xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision fd5cb0acea84d278e04e640d37ca2398f894991f)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.61 2005/01/30 17:23:05 chs Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993, The Regents of the University of California.
6  *
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by Charles D. Cranor,
23  *      Washington University, the University of California, Berkeley and
24  *      its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)vm_pageout.c        8.5 (Berkeley) 2/14/94
42  * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 /*
70  * uvm_pdaemon.c: the page daemon
71  */
72 
73 #include <sys/cdefs.h>
74 __KERNEL_RCSID(0, "$NetBSD: uvm_pdaemon.c,v 1.61 2005/01/30 17:23:05 chs Exp $");
75 
76 #include "opt_uvmhist.h"
77 
78 #include <sys/param.h>
79 #include <sys/proc.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/pool.h>
83 #include <sys/buf.h>
84 #include <sys/vnode.h>
85 
86 #include <uvm/uvm.h>
87 
88 /*
89  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
90  * in a pass thru the inactive list when swap is full.  the value should be
91  * "small"... if it's too large we'll cycle the active pages thru the inactive
92  * queue too quickly to for them to be referenced and avoid being freed.
93  */
94 
95 #define UVMPD_NUMDIRTYREACTS 16
96 
97 
98 /*
99  * local prototypes
100  */
101 
102 void		uvmpd_scan(void);
103 void		uvmpd_scan_inactive(struct pglist *);
104 void		uvmpd_tune(void);
105 
106 /*
107  * XXX hack to avoid hangs when large processes fork.
108  */
109 int uvm_extrapages;
110 
111 /*
112  * uvm_wait: wait (sleep) for the page daemon to free some pages
113  *
114  * => should be called with all locks released
115  * => should _not_ be called by the page daemon (to avoid deadlock)
116  */
117 
118 void
119 uvm_wait(wmsg)
120 	const char *wmsg;
121 {
122 	int timo = 0;
123 	int s = splbio();
124 
125 	/*
126 	 * check for page daemon going to sleep (waiting for itself)
127 	 */
128 
129 	if (curproc == uvm.pagedaemon_proc && uvmexp.paging == 0) {
130 		/*
131 		 * now we have a problem: the pagedaemon wants to go to
132 		 * sleep until it frees more memory.   but how can it
133 		 * free more memory if it is asleep?  that is a deadlock.
134 		 * we have two options:
135 		 *  [1] panic now
136 		 *  [2] put a timeout on the sleep, thus causing the
137 		 *      pagedaemon to only pause (rather than sleep forever)
138 		 *
139 		 * note that option [2] will only help us if we get lucky
140 		 * and some other process on the system breaks the deadlock
141 		 * by exiting or freeing memory (thus allowing the pagedaemon
142 		 * to continue).  for now we panic if DEBUG is defined,
143 		 * otherwise we hope for the best with option [2] (better
144 		 * yet, this should never happen in the first place!).
145 		 */
146 
147 		printf("pagedaemon: deadlock detected!\n");
148 		timo = hz >> 3;		/* set timeout */
149 #if defined(DEBUG)
150 		/* DEBUG: panic so we can debug it */
151 		panic("pagedaemon deadlock");
152 #endif
153 	}
154 
155 	simple_lock(&uvm.pagedaemon_lock);
156 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
157 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, wmsg,
158 	    timo);
159 
160 	splx(s);
161 }
162 
163 
164 /*
165  * uvmpd_tune: tune paging parameters
166  *
167  * => called when ever memory is added (or removed?) to the system
168  * => caller must call with page queues locked
169  */
170 
171 void
172 uvmpd_tune(void)
173 {
174 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
175 
176 	uvmexp.freemin = uvmexp.npages / 20;
177 
178 	/* between 16k and 256k */
179 	/* XXX:  what are these values good for? */
180 	uvmexp.freemin = MAX(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
181 	uvmexp.freemin = MIN(uvmexp.freemin, (256*1024) >> PAGE_SHIFT);
182 
183 	/* Make sure there's always a user page free. */
184 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
185 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
186 
187 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
188 	if (uvmexp.freetarg <= uvmexp.freemin)
189 		uvmexp.freetarg = uvmexp.freemin + 1;
190 
191 	uvmexp.freetarg += uvm_extrapages;
192 	uvm_extrapages = 0;
193 
194 	/* uvmexp.inactarg: computed in main daemon loop */
195 
196 	uvmexp.wiredmax = uvmexp.npages / 3;
197 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
198 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
199 }
200 
201 /*
202  * uvm_pageout: the main loop for the pagedaemon
203  */
204 
205 void
206 uvm_pageout(void *arg)
207 {
208 	int bufcnt, npages = 0;
209 	int extrapages = 0;
210 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
211 
212 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
213 
214 	/*
215 	 * ensure correct priority and set paging parameters...
216 	 */
217 
218 	uvm.pagedaemon_proc = curproc;
219 	uvm_lock_pageq();
220 	npages = uvmexp.npages;
221 	uvmpd_tune();
222 	uvm_unlock_pageq();
223 
224 	/*
225 	 * main loop
226 	 */
227 
228 	for (;;) {
229 		simple_lock(&uvm.pagedaemon_lock);
230 
231 		UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
232 		UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
233 		    &uvm.pagedaemon_lock, FALSE, "pgdaemon", 0);
234 		uvmexp.pdwoke++;
235 		UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
236 
237 		/*
238 		 * now lock page queues and recompute inactive count
239 		 */
240 
241 		uvm_lock_pageq();
242 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
243 			npages = uvmexp.npages;
244 			extrapages = uvm_extrapages;
245 			uvmpd_tune();
246 		}
247 
248 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
249 		if (uvmexp.inactarg <= uvmexp.freetarg) {
250 			uvmexp.inactarg = uvmexp.freetarg + 1;
251 		}
252 
253 		/*
254 		 * Estimate a hint.  Note that bufmem are returned to
255 		 * system only when entire pool page is empty.
256 		 */
257 		bufcnt = uvmexp.freetarg - uvmexp.free;
258 		if (bufcnt < 0)
259 			bufcnt = 0;
260 
261 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d, inact/itarg=%d/%d",
262 		    uvmexp.free, uvmexp.freetarg, uvmexp.inactive,
263 		    uvmexp.inactarg);
264 
265 		/*
266 		 * scan if needed
267 		 */
268 
269 		if (uvmexp.free + uvmexp.paging < uvmexp.freetarg ||
270 		    uvmexp.inactive < uvmexp.inactarg) {
271 			uvmpd_scan();
272 		}
273 
274 		/*
275 		 * if there's any free memory to be had,
276 		 * wake up any waiters.
277 		 */
278 
279 		if (uvmexp.free > uvmexp.reserve_kernel ||
280 		    uvmexp.paging == 0) {
281 			wakeup(&uvmexp.free);
282 		}
283 
284 		/*
285 		 * scan done.  unlock page queues (the only lock we are holding)
286 		 */
287 
288 		uvm_unlock_pageq();
289 
290 		buf_drain(bufcnt << PAGE_SHIFT);
291 
292 		/*
293 		 * drain pool resources now that we're not holding any locks
294 		 */
295 
296 		pool_drain(0);
297 
298 		/*
299 		 * free any cached u-areas we don't need
300 		 */
301 		uvm_uarea_drain(TRUE);
302 
303 	}
304 	/*NOTREACHED*/
305 }
306 
307 
308 /*
309  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
310  */
311 
312 void
313 uvm_aiodone_daemon(void *arg)
314 {
315 	int s, free;
316 	struct buf *bp, *nbp;
317 	UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist);
318 
319 	for (;;) {
320 
321 		/*
322 		 * carefully attempt to go to sleep (without losing "wakeups"!).
323 		 * we need splbio because we want to make sure the aio_done list
324 		 * is totally empty before we go to sleep.
325 		 */
326 
327 		s = splbio();
328 		simple_lock(&uvm.aiodoned_lock);
329 		if (TAILQ_FIRST(&uvm.aio_done) == NULL) {
330 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
331 			UVM_UNLOCK_AND_WAIT(&uvm.aiodoned,
332 			    &uvm.aiodoned_lock, FALSE, "aiodoned", 0);
333 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
334 
335 			/* relock aiodoned_lock, still at splbio */
336 			simple_lock(&uvm.aiodoned_lock);
337 		}
338 
339 		/*
340 		 * check for done aio structures
341 		 */
342 
343 		bp = TAILQ_FIRST(&uvm.aio_done);
344 		if (bp) {
345 			TAILQ_INIT(&uvm.aio_done);
346 		}
347 
348 		simple_unlock(&uvm.aiodoned_lock);
349 		splx(s);
350 
351 		/*
352 		 * process each i/o that's done.
353 		 */
354 
355 		free = uvmexp.free;
356 		while (bp != NULL) {
357 			nbp = TAILQ_NEXT(bp, b_freelist);
358 			(*bp->b_iodone)(bp);
359 			bp = nbp;
360 		}
361 		if (free <= uvmexp.reserve_kernel) {
362 			s = uvm_lock_fpageq();
363 			wakeup(&uvm.pagedaemon);
364 			uvm_unlock_fpageq(s);
365 		} else {
366 			simple_lock(&uvm.pagedaemon_lock);
367 			wakeup(&uvmexp.free);
368 			simple_unlock(&uvm.pagedaemon_lock);
369 		}
370 	}
371 }
372 
373 /*
374  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
375  *
376  * => called with page queues locked
377  * => we work on meeting our free target by converting inactive pages
378  *    into free pages.
379  * => we handle the building of swap-backed clusters
380  * => we return TRUE if we are exiting because we met our target
381  */
382 
383 void
384 uvmpd_scan_inactive(pglst)
385 	struct pglist *pglst;
386 {
387 	int error;
388 	struct vm_page *p, *nextpg = NULL; /* Quell compiler warning */
389 	struct uvm_object *uobj;
390 	struct vm_anon *anon;
391 	struct vm_page *swpps[round_page(MAXPHYS) >> PAGE_SHIFT];
392 	struct simplelock *slock;
393 	int swnpages, swcpages;
394 	int swslot;
395 	int dirtyreacts, t, result;
396 	boolean_t anonunder, fileunder, execunder;
397 	boolean_t anonover, fileover, execover;
398 	boolean_t anonreact, filereact, execreact;
399 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
400 
401 	/*
402 	 * swslot is non-zero if we are building a swap cluster.  we want
403 	 * to stay in the loop while we have a page to scan or we have
404 	 * a swap-cluster to build.
405 	 */
406 
407 	swslot = 0;
408 	swnpages = swcpages = 0;
409 	dirtyreacts = 0;
410 
411 	/*
412 	 * decide which types of pages we want to reactivate instead of freeing
413 	 * to keep usage within the minimum and maximum usage limits.
414 	 */
415 
416 	t = uvmexp.active + uvmexp.inactive + uvmexp.free;
417 	anonunder = (uvmexp.anonpages <= (t * uvmexp.anonmin) >> 8);
418 	fileunder = (uvmexp.filepages <= (t * uvmexp.filemin) >> 8);
419 	execunder = (uvmexp.execpages <= (t * uvmexp.execmin) >> 8);
420 	anonover = uvmexp.anonpages > ((t * uvmexp.anonmax) >> 8);
421 	fileover = uvmexp.filepages > ((t * uvmexp.filemax) >> 8);
422 	execover = uvmexp.execpages > ((t * uvmexp.execmax) >> 8);
423 	anonreact = anonunder || (!anonover && (fileover || execover));
424 	filereact = fileunder || (!fileover && (anonover || execover));
425 	execreact = execunder || (!execover && (anonover || fileover));
426 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
427 		uobj = NULL;
428 		anon = NULL;
429 		if (p) {
430 
431 			/*
432 			 * see if we've met the free target.
433 			 */
434 
435 			if (uvmexp.free + uvmexp.paging >=
436 			    uvmexp.freetarg << 2 ||
437 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
438 				UVMHIST_LOG(pdhist,"  met free target: "
439 					    "exit loop", 0, 0, 0, 0);
440 
441 				if (swslot == 0) {
442 					/* exit now if no swap-i/o pending */
443 					break;
444 				}
445 
446 				/* set p to null to signal final swap i/o */
447 				p = NULL;
448 				nextpg = NULL;
449 			}
450 		}
451 		if (p) {	/* if (we have a new page to consider) */
452 
453 			/*
454 			 * we are below target and have a new page to consider.
455 			 */
456 
457 			uvmexp.pdscans++;
458 			nextpg = TAILQ_NEXT(p, pageq);
459 
460 			/*
461 			 * move referenced pages back to active queue and
462 			 * skip to next page.
463 			 */
464 
465 			if (pmap_clear_reference(p)) {
466 				uvm_pageactivate(p);
467 				uvmexp.pdreact++;
468 				continue;
469 			}
470 			anon = p->uanon;
471 			uobj = p->uobject;
472 
473 			/*
474 			 * enforce the minimum thresholds on different
475 			 * types of memory usage.  if reusing the current
476 			 * page would reduce that type of usage below its
477 			 * minimum, reactivate the page instead and move
478 			 * on to the next page.
479 			 */
480 
481 			if (uobj && UVM_OBJ_IS_VTEXT(uobj) && execreact) {
482 				uvm_pageactivate(p);
483 				uvmexp.pdreexec++;
484 				continue;
485 			}
486 			if (uobj && UVM_OBJ_IS_VNODE(uobj) &&
487 			    !UVM_OBJ_IS_VTEXT(uobj) && filereact) {
488 				uvm_pageactivate(p);
489 				uvmexp.pdrefile++;
490 				continue;
491 			}
492 			if ((anon || UVM_OBJ_IS_AOBJ(uobj)) && anonreact) {
493 				uvm_pageactivate(p);
494 				uvmexp.pdreanon++;
495 				continue;
496 			}
497 
498 			/*
499 			 * first we attempt to lock the object that this page
500 			 * belongs to.  if our attempt fails we skip on to
501 			 * the next page (no harm done).  it is important to
502 			 * "try" locking the object as we are locking in the
503 			 * wrong order (pageq -> object) and we don't want to
504 			 * deadlock.
505 			 *
506 			 * the only time we expect to see an ownerless page
507 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
508 			 * anon has loaned a page from a uvm_object and the
509 			 * uvm_object has dropped the ownership.  in that
510 			 * case, the anon can "take over" the loaned page
511 			 * and make it its own.
512 			 */
513 
514 			/* does the page belong to an object? */
515 			if (uobj != NULL) {
516 				slock = &uobj->vmobjlock;
517 				if (!simple_lock_try(slock)) {
518 					continue;
519 				}
520 				if (p->flags & PG_BUSY) {
521 					simple_unlock(slock);
522 					uvmexp.pdbusy++;
523 					continue;
524 				}
525 				uvmexp.pdobscan++;
526 			} else {
527 				KASSERT(anon != NULL);
528 				slock = &anon->an_lock;
529 				if (!simple_lock_try(slock)) {
530 					continue;
531 				}
532 
533 				/*
534 				 * set PQ_ANON if it isn't set already.
535 				 */
536 
537 				if ((p->pqflags & PQ_ANON) == 0) {
538 					KASSERT(p->loan_count > 0);
539 					p->loan_count--;
540 					p->pqflags |= PQ_ANON;
541 					/* anon now owns it */
542 				}
543 				if (p->flags & PG_BUSY) {
544 					simple_unlock(slock);
545 					uvmexp.pdbusy++;
546 					continue;
547 				}
548 				uvmexp.pdanscan++;
549 			}
550 
551 
552 			/*
553 			 * we now have the object and the page queues locked.
554 			 * if the page is not swap-backed, call the object's
555 			 * pager to flush and free the page.
556 			 */
557 
558 			if ((p->pqflags & PQ_SWAPBACKED) == 0) {
559 				uvm_unlock_pageq();
560 				(void) (uobj->pgops->pgo_put)(uobj, p->offset,
561 				    p->offset + PAGE_SIZE,
562 				    PGO_CLEANIT|PGO_FREE);
563 				uvm_lock_pageq();
564 				if (nextpg &&
565 				    (nextpg->pqflags & PQ_INACTIVE) == 0) {
566 					nextpg = TAILQ_FIRST(pglst);
567 				}
568 				continue;
569 			}
570 
571 			/*
572 			 * the page is swap-backed.  remove all the permissions
573 			 * from the page so we can sync the modified info
574 			 * without any race conditions.  if the page is clean
575 			 * we can free it now and continue.
576 			 */
577 
578 			pmap_page_protect(p, VM_PROT_NONE);
579 			if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
580 				p->flags &= ~(PG_CLEAN);
581 			}
582 			if (p->flags & PG_CLEAN) {
583 				int slot;
584 				int pageidx;
585 
586 				pageidx = p->offset >> PAGE_SHIFT;
587 				uvm_pagefree(p);
588 				uvmexp.pdfreed++;
589 
590 				/*
591 				 * for anons, we need to remove the page
592 				 * from the anon ourselves.  for aobjs,
593 				 * pagefree did that for us.
594 				 */
595 
596 				if (anon) {
597 					KASSERT(anon->an_swslot != 0);
598 					anon->u.an_page = NULL;
599 					slot = anon->an_swslot;
600 				} else {
601 					slot = uao_find_swslot(uobj, pageidx);
602 				}
603 				simple_unlock(slock);
604 
605 				if (slot > 0) {
606 					/* this page is now only in swap. */
607 					simple_lock(&uvm.swap_data_lock);
608 					KASSERT(uvmexp.swpgonly <
609 						uvmexp.swpginuse);
610 					uvmexp.swpgonly++;
611 					simple_unlock(&uvm.swap_data_lock);
612 				}
613 				continue;
614 			}
615 
616 			/*
617 			 * this page is dirty, skip it if we'll have met our
618 			 * free target when all the current pageouts complete.
619 			 */
620 
621 			if (uvmexp.free + uvmexp.paging >
622 			    uvmexp.freetarg << 2) {
623 				simple_unlock(slock);
624 				continue;
625 			}
626 
627 			/*
628 			 * free any swap space allocated to the page since
629 			 * we'll have to write it again with its new data.
630 			 */
631 
632 			if ((p->pqflags & PQ_ANON) && anon->an_swslot) {
633 				uvm_swap_free(anon->an_swslot, 1);
634 				anon->an_swslot = 0;
635 			} else if (p->pqflags & PQ_AOBJ) {
636 				uao_dropswap(uobj, p->offset >> PAGE_SHIFT);
637 			}
638 
639 			/*
640 			 * if all pages in swap are only in swap,
641 			 * the swap space is full and we can't page out
642 			 * any more swap-backed pages.  reactivate this page
643 			 * so that we eventually cycle all pages through
644 			 * the inactive queue.
645 			 */
646 
647 			if (uvm_swapisfull()) {
648 				dirtyreacts++;
649 				uvm_pageactivate(p);
650 				simple_unlock(slock);
651 				continue;
652 			}
653 
654 			/*
655 			 * start new swap pageout cluster (if necessary).
656 			 */
657 
658 			if (swslot == 0) {
659 				/* Even with strange MAXPHYS, the shift
660 				   implicitly rounds down to a page. */
661 				swnpages = MAXPHYS >> PAGE_SHIFT;
662 				swslot = uvm_swap_alloc(&swnpages, TRUE);
663 				if (swslot == 0) {
664 					simple_unlock(slock);
665 					continue;
666 				}
667 				swcpages = 0;
668 			}
669 
670 			/*
671 			 * at this point, we're definitely going reuse this
672 			 * page.  mark the page busy and delayed-free.
673 			 * we should remove the page from the page queues
674 			 * so we don't ever look at it again.
675 			 * adjust counters and such.
676 			 */
677 
678 			p->flags |= PG_BUSY;
679 			UVM_PAGE_OWN(p, "scan_inactive");
680 
681 			p->flags |= PG_PAGEOUT;
682 			uvmexp.paging++;
683 			uvm_pagedequeue(p);
684 
685 			uvmexp.pgswapout++;
686 
687 			/*
688 			 * add the new page to the cluster.
689 			 */
690 
691 			if (anon) {
692 				anon->an_swslot = swslot + swcpages;
693 				simple_unlock(slock);
694 			} else {
695 				result = uao_set_swslot(uobj,
696 				    p->offset >> PAGE_SHIFT, swslot + swcpages);
697 				if (result == -1) {
698 					p->flags &= ~(PG_BUSY|PG_PAGEOUT);
699 					UVM_PAGE_OWN(p, NULL);
700 					uvmexp.paging--;
701 					uvm_pageactivate(p);
702 					simple_unlock(slock);
703 					continue;
704 				}
705 				simple_unlock(slock);
706 			}
707 			swpps[swcpages] = p;
708 			swcpages++;
709 
710 			/*
711 			 * if the cluster isn't full, look for more pages
712 			 * before starting the i/o.
713 			 */
714 
715 			if (swcpages < swnpages) {
716 				continue;
717 			}
718 		}
719 
720 		/*
721 		 * if this is the final pageout we could have a few
722 		 * unused swap blocks.  if so, free them now.
723 		 */
724 
725 		if (swcpages < swnpages) {
726 			uvm_swap_free(swslot + swcpages, (swnpages - swcpages));
727 		}
728 
729 		/*
730 		 * now start the pageout.
731 		 */
732 
733 		uvm_unlock_pageq();
734 		uvmexp.pdpageouts++;
735 		error = uvm_swap_put(swslot, swpps, swcpages, 0);
736 		KASSERT(error == 0);
737 		uvm_lock_pageq();
738 
739 		/*
740 		 * zero swslot to indicate that we are
741 		 * no longer building a swap-backed cluster.
742 		 */
743 
744 		swslot = 0;
745 
746 		/*
747 		 * the pageout is in progress.  bump counters and set up
748 		 * for the next loop.
749 		 */
750 
751 		uvmexp.pdpending++;
752 		if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) {
753 			nextpg = TAILQ_FIRST(pglst);
754 		}
755 	}
756 }
757 
758 /*
759  * uvmpd_scan: scan the page queues and attempt to meet our targets.
760  *
761  * => called with pageq's locked
762  */
763 
764 void
765 uvmpd_scan(void)
766 {
767 	int inactive_shortage, swap_shortage, pages_freed;
768 	struct vm_page *p, *nextpg;
769 	struct uvm_object *uobj;
770 	struct vm_anon *anon;
771 	struct simplelock *slock;
772 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
773 
774 	uvmexp.pdrevs++;
775 	uobj = NULL;
776 	anon = NULL;
777 
778 #ifndef __SWAP_BROKEN
779 
780 	/*
781 	 * swap out some processes if we are below our free target.
782 	 * we need to unlock the page queues for this.
783 	 */
784 
785 	if (uvmexp.free < uvmexp.freetarg && uvmexp.nswapdev != 0) {
786 		uvmexp.pdswout++;
787 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout",
788 		    uvmexp.free, uvmexp.freetarg, 0, 0);
789 		uvm_unlock_pageq();
790 		uvm_swapout_threads();
791 		uvm_lock_pageq();
792 
793 	}
794 #endif
795 
796 	/*
797 	 * now we want to work on meeting our targets.   first we work on our
798 	 * free target by converting inactive pages into free pages.  then
799 	 * we work on meeting our inactive target by converting active pages
800 	 * to inactive ones.
801 	 */
802 
803 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
804 
805 	pages_freed = uvmexp.pdfreed;
806 	uvmpd_scan_inactive(&uvm.page_inactive);
807 	pages_freed = uvmexp.pdfreed - pages_freed;
808 
809 	/*
810 	 * we have done the scan to get free pages.   now we work on meeting
811 	 * our inactive target.
812 	 */
813 
814 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
815 
816 	/*
817 	 * detect if we're not going to be able to page anything out
818 	 * until we free some swap resources from active pages.
819 	 */
820 
821 	swap_shortage = 0;
822 	if (uvmexp.free < uvmexp.freetarg &&
823 	    uvmexp.swpginuse >= uvmexp.swpgavail &&
824 	    !uvm_swapisfull() &&
825 	    pages_freed == 0) {
826 		swap_shortage = uvmexp.freetarg - uvmexp.free;
827 	}
828 
829 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
830 		    inactive_shortage, swap_shortage,0,0);
831 	for (p = TAILQ_FIRST(&uvm.page_active);
832 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
833 	     p = nextpg) {
834 		nextpg = TAILQ_NEXT(p, pageq);
835 		if (p->flags & PG_BUSY) {
836 			continue;
837 		}
838 
839 		/*
840 		 * lock the page's owner.
841 		 */
842 
843 		if (p->uobject != NULL) {
844 			uobj = p->uobject;
845 			slock = &uobj->vmobjlock;
846 			if (!simple_lock_try(slock)) {
847 				continue;
848 			}
849 		} else {
850 			anon = p->uanon;
851 			KASSERT(anon != NULL);
852 			slock = &anon->an_lock;
853 			if (!simple_lock_try(slock)) {
854 				continue;
855 			}
856 
857 			/* take over the page? */
858 			if ((p->pqflags & PQ_ANON) == 0) {
859 				KASSERT(p->loan_count > 0);
860 				p->loan_count--;
861 				p->pqflags |= PQ_ANON;
862 			}
863 		}
864 
865 		/*
866 		 * skip this page if it's busy.
867 		 */
868 
869 		if ((p->flags & PG_BUSY) != 0) {
870 			simple_unlock(slock);
871 			continue;
872 		}
873 
874 		/*
875 		 * if there's a shortage of swap, free any swap allocated
876 		 * to this page so that other pages can be paged out.
877 		 */
878 
879 		if (swap_shortage > 0) {
880 			if ((p->pqflags & PQ_ANON) && anon->an_swslot) {
881 				uvm_swap_free(anon->an_swslot, 1);
882 				anon->an_swslot = 0;
883 				p->flags &= ~PG_CLEAN;
884 				swap_shortage--;
885 			} else if (p->pqflags & PQ_AOBJ) {
886 				int slot = uao_set_swslot(uobj,
887 					p->offset >> PAGE_SHIFT, 0);
888 				if (slot) {
889 					uvm_swap_free(slot, 1);
890 					p->flags &= ~PG_CLEAN;
891 					swap_shortage--;
892 				}
893 			}
894 		}
895 
896 		/*
897 		 * if there's a shortage of inactive pages, deactivate.
898 		 */
899 
900 		if (inactive_shortage > 0) {
901 			/* no need to check wire_count as pg is "active" */
902 			uvm_pagedeactivate(p);
903 			uvmexp.pddeact++;
904 			inactive_shortage--;
905 		}
906 
907 		/*
908 		 * we're done with this page.
909 		 */
910 
911 		simple_unlock(slock);
912 	}
913 }
914