xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 06be8101a16cc95f40783b3cb7afd12112103a9a)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.42 2001/11/10 07:37:01 lukem 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 1.42 2001/11/10 07:37:01 lukem 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 pagedeamon 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 __P((void));
103 boolean_t	uvmpd_scan_inactive __P((struct pglist *));
104 void		uvmpd_tune __P((void));
105 
106 /*
107  * uvm_wait: wait (sleep) for the page daemon to free some pages
108  *
109  * => should be called with all locks released
110  * => should _not_ be called by the page daemon (to avoid deadlock)
111  */
112 
113 void
114 uvm_wait(wmsg)
115 	const char *wmsg;
116 {
117 	int timo = 0;
118 	int s = splbio();
119 
120 	/*
121 	 * check for page daemon going to sleep (waiting for itself)
122 	 */
123 
124 	if (curproc == uvm.pagedaemon_proc && uvmexp.paging == 0) {
125 		/*
126 		 * now we have a problem: the pagedaemon wants to go to
127 		 * sleep until it frees more memory.   but how can it
128 		 * free more memory if it is asleep?  that is a deadlock.
129 		 * we have two options:
130 		 *  [1] panic now
131 		 *  [2] put a timeout on the sleep, thus causing the
132 		 *      pagedaemon to only pause (rather than sleep forever)
133 		 *
134 		 * note that option [2] will only help us if we get lucky
135 		 * and some other process on the system breaks the deadlock
136 		 * by exiting or freeing memory (thus allowing the pagedaemon
137 		 * to continue).  for now we panic if DEBUG is defined,
138 		 * otherwise we hope for the best with option [2] (better
139 		 * yet, this should never happen in the first place!).
140 		 */
141 
142 		printf("pagedaemon: deadlock detected!\n");
143 		timo = hz >> 3;		/* set timeout */
144 #if defined(DEBUG)
145 		/* DEBUG: panic so we can debug it */
146 		panic("pagedaemon deadlock");
147 #endif
148 	}
149 
150 	simple_lock(&uvm.pagedaemon_lock);
151 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
152 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, wmsg,
153 	    timo);
154 
155 	splx(s);
156 }
157 
158 
159 /*
160  * uvmpd_tune: tune paging parameters
161  *
162  * => called when ever memory is added (or removed?) to the system
163  * => caller must call with page queues locked
164  */
165 
166 void
167 uvmpd_tune(void)
168 {
169 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
170 
171 	uvmexp.freemin = uvmexp.npages / 20;
172 
173 	/* between 16k and 256k */
174 	/* XXX:  what are these values good for? */
175 	uvmexp.freemin = MAX(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
176 	uvmexp.freemin = MIN(uvmexp.freemin, (256*1024) >> PAGE_SHIFT);
177 
178 	/* Make sure there's always a user page free. */
179 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
180 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
181 
182 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
183 	if (uvmexp.freetarg <= uvmexp.freemin)
184 		uvmexp.freetarg = uvmexp.freemin + 1;
185 
186 	/* uvmexp.inactarg: computed in main daemon loop */
187 
188 	uvmexp.wiredmax = uvmexp.npages / 3;
189 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
190 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
191 }
192 
193 /*
194  * uvm_pageout: the main loop for the pagedaemon
195  */
196 
197 void
198 uvm_pageout(void *arg)
199 {
200 	int npages = 0;
201 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
202 
203 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
204 
205 	/*
206 	 * ensure correct priority and set paging parameters...
207 	 */
208 
209 	uvm.pagedaemon_proc = curproc;
210 	uvm_lock_pageq();
211 	npages = uvmexp.npages;
212 	uvmpd_tune();
213 	uvm_unlock_pageq();
214 
215 	/*
216 	 * main loop
217 	 */
218 
219 	for (;;) {
220 		simple_lock(&uvm.pagedaemon_lock);
221 
222 		UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
223 		UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
224 		    &uvm.pagedaemon_lock, FALSE, "pgdaemon", 0);
225 		uvmexp.pdwoke++;
226 		UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
227 
228 		/*
229 		 * now lock page queues and recompute inactive count
230 		 */
231 
232 		uvm_lock_pageq();
233 		if (npages != uvmexp.npages) {	/* check for new pages? */
234 			npages = uvmexp.npages;
235 			uvmpd_tune();
236 		}
237 
238 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
239 		if (uvmexp.inactarg <= uvmexp.freetarg) {
240 			uvmexp.inactarg = uvmexp.freetarg + 1;
241 		}
242 
243 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d, inact/itarg=%d/%d",
244 		    uvmexp.free, uvmexp.freetarg, uvmexp.inactive,
245 		    uvmexp.inactarg);
246 
247 		/*
248 		 * scan if needed
249 		 */
250 
251 		if (uvmexp.free + uvmexp.paging < uvmexp.freetarg ||
252 		    uvmexp.inactive < uvmexp.inactarg) {
253 			uvmpd_scan();
254 		}
255 
256 		/*
257 		 * if there's any free memory to be had,
258 		 * wake up any waiters.
259 		 */
260 
261 		if (uvmexp.free > uvmexp.reserve_kernel ||
262 		    uvmexp.paging == 0) {
263 			wakeup(&uvmexp.free);
264 		}
265 
266 		/*
267 		 * scan done.  unlock page queues (the only lock we are holding)
268 		 */
269 
270 		uvm_unlock_pageq();
271 
272 		/*
273 		 * drain pool resources now that we're not holding any locks
274 		 */
275 
276 		pool_drain(0);
277 	}
278 	/*NOTREACHED*/
279 }
280 
281 
282 /*
283  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
284  */
285 
286 void
287 uvm_aiodone_daemon(void *arg)
288 {
289 	int s, free;
290 	struct buf *bp, *nbp;
291 	UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist);
292 
293 	for (;;) {
294 
295 		/*
296 		 * carefully attempt to go to sleep (without losing "wakeups"!).
297 		 * we need splbio because we want to make sure the aio_done list
298 		 * is totally empty before we go to sleep.
299 		 */
300 
301 		s = splbio();
302 		simple_lock(&uvm.aiodoned_lock);
303 		if (TAILQ_FIRST(&uvm.aio_done) == NULL) {
304 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
305 			UVM_UNLOCK_AND_WAIT(&uvm.aiodoned,
306 			    &uvm.aiodoned_lock, FALSE, "aiodoned", 0);
307 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
308 
309 			/* relock aiodoned_lock, still at splbio */
310 			simple_lock(&uvm.aiodoned_lock);
311 		}
312 
313 		/*
314 		 * check for done aio structures
315 		 */
316 
317 		bp = TAILQ_FIRST(&uvm.aio_done);
318 		if (bp) {
319 			TAILQ_INIT(&uvm.aio_done);
320 		}
321 
322 		simple_unlock(&uvm.aiodoned_lock);
323 		splx(s);
324 
325 		/*
326 		 * process each i/o that's done.
327 		 */
328 
329 		free = uvmexp.free;
330 		while (bp != NULL) {
331 			nbp = TAILQ_NEXT(bp, b_freelist);
332 			(*bp->b_iodone)(bp);
333 			bp = nbp;
334 		}
335 		if (free <= uvmexp.reserve_kernel) {
336 			s = uvm_lock_fpageq();
337 			wakeup(&uvm.pagedaemon);
338 			uvm_unlock_fpageq(s);
339 		} else {
340 			simple_lock(&uvm.pagedaemon_lock);
341 			wakeup(&uvmexp.free);
342 			simple_unlock(&uvm.pagedaemon_lock);
343 		}
344 	}
345 }
346 
347 /*
348  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
349  *
350  * => called with page queues locked
351  * => we work on meeting our free target by converting inactive pages
352  *    into free pages.
353  * => we handle the building of swap-backed clusters
354  * => we return TRUE if we are exiting because we met our target
355  */
356 
357 boolean_t
358 uvmpd_scan_inactive(pglst)
359 	struct pglist *pglst;
360 {
361 	int error;
362 	struct vm_page *p, *nextpg;
363 	struct uvm_object *uobj;
364 	struct vm_anon *anon;
365 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT];
366 	struct simplelock *slock;
367 	int swnpages, swcpages;
368 	int swslot;
369 	int dirtyreacts, t, result;
370 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
371 
372 	/*
373 	 * swslot is non-zero if we are building a swap cluster.  we want
374 	 * to stay in the loop while we have a page to scan or we have
375 	 * a swap-cluster to build.
376 	 */
377 
378 	swslot = 0;
379 	swnpages = swcpages = 0;
380 	dirtyreacts = 0;
381 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
382 		uobj = NULL;
383 		anon = NULL;
384 		if (p) {
385 
386 			/*
387 			 * see if we've met the free target.
388 			 */
389 
390 			if (uvmexp.free + uvmexp.paging >=
391 			    uvmexp.freetarg << 2 ||
392 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
393 				UVMHIST_LOG(pdhist,"  met free target: "
394 					    "exit loop", 0, 0, 0, 0);
395 
396 				if (swslot == 0) {
397 					/* exit now if no swap-i/o pending */
398 					break;
399 				}
400 
401 				/* set p to null to signal final swap i/o */
402 				p = NULL;
403 				nextpg = NULL;
404 			}
405 		}
406 		if (p) {	/* if (we have a new page to consider) */
407 
408 			/*
409 			 * we are below target and have a new page to consider.
410 			 */
411 
412 			uvmexp.pdscans++;
413 			nextpg = TAILQ_NEXT(p, pageq);
414 
415 			/*
416 			 * move referenced pages back to active queue and
417 			 * skip to next page.
418 			 */
419 
420 			if (pmap_clear_reference(p)) {
421 				uvm_pageactivate(p);
422 				uvmexp.pdreact++;
423 				continue;
424 			}
425 			anon = p->uanon;
426 			uobj = p->uobject;
427 
428 			/*
429 			 * enforce the minimum thresholds on different
430 			 * types of memory usage.  if reusing the current
431 			 * page would reduce that type of usage below its
432 			 * minimum, reactivate the page instead and move
433 			 * on to the next page.
434 			 */
435 
436 			t = uvmexp.active + uvmexp.inactive + uvmexp.free;
437 			if (anon &&
438 			    uvmexp.anonpages <= (t * uvmexp.anonmin) >> 8) {
439 				uvm_pageactivate(p);
440 				uvmexp.pdreanon++;
441 				continue;
442 			}
443 			if (uobj && UVM_OBJ_IS_VTEXT(uobj) &&
444 			    uvmexp.vtextpages <= (t * uvmexp.vtextmin) >> 8) {
445 				uvm_pageactivate(p);
446 				uvmexp.pdrevtext++;
447 				continue;
448 			}
449 			if (uobj && UVM_OBJ_IS_VNODE(uobj) &&
450 			    !UVM_OBJ_IS_VTEXT(uobj) &&
451 			    uvmexp.vnodepages <= (t * uvmexp.vnodemin) >> 8) {
452 				uvm_pageactivate(p);
453 				uvmexp.pdrevnode++;
454 				continue;
455 			}
456 
457 			/*
458 			 * first we attempt to lock the object that this page
459 			 * belongs to.  if our attempt fails we skip on to
460 			 * the next page (no harm done).  it is important to
461 			 * "try" locking the object as we are locking in the
462 			 * wrong order (pageq -> object) and we don't want to
463 			 * deadlock.
464 			 *
465 			 * the only time we expect to see an ownerless page
466 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
467 			 * anon has loaned a page from a uvm_object and the
468 			 * uvm_object has dropped the ownership.  in that
469 			 * case, the anon can "take over" the loaned page
470 			 * and make it its own.
471 			 */
472 
473 			/* is page part of an anon or ownerless ? */
474 			if ((p->pqflags & PQ_ANON) || uobj == NULL) {
475 				KASSERT(anon != NULL);
476 				slock = &anon->an_lock;
477 				if (!simple_lock_try(slock)) {
478 					/* lock failed, skip this page */
479 					continue;
480 				}
481 
482 				/*
483 				 * if the page is ownerless, claim it in the
484 				 * name of "anon"!
485 				 */
486 
487 				if ((p->pqflags & PQ_ANON) == 0) {
488 					KASSERT(p->loan_count > 0);
489 					p->loan_count--;
490 					p->pqflags |= PQ_ANON;
491 					/* anon now owns it */
492 				}
493 				if (p->flags & PG_BUSY) {
494 					simple_unlock(slock);
495 					uvmexp.pdbusy++;
496 					continue;
497 				}
498 				uvmexp.pdanscan++;
499 			} else {
500 				KASSERT(uobj != NULL);
501 				slock = &uobj->vmobjlock;
502 				if (!simple_lock_try(slock)) {
503 					continue;
504 				}
505 				if (p->flags & PG_BUSY) {
506 					simple_unlock(slock);
507 					uvmexp.pdbusy++;
508 					continue;
509 				}
510 				uvmexp.pdobscan++;
511 			}
512 
513 
514 			/*
515 			 * we now have the object and the page queues locked.
516 			 * if the page is not swap-backed, call the object's
517 			 * pager to flush and free the page.
518 			 */
519 
520 			if ((p->pqflags & PQ_SWAPBACKED) == 0) {
521 				uvm_unlock_pageq();
522 				error = (uobj->pgops->pgo_put)(uobj, p->offset,
523 				    p->offset + PAGE_SIZE,
524 				    PGO_CLEANIT|PGO_FREE);
525 				uvm_lock_pageq();
526 				if (nextpg &&
527 				    (nextpg->flags & PQ_INACTIVE) == 0) {
528 					nextpg = TAILQ_FIRST(pglst);
529 				}
530 				continue;
531 			}
532 
533 			/*
534 			 * the page is swap-backed.  remove all the permissions
535 			 * from the page so we can sync the modified info
536 			 * without any race conditions.  if the page is clean
537 			 * we can free it now and continue.
538 			 */
539 
540 			pmap_page_protect(p, VM_PROT_NONE);
541 			if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
542 				p->flags &= ~(PG_CLEAN);
543 			}
544 			if (p->flags & PG_CLEAN) {
545 				uvm_pagefree(p);
546 				uvmexp.pdfreed++;
547 
548 				/*
549 				 * for anons, we need to remove the page
550 				 * from the anon ourselves.  for aobjs,
551 				 * pagefree did that for us.
552 				 */
553 
554 				if (anon) {
555 					KASSERT(anon->an_swslot != 0);
556 					anon->u.an_page = NULL;
557 				}
558 				simple_unlock(slock);
559 
560 				/* this page is now only in swap. */
561 				simple_lock(&uvm.swap_data_lock);
562 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
563 				uvmexp.swpgonly++;
564 				simple_unlock(&uvm.swap_data_lock);
565 				continue;
566 			}
567 
568 			/*
569 			 * this page is dirty, skip it if we'll have met our
570 			 * free target when all the current pageouts complete.
571 			 */
572 
573 			if (uvmexp.free + uvmexp.paging >
574 			    uvmexp.freetarg << 2) {
575 				simple_unlock(slock);
576 				continue;
577 			}
578 
579 			/*
580 			 * free any swap space allocated to the page since
581 			 * we'll have to write it again with its new data.
582 			 */
583 
584 			if ((p->pqflags & PQ_ANON) && anon->an_swslot) {
585 				uvm_swap_free(anon->an_swslot, 1);
586 				anon->an_swslot = 0;
587 			} else if (p->pqflags & PQ_AOBJ) {
588 				uao_dropswap(uobj, p->offset >> PAGE_SHIFT);
589 			}
590 
591 			/*
592 			 * if all pages in swap are only in swap,
593 			 * the swap space is full and we can't page out
594 			 * any more swap-backed pages.  reactivate this page
595 			 * so that we eventually cycle all pages through
596 			 * the inactive queue.
597 			 */
598 
599 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
600 			if (uvmexp.swpgonly == uvmexp.swpages) {
601 				dirtyreacts++;
602 				uvm_pageactivate(p);
603 				simple_unlock(slock);
604 				continue;
605 			}
606 
607 			/*
608 			 * start new swap pageout cluster (if necessary).
609 			 */
610 
611 			if (swslot == 0) {
612 				swnpages = MAXBSIZE >> PAGE_SHIFT;
613 				swslot = uvm_swap_alloc(&swnpages, TRUE);
614 				if (swslot == 0) {
615 					simple_unlock(slock);
616 					continue;
617 				}
618 				swcpages = 0;
619 			}
620 
621 			/*
622 			 * at this point, we're definitely going reuse this
623 			 * page.  mark the page busy and delayed-free.
624 			 * we should remove the page from the page queues
625 			 * so we don't ever look at it again.
626 			 * adjust counters and such.
627 			 */
628 
629 			p->flags |= PG_BUSY;
630 			UVM_PAGE_OWN(p, "scan_inactive");
631 
632 			p->flags |= PG_PAGEOUT;
633 			uvmexp.paging++;
634 			uvm_pagedequeue(p);
635 
636 			uvmexp.pgswapout++;
637 
638 			/*
639 			 * add the new page to the cluster.
640 			 */
641 
642 			if (anon) {
643 				anon->an_swslot = swslot + swcpages;
644 				simple_unlock(slock);
645 			} else {
646 				result = uao_set_swslot(uobj,
647 				    p->offset >> PAGE_SHIFT, swslot + swcpages);
648 				if (result == -1) {
649 					p->flags &= ~(PG_BUSY|PG_PAGEOUT);
650 					UVM_PAGE_OWN(p, NULL);
651 					uvmexp.paging--;
652 					uvm_pageactivate(p);
653 					simple_unlock(slock);
654 					continue;
655 				}
656 				simple_unlock(slock);
657 			}
658 			swpps[swcpages] = p;
659 			swcpages++;
660 
661 			/*
662 			 * if the cluster isn't full, look for more pages
663 			 * before starting the i/o.
664 			 */
665 
666 			if (swcpages < swnpages) {
667 				continue;
668 			}
669 		}
670 
671 		/*
672 		 * if this is the final pageout we could have a few
673 		 * unused swap blocks.  if so, free them now.
674 		 */
675 
676 		if (swcpages < swnpages) {
677 			uvm_swap_free(swslot + swcpages, (swnpages - swcpages));
678 		}
679 
680 		/*
681 		 * now start the pageout.
682 		 */
683 
684 		uvm_unlock_pageq();
685 		uvmexp.pdpageouts++;
686 		error = uvm_swap_put(swslot, swpps, swcpages, 0);
687 		KASSERT(error == 0);
688 		uvm_lock_pageq();
689 
690 		/*
691 		 * zero swslot to indicate that we are
692 		 * no longer building a swap-backed cluster.
693 		 */
694 
695 		swslot = 0;
696 
697 		/*
698 		 * the pageout is in progress.  bump counters and set up
699 		 * for the next loop.
700 		 */
701 
702 		uvmexp.pdpending++;
703 		if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) {
704 			nextpg = TAILQ_FIRST(pglst);
705 		}
706 	}
707 	return (error);
708 }
709 
710 /*
711  * uvmpd_scan: scan the page queues and attempt to meet our targets.
712  *
713  * => called with pageq's locked
714  */
715 
716 void
717 uvmpd_scan(void)
718 {
719 	int inactive_shortage, swap_shortage, pages_freed;
720 	struct vm_page *p, *nextpg;
721 	struct uvm_object *uobj;
722 	struct vm_anon *anon;
723 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
724 
725 	uvmexp.pdrevs++;
726 	uobj = NULL;
727 	anon = NULL;
728 
729 #ifndef __SWAP_BROKEN
730 
731 	/*
732 	 * swap out some processes if we are below our free target.
733 	 * we need to unlock the page queues for this.
734 	 */
735 
736 	if (uvmexp.free < uvmexp.freetarg && uvmexp.nswapdev != 0) {
737 		uvmexp.pdswout++;
738 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout",
739 		    uvmexp.free, uvmexp.freetarg, 0, 0);
740 		uvm_unlock_pageq();
741 		uvm_swapout_threads();
742 		uvm_lock_pageq();
743 
744 	}
745 #endif
746 
747 	/*
748 	 * now we want to work on meeting our targets.   first we work on our
749 	 * free target by converting inactive pages into free pages.  then
750 	 * we work on meeting our inactive target by converting active pages
751 	 * to inactive ones.
752 	 */
753 
754 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
755 
756 	/*
757 	 * alternate starting queue between swap and object based on the
758 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
759 	 */
760 
761 	pages_freed = uvmexp.pdfreed;
762 	(void) uvmpd_scan_inactive(&uvm.page_inactive);
763 	pages_freed = uvmexp.pdfreed - pages_freed;
764 
765 	/*
766 	 * we have done the scan to get free pages.   now we work on meeting
767 	 * our inactive target.
768 	 */
769 
770 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
771 
772 	/*
773 	 * detect if we're not going to be able to page anything out
774 	 * until we free some swap resources from active pages.
775 	 */
776 
777 	swap_shortage = 0;
778 	if (uvmexp.free < uvmexp.freetarg &&
779 	    uvmexp.swpginuse == uvmexp.swpages &&
780 	    uvmexp.swpgonly < uvmexp.swpages &&
781 	    pages_freed == 0) {
782 		swap_shortage = uvmexp.freetarg - uvmexp.free;
783 	}
784 
785 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
786 		    inactive_shortage, swap_shortage,0,0);
787 	for (p = TAILQ_FIRST(&uvm.page_active);
788 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
789 	     p = nextpg) {
790 		nextpg = TAILQ_NEXT(p, pageq);
791 		if (p->flags & PG_BUSY) {
792 			continue;
793 		}
794 
795 		/*
796 		 * lock the page's owner.
797 		 */
798 		/* is page anon owned or ownerless? */
799 		if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
800 			anon = p->uanon;
801 			KASSERT(anon != NULL);
802 			if (!simple_lock_try(&anon->an_lock)) {
803 				continue;
804 			}
805 
806 			/* take over the page? */
807 			if ((p->pqflags & PQ_ANON) == 0) {
808 				KASSERT(p->loan_count > 0);
809 				p->loan_count--;
810 				p->pqflags |= PQ_ANON;
811 			}
812 		} else {
813 			uobj = p->uobject;
814 			if (!simple_lock_try(&uobj->vmobjlock)) {
815 				continue;
816 			}
817 		}
818 
819 		/*
820 		 * skip this page if it's busy.
821 		 */
822 
823 		if ((p->flags & PG_BUSY) != 0) {
824 			if (p->pqflags & PQ_ANON)
825 				simple_unlock(&anon->an_lock);
826 			else
827 				simple_unlock(&uobj->vmobjlock);
828 			continue;
829 		}
830 
831 		/*
832 		 * if there's a shortage of swap, free any swap allocated
833 		 * to this page so that other pages can be paged out.
834 		 */
835 
836 		if (swap_shortage > 0) {
837 			if ((p->pqflags & PQ_ANON) && anon->an_swslot) {
838 				uvm_swap_free(anon->an_swslot, 1);
839 				anon->an_swslot = 0;
840 				p->flags &= ~PG_CLEAN;
841 				swap_shortage--;
842 			} else if (p->pqflags & PQ_AOBJ) {
843 				int slot = uao_set_swslot(uobj,
844 					p->offset >> PAGE_SHIFT, 0);
845 				if (slot) {
846 					uvm_swap_free(slot, 1);
847 					p->flags &= ~PG_CLEAN;
848 					swap_shortage--;
849 				}
850 			}
851 		}
852 
853 		/*
854 		 * if there's a shortage of inactive pages, deactivate.
855 		 */
856 
857 		if (inactive_shortage > 0) {
858 			/* no need to check wire_count as pg is "active" */
859 			uvm_pagedeactivate(p);
860 			uvmexp.pddeact++;
861 			inactive_shortage--;
862 		}
863 
864 		/*
865 		 * we're done with this page.
866 		 */
867 
868 		if (p->pqflags & PQ_ANON)
869 			simple_unlock(&anon->an_lock);
870 		else
871 			simple_unlock(&uobj->vmobjlock);
872 	}
873 }
874