproc = p;
1542 		info->bigsize = size;
1543 	}
1544 	lwkt_yield();
1545 	return(0);
1546 }
1547 
1548 /*
1549  * This routine tries to maintain the pseudo LRU active queue,
1550  * so that during long periods of time where there is no paging,
1551  * that some statistic accumulation still occurs.  This code
1552  * helps the situation where paging just starts to occur.
1553  */
1554 static void
1555 vm_pageout_page_stats(int q)
1556 {
1557 	static int fullintervalcount = 0;
1558 	struct vm_page marker;
1559 	vm_page_t m;
1560 	int pcount, tpcount;		/* Number of pages to check */
1561 	int page_shortage;
1562 
1563 	page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1564 			 vmstats.v_free_min) -
1565 			(vmstats.v_free_count + vmstats.v_inactive_count +
1566 			 vmstats.v_cache_count);
1567 
1568 	if (page_shortage <= 0)
1569 		return;
1570 
1571 	pcount = vmstats.v_active_count;
1572 	fullintervalcount += vm_pageout_stats_interval;
1573 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1574 		tpcount = (vm_pageout_stats_max * vmstats.v_active_count) /
1575 			  vmstats.v_page_count;
1576 		if (pcount > tpcount)
1577 			pcount = tpcount;
1578 	} else {
1579 		fullintervalcount = 0;
1580 	}
1581 
1582 	bzero(&marker, sizeof(marker));
1583 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1584 	marker.queue = PQ_ACTIVE + q;
1585 	marker.pc = q;
1586 	marker.wire_count = 1;
1587 
1588 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1589 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1590 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1591 
1592 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1593 	       pcount-- > 0)
1594 	{
1595 		int actcount;
1596 
1597 		vm_page_and_queue_spin_lock(m);
1598 		if (m != TAILQ_NEXT(&marker, pageq)) {
1599 			vm_page_and_queue_spin_unlock(m);
1600 			++pcount;
1601 			continue;
1602 		}
1603 		KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1604 		TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1605 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1606 				   &marker, pageq);
1607 
1608 		/*
1609 		 * Ignore markers
1610 		 */
1611 		if (m->flags & PG_MARKER) {
1612 			vm_page_and_queue_spin_unlock(m);
1613 			continue;
1614 		}
1615 
1616 		/*
1617 		 * Ignore pages we can't busy
1618 		 */
1619 		if (vm_page_busy_try(m, TRUE)) {
1620 			vm_page_and_queue_spin_unlock(m);
1621 			continue;
1622 		}
1623 		vm_page_and_queue_spin_unlock(m);
1624 		KKASSERT(m->queue - m->pc == PQ_ACTIVE);
1625 
1626 		/*
1627 		 * We now have a safely busied page, the page and queue
1628 		 * spinlocks have been released.
1629 		 *
1630 		 * Ignore held pages
1631 		 */
1632 		if (m->hold_count) {
1633 			vm_page_wakeup(m);
1634 			continue;
1635 		}
1636 
1637 		/*
1638 		 * Calculate activity
1639 		 */
1640 		actcount = 0;
1641 		if (m->flags & PG_REFERENCED) {
1642 			vm_page_flag_clear(m, PG_REFERENCED);
1643 			actcount += 1;
1644 		}
1645 		actcount += pmap_ts_referenced(m);
1646 
1647 		/*
1648 		 * Update act_count and move page to end of queue.
1649 		 */
1650 		if (actcount) {
1651 			m->act_count += ACT_ADVANCE + actcount;
1652 			if (m->act_count > ACT_MAX)
1653 				m->act_count = ACT_MAX;
1654 			vm_page_and_queue_spin_lock(m);
1655 			if (m->queue - m->pc == PQ_ACTIVE) {
1656 				TAILQ_REMOVE(
1657 					&vm_page_queues[PQ_ACTIVE + q].pl,
1658 					m, pageq);
1659 				TAILQ_INSERT_TAIL(
1660 					&vm_page_queues[PQ_ACTIVE + q].pl,
1661 					m, pageq);
1662 			}
1663 			vm_page_and_queue_spin_unlock(m);
1664 			vm_page_wakeup(m);
1665 			continue;
1666 		}
1667 
1668 		if (m->act_count == 0) {
1669 			/*
1670 			 * We turn off page access, so that we have
1671 			 * more accurate RSS stats.  We don't do this
1672 			 * in the normal page deactivation when the
1673 			 * system is loaded VM wise, because the
1674 			 * cost of the large number of page protect
1675 			 * operations would be higher than the value
1676 			 * of doing the operation.
1677 			 *
1678 			 * We use the marker to save our place so
1679 			 * we can release the spin lock.  both (m)
1680 			 * and (next) will be invalid.
1681 			 */
1682 			vm_page_protect(m, VM_PROT_NONE);
1683 			vm_page_deactivate(m);
1684 		} else {
1685 			m->act_count -= min(m->act_count, ACT_DECLINE);
1686 			vm_page_and_queue_spin_lock(m);
1687 			if (m->queue - m->pc == PQ_ACTIVE) {
1688 				TAILQ_REMOVE(
1689 					&vm_page_queues[PQ_ACTIVE + q].pl,
1690 					m, pageq);
1691 				TAILQ_INSERT_TAIL(
1692 					&vm_page_queues[PQ_ACTIVE + q].pl,
1693 					m, pageq);
1694 			}
1695 			vm_page_and_queue_spin_unlock(m);
1696 		}
1697 		vm_page_wakeup(m);
1698 	}
1699 
1700 	/*
1701 	 * Remove our local marker
1702 	 */
1703 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1704 	TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1705 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1706 }
1707 
1708 static int
1709 vm_pageout_free_page_calc(vm_size_t count)
1710 {
1711 	if (count < vmstats.v_page_count)
1712 		 return 0;
1713 	/*
1714 	 * free_reserved needs to include enough for the largest swap pager
1715 	 * structures plus enough for any pv_entry structs when paging.
1716 	 *
1717 	 * v_free_min		normal allocations
1718 	 * v_free_reserved	system allocations
1719 	 * v_pageout_free_min	allocations by pageout daemon
1720 	 * v_interrupt_free_min	low level allocations (e.g swap structures)
1721 	 */
1722 	if (vmstats.v_page_count > 1024)
1723 		vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1724 	else
1725 		vmstats.v_free_min = 64;
1726 	vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1727 	vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1728 	vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1729 	vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1730 
1731 	return 1;
1732 }
1733 
1734 
1735 /*
1736  * vm_pageout is the high level pageout daemon.
1737  *
1738  * No requirements.
1739  */
1740 static void
1741 vm_pageout_thread(void)
1742 {
1743 	int pass;
1744 	int q;
1745 
1746 	/*
1747 	 * Initialize some paging parameters.
1748 	 */
1749 	curthread->td_flags |= TDF_SYSTHREAD;
1750 
1751 	if (vmstats.v_page_count < 2000)
1752 		vm_pageout_page_count = 8;
1753 
1754 	vm_pageout_free_page_calc(vmstats.v_page_count);
1755 
1756 	/*
1757 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1758 	 * that these are more a measure of the VM cache queue hysteresis
1759 	 * then the VM free queue.  Specifically, v_free_target is the
1760 	 * high water mark (free+cache pages).
1761 	 *
1762 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1763 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1764 	 * be big enough to handle memory needs while the pageout daemon
1765 	 * is signalled and run to free more pages.
1766 	 */
1767 	if (vmstats.v_free_count > 6144)
1768 		vmstats.v_free_target = 4 * vmstats.v_free_min + vmstats.v_free_reserved;
1769 	else
1770 		vmstats.v_free_target = 2 * vmstats.v_free_min + vmstats.v_free_reserved;
1771 
1772 	/*
1773 	 * NOTE: With the new buffer cache b_act_count we want the default
1774 	 *	 inactive target to be a percentage of available memory.
1775 	 *
1776 	 *	 The inactive target essentially determines the minimum
1777 	 *	 number of 'temporary' pages capable of caching one-time-use
1778 	 *	 files when the VM system is otherwise full of pages
1779 	 *	 belonging to multi-time-use files or active program data.
1780 	 *
1781 	 * NOTE: The inactive target is aggressively persued only if the
1782 	 *	 inactive queue becomes too small.  If the inactive queue
1783 	 *	 is large enough to satisfy page movement to free+cache
1784 	 *	 then it is repopulated more slowly from the active queue.
1785 	 *	 This allows a general inactive_target default to be set.
1786 	 *
1787 	 *	 There is an issue here for processes which sit mostly idle
1788 	 *	 'overnight', such as sshd, tcsh, and X.  Any movement from
1789 	 *	 the active queue will eventually cause such pages to
1790 	 *	 recycle eventually causing a lot of paging in the morning.
1791 	 *	 To reduce the incidence of this pages cycled out of the
1792 	 *	 buffer cache are moved directly to the inactive queue if
1793 	 *	 they were only used once or twice.
1794 	 *
1795 	 *	 The vfs.vm_cycle_point sysctl can be used to adjust this.
1796 	 *	 Increasing the value (up to 64) increases the number of
1797 	 *	 buffer recyclements which go directly to the inactive queue.
1798 	 */
1799 	if (vmstats.v_free_count > 2048) {
1800 		vmstats.v_cache_min = vmstats.v_free_target;
1801 		vmstats.v_cache_max = 2 * vmstats.v_cache_min;
1802 	} else {
1803 		vmstats.v_cache_min = 0;
1804 		vmstats.v_cache_max = 0;
1805 	}
1806 	vmstats.v_inactive_target = vmstats.v_free_count / 4;
1807 
1808 	/* XXX does not really belong here */
1809 	if (vm_page_max_wired == 0)
1810 		vm_page_max_wired = vmstats.v_free_count / 3;
1811 
1812 	if (vm_pageout_stats_max == 0)
1813 		vm_pageout_stats_max = vmstats.v_free_target;
1814 
1815 	/*
1816 	 * Set interval in seconds for stats scan.
1817 	 */
1818 	if (vm_pageout_stats_interval == 0)
1819 		vm_pageout_stats_interval = 5;
1820 	if (vm_pageout_full_stats_interval == 0)
1821 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
1822 
1823 
1824 	/*
1825 	 * Set maximum free per pass
1826 	 */
1827 	if (vm_pageout_stats_free_max == 0)
1828 		vm_pageout_stats_free_max = 5;
1829 
1830 	swap_pager_swap_init();
1831 	pass = 0;
1832 
1833 	/*
1834 	 * The pageout daemon is never done, so loop forever.
1835 	 */
1836 	while (TRUE) {
1837 		int error;
1838 		int delta1;
1839 		int delta2;
1840 		int inactive_shortage;
1841 		int active_shortage;
1842 		int vnodes_skipped = 0;
1843 		int recycle_count = 0;
1844 		int tmp;
1845 
1846 		/*
1847 		 * Wait for an action request.  If we timeout check to
1848 		 * see if paging is needed (in case the normal wakeup
1849 		 * code raced us).
1850 		 */
1851 		if (vm_pages_needed == 0) {
1852 			error = tsleep(&vm_pages_needed,
1853 				       0, "psleep",
1854 				       vm_pageout_stats_interval * hz);
1855 			if (error &&
1856 			    vm_paging_needed() == 0 &&
1857 			    vm_pages_needed == 0) {
1858 				for (q = 0; q < PQ_MAXL2_SIZE; ++q)
1859 					vm_pageout_page_stats(q);
1860 				continue;
1861 			}
1862 			vm_pages_needed = 1;
1863 		}
1864 
1865 		mycpu->gd_cnt.v_pdwakeups++;
1866 
1867 		/*
1868 		 * Do whatever cleanup that the pmap code can.
1869 		 */
1870 		pmap_collect();
1871 
1872 		/*
1873 		 * Scan for pageout.  Try to avoid thrashing the system
1874 		 * with activity.
1875 		 *
1876 		 * Calculate our target for the number of free+cache pages we
1877 		 * want to get to.  This is higher then the number that causes
1878 		 * allocations to stall (severe) in order to provide hysteresis,
1879 		 * and if we don't make it all the way but get to the minimum
1880 		 * we're happy.
1881 		 */
1882 		inactive_shortage = vm_paging_target() + vm_pageout_deficit;
1883 		vm_pageout_deficit = 0;
1884 		delta1 = 0;
1885 		for (q = 0; q < PQ_MAXL2_SIZE; ++q) {
1886 			delta1 += vm_pageout_scan_inactive(
1887 					pass, q,
1888 					inactive_shortage / PQ_MAXL2_SIZE + 1,
1889 					&vnodes_skipped);
1890 		}
1891 
1892 		/*
1893 		 * Figure out how many active pages we must deactivate.  If
1894 		 * we were able to reach our target with just the inactive
1895 		 * scan above we limit the number of active pages we
1896 		 * deactivate to reduce unnecessary work.
1897 		 */
1898 		active_shortage = vmstats.v_inactive_target -
1899 				  vmstats.v_inactive_count;
1900 
1901 		tmp = inactive_shortage;
1902 		if (tmp < vmstats.v_inactive_target / 10)
1903 			tmp = vmstats.v_inactive_target / 10;
1904 		inactive_shortage -= delta1;
1905 		if (inactive_shortage <= 0 && active_shortage > tmp * 2)
1906 			active_shortage = tmp * 2;
1907 
1908 		delta2 = 0;
1909 		for (q = 0; q < PQ_MAXL2_SIZE; ++q) {
1910 			delta2 += vm_pageout_scan_active(
1911 					pass, q,
1912 					inactive_shortage / PQ_MAXL2_SIZE + 1,
1913 					active_shortage / PQ_MAXL2_SIZE + 1,
1914 					&recycle_count);
1915 		}
1916 
1917 		/*
1918 		 * Finally free enough cache pages to meet our free page
1919 		 * requirement and take more drastic measures if we are
1920 		 * still in trouble.
1921 		 */
1922 		inactive_shortage -= delta2;
1923 		vm_pageout_scan_cache(inactive_shortage, vnodes_skipped,
1924 				      recycle_count);
1925 
1926 		/*
1927 		 * Wait for more work.
1928 		 */
1929 		if (inactive_shortage > 0) {
1930 			++pass;
1931 			if (swap_pager_full) {
1932 				/*
1933 				 * Running out of memory, catastrophic back-off
1934 				 * to one-second intervals.
1935 				 */
1936 				tsleep(&vm_pages_needed, 0, "pdelay", hz);
1937 			} else if (pass < 10 && vm_pages_needed > 1) {
1938 				/*
1939 				 * Normal operation, additional processes
1940 				 * have already kicked us.  Retry immediately.
1941 				 */
1942 			} else if (pass < 10) {
1943 				/*
1944 				 * Normal operation, fewer processes.  Delay
1945 				 * a bit but allow wakeups.
1946 				 */
1947 				vm_pages_needed = 0;
1948 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1949 				vm_pages_needed = 1;
1950 			} else {
1951 				/*
1952 				 * We've taken too many passes, forced delay.
1953 				 */
1954 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
1955 			}
1956 		} else {
1957 			/*
1958 			 * Interlocked wakeup of waiters (non-optional)
1959 			 */
1960 			pass = 0;
1961 			if (vm_pages_needed && !vm_page_count_min(0)) {
1962 				wakeup(&vmstats.v_free_count);
1963 				vm_pages_needed = 0;
1964 			}
1965 		}
1966 	}
1967 }
1968 
1969 static struct kproc_desc page_kp = {
1970 	"pagedaemon",
1971 	vm_pageout_thread,
1972 	&pagethread
1973 };
1974 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp)
1975 
1976 
1977 /*
1978  * Called after allocating a page out of the cache or free queue
1979  * to possibly wake the pagedaemon up to replentish our supply.
1980  *
1981  * We try to generate some hysteresis by waking the pagedaemon up
1982  * when our free+cache pages go below the free_min+cache_min level.
1983  * The pagedaemon tries to get the count back up to at least the
1984  * minimum, and through to the target level if possible.
1985  *
1986  * If the pagedaemon is already active bump vm_pages_needed as a hint
1987  * that there are even more requests pending.
1988  *
1989  * SMP races ok?
1990  * No requirements.
1991  */
1992 void
1993 pagedaemon_wakeup(void)
1994 {
1995 	if (vm_paging_needed() && curthread != pagethread) {
1996 		if (vm_pages_needed == 0) {
1997 			vm_pages_needed = 1;	/* SMP race ok */
1998 			wakeup(&vm_pages_needed);
1999 		} else if (vm_page_count_min(0)) {
2000 			++vm_pages_needed;	/* SMP race ok */
2001 		}
2002 	}
2003 }
2004 
2005 #if !defined(NO_SWAPPING)
2006 
2007 /*
2008  * SMP races ok?
2009  * No requirements.
2010  */
2011 static void
2012 vm_req_vmdaemon(void)
2013 {
2014 	static int lastrun = 0;
2015 
2016 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2017 		wakeup(&vm_daemon_needed);
2018 		lastrun = ticks;
2019 	}
2020 }
2021 
2022 static int vm_daemon_callback(struct proc *p, void *data __unused);
2023 
2024 /*
2025  * No requirements.
2026  */
2027 static void
2028 vm_daemon(void)
2029 {
2030 	/*
2031 	 * XXX vm_daemon_needed specific token?
2032 	 */
2033 	while (TRUE) {
2034 		tsleep(&vm_daemon_needed, 0, "psleep", 0);
2035 		if (vm_pageout_req_swapout) {
2036 			swapout_procs(vm_pageout_req_swapout);
2037 			vm_pageout_req_swapout = 0;
2038 		}
2039 		/*
2040 		 * scan the processes for exceeding their rlimits or if
2041 		 * process is swapped out -- deactivate pages
2042 		 */
2043 		allproc_scan(vm_daemon_callback, NULL);
2044 	}
2045 }
2046 
2047 /*
2048  * Caller must hold proc_token.
2049  */
2050 static int
2051 vm_daemon_callback(struct proc *p, void *data __unused)
2052 {
2053 	vm_pindex_t limit, size;
2054 
2055 	/*
2056 	 * if this is a system process or if we have already
2057 	 * looked at this process, skip it.
2058 	 */
2059 	if (p->p_flag & (P_SYSTEM | P_WEXIT))
2060 		return (0);
2061 
2062 	/*
2063 	 * if the process is in a non-running type state,
2064 	 * don't touch it.
2065 	 */
2066 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP)
2067 		return (0);
2068 
2069 	/*
2070 	 * get a limit
2071 	 */
2072 	limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2073 			        p->p_rlimit[RLIMIT_RSS].rlim_max));
2074 
2075 	/*
2076 	 * let processes that are swapped out really be
2077 	 * swapped out.  Set the limit to nothing to get as
2078 	 * many pages out to swap as possible.
2079 	 */
2080 	if (p->p_flag & P_SWAPPEDOUT)
2081 		limit = 0;
2082 
2083 	lwkt_gettoken(&p->p_vmspace->vm_map.token);
2084 	size = vmspace_resident_count(p->p_vmspace);
2085 	if (limit >= 0 && size >= limit) {
2086 		vm_pageout_map_deactivate_pages(&p->p_vmspace->vm_map, limit);
2087 	}
2088 	lwkt_reltoken(&p->p_vmspace->vm_map.token);
2089 	return (0);
2090 }
2091 
2092 #endif
2093