xref: /dflybsd-src/sys/vm/vm_pageout.c (revision 617f6bd6553bb7523bb0198e7aaa5d9c36fb6e05)
1 /*
2  * Copyright (c) 1991 Regents of the University of California.
3  * All rights reserved.
4  * Copyright (c) 1994 John S. Dyson
5  * All rights reserved.
6  * Copyright (c) 1994 David Greenman
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
37  *
38  *
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  *
64  * $FreeBSD: src/sys/vm/vm_pageout.c,v 1.151.2.15 2002/12/29 18:21:04 dillon Exp $
65  */
66 
67 /*
68  *	The proverbial page-out daemon.
69  */
70 
71 #include "opt_vm.h"
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/kernel.h>
75 #include <sys/proc.h>
76 #include <sys/kthread.h>
77 #include <sys/resourcevar.h>
78 #include <sys/signalvar.h>
79 #include <sys/vnode.h>
80 #include <sys/vmmeter.h>
81 #include <sys/conf.h>
82 #include <sys/sysctl.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <sys/lock.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_extern.h>
94 
95 #include <sys/thread2.h>
96 #include <sys/spinlock2.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_page(vm_page_t m, int *max_launderp,
105 			   int *vnodes_skippedp, struct vnode **vpfailedp,
106 			   int pass, int vmflush_flags);
107 static int vm_pageout_clean_helper (vm_page_t, int);
108 static int vm_pageout_free_page_calc (vm_size_t count);
109 static void vm_pageout_page_free(vm_page_t m) ;
110 struct thread *emergpager;
111 struct thread *pagethread;
112 static int sequence_emerg_pager;
113 
114 #if !defined(NO_SWAPPING)
115 /* the kernel process "vm_daemon"*/
116 static void vm_daemon (void);
117 static struct	thread *vmthread;
118 
119 static struct kproc_desc vm_kp = {
120 	"vmdaemon",
121 	vm_daemon,
122 	&vmthread
123 };
124 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
125 #endif
126 
127 int vm_pages_needed = 0;	/* Event on which pageout daemon sleeps */
128 int vm_pageout_deficit = 0;	/* Estimated number of pages deficit */
129 int vm_pageout_pages_needed = 0;/* pageout daemon needs pages */
130 int vm_page_free_hysteresis = 16;
131 static int vm_pagedaemon_time;
132 
133 #if !defined(NO_SWAPPING)
134 static int vm_pageout_req_swapout;
135 static int vm_daemon_needed;
136 #endif
137 static int vm_max_launder = 4096;
138 static int vm_emerg_launder = 100;
139 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
140 static int vm_pageout_full_stats_interval = 0;
141 static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
142 static int defer_swap_pageouts=0;
143 static int disable_swap_pageouts=0;
144 static u_int vm_anonmem_decline = ACT_DECLINE;
145 static u_int vm_filemem_decline = ACT_DECLINE * 2;
146 
147 #if defined(NO_SWAPPING)
148 static int vm_swap_enabled=0;
149 static int vm_swap_idle_enabled=0;
150 #else
151 static int vm_swap_enabled=1;
152 static int vm_swap_idle_enabled=0;
153 #endif
154 int vm_pageout_memuse_mode=1;	/* 0-disable, 1-passive, 2-active swp*/
155 
156 SYSCTL_UINT(_vm, VM_PAGEOUT_ALGORITHM, anonmem_decline,
157 	CTLFLAG_RW, &vm_anonmem_decline, 0, "active->inactive anon memory");
158 
159 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, filemem_decline,
160 	CTLFLAG_RW, &vm_filemem_decline, 0, "active->inactive file cache");
161 
162 SYSCTL_INT(_vm, OID_AUTO, page_free_hysteresis,
163 	CTLFLAG_RW, &vm_page_free_hysteresis, 0,
164 	"Free more pages than the minimum required");
165 
166 SYSCTL_INT(_vm, OID_AUTO, max_launder,
167 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
168 SYSCTL_INT(_vm, OID_AUTO, emerg_launder,
169 	CTLFLAG_RW, &vm_emerg_launder, 0, "Emergency pager minimum");
170 
171 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
172 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
173 
174 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
175 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
176 
177 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
178 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
179 
180 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
181 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
182 SYSCTL_INT(_vm, OID_AUTO, pageout_memuse_mode,
183 	CTLFLAG_RW, &vm_pageout_memuse_mode, 0, "memoryuse resource mode");
184 
185 #if defined(NO_SWAPPING)
186 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
187 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
188 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
189 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
190 #else
191 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
192 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
193 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
194 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
195 #endif
196 
197 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
198 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
199 
200 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
201 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
202 
203 static int pageout_lock_miss;
204 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
205 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
206 
207 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
208 
209 #if !defined(NO_SWAPPING)
210 static void vm_req_vmdaemon (void);
211 #endif
212 static void vm_pageout_page_stats(int q);
213 
214 /*
215  * Calculate approximately how many pages on each queue to try to
216  * clean.  An exact calculation creates an edge condition when the
217  * queues are unbalanced so add significant slop.  The queue scans
218  * will stop early when targets are reached and will start where they
219  * left off on the next pass.
220  *
221  * We need to be generous here because there are all sorts of loading
222  * conditions that can cause edge cases if try to average over all queues.
223  * In particular, storage subsystems have become so fast that paging
224  * activity can become quite frantic.  Eventually we will probably need
225  * two paging threads, one for dirty pages and one for clean, to deal
226  * with the bandwidth requirements.
227 
228  * So what we do is calculate a value that can be satisfied nominally by
229  * only having to scan half the queues.
230  */
231 static __inline int
232 PQAVERAGE(int n)
233 {
234 	int avg;
235 
236 	if (n >= 0) {
237 		avg = ((n + (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) + 1);
238 	} else {
239 		avg = ((n - (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) - 1);
240 	}
241 	return avg;
242 }
243 
244 /*
245  * vm_pageout_clean_helper:
246  *
247  * Clean the page and remove it from the laundry.  The page must be busied
248  * by the caller and will be disposed of (put away, flushed) by this routine.
249  */
250 static int
251 vm_pageout_clean_helper(vm_page_t m, int vmflush_flags)
252 {
253 	vm_object_t object;
254 	vm_page_t mc[BLIST_MAX_ALLOC];
255 	int error;
256 	int ib, is, page_base;
257 	vm_pindex_t pindex = m->pindex;
258 
259 	object = m->object;
260 
261 	/*
262 	 * Don't mess with the page if it's held or special.
263 	 *
264 	 * XXX do we really need to check hold_count here?  hold_count
265 	 * isn't supposed to mess with vm_page ops except prevent the
266 	 * page from being reused.
267 	 */
268 	if (m->hold_count != 0 || (m->flags & PG_UNMANAGED)) {
269 		vm_page_wakeup(m);
270 		return 0;
271 	}
272 
273 	/*
274 	 * Place page in cluster.  Align cluster for optimal swap space
275 	 * allocation (whether it is swap or not).  This is typically ~16-32
276 	 * pages, which also tends to align the cluster to multiples of the
277 	 * filesystem block size if backed by a filesystem.
278 	 */
279 	page_base = pindex % BLIST_MAX_ALLOC;
280 	mc[page_base] = m;
281 	ib = page_base - 1;
282 	is = page_base + 1;
283 
284 	/*
285 	 * Scan object for clusterable pages.
286 	 *
287 	 * We can cluster ONLY if: ->> the page is NOT
288 	 * clean, wired, busy, held, or mapped into a
289 	 * buffer, and one of the following:
290 	 * 1) The page is inactive, or a seldom used
291 	 *    active page.
292 	 * -or-
293 	 * 2) we force the issue.
294 	 *
295 	 * During heavy mmap/modification loads the pageout
296 	 * daemon can really fragment the underlying file
297 	 * due to flushing pages out of order and not trying
298 	 * align the clusters (which leave sporatic out-of-order
299 	 * holes).  To solve this problem we do the reverse scan
300 	 * first and attempt to align our cluster, then do a
301 	 * forward scan if room remains.
302 	 */
303 	vm_object_hold(object);
304 
305 	while (ib >= 0) {
306 		vm_page_t p;
307 
308 		p = vm_page_lookup_busy_try(object, pindex - page_base + ib,
309 					    TRUE, &error);
310 		if (error || p == NULL)
311 			break;
312 		if ((p->queue - p->pc) == PQ_CACHE ||
313 		    (p->flags & PG_UNMANAGED)) {
314 			vm_page_wakeup(p);
315 			break;
316 		}
317 		vm_page_test_dirty(p);
318 		if (((p->dirty & p->valid) == 0 &&
319 		     (p->flags & PG_NEED_COMMIT) == 0) ||
320 		    p->wire_count != 0 ||	/* may be held by buf cache */
321 		    p->hold_count != 0) {	/* may be undergoing I/O */
322 			vm_page_wakeup(p);
323 			break;
324 		}
325 		if (p->queue - p->pc != PQ_INACTIVE) {
326 			if (p->queue - p->pc != PQ_ACTIVE ||
327 			    (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
328 				vm_page_wakeup(p);
329 				break;
330 			}
331 		}
332 
333 		/*
334 		 * Try to maintain page groupings in the cluster.
335 		 */
336 		if (m->flags & PG_WINATCFLS)
337 			vm_page_flag_set(p, PG_WINATCFLS);
338 		else
339 			vm_page_flag_clear(p, PG_WINATCFLS);
340 		p->act_count = m->act_count;
341 
342 		mc[ib] = p;
343 		--ib;
344 	}
345 	++ib;	/* fixup */
346 
347 	while (is < BLIST_MAX_ALLOC &&
348 	       pindex - page_base + is < object->size) {
349 		vm_page_t p;
350 
351 		p = vm_page_lookup_busy_try(object, pindex - page_base + is,
352 					    TRUE, &error);
353 		if (error || p == NULL)
354 			break;
355 		if (((p->queue - p->pc) == PQ_CACHE) ||
356 		    (p->flags & PG_UNMANAGED)) {
357 			vm_page_wakeup(p);
358 			break;
359 		}
360 		vm_page_test_dirty(p);
361 		if (((p->dirty & p->valid) == 0 &&
362 		     (p->flags & PG_NEED_COMMIT) == 0) ||
363 		    p->wire_count != 0 ||	/* may be held by buf cache */
364 		    p->hold_count != 0) {	/* may be undergoing I/O */
365 			vm_page_wakeup(p);
366 			break;
367 		}
368 		if (p->queue - p->pc != PQ_INACTIVE) {
369 			if (p->queue - p->pc != PQ_ACTIVE ||
370 			    (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
371 				vm_page_wakeup(p);
372 				break;
373 			}
374 		}
375 
376 		/*
377 		 * Try to maintain page groupings in the cluster.
378 		 */
379 		if (m->flags & PG_WINATCFLS)
380 			vm_page_flag_set(p, PG_WINATCFLS);
381 		else
382 			vm_page_flag_clear(p, PG_WINATCFLS);
383 		p->act_count = m->act_count;
384 
385 		mc[is] = p;
386 		++is;
387 	}
388 
389 	vm_object_drop(object);
390 
391 	/*
392 	 * we allow reads during pageouts...
393 	 */
394 	return vm_pageout_flush(&mc[ib], is - ib, vmflush_flags);
395 }
396 
397 /*
398  * vm_pageout_flush() - launder the given pages
399  *
400  *	The given pages are laundered.  Note that we setup for the start of
401  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
402  *	reference count all in here rather then in the parent.  If we want
403  *	the parent to do more sophisticated things we may have to change
404  *	the ordering.
405  *
406  *	The pages in the array must be busied by the caller and will be
407  *	unbusied by this function.
408  */
409 int
410 vm_pageout_flush(vm_page_t *mc, int count, int vmflush_flags)
411 {
412 	vm_object_t object;
413 	int pageout_status[count];
414 	int numpagedout = 0;
415 	int i;
416 
417 	/*
418 	 * Initiate I/O.  Bump the vm_page_t->busy counter.
419 	 */
420 	for (i = 0; i < count; i++) {
421 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
422 			("vm_pageout_flush page %p index %d/%d: partially "
423 			 "invalid page", mc[i], i, count));
424 		vm_page_io_start(mc[i]);
425 	}
426 
427 	/*
428 	 * We must make the pages read-only.  This will also force the
429 	 * modified bit in the related pmaps to be cleared.  The pager
430 	 * cannot clear the bit for us since the I/O completion code
431 	 * typically runs from an interrupt.  The act of making the page
432 	 * read-only handles the case for us.
433 	 *
434 	 * Then we can unbusy the pages, we still hold a reference by virtue
435 	 * of our soft-busy.
436 	 */
437 	for (i = 0; i < count; i++) {
438 		if (vmflush_flags & VM_PAGER_TRY_TO_CACHE)
439 			vm_page_protect(mc[i], VM_PROT_NONE);
440 		else
441 			vm_page_protect(mc[i], VM_PROT_READ);
442 		vm_page_wakeup(mc[i]);
443 	}
444 
445 	object = mc[0]->object;
446 	vm_object_pip_add(object, count);
447 
448 	vm_pager_put_pages(object, mc, count,
449 			   (vmflush_flags |
450 			    ((object == &kernel_object) ?
451 				VM_PAGER_PUT_SYNC : 0)),
452 			   pageout_status);
453 
454 	for (i = 0; i < count; i++) {
455 		vm_page_t mt = mc[i];
456 
457 		switch (pageout_status[i]) {
458 		case VM_PAGER_OK:
459 			numpagedout++;
460 			break;
461 		case VM_PAGER_PEND:
462 			numpagedout++;
463 			break;
464 		case VM_PAGER_BAD:
465 			/*
466 			 * Page outside of range of object. Right now we
467 			 * essentially lose the changes by pretending it
468 			 * worked.
469 			 */
470 			vm_page_busy_wait(mt, FALSE, "pgbad");
471 			pmap_clear_modify(mt);
472 			vm_page_undirty(mt);
473 			vm_page_wakeup(mt);
474 			break;
475 		case VM_PAGER_ERROR:
476 		case VM_PAGER_FAIL:
477 			/*
478 			 * A page typically cannot be paged out when we
479 			 * have run out of swap.  We leave the page
480 			 * marked inactive and will try to page it out
481 			 * again later.
482 			 *
483 			 * Starvation of the active page list is used to
484 			 * determine when the system is massively memory
485 			 * starved.
486 			 */
487 			break;
488 		case VM_PAGER_AGAIN:
489 			break;
490 		}
491 
492 		/*
493 		 * If not PENDing this was a synchronous operation and we
494 		 * clean up after the I/O.  If it is PENDing the mess is
495 		 * cleaned up asynchronously.
496 		 *
497 		 * Also nominally act on the caller's wishes if the caller
498 		 * wants to try to really clean (cache or free) the page.
499 		 *
500 		 * Also nominally deactivate the page if the system is
501 		 * memory-stressed.
502 		 */
503 		if (pageout_status[i] != VM_PAGER_PEND) {
504 			vm_page_busy_wait(mt, FALSE, "pgouw");
505 			vm_page_io_finish(mt);
506 			if (vmflush_flags & VM_PAGER_TRY_TO_CACHE) {
507 				vm_page_try_to_cache(mt);
508 			} else if (vm_page_count_severe()) {
509 				vm_page_deactivate(mt);
510 				vm_page_wakeup(mt);
511 			} else {
512 				vm_page_wakeup(mt);
513 			}
514 			vm_object_pip_wakeup(object);
515 		}
516 	}
517 	return numpagedout;
518 }
519 
520 #if !defined(NO_SWAPPING)
521 
522 /*
523  * Callback function, page busied for us.  We must dispose of the busy
524  * condition.  Any related pmap pages may be held but will not be locked.
525  */
526 static
527 int
528 vm_pageout_mdp_callback(struct pmap_pgscan_info *info, vm_offset_t va,
529 			vm_page_t p)
530 {
531 	int actcount;
532 	int cleanit = 0;
533 
534 	/*
535 	 * Basic tests - There should never be a marker, and we can stop
536 	 *		 once the RSS is below the required level.
537 	 */
538 	KKASSERT((p->flags & PG_MARKER) == 0);
539 	if (pmap_resident_tlnw_count(info->pmap) <= info->limit) {
540 		vm_page_wakeup(p);
541 		return(-1);
542 	}
543 
544 	mycpu->gd_cnt.v_pdpages++;
545 
546 	if (p->wire_count || p->hold_count || (p->flags & PG_UNMANAGED)) {
547 		vm_page_wakeup(p);
548 		goto done;
549 	}
550 
551 	++info->actioncount;
552 
553 	/*
554 	 * Check if the page has been referened recently.  If it has,
555 	 * activate it and skip.
556 	 */
557 	actcount = pmap_ts_referenced(p);
558 	if (actcount) {
559 		vm_page_flag_set(p, PG_REFERENCED);
560 	} else if (p->flags & PG_REFERENCED) {
561 		actcount = 1;
562 	}
563 
564 	if (actcount) {
565 		if (p->queue - p->pc != PQ_ACTIVE) {
566 			vm_page_and_queue_spin_lock(p);
567 			if (p->queue - p->pc != PQ_ACTIVE) {
568 				vm_page_and_queue_spin_unlock(p);
569 				vm_page_activate(p);
570 			} else {
571 				vm_page_and_queue_spin_unlock(p);
572 			}
573 		} else {
574 			p->act_count += actcount;
575 			if (p->act_count > ACT_MAX)
576 				p->act_count = ACT_MAX;
577 		}
578 		vm_page_flag_clear(p, PG_REFERENCED);
579 		vm_page_wakeup(p);
580 		goto done;
581 	}
582 
583 	/*
584 	 * Remove the page from this particular pmap.  Once we do this, our
585 	 * pmap scans will not see it again (unless it gets faulted in), so
586 	 * we must actively dispose of or deal with the page.
587 	 */
588 	pmap_remove_specific(info->pmap, p);
589 
590 	/*
591 	 * If the page is not mapped to another process (i.e. as would be
592 	 * typical if this were a shared page from a library) then deactivate
593 	 * the page and clean it in two passes only.
594 	 *
595 	 * If the page hasn't been referenced since the last check, remove it
596 	 * from the pmap.  If it is no longer mapped, deactivate it
597 	 * immediately, accelerating the normal decline.
598 	 *
599 	 * Once the page has been removed from the pmap the RSS code no
600 	 * longer tracks it so we have to make sure that it is staged for
601 	 * potential flush action.
602 	 */
603 	if ((p->flags & PG_MAPPED) == 0) {
604 		if (p->queue - p->pc == PQ_ACTIVE) {
605 			vm_page_deactivate(p);
606 		}
607 		if (p->queue - p->pc == PQ_INACTIVE) {
608 			cleanit = 1;
609 		}
610 	}
611 
612 	/*
613 	 * Ok, try to fully clean the page and any nearby pages such that at
614 	 * least the requested page is freed or moved to the cache queue.
615 	 *
616 	 * We usually do this synchronously to allow us to get the page into
617 	 * the CACHE queue quickly, which will prevent memory exhaustion if
618 	 * a process with a memoryuse limit is running away.  However, the
619 	 * sysadmin may desire to set vm.swap_user_async which relaxes this
620 	 * and improves write performance.
621 	 */
622 	if (cleanit) {
623 		int max_launder = 0x7FFF;
624 		int vnodes_skipped = 0;
625 		int vmflush_flags;
626 		struct vnode *vpfailed = NULL;
627 
628 		info->offset = va;
629 
630 		if (vm_pageout_memuse_mode >= 2) {
631 			vmflush_flags = VM_PAGER_TRY_TO_CACHE |
632 					VM_PAGER_ALLOW_ACTIVE;
633 			if (swap_user_async == 0)
634 				vmflush_flags |= VM_PAGER_PUT_SYNC;
635 			vm_page_flag_set(p, PG_WINATCFLS);
636 			info->cleancount +=
637 				vm_pageout_page(p, &max_launder,
638 						&vnodes_skipped,
639 						&vpfailed, 1, vmflush_flags);
640 		} else {
641 			vm_page_wakeup(p);
642 			++info->cleancount;
643 		}
644 	} else {
645 		vm_page_wakeup(p);
646 	}
647 
648 	/*
649 	 * Must be at end to avoid SMP races.
650 	 */
651 done:
652 	lwkt_user_yield();
653 	return 0;
654 }
655 
656 /*
657  * Deactivate some number of pages in a map due to set RLIMIT_RSS limits.
658  * that is relatively difficult to do.  We try to keep track of where we
659  * left off last time to reduce scan overhead.
660  *
661  * Called when vm_pageout_memuse_mode is >= 1.
662  */
663 void
664 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t limit)
665 {
666 	vm_offset_t pgout_offset;
667 	struct pmap_pgscan_info info;
668 	int retries = 3;
669 
670 	pgout_offset = map->pgout_offset;
671 again:
672 #if 0
673 	kprintf("%016jx ", pgout_offset);
674 #endif
675 	if (pgout_offset < VM_MIN_USER_ADDRESS)
676 		pgout_offset = VM_MIN_USER_ADDRESS;
677 	if (pgout_offset >= VM_MAX_USER_ADDRESS)
678 		pgout_offset = 0;
679 	info.pmap = vm_map_pmap(map);
680 	info.limit = limit;
681 	info.beg_addr = pgout_offset;
682 	info.end_addr = VM_MAX_USER_ADDRESS;
683 	info.callback = vm_pageout_mdp_callback;
684 	info.cleancount = 0;
685 	info.actioncount = 0;
686 	info.busycount = 0;
687 
688 	pmap_pgscan(&info);
689 	pgout_offset = info.offset;
690 #if 0
691 	kprintf("%016jx %08lx %08lx\n", pgout_offset,
692 		info.cleancount, info.actioncount);
693 #endif
694 
695 	if (pgout_offset != VM_MAX_USER_ADDRESS &&
696 	    pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
697 		goto again;
698 	} else if (retries &&
699 		   pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
700 		--retries;
701 		goto again;
702 	}
703 	map->pgout_offset = pgout_offset;
704 }
705 #endif
706 
707 /*
708  * Called when the pageout scan wants to free a page.  We no longer
709  * try to cycle the vm_object here with a reference & dealloc, which can
710  * cause a non-trivial object collapse in a critical path.
711  *
712  * It is unclear why we cycled the ref_count in the past, perhaps to try
713  * to optimize shadow chain collapses but I don't quite see why it would
714  * be necessary.  An OBJ_DEAD object should terminate any and all vm_pages
715  * synchronously and not have to be kicked-start.
716  */
717 static void
718 vm_pageout_page_free(vm_page_t m)
719 {
720 	vm_page_protect(m, VM_PROT_NONE);
721 	vm_page_free(m);
722 }
723 
724 /*
725  * vm_pageout_scan does the dirty work for the pageout daemon.
726  */
727 struct vm_pageout_scan_info {
728 	struct proc *bigproc;
729 	vm_offset_t bigsize;
730 };
731 
732 static int vm_pageout_scan_callback(struct proc *p, void *data);
733 
734 /*
735  * Scan inactive queue
736  *
737  * WARNING! Can be called from two pagedaemon threads simultaneously.
738  */
739 static int
740 vm_pageout_scan_inactive(int pass, int q, int avail_shortage,
741 			 int *vnodes_skipped)
742 {
743 	vm_page_t m;
744 	struct vm_page marker;
745 	struct vnode *vpfailed;		/* warning, allowed to be stale */
746 	int maxscan;
747 	int delta = 0;
748 	int max_launder;
749 	int isep;
750 
751 	isep = (curthread == emergpager);
752 
753 	/*
754 	 * Start scanning the inactive queue for pages we can move to the
755 	 * cache or free.  The scan will stop when the target is reached or
756 	 * we have scanned the entire inactive queue.  Note that m->act_count
757 	 * is not used to form decisions for the inactive queue, only for the
758 	 * active queue.
759 	 *
760 	 * max_launder limits the number of dirty pages we flush per scan.
761 	 * For most systems a smaller value (16 or 32) is more robust under
762 	 * extreme memory and disk pressure because any unnecessary writes
763 	 * to disk can result in extreme performance degredation.  However,
764 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
765 	 * used) will die horribly with limited laundering.  If the pageout
766 	 * daemon cannot clean enough pages in the first pass, we let it go
767 	 * all out in succeeding passes.
768 	 *
769 	 * NOTE!  THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
770 	 *	  PAGES.
771 	 */
772 	if ((max_launder = vm_max_launder) <= 1)
773 		max_launder = 1;
774 	if (pass)
775 		max_launder = 10000;
776 
777 	/*
778 	 * Initialize our marker
779 	 */
780 	bzero(&marker, sizeof(marker));
781 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
782 	marker.queue = PQ_INACTIVE + q;
783 	marker.pc = q;
784 	marker.wire_count = 1;
785 
786 	/*
787 	 * Inactive queue scan.
788 	 *
789 	 * NOTE: The vm_page must be spinlocked before the queue to avoid
790 	 *	 deadlocks, so it is easiest to simply iterate the loop
791 	 *	 with the queue unlocked at the top.
792 	 */
793 	vpfailed = NULL;
794 
795 	vm_page_queues_spin_lock(PQ_INACTIVE + q);
796 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
797 	maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt;
798 
799 	/*
800 	 * Queue locked at top of loop to avoid stack marker issues.
801 	 */
802 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
803 	       maxscan-- > 0 && avail_shortage - delta > 0)
804 	{
805 		int count;
806 
807 		KKASSERT(m->queue == PQ_INACTIVE + q);
808 		TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
809 			     &marker, pageq);
810 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
811 				   &marker, pageq);
812 		mycpu->gd_cnt.v_pdpages++;
813 
814 		/*
815 		 * Skip marker pages (atomic against other markers to avoid
816 		 * infinite hop-over scans).
817 		 */
818 		if (m->flags & PG_MARKER)
819 			continue;
820 
821 		/*
822 		 * Try to busy the page.  Don't mess with pages which are
823 		 * already busy or reorder them in the queue.
824 		 */
825 		if (vm_page_busy_try(m, TRUE))
826 			continue;
827 
828 		/*
829 		 * Remaining operations run with the page busy and neither
830 		 * the page or the queue will be spin-locked.
831 		 */
832 		vm_page_queues_spin_unlock(PQ_INACTIVE + q);
833 		KKASSERT(m->queue == PQ_INACTIVE + q);
834 
835 		/*
836 		 * The emergency pager runs when the primary pager gets
837 		 * stuck, which typically means the primary pager deadlocked
838 		 * on a vnode-backed page.  Therefore, the emergency pager
839 		 * must skip any complex objects.
840 		 *
841 		 * We disallow VNODEs unless they are VCHR whos device ops
842 		 * does not flag D_NOEMERGPGR.
843 		 */
844 		if (isep && m->object) {
845 			struct vnode *vp;
846 
847 			switch(m->object->type) {
848 			case OBJT_DEFAULT:
849 			case OBJT_SWAP:
850 				/*
851 				 * Allow anonymous memory and assume that
852 				 * swap devices are not complex, since its
853 				 * kinda worthless if we can't swap out dirty
854 				 * anonymous pages.
855 				 */
856 				break;
857 			case OBJT_VNODE:
858 				/*
859 				 * Allow VCHR device if the D_NOEMERGPGR
860 				 * flag is not set, deny other vnode types
861 				 * as being too complex.
862 				 */
863 				vp = m->object->handle;
864 				if (vp && vp->v_type == VCHR &&
865 				    vp->v_rdev && vp->v_rdev->si_ops &&
866 				    (vp->v_rdev->si_ops->head.flags &
867 				     D_NOEMERGPGR) == 0) {
868 					break;
869 				}
870 				/* Deny - fall through */
871 			default:
872 				/*
873 				 * Deny
874 				 */
875 				vm_page_wakeup(m);
876 				vm_page_queues_spin_lock(PQ_INACTIVE + q);
877 				lwkt_yield();
878 				continue;
879 			}
880 		}
881 
882 		/*
883 		 * Try to pageout the page and perhaps other nearby pages.
884 		 */
885 		count = vm_pageout_page(m, &max_launder, vnodes_skipped,
886 					&vpfailed, pass, 0);
887 		delta += count;
888 
889 		/*
890 		 * Systems with a ton of memory can wind up with huge
891 		 * deactivation counts.  Because the inactive scan is
892 		 * doing a lot of flushing, the combination can result
893 		 * in excessive paging even in situations where other
894 		 * unrelated threads free up sufficient VM.
895 		 *
896 		 * To deal with this we abort the nominal active->inactive
897 		 * scan before we hit the inactive target when free+cache
898 		 * levels have reached a reasonable target.
899 		 *
900 		 * When deciding to stop early we need to add some slop to
901 		 * the test and we need to return full completion to the caller
902 		 * to prevent the caller from thinking there is something
903 		 * wrong and issuing a low-memory+swap warning or pkill.
904 		 *
905 		 * A deficit forces paging regardless of the state of the
906 		 * VM page queues (used for RSS enforcement).
907 		 */
908 		lwkt_yield();
909 		vm_page_queues_spin_lock(PQ_INACTIVE + q);
910 		if (vm_paging_target() < -vm_max_launder) {
911 			/*
912 			 * Stopping early, return full completion to caller.
913 			 */
914 			if (delta < avail_shortage)
915 				delta = avail_shortage;
916 			break;
917 		}
918 	}
919 
920 	/* page queue still spin-locked */
921 	TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
922 	vm_page_queues_spin_unlock(PQ_INACTIVE + q);
923 
924 	return (delta);
925 }
926 
927 /*
928  * Pageout the specified page, return the total number of pages paged out
929  * (this routine may cluster).
930  *
931  * The page must be busied and soft-busied by the caller and will be disposed
932  * of by this function.
933  */
934 static int
935 vm_pageout_page(vm_page_t m, int *max_launderp, int *vnodes_skippedp,
936 		struct vnode **vpfailedp, int pass, int vmflush_flags)
937 {
938 	vm_object_t object;
939 	int actcount;
940 	int count = 0;
941 
942 	/*
943 	 * It is possible for a page to be busied ad-hoc (e.g. the
944 	 * pmap_collect() code) and wired and race against the
945 	 * allocation of a new page.  vm_page_alloc() may be forced
946 	 * to deactivate the wired page in which case it winds up
947 	 * on the inactive queue and must be handled here.  We
948 	 * correct the problem simply by unqueuing the page.
949 	 */
950 	if (m->wire_count) {
951 		vm_page_unqueue_nowakeup(m);
952 		vm_page_wakeup(m);
953 		kprintf("WARNING: pagedaemon: wired page on "
954 			"inactive queue %p\n", m);
955 		return 0;
956 	}
957 
958 	/*
959 	 * A held page may be undergoing I/O, so skip it.
960 	 */
961 	if (m->hold_count) {
962 		vm_page_and_queue_spin_lock(m);
963 		if (m->queue - m->pc == PQ_INACTIVE) {
964 			TAILQ_REMOVE(
965 				&vm_page_queues[m->queue].pl, m, pageq);
966 			TAILQ_INSERT_TAIL(
967 				&vm_page_queues[m->queue].pl, m, pageq);
968 			++vm_swapcache_inactive_heuristic;
969 		}
970 		vm_page_and_queue_spin_unlock(m);
971 		vm_page_wakeup(m);
972 		return 0;
973 	}
974 
975 	if (m->object == NULL || m->object->ref_count == 0) {
976 		/*
977 		 * If the object is not being used, we ignore previous
978 		 * references.
979 		 */
980 		vm_page_flag_clear(m, PG_REFERENCED);
981 		pmap_clear_reference(m);
982 		/* fall through to end */
983 	} else if (((m->flags & PG_REFERENCED) == 0) &&
984 		    (actcount = pmap_ts_referenced(m))) {
985 		/*
986 		 * Otherwise, if the page has been referenced while
987 		 * in the inactive queue, we bump the "activation
988 		 * count" upwards, making it less likely that the
989 		 * page will be added back to the inactive queue
990 		 * prematurely again.  Here we check the page tables
991 		 * (or emulated bits, if any), given the upper level
992 		 * VM system not knowing anything about existing
993 		 * references.
994 		 */
995 		vm_page_activate(m);
996 		m->act_count += (actcount + ACT_ADVANCE);
997 		vm_page_wakeup(m);
998 		return 0;
999 	}
1000 
1001 	/*
1002 	 * (m) is still busied.
1003 	 *
1004 	 * If the upper level VM system knows about any page
1005 	 * references, we activate the page.  We also set the
1006 	 * "activation count" higher than normal so that we will less
1007 	 * likely place pages back onto the inactive queue again.
1008 	 */
1009 	if ((m->flags & PG_REFERENCED) != 0) {
1010 		vm_page_flag_clear(m, PG_REFERENCED);
1011 		actcount = pmap_ts_referenced(m);
1012 		vm_page_activate(m);
1013 		m->act_count += (actcount + ACT_ADVANCE + 1);
1014 		vm_page_wakeup(m);
1015 		return 0;
1016 	}
1017 
1018 	/*
1019 	 * If the upper level VM system doesn't know anything about
1020 	 * the page being dirty, we have to check for it again.  As
1021 	 * far as the VM code knows, any partially dirty pages are
1022 	 * fully dirty.
1023 	 *
1024 	 * Pages marked PG_WRITEABLE may be mapped into the user
1025 	 * address space of a process running on another cpu.  A
1026 	 * user process (without holding the MP lock) running on
1027 	 * another cpu may be able to touch the page while we are
1028 	 * trying to remove it.  vm_page_cache() will handle this
1029 	 * case for us.
1030 	 */
1031 	if (m->dirty == 0) {
1032 		vm_page_test_dirty(m);
1033 	} else {
1034 		vm_page_dirty(m);
1035 	}
1036 
1037 	if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
1038 		/*
1039 		 * Invalid pages can be easily freed
1040 		 */
1041 		vm_pageout_page_free(m);
1042 		mycpu->gd_cnt.v_dfree++;
1043 		++count;
1044 	} else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
1045 		/*
1046 		 * Clean pages can be placed onto the cache queue.
1047 		 * This effectively frees them.
1048 		 */
1049 		vm_page_cache(m);
1050 		++count;
1051 	} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
1052 		/*
1053 		 * Dirty pages need to be paged out, but flushing
1054 		 * a page is extremely expensive verses freeing
1055 		 * a clean page.  Rather then artificially limiting
1056 		 * the number of pages we can flush, we instead give
1057 		 * dirty pages extra priority on the inactive queue
1058 		 * by forcing them to be cycled through the queue
1059 		 * twice before being flushed, after which the
1060 		 * (now clean) page will cycle through once more
1061 		 * before being freed.  This significantly extends
1062 		 * the thrash point for a heavily loaded machine.
1063 		 */
1064 		vm_page_flag_set(m, PG_WINATCFLS);
1065 		vm_page_and_queue_spin_lock(m);
1066 		if (m->queue - m->pc == PQ_INACTIVE) {
1067 			TAILQ_REMOVE(
1068 				&vm_page_queues[m->queue].pl, m, pageq);
1069 			TAILQ_INSERT_TAIL(
1070 				&vm_page_queues[m->queue].pl, m, pageq);
1071 			++vm_swapcache_inactive_heuristic;
1072 		}
1073 		vm_page_and_queue_spin_unlock(m);
1074 		vm_page_wakeup(m);
1075 	} else if (*max_launderp > 0) {
1076 		/*
1077 		 * We always want to try to flush some dirty pages if
1078 		 * we encounter them, to keep the system stable.
1079 		 * Normally this number is small, but under extreme
1080 		 * pressure where there are insufficient clean pages
1081 		 * on the inactive queue, we may have to go all out.
1082 		 */
1083 		int swap_pageouts_ok;
1084 		struct vnode *vp = NULL;
1085 
1086 		swap_pageouts_ok = 0;
1087 		object = m->object;
1088 		if (object &&
1089 		    (object->type != OBJT_SWAP) &&
1090 		    (object->type != OBJT_DEFAULT)) {
1091 			swap_pageouts_ok = 1;
1092 		} else {
1093 			swap_pageouts_ok = !(defer_swap_pageouts ||
1094 					     disable_swap_pageouts);
1095 			swap_pageouts_ok |= (!disable_swap_pageouts &&
1096 					     defer_swap_pageouts &&
1097 					     vm_page_count_min(0));
1098 		}
1099 
1100 		/*
1101 		 * We don't bother paging objects that are "dead".
1102 		 * Those objects are in a "rundown" state.
1103 		 */
1104 		if (!swap_pageouts_ok ||
1105 		    (object == NULL) ||
1106 		    (object->flags & OBJ_DEAD)) {
1107 			vm_page_and_queue_spin_lock(m);
1108 			if (m->queue - m->pc == PQ_INACTIVE) {
1109 				TAILQ_REMOVE(
1110 				    &vm_page_queues[m->queue].pl,
1111 				    m, pageq);
1112 				TAILQ_INSERT_TAIL(
1113 				    &vm_page_queues[m->queue].pl,
1114 				    m, pageq);
1115 				++vm_swapcache_inactive_heuristic;
1116 			}
1117 			vm_page_and_queue_spin_unlock(m);
1118 			vm_page_wakeup(m);
1119 			return 0;
1120 		}
1121 
1122 		/*
1123 		 * (m) is still busied.
1124 		 *
1125 		 * The object is already known NOT to be dead.   It
1126 		 * is possible for the vget() to block the whole
1127 		 * pageout daemon, but the new low-memory handling
1128 		 * code should prevent it.
1129 		 *
1130 		 * The previous code skipped locked vnodes and, worse,
1131 		 * reordered pages in the queue.  This results in
1132 		 * completely non-deterministic operation because,
1133 		 * quite often, a vm_fault has initiated an I/O and
1134 		 * is holding a locked vnode at just the point where
1135 		 * the pageout daemon is woken up.
1136 		 *
1137 		 * We can't wait forever for the vnode lock, we might
1138 		 * deadlock due to a vn_read() getting stuck in
1139 		 * vm_wait while holding this vnode.  We skip the
1140 		 * vnode if we can't get it in a reasonable amount
1141 		 * of time.
1142 		 *
1143 		 * vpfailed is used to (try to) avoid the case where
1144 		 * a large number of pages are associated with a
1145 		 * locked vnode, which could cause the pageout daemon
1146 		 * to stall for an excessive amount of time.
1147 		 */
1148 		if (object->type == OBJT_VNODE) {
1149 			int flags;
1150 
1151 			vp = object->handle;
1152 			flags = LK_EXCLUSIVE;
1153 			if (vp == *vpfailedp)
1154 				flags |= LK_NOWAIT;
1155 			else
1156 				flags |= LK_TIMELOCK;
1157 			vm_page_hold(m);
1158 			vm_page_wakeup(m);
1159 
1160 			/*
1161 			 * We have unbusied (m) temporarily so we can
1162 			 * acquire the vp lock without deadlocking.
1163 			 * (m) is held to prevent destruction.
1164 			 */
1165 			if (vget(vp, flags) != 0) {
1166 				*vpfailedp = vp;
1167 				++pageout_lock_miss;
1168 				if (object->flags & OBJ_MIGHTBEDIRTY)
1169 					    ++*vnodes_skippedp;
1170 				vm_page_unhold(m);
1171 				return 0;
1172 			}
1173 
1174 			/*
1175 			 * The page might have been moved to another
1176 			 * queue during potential blocking in vget()
1177 			 * above.  The page might have been freed and
1178 			 * reused for another vnode.  The object might
1179 			 * have been reused for another vnode.
1180 			 */
1181 			if (m->queue - m->pc != PQ_INACTIVE ||
1182 			    m->object != object ||
1183 			    object->handle != vp) {
1184 				if (object->flags & OBJ_MIGHTBEDIRTY)
1185 					++*vnodes_skippedp;
1186 				vput(vp);
1187 				vm_page_unhold(m);
1188 				return 0;
1189 			}
1190 
1191 			/*
1192 			 * The page may have been busied during the
1193 			 * blocking in vput();  We don't move the
1194 			 * page back onto the end of the queue so that
1195 			 * statistics are more correct if we don't.
1196 			 */
1197 			if (vm_page_busy_try(m, TRUE)) {
1198 				vput(vp);
1199 				vm_page_unhold(m);
1200 				return 0;
1201 			}
1202 			vm_page_unhold(m);
1203 
1204 			/*
1205 			 * (m) is busied again
1206 			 *
1207 			 * We own the busy bit and remove our hold
1208 			 * bit.  If the page is still held it
1209 			 * might be undergoing I/O, so skip it.
1210 			 */
1211 			if (m->hold_count) {
1212 				vm_page_and_queue_spin_lock(m);
1213 				if (m->queue - m->pc == PQ_INACTIVE) {
1214 					TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq);
1215 					TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1216 					++vm_swapcache_inactive_heuristic;
1217 				}
1218 				vm_page_and_queue_spin_unlock(m);
1219 				if (object->flags & OBJ_MIGHTBEDIRTY)
1220 					++*vnodes_skippedp;
1221 				vm_page_wakeup(m);
1222 				vput(vp);
1223 				return 0;
1224 			}
1225 			/* (m) is left busied as we fall through */
1226 		}
1227 
1228 		/*
1229 		 * page is busy and not held here.
1230 		 *
1231 		 * If a page is dirty, then it is either being washed
1232 		 * (but not yet cleaned) or it is still in the
1233 		 * laundry.  If it is still in the laundry, then we
1234 		 * start the cleaning operation.
1235 		 *
1236 		 * decrement inactive_shortage on success to account
1237 		 * for the (future) cleaned page.  Otherwise we
1238 		 * could wind up laundering or cleaning too many
1239 		 * pages.
1240 		 *
1241 		 * NOTE: Cleaning the page here does not cause
1242 		 *	 force_deficit to be adjusted, because the
1243 		 *	 page is not being freed or moved to the
1244 		 *	 cache.
1245 		 */
1246 		count = vm_pageout_clean_helper(m, vmflush_flags);
1247 		*max_launderp -= count;
1248 
1249 		/*
1250 		 * Clean ate busy, page no longer accessible
1251 		 */
1252 		if (vp != NULL)
1253 			vput(vp);
1254 	} else {
1255 		vm_page_wakeup(m);
1256 	}
1257 	return count;
1258 }
1259 
1260 /*
1261  * Scan active queue
1262  *
1263  * WARNING! Can be called from two pagedaemon threads simultaneously.
1264  */
1265 static int
1266 vm_pageout_scan_active(int pass, int q,
1267 		       int avail_shortage, int inactive_shortage,
1268 		       int *recycle_countp)
1269 {
1270 	struct vm_page marker;
1271 	vm_page_t m;
1272 	int actcount;
1273 	int delta = 0;
1274 	int maxscan;
1275 	int isep;
1276 
1277 	isep = (curthread == emergpager);
1278 
1279 	/*
1280 	 * We want to move pages from the active queue to the inactive
1281 	 * queue to get the inactive queue to the inactive target.  If
1282 	 * we still have a page shortage from above we try to directly free
1283 	 * clean pages instead of moving them.
1284 	 *
1285 	 * If we do still have a shortage we keep track of the number of
1286 	 * pages we free or cache (recycle_count) as a measure of thrashing
1287 	 * between the active and inactive queues.
1288 	 *
1289 	 * If we were able to completely satisfy the free+cache targets
1290 	 * from the inactive pool we limit the number of pages we move
1291 	 * from the active pool to the inactive pool to 2x the pages we
1292 	 * had removed from the inactive pool (with a minimum of 1/5 the
1293 	 * inactive target).  If we were not able to completely satisfy
1294 	 * the free+cache targets we go for the whole target aggressively.
1295 	 *
1296 	 * NOTE: Both variables can end up negative.
1297 	 * NOTE: We are still in a critical section.
1298 	 *
1299 	 * NOTE!  THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
1300 	 *	  PAGES.
1301 	 */
1302 
1303 	bzero(&marker, sizeof(marker));
1304 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1305 	marker.queue = PQ_ACTIVE + q;
1306 	marker.pc = q;
1307 	marker.wire_count = 1;
1308 
1309 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1310 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1311 	maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt;
1312 
1313 	/*
1314 	 * Queue locked at top of loop to avoid stack marker issues.
1315 	 */
1316 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1317 	       maxscan-- > 0 && (avail_shortage - delta > 0 ||
1318 				inactive_shortage > 0))
1319 	{
1320 		KKASSERT(m->queue == PQ_ACTIVE + q);
1321 		TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1322 			     &marker, pageq);
1323 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1324 				   &marker, pageq);
1325 
1326 		/*
1327 		 * Skip marker pages (atomic against other markers to avoid
1328 		 * infinite hop-over scans).
1329 		 */
1330 		if (m->flags & PG_MARKER)
1331 			continue;
1332 
1333 		/*
1334 		 * Try to busy the page.  Don't mess with pages which are
1335 		 * already busy or reorder them in the queue.
1336 		 */
1337 		if (vm_page_busy_try(m, TRUE))
1338 			continue;
1339 
1340 		/*
1341 		 * Remaining operations run with the page busy and neither
1342 		 * the page or the queue will be spin-locked.
1343 		 */
1344 		vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1345 		KKASSERT(m->queue == PQ_ACTIVE + q);
1346 
1347 		/*
1348 		 * Don't deactivate pages that are held, even if we can
1349 		 * busy them.  (XXX why not?)
1350 		 */
1351 		if (m->hold_count != 0) {
1352 			vm_page_and_queue_spin_lock(m);
1353 			if (m->queue - m->pc == PQ_ACTIVE) {
1354 				TAILQ_REMOVE(
1355 					&vm_page_queues[PQ_ACTIVE + q].pl,
1356 					m, pageq);
1357 				TAILQ_INSERT_TAIL(
1358 					&vm_page_queues[PQ_ACTIVE + q].pl,
1359 					m, pageq);
1360 			}
1361 			vm_page_and_queue_spin_unlock(m);
1362 			vm_page_wakeup(m);
1363 			goto next;
1364 		}
1365 
1366 		/*
1367 		 * The emergency pager ignores vnode-backed pages as these
1368 		 * are the pages that probably bricked the main pager.
1369 		 */
1370 		if (isep && m->object && m->object->type == OBJT_VNODE) {
1371 			vm_page_and_queue_spin_lock(m);
1372 			if (m->queue - m->pc == PQ_ACTIVE) {
1373 				TAILQ_REMOVE(
1374 					&vm_page_queues[PQ_ACTIVE + q].pl,
1375 					m, pageq);
1376 				TAILQ_INSERT_TAIL(
1377 					&vm_page_queues[PQ_ACTIVE + q].pl,
1378 					m, pageq);
1379 			}
1380 			vm_page_and_queue_spin_unlock(m);
1381 			vm_page_wakeup(m);
1382 			goto next;
1383 		}
1384 
1385 		/*
1386 		 * The count for pagedaemon pages is done after checking the
1387 		 * page for eligibility...
1388 		 */
1389 		mycpu->gd_cnt.v_pdpages++;
1390 
1391 		/*
1392 		 * Check to see "how much" the page has been used and clear
1393 		 * the tracking access bits.  If the object has no references
1394 		 * don't bother paying the expense.
1395 		 */
1396 		actcount = 0;
1397 		if (m->object && m->object->ref_count != 0) {
1398 			if (m->flags & PG_REFERENCED)
1399 				++actcount;
1400 			actcount += pmap_ts_referenced(m);
1401 			if (actcount) {
1402 				m->act_count += ACT_ADVANCE + actcount;
1403 				if (m->act_count > ACT_MAX)
1404 					m->act_count = ACT_MAX;
1405 			}
1406 		}
1407 		vm_page_flag_clear(m, PG_REFERENCED);
1408 
1409 		/*
1410 		 * actcount is only valid if the object ref_count is non-zero.
1411 		 * If the page does not have an object, actcount will be zero.
1412 		 */
1413 		if (actcount && m->object->ref_count != 0) {
1414 			vm_page_and_queue_spin_lock(m);
1415 			if (m->queue - m->pc == PQ_ACTIVE) {
1416 				TAILQ_REMOVE(
1417 					&vm_page_queues[PQ_ACTIVE + q].pl,
1418 					m, pageq);
1419 				TAILQ_INSERT_TAIL(
1420 					&vm_page_queues[PQ_ACTIVE + q].pl,
1421 					m, pageq);
1422 			}
1423 			vm_page_and_queue_spin_unlock(m);
1424 			vm_page_wakeup(m);
1425 		} else {
1426 			switch(m->object->type) {
1427 			case OBJT_DEFAULT:
1428 			case OBJT_SWAP:
1429 				m->act_count -= min(m->act_count,
1430 						    vm_anonmem_decline);
1431 				break;
1432 			default:
1433 				m->act_count -= min(m->act_count,
1434 						    vm_filemem_decline);
1435 				break;
1436 			}
1437 			if (vm_pageout_algorithm ||
1438 			    (m->object == NULL) ||
1439 			    (m->object && (m->object->ref_count == 0)) ||
1440 			    m->act_count < pass + 1
1441 			) {
1442 				/*
1443 				 * Deactivate the page.  If we had a
1444 				 * shortage from our inactive scan try to
1445 				 * free (cache) the page instead.
1446 				 *
1447 				 * Don't just blindly cache the page if
1448 				 * we do not have a shortage from the
1449 				 * inactive scan, that could lead to
1450 				 * gigabytes being moved.
1451 				 */
1452 				--inactive_shortage;
1453 				if (avail_shortage - delta > 0 ||
1454 				    (m->object && (m->object->ref_count == 0)))
1455 				{
1456 					if (avail_shortage - delta > 0)
1457 						++*recycle_countp;
1458 					vm_page_protect(m, VM_PROT_NONE);
1459 					if (m->dirty == 0 &&
1460 					    (m->flags & PG_NEED_COMMIT) == 0 &&
1461 					    avail_shortage - delta > 0) {
1462 						vm_page_cache(m);
1463 					} else {
1464 						vm_page_deactivate(m);
1465 						vm_page_wakeup(m);
1466 					}
1467 				} else {
1468 					vm_page_deactivate(m);
1469 					vm_page_wakeup(m);
1470 				}
1471 				++delta;
1472 			} else {
1473 				vm_page_and_queue_spin_lock(m);
1474 				if (m->queue - m->pc == PQ_ACTIVE) {
1475 					TAILQ_REMOVE(
1476 					    &vm_page_queues[PQ_ACTIVE + q].pl,
1477 					    m, pageq);
1478 					TAILQ_INSERT_TAIL(
1479 					    &vm_page_queues[PQ_ACTIVE + q].pl,
1480 					    m, pageq);
1481 				}
1482 				vm_page_and_queue_spin_unlock(m);
1483 				vm_page_wakeup(m);
1484 			}
1485 		}
1486 next:
1487 		lwkt_yield();
1488 		vm_page_queues_spin_lock(PQ_ACTIVE + q);
1489 	}
1490 
1491 	/*
1492 	 * Clean out our local marker.
1493 	 *
1494 	 * Page queue still spin-locked.
1495 	 */
1496 	TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1497 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1498 
1499 	return (delta);
1500 }
1501 
1502 /*
1503  * The number of actually free pages can drop down to v_free_reserved,
1504  * we try to build the free count back above v_free_min.  Note that
1505  * vm_paging_needed() also returns TRUE if v_free_count is not at
1506  * least v_free_min so that is the minimum we must build the free
1507  * count to.
1508  *
1509  * We use a slightly higher target to improve hysteresis,
1510  * ((v_free_target + v_free_min) / 2).  Since v_free_target
1511  * is usually the same as v_cache_min this maintains about
1512  * half the pages in the free queue as are in the cache queue,
1513  * providing pretty good pipelining for pageout operation.
1514  *
1515  * The system operator can manipulate vm.v_cache_min and
1516  * vm.v_free_target to tune the pageout demon.  Be sure
1517  * to keep vm.v_free_min < vm.v_free_target.
1518  *
1519  * Note that the original paging target is to get at least
1520  * (free_min + cache_min) into (free + cache).  The slightly
1521  * higher target will shift additional pages from cache to free
1522  * without effecting the original paging target in order to
1523  * maintain better hysteresis and not have the free count always
1524  * be dead-on v_free_min.
1525  *
1526  * NOTE: we are still in a critical section.
1527  *
1528  * Pages moved from PQ_CACHE to totally free are not counted in the
1529  * pages_freed counter.
1530  *
1531  * WARNING! Can be called from two pagedaemon threads simultaneously.
1532  */
1533 static void
1534 vm_pageout_scan_cache(int avail_shortage, int pass,
1535 		      int vnodes_skipped, int recycle_count)
1536 {
1537 	static int lastkillticks;
1538 	struct vm_pageout_scan_info info;
1539 	vm_page_t m;
1540 	int isep;
1541 
1542 	isep = (curthread == emergpager);
1543 
1544 	while (vmstats.v_free_count <
1545 	       (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1546 		/*
1547 		 * This steals some code from vm/vm_page.c
1548 		 */
1549 		static int cache_rover = 0;
1550 
1551 		m = vm_page_list_find(PQ_CACHE, cache_rover & PQ_L2_MASK);
1552 		if (m == NULL)
1553 			break;
1554 		/* page is returned removed from its queue and spinlocked */
1555 		if (vm_page_busy_try(m, TRUE)) {
1556 			vm_page_deactivate_locked(m);
1557 			vm_page_spin_unlock(m);
1558 			continue;
1559 		}
1560 		vm_page_spin_unlock(m);
1561 		pagedaemon_wakeup();
1562 		lwkt_yield();
1563 
1564 		/*
1565 		 * Remaining operations run with the page busy and neither
1566 		 * the page or the queue will be spin-locked.
1567 		 */
1568 		if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1569 		    m->hold_count ||
1570 		    m->wire_count) {
1571 			vm_page_deactivate(m);
1572 			vm_page_wakeup(m);
1573 			continue;
1574 		}
1575 		KKASSERT((m->flags & PG_MAPPED) == 0);
1576 		KKASSERT(m->dirty == 0);
1577 		cache_rover += PQ_PRIME2;
1578 		vm_pageout_page_free(m);
1579 		mycpu->gd_cnt.v_dfree++;
1580 	}
1581 
1582 #if !defined(NO_SWAPPING)
1583 	/*
1584 	 * Idle process swapout -- run once per second.
1585 	 */
1586 	if (vm_swap_idle_enabled) {
1587 		static time_t lsec;
1588 		if (time_uptime != lsec) {
1589 			atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_IDLE);
1590 			vm_req_vmdaemon();
1591 			lsec = time_uptime;
1592 		}
1593 	}
1594 #endif
1595 
1596 	/*
1597 	 * If we didn't get enough free pages, and we have skipped a vnode
1598 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1599 	 * if we did not get enough free pages.
1600 	 */
1601 	if (vm_paging_target() > 0) {
1602 		if (vnodes_skipped && vm_page_count_min(0))
1603 			speedup_syncer(NULL);
1604 #if !defined(NO_SWAPPING)
1605 		if (vm_swap_enabled && vm_page_count_target()) {
1606 			atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_NORMAL);
1607 			vm_req_vmdaemon();
1608 		}
1609 #endif
1610 	}
1611 
1612 	/*
1613 	 * Handle catastrophic conditions.  Under good conditions we should
1614 	 * be at the target, well beyond our minimum.  If we could not even
1615 	 * reach our minimum the system is under heavy stress.  But just being
1616 	 * under heavy stress does not trigger process killing.
1617 	 *
1618 	 * We consider ourselves to have run out of memory if the swap pager
1619 	 * is full and avail_shortage is still positive.  The secondary check
1620 	 * ensures that we do not kill processes if the instantanious
1621 	 * availability is good, even if the pageout demon pass says it
1622 	 * couldn't get to the target.
1623 	 *
1624 	 * NOTE!  THE EMERGENCY PAGER (isep) DOES NOT HANDLE SWAP FULL
1625 	 *	  SITUATIONS.
1626 	 */
1627 	if (swap_pager_almost_full &&
1628 	    pass > 0 &&
1629 	    isep == 0 &&
1630 	    (vm_page_count_min(recycle_count) || avail_shortage > 0)) {
1631 		kprintf("Warning: system low on memory+swap "
1632 			"shortage %d for %d ticks!\n",
1633 			avail_shortage, ticks - swap_fail_ticks);
1634 		if (bootverbose)
1635 		kprintf("Metrics: spaf=%d spf=%d pass=%d avail=%d target=%d last=%u\n",
1636 			swap_pager_almost_full,
1637 			swap_pager_full,
1638 			pass,
1639 			avail_shortage,
1640 			vm_paging_target(),
1641 			(unsigned int)(ticks - lastkillticks));
1642 	}
1643 	if (swap_pager_full &&
1644 	    pass > 1 &&
1645 	    isep == 0 &&
1646 	    avail_shortage > 0 &&
1647 	    vm_paging_target() > 0 &&
1648 	    (unsigned int)(ticks - lastkillticks) >= hz) {
1649 		/*
1650 		 * Kill something, maximum rate once per second to give
1651 		 * the process time to free up sufficient memory.
1652 		 */
1653 		lastkillticks = ticks;
1654 		info.bigproc = NULL;
1655 		info.bigsize = 0;
1656 		allproc_scan(vm_pageout_scan_callback, &info, 0);
1657 		if (info.bigproc != NULL) {
1658 			kprintf("Try to kill process %d %s\n",
1659 				info.bigproc->p_pid, info.bigproc->p_comm);
1660 			info.bigproc->p_nice = PRIO_MIN;
1661 			info.bigproc->p_usched->resetpriority(
1662 				FIRST_LWP_IN_PROC(info.bigproc));
1663 			atomic_set_int(&info.bigproc->p_flags, P_LOWMEMKILL);
1664 			killproc(info.bigproc, "out of swap space");
1665 			wakeup(&vmstats.v_free_count);
1666 			PRELE(info.bigproc);
1667 		}
1668 	}
1669 }
1670 
1671 static int
1672 vm_pageout_scan_callback(struct proc *p, void *data)
1673 {
1674 	struct vm_pageout_scan_info *info = data;
1675 	vm_offset_t size;
1676 
1677 	/*
1678 	 * Never kill system processes or init.  If we have configured swap
1679 	 * then try to avoid killing low-numbered pids.
1680 	 */
1681 	if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1682 	    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1683 		return (0);
1684 	}
1685 
1686 	lwkt_gettoken(&p->p_token);
1687 
1688 	/*
1689 	 * if the process is in a non-running type state,
1690 	 * don't touch it.
1691 	 */
1692 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
1693 		lwkt_reltoken(&p->p_token);
1694 		return (0);
1695 	}
1696 
1697 	/*
1698 	 * Get the approximate process size.  Note that anonymous pages
1699 	 * with backing swap will be counted twice, but there should not
1700 	 * be too many such pages due to the stress the VM system is
1701 	 * under at this point.
1702 	 */
1703 	size = vmspace_anonymous_count(p->p_vmspace) +
1704 		vmspace_swap_count(p->p_vmspace);
1705 
1706 	/*
1707 	 * If the this process is bigger than the biggest one
1708 	 * remember it.
1709 	 */
1710 	if (info->bigsize < size) {
1711 		if (info->bigproc)
1712 			PRELE(info->bigproc);
1713 		PHOLD(p);
1714 		info->bigproc = p;
1715 		info->bigsize = size;
1716 	}
1717 	lwkt_reltoken(&p->p_token);
1718 	lwkt_yield();
1719 
1720 	return(0);
1721 }
1722 
1723 /*
1724  * This routine tries to maintain the pseudo LRU active queue,
1725  * so that during long periods of time where there is no paging,
1726  * that some statistic accumulation still occurs.  This code
1727  * helps the situation where paging just starts to occur.
1728  */
1729 static void
1730 vm_pageout_page_stats(int q)
1731 {
1732 	static int fullintervalcount = 0;
1733 	struct vm_page marker;
1734 	vm_page_t m;
1735 	int pcount, tpcount;		/* Number of pages to check */
1736 	int page_shortage;
1737 
1738 	page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1739 			 vmstats.v_free_min) -
1740 			(vmstats.v_free_count + vmstats.v_inactive_count +
1741 			 vmstats.v_cache_count);
1742 
1743 	if (page_shortage <= 0)
1744 		return;
1745 
1746 	pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1747 	fullintervalcount += vm_pageout_stats_interval;
1748 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1749 		tpcount = (vm_pageout_stats_max * pcount) /
1750 			  vmstats.v_page_count + 1;
1751 		if (pcount > tpcount)
1752 			pcount = tpcount;
1753 	} else {
1754 		fullintervalcount = 0;
1755 	}
1756 
1757 	bzero(&marker, sizeof(marker));
1758 	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
1759 	marker.queue = PQ_ACTIVE + q;
1760 	marker.pc = q;
1761 	marker.wire_count = 1;
1762 
1763 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1764 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1765 
1766 	/*
1767 	 * Queue locked at top of loop to avoid stack marker issues.
1768 	 */
1769 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1770 	       pcount-- > 0)
1771 	{
1772 		int actcount;
1773 
1774 		KKASSERT(m->queue == PQ_ACTIVE + q);
1775 		TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1776 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1777 				   &marker, pageq);
1778 
1779 		/*
1780 		 * Skip marker pages (atomic against other markers to avoid
1781 		 * infinite hop-over scans).
1782 		 */
1783 		if (m->flags & PG_MARKER)
1784 			continue;
1785 
1786 		/*
1787 		 * Ignore pages we can't busy
1788 		 */
1789 		if (vm_page_busy_try(m, TRUE))
1790 			continue;
1791 
1792 		/*
1793 		 * Remaining operations run with the page busy and neither
1794 		 * the page or the queue will be spin-locked.
1795 		 */
1796 		vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1797 		KKASSERT(m->queue == PQ_ACTIVE + q);
1798 
1799 		/*
1800 		 * We now have a safely busied page, the page and queue
1801 		 * spinlocks have been released.
1802 		 *
1803 		 * Ignore held pages
1804 		 */
1805 		if (m->hold_count) {
1806 			vm_page_wakeup(m);
1807 			goto next;
1808 		}
1809 
1810 		/*
1811 		 * Calculate activity
1812 		 */
1813 		actcount = 0;
1814 		if (m->flags & PG_REFERENCED) {
1815 			vm_page_flag_clear(m, PG_REFERENCED);
1816 			actcount += 1;
1817 		}
1818 		actcount += pmap_ts_referenced(m);
1819 
1820 		/*
1821 		 * Update act_count and move page to end of queue.
1822 		 */
1823 		if (actcount) {
1824 			m->act_count += ACT_ADVANCE + actcount;
1825 			if (m->act_count > ACT_MAX)
1826 				m->act_count = ACT_MAX;
1827 			vm_page_and_queue_spin_lock(m);
1828 			if (m->queue - m->pc == PQ_ACTIVE) {
1829 				TAILQ_REMOVE(
1830 					&vm_page_queues[PQ_ACTIVE + q].pl,
1831 					m, pageq);
1832 				TAILQ_INSERT_TAIL(
1833 					&vm_page_queues[PQ_ACTIVE + q].pl,
1834 					m, pageq);
1835 			}
1836 			vm_page_and_queue_spin_unlock(m);
1837 			vm_page_wakeup(m);
1838 			goto next;
1839 		}
1840 
1841 		if (m->act_count == 0) {
1842 			/*
1843 			 * We turn off page access, so that we have
1844 			 * more accurate RSS stats.  We don't do this
1845 			 * in the normal page deactivation when the
1846 			 * system is loaded VM wise, because the
1847 			 * cost of the large number of page protect
1848 			 * operations would be higher than the value
1849 			 * of doing the operation.
1850 			 *
1851 			 * We use the marker to save our place so
1852 			 * we can release the spin lock.  both (m)
1853 			 * and (next) will be invalid.
1854 			 */
1855 			vm_page_protect(m, VM_PROT_NONE);
1856 			vm_page_deactivate(m);
1857 		} else {
1858 			m->act_count -= min(m->act_count, ACT_DECLINE);
1859 			vm_page_and_queue_spin_lock(m);
1860 			if (m->queue - m->pc == PQ_ACTIVE) {
1861 				TAILQ_REMOVE(
1862 					&vm_page_queues[PQ_ACTIVE + q].pl,
1863 					m, pageq);
1864 				TAILQ_INSERT_TAIL(
1865 					&vm_page_queues[PQ_ACTIVE + q].pl,
1866 					m, pageq);
1867 			}
1868 			vm_page_and_queue_spin_unlock(m);
1869 		}
1870 		vm_page_wakeup(m);
1871 next:
1872 		vm_page_queues_spin_lock(PQ_ACTIVE + q);
1873 	}
1874 
1875 	/*
1876 	 * Remove our local marker
1877 	 *
1878 	 * Page queue still spin-locked.
1879 	 */
1880 	TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1881 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1882 }
1883 
1884 static int
1885 vm_pageout_free_page_calc(vm_size_t count)
1886 {
1887 	if (count < vmstats.v_page_count)
1888 		 return 0;
1889 	/*
1890 	 * free_reserved needs to include enough for the largest swap pager
1891 	 * structures plus enough for any pv_entry structs when paging.
1892 	 *
1893 	 * v_free_min		normal allocations
1894 	 * v_free_reserved	system allocations
1895 	 * v_pageout_free_min	allocations by pageout daemon
1896 	 * v_interrupt_free_min	low level allocations (e.g swap structures)
1897 	 */
1898 	if (vmstats.v_page_count > 1024)
1899 		vmstats.v_free_min = 64 + (vmstats.v_page_count - 1024) / 200;
1900 	else
1901 		vmstats.v_free_min = 64;
1902 
1903 	/*
1904 	 * Make sure the vmmeter slop can't blow out our global minimums.
1905 	 *
1906 	 * However, to accomodate weird configurations (vkernels with many
1907 	 * cpus and little memory, or artifically reduced hw.physmem), do
1908 	 * not allow v_free_min to exceed 1/20 of ram or the pageout demon
1909 	 * will go out of control.
1910 	 */
1911 	if (vmstats.v_free_min < VMMETER_SLOP_COUNT * ncpus * 10)
1912 		vmstats.v_free_min = VMMETER_SLOP_COUNT * ncpus * 10;
1913 	if (vmstats.v_free_min > vmstats.v_page_count / 20)
1914 		vmstats.v_free_min = vmstats.v_page_count / 20;
1915 
1916 	vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
1917 	vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
1918 	vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
1919 	vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
1920 
1921 	return 1;
1922 }
1923 
1924 
1925 /*
1926  * vm_pageout is the high level pageout daemon.  TWO kernel threads run
1927  * this daemon, the primary pageout daemon and the emergency pageout daemon.
1928  *
1929  * The emergency pageout daemon takes over when the primary pageout daemon
1930  * deadlocks.  The emergency pageout daemon ONLY pages out to swap, thus
1931  * avoiding the many low-memory deadlocks which can occur when paging out
1932  * to VFS's.
1933  */
1934 static void
1935 vm_pageout_thread(void)
1936 {
1937 	int pass;
1938 	int q;
1939 	int q1iterator = 0;
1940 	int q2iterator = 0;
1941 	int isep;
1942 	int emrunning;
1943 
1944 	curthread->td_flags |= TDF_SYSTHREAD;
1945 
1946 	/*
1947 	 * We only need to setup once.
1948 	 */
1949 	isep = 0;
1950 	emrunning = 0;
1951 	if (curthread == emergpager) {
1952 		isep = 1;
1953 		goto skip_setup;
1954 	}
1955 
1956 	/*
1957 	 * Initialize some paging parameters.
1958 	 */
1959 	vm_pageout_free_page_calc(vmstats.v_page_count);
1960 
1961 	/*
1962 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
1963 	 * that these are more a measure of the VM cache queue hysteresis
1964 	 * then the VM free queue.  Specifically, v_free_target is the
1965 	 * high water mark (free+cache pages).
1966 	 *
1967 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
1968 	 * low water mark, while v_free_min is the stop.  v_cache_min must
1969 	 * be big enough to handle memory needs while the pageout daemon
1970 	 * is signalled and run to free more pages.
1971 	 */
1972 	if (vmstats.v_free_count > 6144)
1973 		vmstats.v_free_target = 4 * vmstats.v_free_min +
1974 					vmstats.v_free_reserved;
1975 	else
1976 		vmstats.v_free_target = 2 * vmstats.v_free_min +
1977 					vmstats.v_free_reserved;
1978 
1979 	/*
1980 	 * NOTE: With the new buffer cache b_act_count we want the default
1981 	 *	 inactive target to be a percentage of available memory.
1982 	 *
1983 	 *	 The inactive target essentially determines the minimum
1984 	 *	 number of 'temporary' pages capable of caching one-time-use
1985 	 *	 files when the VM system is otherwise full of pages
1986 	 *	 belonging to multi-time-use files or active program data.
1987 	 *
1988 	 * NOTE: The inactive target is aggressively persued only if the
1989 	 *	 inactive queue becomes too small.  If the inactive queue
1990 	 *	 is large enough to satisfy page movement to free+cache
1991 	 *	 then it is repopulated more slowly from the active queue.
1992 	 *	 This allows a general inactive_target default to be set.
1993 	 *
1994 	 *	 There is an issue here for processes which sit mostly idle
1995 	 *	 'overnight', such as sshd, tcsh, and X.  Any movement from
1996 	 *	 the active queue will eventually cause such pages to
1997 	 *	 recycle eventually causing a lot of paging in the morning.
1998 	 *	 To reduce the incidence of this pages cycled out of the
1999 	 *	 buffer cache are moved directly to the inactive queue if
2000 	 *	 they were only used once or twice.
2001 	 *
2002 	 *	 The vfs.vm_cycle_point sysctl can be used to adjust this.
2003 	 *	 Increasing the value (up to 64) increases the number of
2004 	 *	 buffer recyclements which go directly to the inactive queue.
2005 	 */
2006 	if (vmstats.v_free_count > 2048) {
2007 		vmstats.v_cache_min = vmstats.v_free_target;
2008 		vmstats.v_cache_max = 2 * vmstats.v_cache_min;
2009 	} else {
2010 		vmstats.v_cache_min = 0;
2011 		vmstats.v_cache_max = 0;
2012 	}
2013 	vmstats.v_inactive_target = vmstats.v_free_count / 4;
2014 
2015 	/* XXX does not really belong here */
2016 	if (vm_page_max_wired == 0)
2017 		vm_page_max_wired = vmstats.v_free_count / 3;
2018 
2019 	if (vm_pageout_stats_max == 0)
2020 		vm_pageout_stats_max = vmstats.v_free_target;
2021 
2022 	/*
2023 	 * Set interval in seconds for stats scan.
2024 	 */
2025 	if (vm_pageout_stats_interval == 0)
2026 		vm_pageout_stats_interval = 5;
2027 	if (vm_pageout_full_stats_interval == 0)
2028 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
2029 
2030 
2031 	/*
2032 	 * Set maximum free per pass
2033 	 */
2034 	if (vm_pageout_stats_free_max == 0)
2035 		vm_pageout_stats_free_max = 5;
2036 
2037 	swap_pager_swap_init();
2038 	pass = 0;
2039 
2040 	atomic_swap_int(&sequence_emerg_pager, 1);
2041 	wakeup(&sequence_emerg_pager);
2042 
2043 skip_setup:
2044 	/*
2045 	 * Sequence emergency pager startup
2046 	 */
2047 	if (isep) {
2048 		while (sequence_emerg_pager == 0)
2049 			tsleep(&sequence_emerg_pager, 0, "pstartup", hz);
2050 	}
2051 
2052 	/*
2053 	 * The pageout daemon is never done, so loop forever.
2054 	 *
2055 	 * WARNING!  This code is being executed by two kernel threads
2056 	 *	     potentially simultaneously.
2057 	 */
2058 	while (TRUE) {
2059 		int error;
2060 		int avail_shortage;
2061 		int inactive_shortage;
2062 		int vnodes_skipped = 0;
2063 		int recycle_count = 0;
2064 		int tmp;
2065 
2066 		/*
2067 		 * Wait for an action request.  If we timeout check to
2068 		 * see if paging is needed (in case the normal wakeup
2069 		 * code raced us).
2070 		 */
2071 		if (isep) {
2072 			/*
2073 			 * Emergency pagedaemon monitors the primary
2074 			 * pagedaemon while vm_pages_needed != 0.
2075 			 *
2076 			 * The emergency pagedaemon only runs if VM paging
2077 			 * is needed and the primary pagedaemon has not
2078 			 * updated vm_pagedaemon_time for more than 2 seconds.
2079 			 */
2080 			if (vm_pages_needed)
2081 				tsleep(&vm_pagedaemon_time, 0, "psleep", hz);
2082 			else
2083 				tsleep(&vm_pagedaemon_time, 0, "psleep", hz*10);
2084 			if (vm_pages_needed == 0) {
2085 				pass = 0;
2086 				continue;
2087 			}
2088 			if ((int)(ticks - vm_pagedaemon_time) < hz * 2) {
2089 				if (emrunning) {
2090 					emrunning = 0;
2091 					kprintf("Emergency pager finished\n");
2092 				}
2093 				pass = 0;
2094 				continue;
2095 			}
2096 			if (emrunning == 0) {
2097 				emrunning = 1;
2098 				kprintf("Emergency pager running\n");
2099 			}
2100 		} else {
2101 			/*
2102 			 * Primary pagedaemon
2103 			 */
2104 			if (vm_pages_needed == 0) {
2105 				error = tsleep(&vm_pages_needed,
2106 					       0, "psleep",
2107 					       vm_pageout_stats_interval * hz);
2108 				if (error &&
2109 				    vm_paging_needed() == 0 &&
2110 				    vm_pages_needed == 0) {
2111 					for (q = 0; q < PQ_L2_SIZE; ++q)
2112 						vm_pageout_page_stats(q);
2113 					continue;
2114 				}
2115 				vm_pagedaemon_time = ticks;
2116 				vm_pages_needed = 1;
2117 
2118 				/*
2119 				 * Wake the emergency pagedaemon up so it
2120 				 * can monitor us.  It will automatically
2121 				 * go back into a long sleep when
2122 				 * vm_pages_needed returns to 0.
2123 				 */
2124 				wakeup(&vm_pagedaemon_time);
2125 			}
2126 		}
2127 
2128 		mycpu->gd_cnt.v_pdwakeups++;
2129 
2130 		/*
2131 		 * Scan for INACTIVE->CLEAN/PAGEOUT
2132 		 *
2133 		 * This routine tries to avoid thrashing the system with
2134 		 * unnecessary activity.
2135 		 *
2136 		 * Calculate our target for the number of free+cache pages we
2137 		 * want to get to.  This is higher then the number that causes
2138 		 * allocations to stall (severe) in order to provide hysteresis,
2139 		 * and if we don't make it all the way but get to the minimum
2140 		 * we're happy.  Goose it a bit if there are multiple requests
2141 		 * for memory.
2142 		 *
2143 		 * Don't reduce avail_shortage inside the loop or the
2144 		 * PQAVERAGE() calculation will break.
2145 		 *
2146 		 * NOTE! deficit is differentiated from avail_shortage as
2147 		 *	 REQUIRING at least (deficit) pages to be cleaned,
2148 		 *	 even if the page queues are in good shape.  This
2149 		 *	 is used primarily for handling per-process
2150 		 *	 RLIMIT_RSS and may also see small values when
2151 		 *	 processes block due to low memory.
2152 		 */
2153 		vmstats_rollup();
2154 		if (isep == 0)
2155 			vm_pagedaemon_time = ticks;
2156 		avail_shortage = vm_paging_target() + vm_pageout_deficit;
2157 		vm_pageout_deficit = 0;
2158 
2159 		if (avail_shortage > 0) {
2160 			int delta = 0;
2161 
2162 			for (q = 0; q < PQ_L2_SIZE; ++q) {
2163 				delta += vm_pageout_scan_inactive(
2164 					    pass,
2165 					    (q + q1iterator) & PQ_L2_MASK,
2166 					    PQAVERAGE(avail_shortage),
2167 					    &vnodes_skipped);
2168 				if (avail_shortage - delta <= 0)
2169 					break;
2170 			}
2171 			avail_shortage -= delta;
2172 			q1iterator = q + 1;
2173 		}
2174 
2175 		/*
2176 		 * Figure out how many active pages we must deactivate.  If
2177 		 * we were able to reach our target with just the inactive
2178 		 * scan above we limit the number of active pages we
2179 		 * deactivate to reduce unnecessary work.
2180 		 */
2181 		vmstats_rollup();
2182 		if (isep == 0)
2183 			vm_pagedaemon_time = ticks;
2184 		inactive_shortage = vmstats.v_inactive_target -
2185 				    vmstats.v_inactive_count;
2186 
2187 		/*
2188 		 * If we were unable to free sufficient inactive pages to
2189 		 * satisfy the free/cache queue requirements then simply
2190 		 * reaching the inactive target may not be good enough.
2191 		 * Try to deactivate pages in excess of the target based
2192 		 * on the shortfall.
2193 		 *
2194 		 * However to prevent thrashing the VM system do not
2195 		 * deactivate more than an additional 1/10 the inactive
2196 		 * target's worth of active pages.
2197 		 */
2198 		if (avail_shortage > 0) {
2199 			tmp = avail_shortage * 2;
2200 			if (tmp > vmstats.v_inactive_target / 10)
2201 				tmp = vmstats.v_inactive_target / 10;
2202 			inactive_shortage += tmp;
2203 		}
2204 
2205 		/*
2206 		 * Only trigger a pmap cleanup on inactive shortage.
2207 		 */
2208 		if (isep == 0 && inactive_shortage > 0) {
2209 			pmap_collect();
2210 		}
2211 
2212 		/*
2213 		 * Scan for ACTIVE->INACTIVE
2214 		 *
2215 		 * Only trigger on inactive shortage.  Triggering on
2216 		 * avail_shortage can starve the active queue with
2217 		 * unnecessary active->inactive transitions and destroy
2218 		 * performance.
2219 		 *
2220 		 * If this is the emergency pager, always try to move
2221 		 * a few pages from active to inactive because the inactive
2222 		 * queue might have enough pages, but not enough anonymous
2223 		 * pages.
2224 		 */
2225 		if (isep && inactive_shortage < vm_emerg_launder)
2226 			inactive_shortage = vm_emerg_launder;
2227 
2228 		if (/*avail_shortage > 0 ||*/ inactive_shortage > 0) {
2229 			int delta = 0;
2230 
2231 			for (q = 0; q < PQ_L2_SIZE; ++q) {
2232 				delta += vm_pageout_scan_active(
2233 						pass,
2234 						(q + q2iterator) & PQ_L2_MASK,
2235 						PQAVERAGE(avail_shortage),
2236 						PQAVERAGE(inactive_shortage),
2237 						&recycle_count);
2238 				if (inactive_shortage - delta <= 0 &&
2239 				    avail_shortage - delta <= 0) {
2240 					break;
2241 				}
2242 			}
2243 			inactive_shortage -= delta;
2244 			avail_shortage -= delta;
2245 			q2iterator = q + 1;
2246 		}
2247 
2248 		/*
2249 		 * Scan for CACHE->FREE
2250 		 *
2251 		 * Finally free enough cache pages to meet our free page
2252 		 * requirement and take more drastic measures if we are
2253 		 * still in trouble.
2254 		 */
2255 		vmstats_rollup();
2256 		if (isep == 0)
2257 			vm_pagedaemon_time = ticks;
2258 		vm_pageout_scan_cache(avail_shortage, pass,
2259 				      vnodes_skipped, recycle_count);
2260 
2261 		/*
2262 		 * Wait for more work.
2263 		 */
2264 		if (avail_shortage > 0) {
2265 			++pass;
2266 			if (pass < 10 && vm_pages_needed > 1) {
2267 				/*
2268 				 * Normal operation, additional processes
2269 				 * have already kicked us.  Retry immediately
2270 				 * unless swap space is completely full in
2271 				 * which case delay a bit.
2272 				 */
2273 				if (swap_pager_full) {
2274 					tsleep(&vm_pages_needed, 0, "pdelay",
2275 						hz / 5);
2276 				} /* else immediate retry */
2277 			} else if (pass < 10) {
2278 				/*
2279 				 * Normal operation, fewer processes.  Delay
2280 				 * a bit but allow wakeups.  vm_pages_needed
2281 				 * is only adjusted against the primary
2282 				 * pagedaemon here.
2283 				 */
2284 				if (isep == 0)
2285 					vm_pages_needed = 0;
2286 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2287 				if (isep == 0)
2288 					vm_pages_needed = 1;
2289 			} else if (swap_pager_full == 0) {
2290 				/*
2291 				 * We've taken too many passes, forced delay.
2292 				 */
2293 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2294 			} else {
2295 				/*
2296 				 * Running out of memory, catastrophic
2297 				 * back-off to one-second intervals.
2298 				 */
2299 				tsleep(&vm_pages_needed, 0, "pdelay", hz);
2300 			}
2301 		} else if (vm_pages_needed) {
2302 			/*
2303 			 * Interlocked wakeup of waiters (non-optional).
2304 			 *
2305 			 * Similar to vm_page_free_wakeup() in vm_page.c,
2306 			 * wake
2307 			 */
2308 			pass = 0;
2309 			if (!vm_page_count_min(vm_page_free_hysteresis) ||
2310 			    !vm_page_count_target()) {
2311 				vm_pages_needed = 0;
2312 				wakeup(&vmstats.v_free_count);
2313 			}
2314 		} else {
2315 			pass = 0;
2316 		}
2317 	}
2318 }
2319 
2320 static struct kproc_desc pg1_kp = {
2321 	"pagedaemon",
2322 	vm_pageout_thread,
2323 	&pagethread
2324 };
2325 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &pg1_kp);
2326 
2327 static struct kproc_desc pg2_kp = {
2328 	"emergpager",
2329 	vm_pageout_thread,
2330 	&emergpager
2331 };
2332 SYSINIT(emergpager, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, kproc_start, &pg2_kp);
2333 
2334 
2335 /*
2336  * Called after allocating a page out of the cache or free queue
2337  * to possibly wake the pagedaemon up to replentish our supply.
2338  *
2339  * We try to generate some hysteresis by waking the pagedaemon up
2340  * when our free+cache pages go below the free_min+cache_min level.
2341  * The pagedaemon tries to get the count back up to at least the
2342  * minimum, and through to the target level if possible.
2343  *
2344  * If the pagedaemon is already active bump vm_pages_needed as a hint
2345  * that there are even more requests pending.
2346  *
2347  * SMP races ok?
2348  * No requirements.
2349  */
2350 void
2351 pagedaemon_wakeup(void)
2352 {
2353 	if (vm_paging_needed() && curthread != pagethread) {
2354 		if (vm_pages_needed == 0) {
2355 			vm_pages_needed = 1;	/* SMP race ok */
2356 			wakeup(&vm_pages_needed);
2357 		} else if (vm_page_count_min(0)) {
2358 			++vm_pages_needed;	/* SMP race ok */
2359 		}
2360 	}
2361 }
2362 
2363 #if !defined(NO_SWAPPING)
2364 
2365 /*
2366  * SMP races ok?
2367  * No requirements.
2368  */
2369 static void
2370 vm_req_vmdaemon(void)
2371 {
2372 	static int lastrun = 0;
2373 
2374 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2375 		wakeup(&vm_daemon_needed);
2376 		lastrun = ticks;
2377 	}
2378 }
2379 
2380 static int vm_daemon_callback(struct proc *p, void *data __unused);
2381 
2382 /*
2383  * No requirements.
2384  */
2385 static void
2386 vm_daemon(void)
2387 {
2388 	int req_swapout;
2389 
2390 	while (TRUE) {
2391 		tsleep(&vm_daemon_needed, 0, "psleep", 0);
2392 		req_swapout = atomic_swap_int(&vm_pageout_req_swapout, 0);
2393 
2394 		/*
2395 		 * forced swapouts
2396 		 */
2397 		if (req_swapout)
2398 			swapout_procs(vm_pageout_req_swapout);
2399 
2400 		/*
2401 		 * scan the processes for exceeding their rlimits or if
2402 		 * process is swapped out -- deactivate pages
2403 		 */
2404 		allproc_scan(vm_daemon_callback, NULL, 0);
2405 	}
2406 }
2407 
2408 static int
2409 vm_daemon_callback(struct proc *p, void *data __unused)
2410 {
2411 	struct vmspace *vm;
2412 	vm_pindex_t limit, size;
2413 
2414 	/*
2415 	 * if this is a system process or if we have already
2416 	 * looked at this process, skip it.
2417 	 */
2418 	lwkt_gettoken(&p->p_token);
2419 
2420 	if (p->p_flags & (P_SYSTEM | P_WEXIT)) {
2421 		lwkt_reltoken(&p->p_token);
2422 		return (0);
2423 	}
2424 
2425 	/*
2426 	 * if the process is in a non-running type state,
2427 	 * don't touch it.
2428 	 */
2429 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
2430 		lwkt_reltoken(&p->p_token);
2431 		return (0);
2432 	}
2433 
2434 	/*
2435 	 * get a limit
2436 	 */
2437 	limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2438 			        p->p_rlimit[RLIMIT_RSS].rlim_max));
2439 
2440 	/*
2441 	 * let processes that are swapped out really be
2442 	 * swapped out.  Set the limit to nothing to get as
2443 	 * many pages out to swap as possible.
2444 	 */
2445 	if (p->p_flags & P_SWAPPEDOUT)
2446 		limit = 0;
2447 
2448 	vm = p->p_vmspace;
2449 	vmspace_hold(vm);
2450 	size = pmap_resident_tlnw_count(&vm->vm_pmap);
2451 	if (limit >= 0 && size > 4096 &&
2452 	    size - 4096 >= limit && vm_pageout_memuse_mode >= 1) {
2453 		vm_pageout_map_deactivate_pages(&vm->vm_map, limit);
2454 	}
2455 	vmspace_drop(vm);
2456 
2457 	lwkt_reltoken(&p->p_token);
2458 
2459 	return (0);
2460 }
2461 
2462 #endif
2463