xref: /dflybsd-src/sys/vm/vm_pageout.c (revision 872a09d51adf63b4bdae6adb1d96a53f76e161e2)
1 /*
2  * Copyright (c) 2003-2020 The DragonFly Project.  All rights reserved.
3  *
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  * Copyright (c) 1991 Regents of the University of California.
35  * All rights reserved.
36  * Copyright (c) 1994 John S. Dyson
37  * All rights reserved.
38  * Copyright (c) 1994 David Greenman
39  * All rights reserved.
40  *
41  * This code is derived from software contributed to Berkeley by
42  * The Mach Operating System project at Carnegie-Mellon University.
43  *
44  * Redistribution and use in source and binary forms, with or without
45  * modification, are permitted provided that the following conditions
46  * are met:
47  * 1. Redistributions of source code must retain the above copyright
48  *    notice, this list of conditions and the following disclaimer.
49  * 2. Redistributions in binary form must reproduce the above copyright
50  *    notice, this list of conditions and the following disclaimer in the
51  *    documentation and/or other materials provided with the distribution.
52  * 3. Neither the name of the University nor the names of its contributors
53  *    may be used to endorse or promote products derived from this software
54  *    without specific prior written permission.
55  *
56  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
57  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
58  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
59  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
60  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
61  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
62  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
63  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
64  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
65  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66  * SUCH DAMAGE.
67  *
68  *	from: @(#)vm_pageout.c	7.4 (Berkeley) 5/7/91
69  *
70  *
71  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
72  * All rights reserved.
73  *
74  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
75  *
76  * Permission to use, copy, modify and distribute this software and
77  * its documentation is hereby granted, provided that both the copyright
78  * notice and this permission notice appear in all copies of the
79  * software, derivative works or modified versions, and any portions
80  * thereof, and that both notices appear in supporting documentation.
81  *
82  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
83  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
84  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
85  *
86  * Carnegie Mellon requests users of this software to return to
87  *
88  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
89  *  School of Computer Science
90  *  Carnegie Mellon University
91  *  Pittsburgh PA 15213-3890
92  *
93  * any improvements or extensions that they make and grant Carnegie the
94  * rights to redistribute these changes.
95  */
96 
97 /*
98  * The proverbial page-out daemon, rewritten many times over the decades.
99  */
100 
101 #include "opt_vm.h"
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/kernel.h>
105 #include <sys/proc.h>
106 #include <sys/kthread.h>
107 #include <sys/resourcevar.h>
108 #include <sys/signalvar.h>
109 #include <sys/vnode.h>
110 #include <sys/vmmeter.h>
111 #include <sys/conf.h>
112 #include <sys/sysctl.h>
113 
114 #include <vm/vm.h>
115 #include <vm/vm_param.h>
116 #include <sys/lock.h>
117 #include <vm/vm_object.h>
118 #include <vm/vm_page.h>
119 #include <vm/vm_map.h>
120 #include <vm/vm_pageout.h>
121 #include <vm/vm_pager.h>
122 #include <vm/swap_pager.h>
123 #include <vm/vm_extern.h>
124 
125 #include <sys/spinlock2.h>
126 #include <vm/vm_page2.h>
127 
128 /*
129  * System initialization
130  */
131 
132 /* the kernel process "vm_pageout"*/
133 static int vm_pageout_page(vm_page_t m, long *max_launderp,
134 			   long *vnodes_skippedp, struct vnode **vpfailedp,
135 			   int pass, int vmflush_flags);
136 static int vm_pageout_clean_helper (vm_page_t, int);
137 static void vm_pageout_free_page_calc (vm_size_t count);
138 static void vm_pageout_page_free(vm_page_t m) ;
139 struct thread *emergpager;
140 struct thread *pagethread;
141 static int sequence_emerg_pager;
142 
143 #if !defined(NO_SWAPPING)
144 /* the kernel process "vm_daemon"*/
145 static void vm_daemon (void);
146 static struct	thread *vmthread;
147 
148 static struct kproc_desc vm_kp = {
149 	"vmdaemon",
150 	vm_daemon,
151 	&vmthread
152 };
153 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp);
154 #endif
155 
156 int vm_pages_needed = 0;	/* Event on which pageout daemon sleeps */
157 int vm_pageout_deficit = 0;	/* Estimated number of pages deficit */
158 int vm_pageout_pages_needed = 0;/* pageout daemon needs pages */
159 int vm_page_free_hysteresis = 16;
160 static int vm_pagedaemon_time;
161 
162 #if !defined(NO_SWAPPING)
163 static int vm_pageout_req_swapout;
164 static int vm_daemon_needed;
165 #endif
166 __read_mostly static int vm_max_launder = 4096;
167 __read_mostly static int vm_emerg_launder = 100;
168 __read_mostly static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0;
169 __read_mostly static int vm_pageout_full_stats_interval = 0;
170 __read_mostly static int vm_pageout_stats_free_max=0, vm_pageout_algorithm=0;
171 __read_mostly static int defer_swap_pageouts=0;
172 __read_mostly static int disable_swap_pageouts=0;
173 __read_mostly static u_int vm_anonmem_decline = ACT_DECLINE;
174 __read_mostly static u_int vm_filemem_decline = ACT_DECLINE * 2;
175 __read_mostly static int vm_pageout_debug;
176 
177 #if defined(NO_SWAPPING)
178 __read_mostly static int vm_swap_enabled=0;
179 __read_mostly static int vm_swap_idle_enabled=0;
180 #else
181 __read_mostly static int vm_swap_enabled=1;
182 __read_mostly static int vm_swap_idle_enabled=0;
183 #endif
184 
185 /* 0-disable, 1-passive, 2-active swp*/
186 __read_mostly int vm_pageout_memuse_mode=2;
187 __read_mostly int vm_pageout_allow_active=1;
188 
189 SYSCTL_UINT(_vm, VM_PAGEOUT_ALGORITHM, anonmem_decline,
190 	CTLFLAG_RW, &vm_anonmem_decline, 0, "active->inactive anon memory");
191 
192 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, filemem_decline,
193 	CTLFLAG_RW, &vm_filemem_decline, 0, "active->inactive file cache");
194 
195 SYSCTL_INT(_vm, OID_AUTO, page_free_hysteresis,
196 	CTLFLAG_RW, &vm_page_free_hysteresis, 0,
197 	"Free more pages than the minimum required");
198 
199 SYSCTL_INT(_vm, OID_AUTO, max_launder,
200 	CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout");
201 SYSCTL_INT(_vm, OID_AUTO, emerg_launder,
202 	CTLFLAG_RW, &vm_emerg_launder, 0, "Emergency pager minimum");
203 
204 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max,
205 	CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length");
206 
207 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval,
208 	CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan");
209 
210 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval,
211 	CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan");
212 
213 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_free_max,
214 	CTLFLAG_RW, &vm_pageout_stats_free_max, 0, "Not implemented");
215 SYSCTL_INT(_vm, OID_AUTO, pageout_memuse_mode,
216 	CTLFLAG_RW, &vm_pageout_memuse_mode, 0, "memoryuse resource mode");
217 SYSCTL_INT(_vm, OID_AUTO, pageout_allow_active,
218 	CTLFLAG_RW, &vm_pageout_allow_active, 0, "allow inactive+active");
219 SYSCTL_INT(_vm, OID_AUTO, pageout_debug,
220 	CTLFLAG_RW, &vm_pageout_debug, 0, "debug pageout pages (count)");
221 
222 
223 #if defined(NO_SWAPPING)
224 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
225 	CTLFLAG_RD, &vm_swap_enabled, 0, "");
226 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
227 	CTLFLAG_RD, &vm_swap_idle_enabled, 0, "");
228 #else
229 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled,
230 	CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout");
231 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled,
232 	CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria");
233 #endif
234 
235 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts,
236 	CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem");
237 
238 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts,
239 	CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages");
240 
241 static int pageout_lock_miss;
242 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss,
243 	CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout");
244 
245 int vm_page_max_wired;		/* XXX max # of wired pages system-wide */
246 
247 #if !defined(NO_SWAPPING)
248 static void vm_req_vmdaemon (void);
249 #endif
250 static void vm_pageout_page_stats(int q);
251 
252 /*
253  * Calculate approximately how many pages on each queue to try to
254  * clean.  An exact calculation creates an edge condition when the
255  * queues are unbalanced so add significant slop.  The queue scans
256  * will stop early when targets are reached and will start where they
257  * left off on the next pass.
258  *
259  * We need to be generous here because there are all sorts of loading
260  * conditions that can cause edge cases if try to average over all queues.
261  * In particular, storage subsystems have become so fast that paging
262  * activity can become quite frantic.  Eventually we will probably need
263  * two paging threads, one for dirty pages and one for clean, to deal
264  * with the bandwidth requirements.
265 
266  * So what we do is calculate a value that can be satisfied nominally by
267  * only having to scan half the queues.
268  */
269 static __inline long
270 PQAVERAGE(long n)
271 {
272 	long avg;
273 
274 	if (n >= 0) {
275 		avg = ((n + (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) + 1);
276 	} else {
277 		avg = ((n - (PQ_L2_SIZE - 1)) / (PQ_L2_SIZE / 2) - 1);
278 	}
279 	return avg;
280 }
281 
282 /*
283  * vm_pageout_clean_helper:
284  *
285  * Clean the page and remove it from the laundry.  The page must be busied
286  * by the caller and will be disposed of (put away, flushed) by this routine.
287  */
288 static int
289 vm_pageout_clean_helper(vm_page_t m, int vmflush_flags)
290 {
291 	vm_object_t object;
292 	vm_page_t mc[BLIST_MAX_ALLOC];
293 	int error;
294 	int ib, is, page_base;
295 	vm_pindex_t pindex = m->pindex;
296 
297 	object = m->object;
298 
299 	/*
300 	 * Don't mess with the page if it's held or special.  Theoretically
301 	 * we can pageout held pages but there is no real need to press our
302 	 * luck, so don't.
303 	 */
304 	if (m->hold_count != 0 || (m->flags & PG_UNQUEUED)) {
305 		vm_page_wakeup(m);
306 		return 0;
307 	}
308 
309 	/*
310 	 * Place page in cluster.  Align cluster for optimal swap space
311 	 * allocation (whether it is swap or not).  This is typically ~16-32
312 	 * pages, which also tends to align the cluster to multiples of the
313 	 * filesystem block size if backed by a filesystem.
314 	 */
315 	page_base = pindex % BLIST_MAX_ALLOC;
316 	mc[page_base] = m;
317 	ib = page_base - 1;
318 	is = page_base + 1;
319 
320 	/*
321 	 * Scan object for clusterable pages.
322 	 *
323 	 * We can cluster ONLY if: ->> the page is NOT
324 	 * clean, wired, busy, held, or mapped into a
325 	 * buffer, and one of the following:
326 	 * 1) The page is inactive, or a seldom used
327 	 *    active page.
328 	 * -or-
329 	 * 2) we force the issue.
330 	 *
331 	 * During heavy mmap/modification loads the pageout
332 	 * daemon can really fragment the underlying file
333 	 * due to flushing pages out of order and not trying
334 	 * align the clusters (which leave sporatic out-of-order
335 	 * holes).  To solve this problem we do the reverse scan
336 	 * first and attempt to align our cluster, then do a
337 	 * forward scan if room remains.
338 	 */
339 	vm_object_hold(object);
340 
341 	while (ib >= 0) {
342 		vm_page_t p;
343 
344 		p = vm_page_lookup_busy_try(object, pindex - page_base + ib,
345 					    TRUE, &error);
346 		if (error || p == NULL)
347 			break;
348 		if ((p->queue - p->pc) == PQ_CACHE ||
349 		    (p->flags & PG_UNQUEUED)) {
350 			vm_page_wakeup(p);
351 			break;
352 		}
353 		vm_page_test_dirty(p);
354 		if (((p->dirty & p->valid) == 0 &&
355 		     (p->flags & PG_NEED_COMMIT) == 0) ||
356 		    p->wire_count != 0 ||	/* may be held by buf cache */
357 		    p->hold_count != 0) {	/* may be undergoing I/O */
358 			vm_page_wakeup(p);
359 			break;
360 		}
361 		if (p->queue - p->pc != PQ_INACTIVE) {
362 			if (p->queue - p->pc != PQ_ACTIVE ||
363 			    (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
364 				vm_page_wakeup(p);
365 				break;
366 			}
367 		}
368 
369 		/*
370 		 * Try to maintain page groupings in the cluster.
371 		 */
372 		if (m->flags & PG_WINATCFLS)
373 			vm_page_flag_set(p, PG_WINATCFLS);
374 		else
375 			vm_page_flag_clear(p, PG_WINATCFLS);
376 		p->act_count = m->act_count;
377 
378 		mc[ib] = p;
379 		--ib;
380 	}
381 	++ib;	/* fixup */
382 
383 	while (is < BLIST_MAX_ALLOC &&
384 	       pindex - page_base + is < object->size) {
385 		vm_page_t p;
386 
387 		p = vm_page_lookup_busy_try(object, pindex - page_base + is,
388 					    TRUE, &error);
389 		if (error || p == NULL)
390 			break;
391 		if (((p->queue - p->pc) == PQ_CACHE) ||
392 		    (p->flags & PG_UNQUEUED)) {
393 			vm_page_wakeup(p);
394 			break;
395 		}
396 		vm_page_test_dirty(p);
397 		if (((p->dirty & p->valid) == 0 &&
398 		     (p->flags & PG_NEED_COMMIT) == 0) ||
399 		    p->wire_count != 0 ||	/* may be held by buf cache */
400 		    p->hold_count != 0) {	/* may be undergoing I/O */
401 			vm_page_wakeup(p);
402 			break;
403 		}
404 		if (p->queue - p->pc != PQ_INACTIVE) {
405 			if (p->queue - p->pc != PQ_ACTIVE ||
406 			    (vmflush_flags & VM_PAGER_ALLOW_ACTIVE) == 0) {
407 				vm_page_wakeup(p);
408 				break;
409 			}
410 		}
411 
412 		/*
413 		 * Try to maintain page groupings in the cluster.
414 		 */
415 		if (m->flags & PG_WINATCFLS)
416 			vm_page_flag_set(p, PG_WINATCFLS);
417 		else
418 			vm_page_flag_clear(p, PG_WINATCFLS);
419 		p->act_count = m->act_count;
420 
421 		mc[is] = p;
422 		++is;
423 	}
424 
425 	vm_object_drop(object);
426 
427 	/*
428 	 * we allow reads during pageouts...
429 	 */
430 	return vm_pageout_flush(&mc[ib], is - ib, vmflush_flags);
431 }
432 
433 /*
434  * vm_pageout_flush() - launder the given pages
435  *
436  *	The given pages are laundered.  Note that we setup for the start of
437  *	I/O ( i.e. busy the page ), mark it read-only, and bump the object
438  *	reference count all in here rather then in the parent.  If we want
439  *	the parent to do more sophisticated things we may have to change
440  *	the ordering.
441  *
442  *	The pages in the array must be busied by the caller and will be
443  *	unbusied by this function.
444  */
445 int
446 vm_pageout_flush(vm_page_t *mc, int count, int vmflush_flags)
447 {
448 	vm_object_t object;
449 	int pageout_status[count];
450 	int numpagedout = 0;
451 	int i;
452 	int dodebug;
453 
454 	if (vm_pageout_debug > 0) {
455 		--vm_pageout_debug;
456 		dodebug = 1;
457 	} else {
458 		dodebug = 0;
459 	}
460 
461 	/*
462 	 * Initiate I/O.  Bump the vm_page_t->busy counter.
463 	 */
464 	for (i = 0; i < count; i++) {
465 		KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL,
466 			("vm_pageout_flush page %p index %d/%d: partially "
467 			 "invalid page", mc[i], i, count));
468 		vm_page_io_start(mc[i]);
469 	}
470 
471 	/*
472 	 * We must make the pages read-only.  This will also force the
473 	 * modified bit in the related pmaps to be cleared.  The pager
474 	 * cannot clear the bit for us since the I/O completion code
475 	 * typically runs from an interrupt.  The act of making the page
476 	 * read-only handles the case for us.
477 	 *
478 	 * Then we can unbusy the pages, we still hold a reference by virtue
479 	 * of our soft-busy.
480 	 */
481 	if (dodebug)
482 		kprintf("pageout(%d): ", count);
483 	for (i = 0; i < count; i++) {
484 		if (vmflush_flags & VM_PAGER_TRY_TO_CACHE)
485 			vm_page_protect(mc[i], VM_PROT_NONE);
486 		else
487 			vm_page_protect(mc[i], VM_PROT_READ);
488 		vm_page_wakeup(mc[i]);
489 		if (dodebug)
490 			kprintf(" %p", mc[i]);
491 	}
492 	if (dodebug)
493 		kprintf("\n");
494 
495 	object = mc[0]->object;
496 	vm_object_pip_add(object, count);
497 
498 	vm_pager_put_pages(object, mc, count,
499 			   (vmflush_flags |
500 			    ((object == &kernel_object) ?
501 				VM_PAGER_PUT_SYNC : 0)),
502 			   pageout_status);
503 
504 	if (dodebug)
505 		kprintf("result: ");
506 	for (i = 0; i < count; i++) {
507 		vm_page_t mt = mc[i];
508 
509 		if (dodebug)
510 			kprintf("  S%d", pageout_status[i]);
511 
512 		switch (pageout_status[i]) {
513 		case VM_PAGER_OK:
514 			numpagedout++;
515 			break;
516 		case VM_PAGER_PEND:
517 			numpagedout++;
518 			break;
519 		case VM_PAGER_BAD:
520 			/*
521 			 * Page outside of range of object. Right now we
522 			 * essentially lose the changes by pretending it
523 			 * worked.
524 			 */
525 			vm_page_busy_wait(mt, FALSE, "pgbad");
526 			pmap_clear_modify(mt);
527 			vm_page_undirty(mt);
528 			vm_page_wakeup(mt);
529 			break;
530 		case VM_PAGER_ERROR:
531 		case VM_PAGER_FAIL:
532 			/*
533 			 * A page typically cannot be paged out when we
534 			 * have run out of swap.  We leave the page
535 			 * marked inactive and will try to page it out
536 			 * again later.
537 			 *
538 			 * Starvation of the active page list is used to
539 			 * determine when the system is massively memory
540 			 * starved.
541 			 */
542 			break;
543 		case VM_PAGER_AGAIN:
544 			break;
545 		}
546 
547 		/*
548 		 * If not PENDing this was a synchronous operation and we
549 		 * clean up after the I/O.  If it is PENDing the mess is
550 		 * cleaned up asynchronously.
551 		 *
552 		 * Also nominally act on the caller's wishes if the caller
553 		 * wants to try to really clean (cache or free) the page.
554 		 *
555 		 * Also nominally deactivate the page if the system is
556 		 * memory-stressed.
557 		 */
558 		if (pageout_status[i] != VM_PAGER_PEND) {
559 			vm_page_busy_wait(mt, FALSE, "pgouw");
560 			vm_page_io_finish(mt);
561 			if (vmflush_flags & VM_PAGER_TRY_TO_CACHE) {
562 				vm_page_try_to_cache(mt);
563 				if (dodebug)
564 				kprintf("A[pq_cache=%d]",
565 					 ((mt->queue - mt->pc) == PQ_CACHE));
566 			} else if (vm_page_count_severe()) {
567 				vm_page_deactivate(mt);
568 				vm_page_wakeup(mt);
569 				if (dodebug)
570 				kprintf("B");
571 			} else {
572 				vm_page_wakeup(mt);
573 				if (dodebug)
574 				kprintf("C");
575 			}
576 			vm_object_pip_wakeup(object);
577 		}
578 	}
579 	if (dodebug)
580 		kprintf("(%d paged out)\n", numpagedout);
581 	return numpagedout;
582 }
583 
584 #if !defined(NO_SWAPPING)
585 
586 /*
587  * Callback function, page busied for us.  We must dispose of the busy
588  * condition.  Any related pmap pages may be held but will not be locked.
589  */
590 static
591 int
592 vm_pageout_mdp_callback(struct pmap_pgscan_info *info, vm_offset_t va,
593 			vm_page_t p)
594 {
595 	int actcount;
596 	int cleanit = 0;
597 
598 	/*
599 	 * Basic tests - There should never be a marker, and we can stop
600 	 *		 once the RSS is below the required level.
601 	 */
602 	KKASSERT((p->flags & PG_MARKER) == 0);
603 	if (pmap_resident_tlnw_count(info->pmap) <= info->limit) {
604 		vm_page_wakeup(p);
605 		return(-1);
606 	}
607 
608 	mycpu->gd_cnt.v_pdpages++;
609 
610 	if (p->wire_count || p->hold_count || (p->flags & PG_UNQUEUED)) {
611 		vm_page_wakeup(p);
612 		goto done;
613 	}
614 
615 	++info->actioncount;
616 
617 	/*
618 	 * Check if the page has been referened recently.  If it has,
619 	 * activate it and skip.
620 	 */
621 	actcount = pmap_ts_referenced(p);
622 	if (actcount) {
623 		vm_page_flag_set(p, PG_REFERENCED);
624 	} else if (p->flags & PG_REFERENCED) {
625 		actcount = 1;
626 	}
627 
628 	if (actcount) {
629 		if (p->queue - p->pc != PQ_ACTIVE) {
630 			vm_page_and_queue_spin_lock(p);
631 			if (p->queue - p->pc != PQ_ACTIVE) {
632 				vm_page_and_queue_spin_unlock(p);
633 				vm_page_activate(p);
634 			} else {
635 				vm_page_and_queue_spin_unlock(p);
636 			}
637 		} else {
638 			p->act_count += actcount;
639 			if (p->act_count > ACT_MAX)
640 				p->act_count = ACT_MAX;
641 		}
642 		vm_page_flag_clear(p, PG_REFERENCED);
643 		vm_page_wakeup(p);
644 		goto done;
645 	}
646 
647 	/*
648 	 * Remove the page from this particular pmap.  Once we do this, our
649 	 * pmap scans will not see it again (unless it gets faulted in), so
650 	 * we must actively dispose of or deal with the page.
651 	 */
652 	pmap_remove_specific(info->pmap, p);
653 
654 	/*
655 	 * If the page is not mapped to another process (i.e. as would be
656 	 * typical if this were a shared page from a library) then deactivate
657 	 * the page and clean it in two passes only.
658 	 *
659 	 * If the page hasn't been referenced since the last check, remove it
660 	 * from the pmap.  If it is no longer mapped, deactivate it
661 	 * immediately, accelerating the normal decline.
662 	 *
663 	 * Once the page has been removed from the pmap the RSS code no
664 	 * longer tracks it so we have to make sure that it is staged for
665 	 * potential flush action.
666 	 */
667 	if ((p->flags & PG_MAPPED) == 0 ||
668 	    (pmap_mapped_sync(p) & PG_MAPPED) == 0) {
669 		if (p->queue - p->pc == PQ_ACTIVE) {
670 			vm_page_deactivate(p);
671 		}
672 		if (p->queue - p->pc == PQ_INACTIVE) {
673 			cleanit = 1;
674 		}
675 	}
676 
677 	/*
678 	 * Ok, try to fully clean the page and any nearby pages such that at
679 	 * least the requested page is freed or moved to the cache queue.
680 	 *
681 	 * We usually do this synchronously to allow us to get the page into
682 	 * the CACHE queue quickly, which will prevent memory exhaustion if
683 	 * a process with a memoryuse limit is running away.  However, the
684 	 * sysadmin may desire to set vm.swap_user_async which relaxes this
685 	 * and improves write performance.
686 	 */
687 	if (cleanit) {
688 		long max_launder = 0x7FFF;
689 		long vnodes_skipped = 0;
690 		int vmflush_flags;
691 		struct vnode *vpfailed = NULL;
692 
693 		info->offset = va;
694 
695 		if (vm_pageout_memuse_mode >= 2) {
696 			vmflush_flags = VM_PAGER_TRY_TO_CACHE |
697 					VM_PAGER_ALLOW_ACTIVE;
698 			if (swap_user_async == 0)
699 				vmflush_flags |= VM_PAGER_PUT_SYNC;
700 			vm_page_flag_set(p, PG_WINATCFLS);
701 			info->cleancount +=
702 				vm_pageout_page(p, &max_launder,
703 						&vnodes_skipped,
704 						&vpfailed, 1, vmflush_flags);
705 		} else {
706 			vm_page_wakeup(p);
707 			++info->cleancount;
708 		}
709 	} else {
710 		vm_page_wakeup(p);
711 	}
712 
713 	/*
714 	 * Must be at end to avoid SMP races.
715 	 */
716 done:
717 	lwkt_user_yield();
718 	return 0;
719 }
720 
721 /*
722  * Deactivate some number of pages in a map due to set RLIMIT_RSS limits.
723  * that is relatively difficult to do.  We try to keep track of where we
724  * left off last time to reduce scan overhead.
725  *
726  * Called when vm_pageout_memuse_mode is >= 1.
727  */
728 void
729 vm_pageout_map_deactivate_pages(vm_map_t map, vm_pindex_t limit)
730 {
731 	vm_offset_t pgout_offset;
732 	struct pmap_pgscan_info info;
733 	int retries = 3;
734 
735 	pgout_offset = map->pgout_offset;
736 again:
737 #if 0
738 	kprintf("%016jx ", pgout_offset);
739 #endif
740 	if (pgout_offset < VM_MIN_USER_ADDRESS)
741 		pgout_offset = VM_MIN_USER_ADDRESS;
742 	if (pgout_offset >= VM_MAX_USER_ADDRESS)
743 		pgout_offset = 0;
744 	info.pmap = vm_map_pmap(map);
745 	info.limit = limit;
746 	info.beg_addr = pgout_offset;
747 	info.end_addr = VM_MAX_USER_ADDRESS;
748 	info.callback = vm_pageout_mdp_callback;
749 	info.cleancount = 0;
750 	info.actioncount = 0;
751 	info.busycount = 0;
752 
753 	pmap_pgscan(&info);
754 	pgout_offset = info.offset;
755 #if 0
756 	kprintf("%016jx %08lx %08lx\n", pgout_offset,
757 		info.cleancount, info.actioncount);
758 #endif
759 
760 	if (pgout_offset != VM_MAX_USER_ADDRESS &&
761 	    pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
762 		goto again;
763 	} else if (retries &&
764 		   pmap_resident_tlnw_count(vm_map_pmap(map)) > limit) {
765 		--retries;
766 		goto again;
767 	}
768 	map->pgout_offset = pgout_offset;
769 }
770 #endif
771 
772 /*
773  * Called when the pageout scan wants to free a page.  We no longer
774  * try to cycle the vm_object here with a reference & dealloc, which can
775  * cause a non-trivial object collapse in a critical path.
776  *
777  * It is unclear why we cycled the ref_count in the past, perhaps to try
778  * to optimize shadow chain collapses but I don't quite see why it would
779  * be necessary.  An OBJ_DEAD object should terminate any and all vm_pages
780  * synchronously and not have to be kicked-start.
781  */
782 static void
783 vm_pageout_page_free(vm_page_t m)
784 {
785 	vm_page_protect(m, VM_PROT_NONE);
786 	vm_page_free(m);
787 }
788 
789 /*
790  * vm_pageout_scan does the dirty work for the pageout daemon.
791  */
792 struct vm_pageout_scan_info {
793 	struct proc *bigproc;
794 	vm_offset_t bigsize;
795 };
796 
797 static int vm_pageout_scan_callback(struct proc *p, void *data);
798 
799 /*
800  * Scan inactive queue
801  *
802  * WARNING! Can be called from two pagedaemon threads simultaneously.
803  */
804 static int
805 vm_pageout_scan_inactive(int pass, int q, long avail_shortage,
806 			 long *vnodes_skipped)
807 {
808 	vm_page_t m;
809 	struct vm_page marker;
810 	struct vnode *vpfailed;		/* warning, allowed to be stale */
811 	long maxscan;
812 	long delta = 0;
813 	long max_launder;
814 	int isep;
815 	int vmflush_flags;
816 
817 	isep = (curthread == emergpager);
818 
819 	/*
820 	 * Start scanning the inactive queue for pages we can move to the
821 	 * cache or free.  The scan will stop when the target is reached or
822 	 * we have scanned the entire inactive queue.  Note that m->act_count
823 	 * is not used to form decisions for the inactive queue, only for the
824 	 * active queue.
825 	 *
826 	 * max_launder limits the number of dirty pages we flush per scan.
827 	 * For most systems a smaller value (16 or 32) is more robust under
828 	 * extreme memory and disk pressure because any unnecessary writes
829 	 * to disk can result in extreme performance degredation.  However,
830 	 * systems with excessive dirty pages (especially when MAP_NOSYNC is
831 	 * used) will die horribly with limited laundering.  If the pageout
832 	 * daemon cannot clean enough pages in the first pass, we let it go
833 	 * all out in succeeding passes.
834 	 *
835 	 * NOTE!  THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
836 	 *	  PAGES.
837 	 */
838 	if ((max_launder = vm_max_launder) <= 1)
839 		max_launder = 1;
840 	if (pass)
841 		max_launder = 10000;
842 
843 	/*
844 	 * Initialize our marker
845 	 */
846 	bzero(&marker, sizeof(marker));
847 	marker.flags = PG_FICTITIOUS | PG_MARKER;
848 	marker.busy_count = PBUSY_LOCKED;
849 	marker.queue = PQ_INACTIVE + q;
850 	marker.pc = q;
851 	marker.wire_count = 1;
852 
853 	/*
854 	 * Inactive queue scan.
855 	 *
856 	 * We pick off approximately 1/10 of each queue.  Each queue is
857 	 * effectively organized LRU so scanning the entire queue would
858 	 * improperly pick up pages that might still be in regular use.
859 	 *
860 	 * NOTE: The vm_page must be spinlocked before the queue to avoid
861 	 *	 deadlocks, so it is easiest to simply iterate the loop
862 	 *	 with the queue unlocked at the top.
863 	 */
864 	vpfailed = NULL;
865 
866 	vm_page_queues_spin_lock(PQ_INACTIVE + q);
867 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
868 	maxscan = vm_page_queues[PQ_INACTIVE + q].lcnt / 10 + 1;
869 
870 	/*
871 	 * Queue locked at top of loop to avoid stack marker issues.
872 	 */
873 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
874 	       maxscan-- > 0 && avail_shortage - delta > 0)
875 	{
876 		int count;
877 
878 		KKASSERT(m->queue == PQ_INACTIVE + q);
879 		TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl,
880 			     &marker, pageq);
881 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE + q].pl, m,
882 				   &marker, pageq);
883 		mycpu->gd_cnt.v_pdpages++;
884 
885 		/*
886 		 * Skip marker pages (atomic against other markers to avoid
887 		 * infinite hop-over scans).
888 		 */
889 		if (m->flags & PG_MARKER)
890 			continue;
891 
892 		/*
893 		 * Try to busy the page.  Don't mess with pages which are
894 		 * already busy or reorder them in the queue.
895 		 */
896 		if (vm_page_busy_try(m, TRUE))
897 			continue;
898 
899 		/*
900 		 * Remaining operations run with the page busy and neither
901 		 * the page or the queue will be spin-locked.
902 		 */
903 		KKASSERT(m->queue == PQ_INACTIVE + q);
904 		vm_page_queues_spin_unlock(PQ_INACTIVE + q);
905 
906 		/*
907 		 * The emergency pager runs when the primary pager gets
908 		 * stuck, which typically means the primary pager deadlocked
909 		 * on a vnode-backed page.  Therefore, the emergency pager
910 		 * must skip any complex objects.
911 		 *
912 		 * We disallow VNODEs unless they are VCHR whos device ops
913 		 * does not flag D_NOEMERGPGR.
914 		 */
915 		if (isep && m->object) {
916 			struct vnode *vp;
917 
918 			switch(m->object->type) {
919 			case OBJT_DEFAULT:
920 			case OBJT_SWAP:
921 				/*
922 				 * Allow anonymous memory and assume that
923 				 * swap devices are not complex, since its
924 				 * kinda worthless if we can't swap out dirty
925 				 * anonymous pages.
926 				 */
927 				break;
928 			case OBJT_VNODE:
929 				/*
930 				 * Allow VCHR device if the D_NOEMERGPGR
931 				 * flag is not set, deny other vnode types
932 				 * as being too complex.
933 				 */
934 				vp = m->object->handle;
935 				if (vp && vp->v_type == VCHR &&
936 				    vp->v_rdev && vp->v_rdev->si_ops &&
937 				    (vp->v_rdev->si_ops->head.flags &
938 				     D_NOEMERGPGR) == 0) {
939 					break;
940 				}
941 				/* Deny - fall through */
942 			default:
943 				/*
944 				 * Deny
945 				 */
946 				vm_page_wakeup(m);
947 				vm_page_queues_spin_lock(PQ_INACTIVE + q);
948 				lwkt_yield();
949 				continue;
950 			}
951 		}
952 
953 		/*
954 		 * Try to pageout the page and perhaps other nearby pages.
955 		 * We want to get the pages into the cache eventually (
956 		 * first or second pass).  Otherwise the pages can wind up
957 		 * just cycling in the inactive queue, getting flushed over
958 		 * and over again.
959 		 */
960 		if (vm_pageout_memuse_mode >= 2)
961 			vm_page_flag_set(m, PG_WINATCFLS);
962 
963 		vmflush_flags = 0;
964 		if (vm_pageout_allow_active)
965 			vmflush_flags |= VM_PAGER_ALLOW_ACTIVE;
966 		if (m->flags & PG_WINATCFLS)
967 			vmflush_flags |= VM_PAGER_TRY_TO_CACHE;
968 		count = vm_pageout_page(m, &max_launder, vnodes_skipped,
969 					&vpfailed, pass, vmflush_flags);
970 		delta += count;
971 
972 		/*
973 		 * Systems with a ton of memory can wind up with huge
974 		 * deactivation counts.  Because the inactive scan is
975 		 * doing a lot of flushing, the combination can result
976 		 * in excessive paging even in situations where other
977 		 * unrelated threads free up sufficient VM.
978 		 *
979 		 * To deal with this we abort the nominal active->inactive
980 		 * scan before we hit the inactive target when free+cache
981 		 * levels have reached a reasonable target.
982 		 *
983 		 * When deciding to stop early we need to add some slop to
984 		 * the test and we need to return full completion to the caller
985 		 * to prevent the caller from thinking there is something
986 		 * wrong and issuing a low-memory+swap warning or pkill.
987 		 *
988 		 * A deficit forces paging regardless of the state of the
989 		 * VM page queues (used for RSS enforcement).
990 		 */
991 		lwkt_yield();
992 		vm_page_queues_spin_lock(PQ_INACTIVE + q);
993 		if (vm_paging_target() < -vm_max_launder) {
994 			/*
995 			 * Stopping early, return full completion to caller.
996 			 */
997 			if (delta < avail_shortage)
998 				delta = avail_shortage;
999 			break;
1000 		}
1001 	}
1002 
1003 	/* page queue still spin-locked */
1004 	TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE + q].pl, &marker, pageq);
1005 	vm_page_queues_spin_unlock(PQ_INACTIVE + q);
1006 
1007 	return (delta);
1008 }
1009 
1010 /*
1011  * Pageout the specified page, return the total number of pages paged out
1012  * (this routine may cluster).
1013  *
1014  * The page must be busied and soft-busied by the caller and will be disposed
1015  * of by this function.
1016  */
1017 static int
1018 vm_pageout_page(vm_page_t m, long *max_launderp, long *vnodes_skippedp,
1019 		struct vnode **vpfailedp, int pass, int vmflush_flags)
1020 {
1021 	vm_object_t object;
1022 	int actcount;
1023 	int count = 0;
1024 
1025 	/*
1026 	 * Wiring no longer removes a page from its queue.  The last unwiring
1027 	 * will requeue the page.  Obviously wired pages cannot be paged out
1028 	 * so unqueue it and return.
1029 	 */
1030 	if (m->wire_count) {
1031 		vm_page_unqueue_nowakeup(m);
1032 		vm_page_wakeup(m);
1033 		return 0;
1034 	}
1035 
1036 	/*
1037 	 * A held page may be undergoing I/O, so skip it.
1038 	 */
1039 	if (m->hold_count) {
1040 		vm_page_and_queue_spin_lock(m);
1041 		if (m->queue - m->pc == PQ_INACTIVE) {
1042 			TAILQ_REMOVE(
1043 				&vm_page_queues[m->queue].pl, m, pageq);
1044 			TAILQ_INSERT_TAIL(
1045 				&vm_page_queues[m->queue].pl, m, pageq);
1046 		}
1047 		vm_page_and_queue_spin_unlock(m);
1048 		vm_page_wakeup(m);
1049 		return 0;
1050 	}
1051 
1052 	if (m->object == NULL || m->object->ref_count == 0) {
1053 		/*
1054 		 * If the object is not being used, we ignore previous
1055 		 * references.
1056 		 */
1057 		vm_page_flag_clear(m, PG_REFERENCED);
1058 		pmap_clear_reference(m);
1059 		/* fall through to end */
1060 	} else if (((m->flags & PG_REFERENCED) == 0) &&
1061 		    (actcount = pmap_ts_referenced(m))) {
1062 		/*
1063 		 * Otherwise, if the page has been referenced while
1064 		 * in the inactive queue, we bump the "activation
1065 		 * count" upwards, making it less likely that the
1066 		 * page will be added back to the inactive queue
1067 		 * prematurely again.  Here we check the page tables
1068 		 * (or emulated bits, if any), given the upper level
1069 		 * VM system not knowing anything about existing
1070 		 * references.
1071 		 */
1072 		vm_page_activate(m);
1073 		m->act_count += (actcount + ACT_ADVANCE);
1074 		vm_page_wakeup(m);
1075 		return 0;
1076 	}
1077 
1078 	/*
1079 	 * (m) is still busied.
1080 	 *
1081 	 * If the upper level VM system knows about any page
1082 	 * references, we activate the page.  We also set the
1083 	 * "activation count" higher than normal so that we will less
1084 	 * likely place pages back onto the inactive queue again.
1085 	 */
1086 	if ((m->flags & PG_REFERENCED) != 0) {
1087 		vm_page_flag_clear(m, PG_REFERENCED);
1088 		actcount = pmap_ts_referenced(m);
1089 		vm_page_activate(m);
1090 		m->act_count += (actcount + ACT_ADVANCE + 1);
1091 		vm_page_wakeup(m);
1092 		return 0;
1093 	}
1094 
1095 	/*
1096 	 * If the upper level VM system doesn't know anything about
1097 	 * the page being dirty, we have to check for it again.  As
1098 	 * far as the VM code knows, any partially dirty pages are
1099 	 * fully dirty.
1100 	 *
1101 	 * Pages marked PG_WRITEABLE may be mapped into the user
1102 	 * address space of a process running on another cpu.  A
1103 	 * user process (without holding the MP lock) running on
1104 	 * another cpu may be able to touch the page while we are
1105 	 * trying to remove it.  vm_page_cache() will handle this
1106 	 * case for us.
1107 	 */
1108 	if (m->dirty == 0) {
1109 		vm_page_test_dirty(m);
1110 	} else {
1111 		vm_page_dirty(m);
1112 	}
1113 
1114 	if (m->valid == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
1115 		/*
1116 		 * Invalid pages can be easily freed
1117 		 */
1118 		vm_pageout_page_free(m);
1119 		mycpu->gd_cnt.v_dfree++;
1120 		++count;
1121 	} else if (m->dirty == 0 && (m->flags & PG_NEED_COMMIT) == 0) {
1122 		/*
1123 		 * Clean pages can be placed onto the cache queue.
1124 		 * This effectively frees them.
1125 		 */
1126 		vm_page_cache(m);
1127 		++count;
1128 	} else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) {
1129 		/*
1130 		 * Dirty pages need to be paged out, but flushing
1131 		 * a page is extremely expensive verses freeing
1132 		 * a clean page.  Rather then artificially limiting
1133 		 * the number of pages we can flush, we instead give
1134 		 * dirty pages extra priority on the inactive queue
1135 		 * by forcing them to be cycled through the queue
1136 		 * twice before being flushed, after which the
1137 		 * (now clean) page will cycle through once more
1138 		 * before being freed.  This significantly extends
1139 		 * the thrash point for a heavily loaded machine.
1140 		 */
1141 		vm_page_flag_set(m, PG_WINATCFLS);
1142 		vm_page_and_queue_spin_lock(m);
1143 		if (m->queue - m->pc == PQ_INACTIVE) {
1144 			TAILQ_REMOVE(
1145 				&vm_page_queues[m->queue].pl, m, pageq);
1146 			TAILQ_INSERT_TAIL(
1147 				&vm_page_queues[m->queue].pl, m, pageq);
1148 		}
1149 		vm_page_and_queue_spin_unlock(m);
1150 		vm_page_wakeup(m);
1151 	} else if (*max_launderp > 0) {
1152 		/*
1153 		 * We always want to try to flush some dirty pages if
1154 		 * we encounter them, to keep the system stable.
1155 		 * Normally this number is small, but under extreme
1156 		 * pressure where there are insufficient clean pages
1157 		 * on the inactive queue, we may have to go all out.
1158 		 */
1159 		int swap_pageouts_ok;
1160 		struct vnode *vp = NULL;
1161 
1162 		swap_pageouts_ok = 0;
1163 		object = m->object;
1164 		if (object &&
1165 		    (object->type != OBJT_SWAP) &&
1166 		    (object->type != OBJT_DEFAULT)) {
1167 			swap_pageouts_ok = 1;
1168 		} else {
1169 			swap_pageouts_ok = !(defer_swap_pageouts ||
1170 					     disable_swap_pageouts);
1171 			swap_pageouts_ok |= (!disable_swap_pageouts &&
1172 					     defer_swap_pageouts &&
1173 					     vm_page_count_min(0));
1174 		}
1175 
1176 		/*
1177 		 * We don't bother paging objects that are "dead".
1178 		 * Those objects are in a "rundown" state.
1179 		 */
1180 		if (!swap_pageouts_ok ||
1181 		    (object == NULL) ||
1182 		    (object->flags & OBJ_DEAD)) {
1183 			vm_page_and_queue_spin_lock(m);
1184 			if (m->queue - m->pc == PQ_INACTIVE) {
1185 				TAILQ_REMOVE(
1186 				    &vm_page_queues[m->queue].pl,
1187 				    m, pageq);
1188 				TAILQ_INSERT_TAIL(
1189 				    &vm_page_queues[m->queue].pl,
1190 				    m, pageq);
1191 			}
1192 			vm_page_and_queue_spin_unlock(m);
1193 			vm_page_wakeup(m);
1194 			return 0;
1195 		}
1196 
1197 		/*
1198 		 * (m) is still busied.
1199 		 *
1200 		 * The object is already known NOT to be dead.   It
1201 		 * is possible for the vget() to block the whole
1202 		 * pageout daemon, but the new low-memory handling
1203 		 * code should prevent it.
1204 		 *
1205 		 * The previous code skipped locked vnodes and, worse,
1206 		 * reordered pages in the queue.  This results in
1207 		 * completely non-deterministic operation because,
1208 		 * quite often, a vm_fault has initiated an I/O and
1209 		 * is holding a locked vnode at just the point where
1210 		 * the pageout daemon is woken up.
1211 		 *
1212 		 * We can't wait forever for the vnode lock, we might
1213 		 * deadlock due to a vn_read() getting stuck in
1214 		 * vm_wait while holding this vnode.  We skip the
1215 		 * vnode if we can't get it in a reasonable amount
1216 		 * of time.
1217 		 *
1218 		 * vpfailed is used to (try to) avoid the case where
1219 		 * a large number of pages are associated with a
1220 		 * locked vnode, which could cause the pageout daemon
1221 		 * to stall for an excessive amount of time.
1222 		 */
1223 		if (object->type == OBJT_VNODE) {
1224 			int flags;
1225 
1226 			vp = object->handle;
1227 			flags = LK_EXCLUSIVE;
1228 			if (vp == *vpfailedp)
1229 				flags |= LK_NOWAIT;
1230 			else
1231 				flags |= LK_TIMELOCK;
1232 			vm_page_hold(m);
1233 			vm_page_wakeup(m);
1234 
1235 			/*
1236 			 * We have unbusied (m) temporarily so we can
1237 			 * acquire the vp lock without deadlocking.
1238 			 * (m) is held to prevent destruction.
1239 			 */
1240 			if (vget(vp, flags) != 0) {
1241 				*vpfailedp = vp;
1242 				++pageout_lock_miss;
1243 				if (object->flags & OBJ_MIGHTBEDIRTY)
1244 					    ++*vnodes_skippedp;
1245 				vm_page_unhold(m);
1246 				return 0;
1247 			}
1248 
1249 			/*
1250 			 * The page might have been moved to another
1251 			 * queue during potential blocking in vget()
1252 			 * above.  The page might have been freed and
1253 			 * reused for another vnode.  The object might
1254 			 * have been reused for another vnode.
1255 			 */
1256 			if (m->queue - m->pc != PQ_INACTIVE ||
1257 			    m->object != object ||
1258 			    object->handle != vp) {
1259 				if (object->flags & OBJ_MIGHTBEDIRTY)
1260 					++*vnodes_skippedp;
1261 				vput(vp);
1262 				vm_page_unhold(m);
1263 				return 0;
1264 			}
1265 
1266 			/*
1267 			 * The page may have been busied during the
1268 			 * blocking in vput();  We don't move the
1269 			 * page back onto the end of the queue so that
1270 			 * statistics are more correct if we don't.
1271 			 */
1272 			if (vm_page_busy_try(m, TRUE)) {
1273 				vput(vp);
1274 				vm_page_unhold(m);
1275 				return 0;
1276 			}
1277 			vm_page_unhold(m);
1278 
1279 			/*
1280 			 * If it was wired while we didn't own it.
1281 			 */
1282 			if (m->wire_count) {
1283 				vm_page_unqueue_nowakeup(m);
1284 				vput(vp);
1285 				vm_page_wakeup(m);
1286 				return 0;
1287 			}
1288 
1289 			/*
1290 			 * (m) is busied again
1291 			 *
1292 			 * We own the busy bit and remove our hold
1293 			 * bit.  If the page is still held it
1294 			 * might be undergoing I/O, so skip it.
1295 			 */
1296 			if (m->hold_count) {
1297 				vm_page_and_queue_spin_lock(m);
1298 				if (m->queue - m->pc == PQ_INACTIVE) {
1299 					TAILQ_REMOVE(&vm_page_queues[m->queue].pl, m, pageq);
1300 					TAILQ_INSERT_TAIL(&vm_page_queues[m->queue].pl, m, pageq);
1301 				}
1302 				vm_page_and_queue_spin_unlock(m);
1303 				if (object->flags & OBJ_MIGHTBEDIRTY)
1304 					++*vnodes_skippedp;
1305 				vm_page_wakeup(m);
1306 				vput(vp);
1307 				return 0;
1308 			}
1309 			/* (m) is left busied as we fall through */
1310 		}
1311 
1312 		/*
1313 		 * page is busy and not held here.
1314 		 *
1315 		 * If a page is dirty, then it is either being washed
1316 		 * (but not yet cleaned) or it is still in the
1317 		 * laundry.  If it is still in the laundry, then we
1318 		 * start the cleaning operation.
1319 		 *
1320 		 * decrement inactive_shortage on success to account
1321 		 * for the (future) cleaned page.  Otherwise we
1322 		 * could wind up laundering or cleaning too many
1323 		 * pages.
1324 		 *
1325 		 * NOTE: Cleaning the page here does not cause
1326 		 *	 force_deficit to be adjusted, because the
1327 		 *	 page is not being freed or moved to the
1328 		 *	 cache.
1329 		 */
1330 		count = vm_pageout_clean_helper(m, vmflush_flags);
1331 		*max_launderp -= count;
1332 
1333 		/*
1334 		 * Clean ate busy, page no longer accessible
1335 		 */
1336 		if (vp != NULL)
1337 			vput(vp);
1338 	} else {
1339 		vm_page_wakeup(m);
1340 	}
1341 	return count;
1342 }
1343 
1344 /*
1345  * Scan active queue
1346  *
1347  * WARNING! Can be called from two pagedaemon threads simultaneously.
1348  */
1349 static int
1350 vm_pageout_scan_active(int pass, int q,
1351 		       long avail_shortage, long inactive_shortage,
1352 		       long *recycle_countp)
1353 {
1354 	struct vm_page marker;
1355 	vm_page_t m;
1356 	int actcount;
1357 	long delta = 0;
1358 	long maxscan;
1359 	int isep;
1360 
1361 	isep = (curthread == emergpager);
1362 
1363 	/*
1364 	 * We want to move pages from the active queue to the inactive
1365 	 * queue to get the inactive queue to the inactive target.  If
1366 	 * we still have a page shortage from above we try to directly free
1367 	 * clean pages instead of moving them.
1368 	 *
1369 	 * If we do still have a shortage we keep track of the number of
1370 	 * pages we free or cache (recycle_count) as a measure of thrashing
1371 	 * between the active and inactive queues.
1372 	 *
1373 	 * If we were able to completely satisfy the free+cache targets
1374 	 * from the inactive pool we limit the number of pages we move
1375 	 * from the active pool to the inactive pool to 2x the pages we
1376 	 * had removed from the inactive pool (with a minimum of 1/5 the
1377 	 * inactive target).  If we were not able to completely satisfy
1378 	 * the free+cache targets we go for the whole target aggressively.
1379 	 *
1380 	 * NOTE: Both variables can end up negative.
1381 	 * NOTE: We are still in a critical section.
1382 	 *
1383 	 * NOTE!  THE EMERGENCY PAGER (isep) DOES NOT LAUNDER VNODE-BACKED
1384 	 *	  PAGES.
1385 	 */
1386 
1387 	bzero(&marker, sizeof(marker));
1388 	marker.flags = PG_FICTITIOUS | PG_MARKER;
1389 	marker.busy_count = PBUSY_LOCKED;
1390 	marker.queue = PQ_ACTIVE + q;
1391 	marker.pc = q;
1392 	marker.wire_count = 1;
1393 
1394 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1395 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1396 	maxscan = vm_page_queues[PQ_ACTIVE + q].lcnt / 10 + 1;
1397 
1398 	/*
1399 	 * Queue locked at top of loop to avoid stack marker issues.
1400 	 */
1401 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1402 	       maxscan-- > 0 && (avail_shortage - delta > 0 ||
1403 				inactive_shortage > 0))
1404 	{
1405 		KKASSERT(m->queue == PQ_ACTIVE + q);
1406 		TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl,
1407 			     &marker, pageq);
1408 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1409 				   &marker, pageq);
1410 
1411 		/*
1412 		 * Skip marker pages (atomic against other markers to avoid
1413 		 * infinite hop-over scans).
1414 		 */
1415 		if (m->flags & PG_MARKER)
1416 			continue;
1417 
1418 		/*
1419 		 * Try to busy the page.  Don't mess with pages which are
1420 		 * already busy or reorder them in the queue.
1421 		 */
1422 		if (vm_page_busy_try(m, TRUE))
1423 			continue;
1424 
1425 		/*
1426 		 * Remaining operations run with the page busy and neither
1427 		 * the page or the queue will be spin-locked.
1428 		 */
1429 		KKASSERT(m->queue == PQ_ACTIVE + q);
1430 		vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1431 
1432 #if 0
1433 		/*
1434 		 * Don't deactivate pages that are held, even if we can
1435 		 * busy them.  (XXX why not?)
1436 		 */
1437 		if (m->hold_count) {
1438 			vm_page_and_queue_spin_lock(m);
1439 			if (m->queue - m->pc == PQ_ACTIVE) {
1440 				TAILQ_REMOVE(
1441 					&vm_page_queues[PQ_ACTIVE + q].pl,
1442 					m, pageq);
1443 				TAILQ_INSERT_TAIL(
1444 					&vm_page_queues[PQ_ACTIVE + q].pl,
1445 					m, pageq);
1446 			}
1447 			vm_page_and_queue_spin_unlock(m);
1448 			vm_page_wakeup(m);
1449 			goto next;
1450 		}
1451 #endif
1452 		/*
1453 		 * We can just remove wired pages from the queue
1454 		 */
1455 		if (m->wire_count) {
1456 			vm_page_unqueue_nowakeup(m);
1457 			vm_page_wakeup(m);
1458 			goto next;
1459 		}
1460 
1461 		/*
1462 		 * The emergency pager ignores vnode-backed pages as these
1463 		 * are the pages that probably bricked the main pager.
1464 		 */
1465 		if (isep && m->object && m->object->type == OBJT_VNODE) {
1466 			vm_page_and_queue_spin_lock(m);
1467 			if (m->queue - m->pc == PQ_ACTIVE) {
1468 				TAILQ_REMOVE(
1469 					&vm_page_queues[PQ_ACTIVE + q].pl,
1470 					m, pageq);
1471 				TAILQ_INSERT_TAIL(
1472 					&vm_page_queues[PQ_ACTIVE + q].pl,
1473 					m, pageq);
1474 			}
1475 			vm_page_and_queue_spin_unlock(m);
1476 			vm_page_wakeup(m);
1477 			goto next;
1478 		}
1479 
1480 		/*
1481 		 * The count for pagedaemon pages is done after checking the
1482 		 * page for eligibility...
1483 		 */
1484 		mycpu->gd_cnt.v_pdpages++;
1485 
1486 		/*
1487 		 * Check to see "how much" the page has been used and clear
1488 		 * the tracking access bits.  If the object has no references
1489 		 * don't bother paying the expense.
1490 		 */
1491 		actcount = 0;
1492 		if (m->object && m->object->ref_count != 0) {
1493 			if (m->flags & PG_REFERENCED)
1494 				++actcount;
1495 			actcount += pmap_ts_referenced(m);
1496 			if (actcount) {
1497 				m->act_count += ACT_ADVANCE + actcount;
1498 				if (m->act_count > ACT_MAX)
1499 					m->act_count = ACT_MAX;
1500 			}
1501 		}
1502 		vm_page_flag_clear(m, PG_REFERENCED);
1503 
1504 		/*
1505 		 * actcount is only valid if the object ref_count is non-zero.
1506 		 * If the page does not have an object, actcount will be zero.
1507 		 */
1508 		if (actcount && m->object->ref_count != 0) {
1509 			vm_page_and_queue_spin_lock(m);
1510 			if (m->queue - m->pc == PQ_ACTIVE) {
1511 				TAILQ_REMOVE(
1512 					&vm_page_queues[PQ_ACTIVE + q].pl,
1513 					m, pageq);
1514 				TAILQ_INSERT_TAIL(
1515 					&vm_page_queues[PQ_ACTIVE + q].pl,
1516 					m, pageq);
1517 			}
1518 			vm_page_and_queue_spin_unlock(m);
1519 			vm_page_wakeup(m);
1520 		} else {
1521 			switch(m->object->type) {
1522 			case OBJT_DEFAULT:
1523 			case OBJT_SWAP:
1524 				m->act_count -= min(m->act_count,
1525 						    vm_anonmem_decline);
1526 				break;
1527 			default:
1528 				m->act_count -= min(m->act_count,
1529 						    vm_filemem_decline);
1530 				break;
1531 			}
1532 			if (vm_pageout_algorithm ||
1533 			    (m->object == NULL) ||
1534 			    (m->object && (m->object->ref_count == 0)) ||
1535 			    m->act_count < pass + 1
1536 			) {
1537 				/*
1538 				 * Deactivate the page.  If we had a
1539 				 * shortage from our inactive scan try to
1540 				 * free (cache) the page instead.
1541 				 *
1542 				 * Don't just blindly cache the page if
1543 				 * we do not have a shortage from the
1544 				 * inactive scan, that could lead to
1545 				 * gigabytes being moved.
1546 				 */
1547 				--inactive_shortage;
1548 				if (avail_shortage - delta > 0 ||
1549 				    (m->object && (m->object->ref_count == 0)))
1550 				{
1551 					if (avail_shortage - delta > 0)
1552 						++*recycle_countp;
1553 					vm_page_protect(m, VM_PROT_NONE);
1554 					if (m->dirty == 0 &&
1555 					    (m->flags & PG_NEED_COMMIT) == 0 &&
1556 					    avail_shortage - delta > 0) {
1557 						vm_page_cache(m);
1558 					} else {
1559 						vm_page_deactivate(m);
1560 						vm_page_wakeup(m);
1561 					}
1562 				} else {
1563 					vm_page_deactivate(m);
1564 					vm_page_wakeup(m);
1565 				}
1566 				++delta;
1567 			} else {
1568 				vm_page_and_queue_spin_lock(m);
1569 				if (m->queue - m->pc == PQ_ACTIVE) {
1570 					TAILQ_REMOVE(
1571 					    &vm_page_queues[PQ_ACTIVE + q].pl,
1572 					    m, pageq);
1573 					TAILQ_INSERT_TAIL(
1574 					    &vm_page_queues[PQ_ACTIVE + q].pl,
1575 					    m, pageq);
1576 				}
1577 				vm_page_and_queue_spin_unlock(m);
1578 				vm_page_wakeup(m);
1579 			}
1580 		}
1581 next:
1582 		lwkt_yield();
1583 		vm_page_queues_spin_lock(PQ_ACTIVE + q);
1584 	}
1585 
1586 	/*
1587 	 * Clean out our local marker.
1588 	 *
1589 	 * Page queue still spin-locked.
1590 	 */
1591 	TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1592 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1593 
1594 	return (delta);
1595 }
1596 
1597 /*
1598  * The number of actually free pages can drop down to v_free_reserved,
1599  * we try to build the free count back above v_free_min.  Note that
1600  * vm_paging_needed() also returns TRUE if v_free_count is not at
1601  * least v_free_min so that is the minimum we must build the free
1602  * count to.
1603  *
1604  * We use a slightly higher target to improve hysteresis,
1605  * ((v_free_target + v_free_min) / 2).  Since v_free_target
1606  * is usually the same as v_cache_min this maintains about
1607  * half the pages in the free queue as are in the cache queue,
1608  * providing pretty good pipelining for pageout operation.
1609  *
1610  * The system operator can manipulate vm.v_cache_min and
1611  * vm.v_free_target to tune the pageout demon.  Be sure
1612  * to keep vm.v_free_min < vm.v_free_target.
1613  *
1614  * Note that the original paging target is to get at least
1615  * (free_min + cache_min) into (free + cache).  The slightly
1616  * higher target will shift additional pages from cache to free
1617  * without effecting the original paging target in order to
1618  * maintain better hysteresis and not have the free count always
1619  * be dead-on v_free_min.
1620  *
1621  * NOTE: we are still in a critical section.
1622  *
1623  * Pages moved from PQ_CACHE to totally free are not counted in the
1624  * pages_freed counter.
1625  *
1626  * WARNING! Can be called from two pagedaemon threads simultaneously.
1627  */
1628 static void
1629 vm_pageout_scan_cache(long avail_shortage, int pass,
1630 		      long vnodes_skipped, long recycle_count)
1631 {
1632 	static int lastkillticks;
1633 	struct vm_pageout_scan_info info;
1634 	vm_page_t m;
1635 	int isep;
1636 
1637 	isep = (curthread == emergpager);
1638 
1639 	while (vmstats.v_free_count <
1640 	       (vmstats.v_free_min + vmstats.v_free_target) / 2) {
1641 		/*
1642 		 * This steals some code from vm/vm_page.c
1643 		 *
1644 		 * Create two rovers and adjust the code to reduce
1645 		 * chances of them winding up at the same index (which
1646 		 * can cause a lot of contention).
1647 		 */
1648 		static int cache_rover[2] = { 0, PQ_L2_MASK / 2 };
1649 
1650 		if (((cache_rover[0] ^ cache_rover[1]) & PQ_L2_MASK) == 0)
1651 			goto next_rover;
1652 
1653 		m = vm_page_list_find(PQ_CACHE, cache_rover[isep] & PQ_L2_MASK);
1654 		if (m == NULL)
1655 			break;
1656 		/*
1657 		 * page is returned removed from its queue and spinlocked
1658 		 *
1659 		 * If the busy attempt fails we can still deactivate the page.
1660 		 */
1661 		if (vm_page_busy_try(m, TRUE)) {
1662 			vm_page_deactivate_locked(m);
1663 			vm_page_spin_unlock(m);
1664 			continue;
1665 		}
1666 		vm_page_spin_unlock(m);
1667 		pagedaemon_wakeup();
1668 		lwkt_yield();
1669 
1670 		/*
1671 		 * Remaining operations run with the page busy and neither
1672 		 * the page or the queue will be spin-locked.
1673 		 */
1674 		if ((m->flags & (PG_UNQUEUED | PG_NEED_COMMIT)) ||
1675 		    m->hold_count ||
1676 		    m->wire_count) {
1677 			vm_page_deactivate(m);
1678 			vm_page_wakeup(m);
1679 			continue;
1680 		}
1681 		pmap_mapped_sync(m);
1682 		KKASSERT((m->flags & PG_MAPPED) == 0);
1683 		KKASSERT(m->dirty == 0);
1684 		vm_pageout_page_free(m);
1685 		mycpu->gd_cnt.v_dfree++;
1686 next_rover:
1687 		if (isep)
1688 			cache_rover[1] -= PQ_PRIME2;
1689 		else
1690 			cache_rover[0] += PQ_PRIME2;
1691 	}
1692 
1693 #if !defined(NO_SWAPPING)
1694 	/*
1695 	 * Idle process swapout -- run once per second.
1696 	 */
1697 	if (vm_swap_idle_enabled) {
1698 		static time_t lsec;
1699 		if (time_uptime != lsec) {
1700 			atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_IDLE);
1701 			vm_req_vmdaemon();
1702 			lsec = time_uptime;
1703 		}
1704 	}
1705 #endif
1706 
1707 	/*
1708 	 * If we didn't get enough free pages, and we have skipped a vnode
1709 	 * in a writeable object, wakeup the sync daemon.  And kick swapout
1710 	 * if we did not get enough free pages.
1711 	 */
1712 	if (vm_paging_target() > 0) {
1713 		if (vnodes_skipped && vm_page_count_min(0))
1714 			speedup_syncer(NULL);
1715 #if !defined(NO_SWAPPING)
1716 		if (vm_swap_enabled && vm_page_count_target()) {
1717 			atomic_set_int(&vm_pageout_req_swapout, VM_SWAP_NORMAL);
1718 			vm_req_vmdaemon();
1719 		}
1720 #endif
1721 	}
1722 
1723 	/*
1724 	 * Handle catastrophic conditions.  Under good conditions we should
1725 	 * be at the target, well beyond our minimum.  If we could not even
1726 	 * reach our minimum the system is under heavy stress.  But just being
1727 	 * under heavy stress does not trigger process killing.
1728 	 *
1729 	 * We consider ourselves to have run out of memory if the swap pager
1730 	 * is full and avail_shortage is still positive.  The secondary check
1731 	 * ensures that we do not kill processes if the instantanious
1732 	 * availability is good, even if the pageout demon pass says it
1733 	 * couldn't get to the target.
1734 	 *
1735 	 * NOTE!  THE EMERGENCY PAGER (isep) DOES NOT HANDLE SWAP FULL
1736 	 *	  SITUATIONS.
1737 	 */
1738 	if (swap_pager_almost_full &&
1739 	    pass > 0 &&
1740 	    isep == 0 &&
1741 	    (vm_page_count_min(recycle_count) || avail_shortage > 0)) {
1742 		kprintf("Warning: system low on memory+swap "
1743 			"shortage %ld for %d ticks!\n",
1744 			avail_shortage, ticks - swap_fail_ticks);
1745 		if (bootverbose)
1746 		kprintf("Metrics: spaf=%d spf=%d pass=%d "
1747 			"avail=%ld target=%ld last=%u\n",
1748 			swap_pager_almost_full,
1749 			swap_pager_full,
1750 			pass,
1751 			avail_shortage,
1752 			vm_paging_target(),
1753 			(unsigned int)(ticks - lastkillticks));
1754 	}
1755 	if (swap_pager_full &&
1756 	    pass > 1 &&
1757 	    isep == 0 &&
1758 	    avail_shortage > 0 &&
1759 	    vm_paging_target() > 0 &&
1760 	    (unsigned int)(ticks - lastkillticks) >= hz) {
1761 		/*
1762 		 * Kill something, maximum rate once per second to give
1763 		 * the process time to free up sufficient memory.
1764 		 */
1765 		lastkillticks = ticks;
1766 		info.bigproc = NULL;
1767 		info.bigsize = 0;
1768 		allproc_scan(vm_pageout_scan_callback, &info, 0);
1769 		if (info.bigproc != NULL) {
1770 			kprintf("Try to kill process %d %s\n",
1771 				info.bigproc->p_pid, info.bigproc->p_comm);
1772 			info.bigproc->p_nice = PRIO_MIN;
1773 			info.bigproc->p_usched->resetpriority(
1774 				FIRST_LWP_IN_PROC(info.bigproc));
1775 			atomic_set_int(&info.bigproc->p_flags, P_LOWMEMKILL);
1776 			killproc(info.bigproc, "out of swap space");
1777 			wakeup(&vmstats.v_free_count);
1778 			PRELE(info.bigproc);
1779 		}
1780 	}
1781 }
1782 
1783 static int
1784 vm_pageout_scan_callback(struct proc *p, void *data)
1785 {
1786 	struct vm_pageout_scan_info *info = data;
1787 	vm_offset_t size;
1788 
1789 	/*
1790 	 * Never kill system processes or init.  If we have configured swap
1791 	 * then try to avoid killing low-numbered pids.
1792 	 */
1793 	if ((p->p_flags & P_SYSTEM) || (p->p_pid == 1) ||
1794 	    ((p->p_pid < 48) && (vm_swap_size != 0))) {
1795 		return (0);
1796 	}
1797 
1798 	lwkt_gettoken(&p->p_token);
1799 
1800 	/*
1801 	 * if the process is in a non-running type state,
1802 	 * don't touch it.
1803 	 */
1804 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
1805 		lwkt_reltoken(&p->p_token);
1806 		return (0);
1807 	}
1808 
1809 	/*
1810 	 * Get the approximate process size.  Note that anonymous pages
1811 	 * with backing swap will be counted twice, but there should not
1812 	 * be too many such pages due to the stress the VM system is
1813 	 * under at this point.
1814 	 */
1815 	size = vmspace_anonymous_count(p->p_vmspace) +
1816 		vmspace_swap_count(p->p_vmspace);
1817 
1818 	/*
1819 	 * If the this process is bigger than the biggest one
1820 	 * remember it.
1821 	 */
1822 	if (info->bigsize < size) {
1823 		if (info->bigproc)
1824 			PRELE(info->bigproc);
1825 		PHOLD(p);
1826 		info->bigproc = p;
1827 		info->bigsize = size;
1828 	}
1829 	lwkt_reltoken(&p->p_token);
1830 	lwkt_yield();
1831 
1832 	return(0);
1833 }
1834 
1835 /*
1836  * This old guy slowly walks PQ_HOLD looking for pages which need to be
1837  * moved back to PQ_FREE.  It is possible for pages to accumulate here
1838  * when vm_page_free() races against vm_page_unhold(), resulting in a
1839  * page being left on a PQ_HOLD queue with hold_count == 0.
1840  *
1841  * It is easier to handle this edge condition here, in non-critical code,
1842  * rather than enforce a spin-lock for every 1->0 transition in
1843  * vm_page_unhold().
1844  *
1845  * NOTE: TAILQ_FOREACH becomes invalid the instant we unlock the queue.
1846  */
1847 static void
1848 vm_pageout_scan_hold(int q)
1849 {
1850 	vm_page_t m;
1851 
1852 	vm_page_queues_spin_lock(PQ_HOLD + q);
1853 	TAILQ_FOREACH(m, &vm_page_queues[PQ_HOLD + q].pl, pageq) {
1854 		if (m->flags & PG_MARKER)
1855 			continue;
1856 
1857 		/*
1858 		 * Process one page and return
1859 		 */
1860 		if (m->hold_count)
1861 			break;
1862 		kprintf("DEBUG: pageout HOLD->FREE %p\n", m);
1863 		vm_page_hold(m);
1864 		vm_page_queues_spin_unlock(PQ_HOLD + q);
1865 		vm_page_unhold(m);	/* reprocess */
1866 		return;
1867 	}
1868 	vm_page_queues_spin_unlock(PQ_HOLD + q);
1869 }
1870 
1871 /*
1872  * This routine tries to maintain the pseudo LRU active queue,
1873  * so that during long periods of time where there is no paging,
1874  * that some statistic accumulation still occurs.  This code
1875  * helps the situation where paging just starts to occur.
1876  */
1877 static void
1878 vm_pageout_page_stats(int q)
1879 {
1880 	static int fullintervalcount = 0;
1881 	struct vm_page marker;
1882 	vm_page_t m;
1883 	long pcount, tpcount;		/* Number of pages to check */
1884 	long page_shortage;
1885 
1886 	page_shortage = (vmstats.v_inactive_target + vmstats.v_cache_max +
1887 			 vmstats.v_free_min) -
1888 			(vmstats.v_free_count + vmstats.v_inactive_count +
1889 			 vmstats.v_cache_count);
1890 
1891 	if (page_shortage <= 0)
1892 		return;
1893 
1894 	pcount = vm_page_queues[PQ_ACTIVE + q].lcnt;
1895 	fullintervalcount += vm_pageout_stats_interval;
1896 	if (fullintervalcount < vm_pageout_full_stats_interval) {
1897 		tpcount = (vm_pageout_stats_max * pcount) /
1898 			  vmstats.v_page_count + 1;
1899 		if (pcount > tpcount)
1900 			pcount = tpcount;
1901 	} else {
1902 		fullintervalcount = 0;
1903 	}
1904 
1905 	bzero(&marker, sizeof(marker));
1906 	marker.flags = PG_FICTITIOUS | PG_MARKER;
1907 	marker.busy_count = PBUSY_LOCKED;
1908 	marker.queue = PQ_ACTIVE + q;
1909 	marker.pc = q;
1910 	marker.wire_count = 1;
1911 
1912 	vm_page_queues_spin_lock(PQ_ACTIVE + q);
1913 	TAILQ_INSERT_HEAD(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1914 
1915 	/*
1916 	 * Queue locked at top of loop to avoid stack marker issues.
1917 	 */
1918 	while ((m = TAILQ_NEXT(&marker, pageq)) != NULL &&
1919 	       pcount-- > 0)
1920 	{
1921 		int actcount;
1922 
1923 		KKASSERT(m->queue == PQ_ACTIVE + q);
1924 		TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
1925 		TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE + q].pl, m,
1926 				   &marker, pageq);
1927 
1928 		/*
1929 		 * Skip marker pages (atomic against other markers to avoid
1930 		 * infinite hop-over scans).
1931 		 */
1932 		if (m->flags & PG_MARKER)
1933 			continue;
1934 
1935 		/*
1936 		 * Ignore pages we can't busy
1937 		 */
1938 		if (vm_page_busy_try(m, TRUE))
1939 			continue;
1940 
1941 		/*
1942 		 * Remaining operations run with the page busy and neither
1943 		 * the page or the queue will be spin-locked.
1944 		 */
1945 		KKASSERT(m->queue == PQ_ACTIVE + q);
1946 		vm_page_queues_spin_unlock(PQ_ACTIVE + q);
1947 
1948 		/*
1949 		 * We can just remove wired pages from the queue
1950 		 */
1951 		if (m->wire_count) {
1952 			vm_page_unqueue_nowakeup(m);
1953 			vm_page_wakeup(m);
1954 			goto next;
1955 		}
1956 
1957 
1958 		/*
1959 		 * We now have a safely busied page, the page and queue
1960 		 * spinlocks have been released.
1961 		 *
1962 		 * Ignore held and wired pages
1963 		 */
1964 		if (m->hold_count || m->wire_count) {
1965 			vm_page_wakeup(m);
1966 			goto next;
1967 		}
1968 
1969 		/*
1970 		 * Calculate activity
1971 		 */
1972 		actcount = 0;
1973 		if (m->flags & PG_REFERENCED) {
1974 			vm_page_flag_clear(m, PG_REFERENCED);
1975 			actcount += 1;
1976 		}
1977 		actcount += pmap_ts_referenced(m);
1978 
1979 		/*
1980 		 * Update act_count and move page to end of queue.
1981 		 */
1982 		if (actcount) {
1983 			m->act_count += ACT_ADVANCE + actcount;
1984 			if (m->act_count > ACT_MAX)
1985 				m->act_count = ACT_MAX;
1986 			vm_page_and_queue_spin_lock(m);
1987 			if (m->queue - m->pc == PQ_ACTIVE) {
1988 				TAILQ_REMOVE(
1989 					&vm_page_queues[PQ_ACTIVE + q].pl,
1990 					m, pageq);
1991 				TAILQ_INSERT_TAIL(
1992 					&vm_page_queues[PQ_ACTIVE + q].pl,
1993 					m, pageq);
1994 			}
1995 			vm_page_and_queue_spin_unlock(m);
1996 			vm_page_wakeup(m);
1997 			goto next;
1998 		}
1999 
2000 		if (m->act_count == 0) {
2001 			/*
2002 			 * We turn off page access, so that we have
2003 			 * more accurate RSS stats.  We don't do this
2004 			 * in the normal page deactivation when the
2005 			 * system is loaded VM wise, because the
2006 			 * cost of the large number of page protect
2007 			 * operations would be higher than the value
2008 			 * of doing the operation.
2009 			 *
2010 			 * We use the marker to save our place so
2011 			 * we can release the spin lock.  both (m)
2012 			 * and (next) will be invalid.
2013 			 */
2014 			vm_page_protect(m, VM_PROT_NONE);
2015 			vm_page_deactivate(m);
2016 		} else {
2017 			m->act_count -= min(m->act_count, ACT_DECLINE);
2018 			vm_page_and_queue_spin_lock(m);
2019 			if (m->queue - m->pc == PQ_ACTIVE) {
2020 				TAILQ_REMOVE(
2021 					&vm_page_queues[PQ_ACTIVE + q].pl,
2022 					m, pageq);
2023 				TAILQ_INSERT_TAIL(
2024 					&vm_page_queues[PQ_ACTIVE + q].pl,
2025 					m, pageq);
2026 			}
2027 			vm_page_and_queue_spin_unlock(m);
2028 		}
2029 		vm_page_wakeup(m);
2030 next:
2031 		vm_page_queues_spin_lock(PQ_ACTIVE + q);
2032 	}
2033 
2034 	/*
2035 	 * Remove our local marker
2036 	 *
2037 	 * Page queue still spin-locked.
2038 	 */
2039 	TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE + q].pl, &marker, pageq);
2040 	vm_page_queues_spin_unlock(PQ_ACTIVE + q);
2041 }
2042 
2043 static void
2044 vm_pageout_free_page_calc(vm_size_t count)
2045 {
2046 	/*
2047 	 * v_free_min		normal allocations
2048 	 * v_free_reserved	system allocations
2049 	 * v_pageout_free_min	allocations by pageout daemon
2050 	 * v_interrupt_free_min	low level allocations (e.g swap structures)
2051 	 *
2052 	 * v_free_min is used to generate several other baselines, and they
2053 	 * can get pretty silly on systems with a lot of memory.
2054 	 */
2055 	vmstats.v_free_min = 64 + vmstats.v_page_count / 200;
2056 	vmstats.v_free_reserved = vmstats.v_free_min * 4 / 8 + 7;
2057 	vmstats.v_free_severe = vmstats.v_free_min * 4 / 8 + 0;
2058 	vmstats.v_pageout_free_min = vmstats.v_free_min * 2 / 8 + 7;
2059 	vmstats.v_interrupt_free_min = vmstats.v_free_min * 1 / 8 + 7;
2060 }
2061 
2062 
2063 /*
2064  * vm_pageout is the high level pageout daemon.  TWO kernel threads run
2065  * this daemon, the primary pageout daemon and the emergency pageout daemon.
2066  *
2067  * The emergency pageout daemon takes over when the primary pageout daemon
2068  * deadlocks.  The emergency pageout daemon ONLY pages out to swap, thus
2069  * avoiding the many low-memory deadlocks which can occur when paging out
2070  * to VFS's.
2071  */
2072 static void
2073 vm_pageout_thread(void)
2074 {
2075 	int pass;
2076 	int q;
2077 	int q1iterator = 0;
2078 	int q2iterator = 0;
2079 	int q3iterator = 0;
2080 	int isep;
2081 
2082 	curthread->td_flags |= TDF_SYSTHREAD;
2083 
2084 	/*
2085 	 * We only need to setup once.
2086 	 */
2087 	isep = 0;
2088 	if (curthread == emergpager) {
2089 		isep = 1;
2090 		goto skip_setup;
2091 	}
2092 
2093 	/*
2094 	 * Initialize some paging parameters.
2095 	 */
2096 	vm_pageout_free_page_calc(vmstats.v_page_count);
2097 
2098 	/*
2099 	 * v_free_target and v_cache_min control pageout hysteresis.  Note
2100 	 * that these are more a measure of the VM cache queue hysteresis
2101 	 * then the VM free queue.  Specifically, v_free_target is the
2102 	 * high water mark (free+cache pages).
2103 	 *
2104 	 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the
2105 	 * low water mark, while v_free_min is the stop.  v_cache_min must
2106 	 * be big enough to handle memory needs while the pageout daemon
2107 	 * is signalled and run to free more pages.
2108 	 */
2109 	vmstats.v_free_target = 4 * vmstats.v_free_min +
2110 				vmstats.v_free_reserved;
2111 
2112 	/*
2113 	 * NOTE: With the new buffer cache b_act_count we want the default
2114 	 *	 inactive target to be a percentage of available memory.
2115 	 *
2116 	 *	 The inactive target essentially determines the minimum
2117 	 *	 number of 'temporary' pages capable of caching one-time-use
2118 	 *	 files when the VM system is otherwise full of pages
2119 	 *	 belonging to multi-time-use files or active program data.
2120 	 *
2121 	 * NOTE: The inactive target is aggressively persued only if the
2122 	 *	 inactive queue becomes too small.  If the inactive queue
2123 	 *	 is large enough to satisfy page movement to free+cache
2124 	 *	 then it is repopulated more slowly from the active queue.
2125 	 *	 This allows a general inactive_target default to be set.
2126 	 *
2127 	 *	 There is an issue here for processes which sit mostly idle
2128 	 *	 'overnight', such as sshd, tcsh, and X.  Any movement from
2129 	 *	 the active queue will eventually cause such pages to
2130 	 *	 recycle eventually causing a lot of paging in the morning.
2131 	 *	 To reduce the incidence of this pages cycled out of the
2132 	 *	 buffer cache are moved directly to the inactive queue if
2133 	 *	 they were only used once or twice.
2134 	 *
2135 	 *	 The vfs.vm_cycle_point sysctl can be used to adjust this.
2136 	 *	 Increasing the value (up to 64) increases the number of
2137 	 *	 buffer recyclements which go directly to the inactive queue.
2138 	 */
2139 	if (vmstats.v_free_count > 2048) {
2140 		vmstats.v_cache_min = vmstats.v_free_target;
2141 		vmstats.v_cache_max = 2 * vmstats.v_cache_min;
2142 	} else {
2143 		vmstats.v_cache_min = 0;
2144 		vmstats.v_cache_max = 0;
2145 	}
2146 	vmstats.v_inactive_target = vmstats.v_free_count / 4;
2147 
2148 	/* XXX does not really belong here */
2149 	if (vm_page_max_wired == 0)
2150 		vm_page_max_wired = vmstats.v_free_count / 3;
2151 
2152 	if (vm_pageout_stats_max == 0)
2153 		vm_pageout_stats_max = vmstats.v_free_target;
2154 
2155 	/*
2156 	 * Set interval in seconds for stats scan.
2157 	 */
2158 	if (vm_pageout_stats_interval == 0)
2159 		vm_pageout_stats_interval = 5;
2160 	if (vm_pageout_full_stats_interval == 0)
2161 		vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4;
2162 
2163 
2164 	/*
2165 	 * Set maximum free per pass
2166 	 */
2167 	if (vm_pageout_stats_free_max == 0)
2168 		vm_pageout_stats_free_max = 5;
2169 
2170 	swap_pager_swap_init();
2171 	pass = 0;
2172 
2173 	atomic_swap_int(&sequence_emerg_pager, 1);
2174 	wakeup(&sequence_emerg_pager);
2175 
2176 skip_setup:
2177 	/*
2178 	 * Sequence emergency pager startup
2179 	 */
2180 	if (isep) {
2181 		while (sequence_emerg_pager == 0)
2182 			tsleep(&sequence_emerg_pager, 0, "pstartup", hz);
2183 	}
2184 
2185 	/*
2186 	 * The pageout daemon is never done, so loop forever.
2187 	 *
2188 	 * WARNING!  This code is being executed by two kernel threads
2189 	 *	     potentially simultaneously.
2190 	 */
2191 	while (TRUE) {
2192 		int error;
2193 		long avail_shortage;
2194 		long inactive_shortage;
2195 		long vnodes_skipped = 0;
2196 		long recycle_count = 0;
2197 		long tmp;
2198 
2199 		/*
2200 		 * Wait for an action request.  If we timeout check to
2201 		 * see if paging is needed (in case the normal wakeup
2202 		 * code raced us).
2203 		 */
2204 		if (isep) {
2205 			/*
2206 			 * Emergency pagedaemon monitors the primary
2207 			 * pagedaemon while vm_pages_needed != 0.
2208 			 *
2209 			 * The emergency pagedaemon only runs if VM paging
2210 			 * is needed and the primary pagedaemon has not
2211 			 * updated vm_pagedaemon_time for more than 2 seconds.
2212 			 */
2213 			if (vm_pages_needed)
2214 				tsleep(&vm_pagedaemon_time, 0, "psleep", hz);
2215 			else
2216 				tsleep(&vm_pagedaemon_time, 0, "psleep", hz*10);
2217 			if (vm_pages_needed == 0) {
2218 				pass = 0;
2219 				continue;
2220 			}
2221 			if ((int)(ticks - vm_pagedaemon_time) < hz * 2) {
2222 				pass = 0;
2223 				continue;
2224 			}
2225 		} else {
2226 			/*
2227 			 * Primary pagedaemon
2228 			 *
2229 			 * NOTE: We unconditionally cleanup PQ_HOLD even
2230 			 *	 when there is no work to do.
2231 			 */
2232 			vm_pageout_scan_hold(q3iterator & PQ_L2_MASK);
2233 			++q3iterator;
2234 
2235 			if (vm_pages_needed == 0) {
2236 				error = tsleep(&vm_pages_needed,
2237 					       0, "psleep",
2238 					       vm_pageout_stats_interval * hz);
2239 				if (error &&
2240 				    vm_paging_needed(0) == 0 &&
2241 				    vm_pages_needed == 0) {
2242 					for (q = 0; q < PQ_L2_SIZE; ++q)
2243 						vm_pageout_page_stats(q);
2244 					continue;
2245 				}
2246 				vm_pagedaemon_time = ticks;
2247 				vm_pages_needed = 1;
2248 
2249 				/*
2250 				 * Wake the emergency pagedaemon up so it
2251 				 * can monitor us.  It will automatically
2252 				 * go back into a long sleep when
2253 				 * vm_pages_needed returns to 0.
2254 				 */
2255 				wakeup(&vm_pagedaemon_time);
2256 			}
2257 		}
2258 
2259 		mycpu->gd_cnt.v_pdwakeups++;
2260 
2261 		/*
2262 		 * Scan for INACTIVE->CLEAN/PAGEOUT
2263 		 *
2264 		 * This routine tries to avoid thrashing the system with
2265 		 * unnecessary activity.
2266 		 *
2267 		 * Calculate our target for the number of free+cache pages we
2268 		 * want to get to.  This is higher then the number that causes
2269 		 * allocations to stall (severe) in order to provide hysteresis,
2270 		 * and if we don't make it all the way but get to the minimum
2271 		 * we're happy.  Goose it a bit if there are multiple requests
2272 		 * for memory.
2273 		 *
2274 		 * Don't reduce avail_shortage inside the loop or the
2275 		 * PQAVERAGE() calculation will break.
2276 		 *
2277 		 * NOTE! deficit is differentiated from avail_shortage as
2278 		 *	 REQUIRING at least (deficit) pages to be cleaned,
2279 		 *	 even if the page queues are in good shape.  This
2280 		 *	 is used primarily for handling per-process
2281 		 *	 RLIMIT_RSS and may also see small values when
2282 		 *	 processes block due to low memory.
2283 		 */
2284 		vmstats_rollup();
2285 		if (isep == 0)
2286 			vm_pagedaemon_time = ticks;
2287 		avail_shortage = vm_paging_target() + vm_pageout_deficit;
2288 		vm_pageout_deficit = 0;
2289 
2290 		if (avail_shortage > 0) {
2291 			long delta = 0;
2292 			int qq;
2293 
2294 			qq = q1iterator;
2295 			for (q = 0; q < PQ_L2_SIZE; ++q) {
2296 				delta += vm_pageout_scan_inactive(
2297 					    pass,
2298 					    qq & PQ_L2_MASK,
2299 					    PQAVERAGE(avail_shortage),
2300 					    &vnodes_skipped);
2301 				if (isep)
2302 					--qq;
2303 				else
2304 					++qq;
2305 				if (avail_shortage - delta <= 0)
2306 					break;
2307 
2308 				/*
2309 				 * It is possible for avail_shortage to be
2310 				 * very large.  If a large program exits or
2311 				 * frees a ton of memory all at once, we do
2312 				 * not have to continue deactivations.
2313 				 *
2314 				 * (We will still run the active->inactive
2315 				 * target, however).
2316 				 */
2317 				if (!vm_page_count_target() &&
2318 				    !vm_page_count_min(
2319 						vm_page_free_hysteresis)) {
2320 					avail_shortage = 0;
2321 					break;
2322 				}
2323 			}
2324 			avail_shortage -= delta;
2325 			q1iterator = qq;
2326 		}
2327 
2328 		/*
2329 		 * Figure out how many active pages we must deactivate.  If
2330 		 * we were able to reach our target with just the inactive
2331 		 * scan above we limit the number of active pages we
2332 		 * deactivate to reduce unnecessary work.
2333 		 */
2334 		vmstats_rollup();
2335 		if (isep == 0)
2336 			vm_pagedaemon_time = ticks;
2337 		inactive_shortage = vmstats.v_inactive_target -
2338 				    vmstats.v_inactive_count;
2339 
2340 		/*
2341 		 * If we were unable to free sufficient inactive pages to
2342 		 * satisfy the free/cache queue requirements then simply
2343 		 * reaching the inactive target may not be good enough.
2344 		 * Try to deactivate pages in excess of the target based
2345 		 * on the shortfall.
2346 		 *
2347 		 * However to prevent thrashing the VM system do not
2348 		 * deactivate more than an additional 1/10 the inactive
2349 		 * target's worth of active pages.
2350 		 */
2351 		if (avail_shortage > 0) {
2352 			tmp = avail_shortage * 2;
2353 			if (tmp > vmstats.v_inactive_target / 10)
2354 				tmp = vmstats.v_inactive_target / 10;
2355 			inactive_shortage += tmp;
2356 		}
2357 
2358 		/*
2359 		 * Only trigger a pmap cleanup on inactive shortage.
2360 		 */
2361 		if (isep == 0 && inactive_shortage > 0) {
2362 			pmap_collect();
2363 		}
2364 
2365 		/*
2366 		 * Scan for ACTIVE->INACTIVE
2367 		 *
2368 		 * Only trigger on inactive shortage.  Triggering on
2369 		 * avail_shortage can starve the active queue with
2370 		 * unnecessary active->inactive transitions and destroy
2371 		 * performance.
2372 		 *
2373 		 * If this is the emergency pager, always try to move
2374 		 * a few pages from active to inactive because the inactive
2375 		 * queue might have enough pages, but not enough anonymous
2376 		 * pages.
2377 		 */
2378 		if (isep && inactive_shortage < vm_emerg_launder)
2379 			inactive_shortage = vm_emerg_launder;
2380 
2381 		if (/*avail_shortage > 0 ||*/ inactive_shortage > 0) {
2382 			long delta = 0;
2383 			int qq;
2384 
2385 			qq = q2iterator;
2386 			for (q = 0; q < PQ_L2_SIZE; ++q) {
2387 				delta += vm_pageout_scan_active(
2388 						pass,
2389 						qq & PQ_L2_MASK,
2390 						PQAVERAGE(avail_shortage),
2391 						PQAVERAGE(inactive_shortage),
2392 						&recycle_count);
2393 				if (isep)
2394 					--qq;
2395 				else
2396 					++qq;
2397 				if (inactive_shortage - delta <= 0 &&
2398 				    avail_shortage - delta <= 0) {
2399 					break;
2400 				}
2401 
2402 				/*
2403 				 * inactive_shortage can be a very large
2404 				 * number.  This is intended to break out
2405 				 * early if our inactive_target has been
2406 				 * reached due to other system activity.
2407 				 */
2408 				if (vmstats.v_inactive_count >
2409 				    vmstats.v_inactive_target) {
2410 					inactive_shortage = 0;
2411 					break;
2412 				}
2413 			}
2414 			inactive_shortage -= delta;
2415 			avail_shortage -= delta;
2416 			q2iterator = qq;
2417 		}
2418 
2419 		/*
2420 		 * Scan for CACHE->FREE
2421 		 *
2422 		 * Finally free enough cache pages to meet our free page
2423 		 * requirement and take more drastic measures if we are
2424 		 * still in trouble.
2425 		 */
2426 		vmstats_rollup();
2427 		if (isep == 0)
2428 			vm_pagedaemon_time = ticks;
2429 		vm_pageout_scan_cache(avail_shortage, pass,
2430 				      vnodes_skipped, recycle_count);
2431 
2432 		/*
2433 		 * This is a bit sophisticated because we do not necessarily
2434 		 * want to force paging until our targets are reached if we
2435 		 * were able to successfully retire the shortage we calculated.
2436 		 */
2437 		if (avail_shortage > 0) {
2438 			/*
2439 			 * If we did not retire enough pages continue the
2440 			 * pageout operation until we are able to.
2441 			 */
2442 			++pass;
2443 
2444 			if (pass < 10 && vm_pages_needed > 1) {
2445 				/*
2446 				 * Normal operation, additional processes
2447 				 * have already kicked us.  Retry immediately
2448 				 * unless swap space is completely full in
2449 				 * which case delay a bit.
2450 				 */
2451 				if (swap_pager_full) {
2452 					tsleep(&vm_pages_needed, 0, "pdelay",
2453 						hz / 5);
2454 				} /* else immediate retry */
2455 			} else if (pass < 10) {
2456 				/*
2457 				 * Do a short sleep for the first 10 passes,
2458 				 * allow the sleep to be woken up by resetting
2459 				 * vm_pages_needed to 1 (NOTE: we are still
2460 				 * active paging!).
2461 				 */
2462 				if (isep == 0)
2463 					vm_pages_needed = 1;
2464 				tsleep(&vm_pages_needed, 0, "pdelay", 2);
2465 			} else if (swap_pager_full == 0) {
2466 				/*
2467 				 * We've taken too many passes, force a
2468 				 * longer delay.
2469 				 */
2470 				tsleep(&vm_pages_needed, 0, "pdelay", hz / 10);
2471 			} else {
2472 				/*
2473 				 * Running out of memory, catastrophic
2474 				 * back-off to one-second intervals.
2475 				 */
2476 				tsleep(&vm_pages_needed, 0, "pdelay", hz);
2477 			}
2478 		} else if (vm_pages_needed) {
2479 			/*
2480 			 * We retired our calculated shortage but we may have
2481 			 * to continue paging if threads drain memory too far
2482 			 * below our target.
2483 			 *
2484 			 * Similar to vm_page_free_wakeup() in vm_page.c.
2485 			 */
2486 			pass = 0;
2487 			if (!vm_paging_needed(0)) {
2488 				/* still more than half-way to our target */
2489 				vm_pages_needed = 0;
2490 				wakeup(&vmstats.v_free_count);
2491 			} else
2492 			if (!vm_page_count_min(vm_page_free_hysteresis)) {
2493 				/*
2494 				 * Continue operations with wakeup
2495 				 * (set variable to avoid overflow)
2496 				 */
2497 				vm_pages_needed = 2;
2498 				wakeup(&vmstats.v_free_count);
2499 			} else {
2500 				/*
2501 				 * No wakeup() needed, continue operations.
2502 				 * (set variable to avoid overflow)
2503 				 */
2504 				vm_pages_needed = 2;
2505 			}
2506 		} else {
2507 			/*
2508 			 * Turn paging back on immediately if we are under
2509 			 * minimum.
2510 			 */
2511 			pass = 0;
2512 		}
2513 	}
2514 }
2515 
2516 static struct kproc_desc pg1_kp = {
2517 	"pagedaemon",
2518 	vm_pageout_thread,
2519 	&pagethread
2520 };
2521 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &pg1_kp);
2522 
2523 static struct kproc_desc pg2_kp = {
2524 	"emergpager",
2525 	vm_pageout_thread,
2526 	&emergpager
2527 };
2528 SYSINIT(emergpager, SI_SUB_KTHREAD_PAGE, SI_ORDER_ANY, kproc_start, &pg2_kp);
2529 
2530 
2531 /*
2532  * Called after allocating a page out of the cache or free queue
2533  * to possibly wake the pagedaemon up to replentish our supply.
2534  *
2535  * We try to generate some hysteresis by waking the pagedaemon up
2536  * when our free+cache pages go below the free_min+cache_min level.
2537  * The pagedaemon tries to get the count back up to at least the
2538  * minimum, and through to the target level if possible.
2539  *
2540  * If the pagedaemon is already active bump vm_pages_needed as a hint
2541  * that there are even more requests pending.
2542  *
2543  * SMP races ok?
2544  * No requirements.
2545  */
2546 void
2547 pagedaemon_wakeup(void)
2548 {
2549 	if (vm_paging_needed(0) && curthread != pagethread) {
2550 		if (vm_pages_needed <= 1) {
2551 			vm_pages_needed = 1;		/* SMP race ok */
2552 			wakeup(&vm_pages_needed);	/* tickle pageout */
2553 		} else if (vm_page_count_min(0)) {
2554 			++vm_pages_needed;		/* SMP race ok */
2555 			/* a wakeup() would be wasted here */
2556 		}
2557 	}
2558 }
2559 
2560 #if !defined(NO_SWAPPING)
2561 
2562 /*
2563  * SMP races ok?
2564  * No requirements.
2565  */
2566 static void
2567 vm_req_vmdaemon(void)
2568 {
2569 	static int lastrun = 0;
2570 
2571 	if ((ticks > (lastrun + hz)) || (ticks < lastrun)) {
2572 		wakeup(&vm_daemon_needed);
2573 		lastrun = ticks;
2574 	}
2575 }
2576 
2577 static int vm_daemon_callback(struct proc *p, void *data __unused);
2578 
2579 /*
2580  * No requirements.
2581  */
2582 static void
2583 vm_daemon(void)
2584 {
2585 	int req_swapout;
2586 
2587 	while (TRUE) {
2588 		tsleep(&vm_daemon_needed, 0, "psleep", 0);
2589 		req_swapout = atomic_swap_int(&vm_pageout_req_swapout, 0);
2590 
2591 		/*
2592 		 * forced swapouts
2593 		 */
2594 		if (req_swapout)
2595 			swapout_procs(vm_pageout_req_swapout);
2596 
2597 		/*
2598 		 * scan the processes for exceeding their rlimits or if
2599 		 * process is swapped out -- deactivate pages
2600 		 */
2601 		allproc_scan(vm_daemon_callback, NULL, 0);
2602 	}
2603 }
2604 
2605 static int
2606 vm_daemon_callback(struct proc *p, void *data __unused)
2607 {
2608 	struct vmspace *vm;
2609 	vm_pindex_t limit, size;
2610 
2611 	/*
2612 	 * if this is a system process or if we have already
2613 	 * looked at this process, skip it.
2614 	 */
2615 	lwkt_gettoken(&p->p_token);
2616 
2617 	if (p->p_flags & (P_SYSTEM | P_WEXIT)) {
2618 		lwkt_reltoken(&p->p_token);
2619 		return (0);
2620 	}
2621 
2622 	/*
2623 	 * if the process is in a non-running type state,
2624 	 * don't touch it.
2625 	 */
2626 	if (p->p_stat != SACTIVE && p->p_stat != SSTOP && p->p_stat != SCORE) {
2627 		lwkt_reltoken(&p->p_token);
2628 		return (0);
2629 	}
2630 
2631 	/*
2632 	 * get a limit
2633 	 */
2634 	limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
2635 			        p->p_rlimit[RLIMIT_RSS].rlim_max));
2636 
2637 	/*
2638 	 * let processes that are swapped out really be
2639 	 * swapped out.  Set the limit to nothing to get as
2640 	 * many pages out to swap as possible.
2641 	 */
2642 	if (p->p_flags & P_SWAPPEDOUT)
2643 		limit = 0;
2644 
2645 	vm = p->p_vmspace;
2646 	vmspace_hold(vm);
2647 	size = pmap_resident_tlnw_count(&vm->vm_pmap);
2648 	if (limit >= 0 && size > 4096 &&
2649 	    size - 4096 >= limit && vm_pageout_memuse_mode >= 1) {
2650 		vm_pageout_map_deactivate_pages(&vm->vm_map, limit);
2651 	}
2652 	vmspace_drop(vm);
2653 
2654 	lwkt_reltoken(&p->p_token);
2655 
2656 	return (0);
2657 }
2658 
2659 #endif
2660