xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 220b5c059a84c51ea44107ea8951a57ffaecdc8c)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.43 2001/12/09 03:07:19 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 1.43 2001/12/09 03:07:19 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 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 	boolean_t anonunder, fileunder, execunder;
371 	boolean_t anonover, fileover, execover;
372 	boolean_t anonreact, filereact, execreact;
373 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
374 
375 	/*
376 	 * swslot is non-zero if we are building a swap cluster.  we want
377 	 * to stay in the loop while we have a page to scan or we have
378 	 * a swap-cluster to build.
379 	 */
380 
381 	swslot = 0;
382 	swnpages = swcpages = 0;
383 	dirtyreacts = 0;
384 
385 	/*
386 	 * decide which types of pages we want to reactivate instead of freeing
387 	 * to keep usage within the minimum and maximum usage limits.
388 	 */
389 
390 	t = uvmexp.active + uvmexp.inactive + uvmexp.free;
391 	anonunder = (uvmexp.anonpages <= (t * uvmexp.anonmin) >> 8);
392 	fileunder = (uvmexp.filepages <= (t * uvmexp.filemin) >> 8);
393 	execunder = (uvmexp.execpages <= (t * uvmexp.execmin) >> 8);
394 	anonover = uvmexp.anonpages > ((t * uvmexp.anonmax) >> 8);
395 	fileover = uvmexp.filepages > ((t * uvmexp.filemax) >> 8);
396 	execover = uvmexp.execpages > ((t * uvmexp.execmax) >> 8);
397 	anonreact = anonunder || (!anonover && (fileover || execover));
398 	filereact = fileunder || (!fileover && (anonover || execover));
399 	execreact = execunder || (!execover && (anonover || fileover));
400 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
401 		uobj = NULL;
402 		anon = NULL;
403 		if (p) {
404 
405 			/*
406 			 * see if we've met the free target.
407 			 */
408 
409 			if (uvmexp.free + uvmexp.paging >=
410 			    uvmexp.freetarg << 2 ||
411 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
412 				UVMHIST_LOG(pdhist,"  met free target: "
413 					    "exit loop", 0, 0, 0, 0);
414 
415 				if (swslot == 0) {
416 					/* exit now if no swap-i/o pending */
417 					break;
418 				}
419 
420 				/* set p to null to signal final swap i/o */
421 				p = NULL;
422 				nextpg = NULL;
423 			}
424 		}
425 		if (p) {	/* if (we have a new page to consider) */
426 
427 			/*
428 			 * we are below target and have a new page to consider.
429 			 */
430 
431 			uvmexp.pdscans++;
432 			nextpg = TAILQ_NEXT(p, pageq);
433 
434 			/*
435 			 * move referenced pages back to active queue and
436 			 * skip to next page.
437 			 */
438 
439 			if (pmap_clear_reference(p)) {
440 				uvm_pageactivate(p);
441 				uvmexp.pdreact++;
442 				continue;
443 			}
444 			anon = p->uanon;
445 			uobj = p->uobject;
446 
447 			/*
448 			 * enforce the minimum thresholds on different
449 			 * types of memory usage.  if reusing the current
450 			 * page would reduce that type of usage below its
451 			 * minimum, reactivate the page instead and move
452 			 * on to the next page.
453 			 */
454 
455 			if (anon && anonreact) {
456 				uvm_pageactivate(p);
457 				uvmexp.pdreanon++;
458 				continue;
459 			}
460 			if (uobj && UVM_OBJ_IS_VTEXT(uobj) && execreact) {
461 				uvm_pageactivate(p);
462 				uvmexp.pdreexec++;
463 				continue;
464 			}
465 			if (uobj && UVM_OBJ_IS_VNODE(uobj) &&
466 			    !UVM_OBJ_IS_VTEXT(uobj) && filereact) {
467 				uvm_pageactivate(p);
468 				uvmexp.pdrefile++;
469 				continue;
470 			}
471 
472 			/*
473 			 * first we attempt to lock the object that this page
474 			 * belongs to.  if our attempt fails we skip on to
475 			 * the next page (no harm done).  it is important to
476 			 * "try" locking the object as we are locking in the
477 			 * wrong order (pageq -> object) and we don't want to
478 			 * deadlock.
479 			 *
480 			 * the only time we expect to see an ownerless page
481 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
482 			 * anon has loaned a page from a uvm_object and the
483 			 * uvm_object has dropped the ownership.  in that
484 			 * case, the anon can "take over" the loaned page
485 			 * and make it its own.
486 			 */
487 
488 			/* is page part of an anon or ownerless ? */
489 			if ((p->pqflags & PQ_ANON) || uobj == NULL) {
490 				KASSERT(anon != NULL);
491 				slock = &anon->an_lock;
492 				if (!simple_lock_try(slock)) {
493 					/* lock failed, skip this page */
494 					continue;
495 				}
496 
497 				/*
498 				 * if the page is ownerless, claim it in the
499 				 * name of "anon"!
500 				 */
501 
502 				if ((p->pqflags & PQ_ANON) == 0) {
503 					KASSERT(p->loan_count > 0);
504 					p->loan_count--;
505 					p->pqflags |= PQ_ANON;
506 					/* anon now owns it */
507 				}
508 				if (p->flags & PG_BUSY) {
509 					simple_unlock(slock);
510 					uvmexp.pdbusy++;
511 					continue;
512 				}
513 				uvmexp.pdanscan++;
514 			} else {
515 				KASSERT(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 			}
527 
528 
529 			/*
530 			 * we now have the object and the page queues locked.
531 			 * if the page is not swap-backed, call the object's
532 			 * pager to flush and free the page.
533 			 */
534 
535 			if ((p->pqflags & PQ_SWAPBACKED) == 0) {
536 				uvm_unlock_pageq();
537 				error = (uobj->pgops->pgo_put)(uobj, p->offset,
538 				    p->offset + PAGE_SIZE,
539 				    PGO_CLEANIT|PGO_FREE);
540 				uvm_lock_pageq();
541 				if (nextpg &&
542 				    (nextpg->flags & PQ_INACTIVE) == 0) {
543 					nextpg = TAILQ_FIRST(pglst);
544 				}
545 				continue;
546 			}
547 
548 			/*
549 			 * the page is swap-backed.  remove all the permissions
550 			 * from the page so we can sync the modified info
551 			 * without any race conditions.  if the page is clean
552 			 * we can free it now and continue.
553 			 */
554 
555 			pmap_page_protect(p, VM_PROT_NONE);
556 			if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
557 				p->flags &= ~(PG_CLEAN);
558 			}
559 			if (p->flags & PG_CLEAN) {
560 				uvm_pagefree(p);
561 				uvmexp.pdfreed++;
562 
563 				/*
564 				 * for anons, we need to remove the page
565 				 * from the anon ourselves.  for aobjs,
566 				 * pagefree did that for us.
567 				 */
568 
569 				if (anon) {
570 					KASSERT(anon->an_swslot != 0);
571 					anon->u.an_page = NULL;
572 				}
573 				simple_unlock(slock);
574 
575 				/* this page is now only in swap. */
576 				simple_lock(&uvm.swap_data_lock);
577 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
578 				uvmexp.swpgonly++;
579 				simple_unlock(&uvm.swap_data_lock);
580 				continue;
581 			}
582 
583 			/*
584 			 * this page is dirty, skip it if we'll have met our
585 			 * free target when all the current pageouts complete.
586 			 */
587 
588 			if (uvmexp.free + uvmexp.paging >
589 			    uvmexp.freetarg << 2) {
590 				simple_unlock(slock);
591 				continue;
592 			}
593 
594 			/*
595 			 * free any swap space allocated to the page since
596 			 * we'll have to write it again with its new data.
597 			 */
598 
599 			if ((p->pqflags & PQ_ANON) && anon->an_swslot) {
600 				uvm_swap_free(anon->an_swslot, 1);
601 				anon->an_swslot = 0;
602 			} else if (p->pqflags & PQ_AOBJ) {
603 				uao_dropswap(uobj, p->offset >> PAGE_SHIFT);
604 			}
605 
606 			/*
607 			 * if all pages in swap are only in swap,
608 			 * the swap space is full and we can't page out
609 			 * any more swap-backed pages.  reactivate this page
610 			 * so that we eventually cycle all pages through
611 			 * the inactive queue.
612 			 */
613 
614 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
615 			if (uvmexp.swpgonly == uvmexp.swpages) {
616 				dirtyreacts++;
617 				uvm_pageactivate(p);
618 				simple_unlock(slock);
619 				continue;
620 			}
621 
622 			/*
623 			 * start new swap pageout cluster (if necessary).
624 			 */
625 
626 			if (swslot == 0) {
627 				swnpages = MAXBSIZE >> PAGE_SHIFT;
628 				swslot = uvm_swap_alloc(&swnpages, TRUE);
629 				if (swslot == 0) {
630 					simple_unlock(slock);
631 					continue;
632 				}
633 				swcpages = 0;
634 			}
635 
636 			/*
637 			 * at this point, we're definitely going reuse this
638 			 * page.  mark the page busy and delayed-free.
639 			 * we should remove the page from the page queues
640 			 * so we don't ever look at it again.
641 			 * adjust counters and such.
642 			 */
643 
644 			p->flags |= PG_BUSY;
645 			UVM_PAGE_OWN(p, "scan_inactive");
646 
647 			p->flags |= PG_PAGEOUT;
648 			uvmexp.paging++;
649 			uvm_pagedequeue(p);
650 
651 			uvmexp.pgswapout++;
652 
653 			/*
654 			 * add the new page to the cluster.
655 			 */
656 
657 			if (anon) {
658 				anon->an_swslot = swslot + swcpages;
659 				simple_unlock(slock);
660 			} else {
661 				result = uao_set_swslot(uobj,
662 				    p->offset >> PAGE_SHIFT, swslot + swcpages);
663 				if (result == -1) {
664 					p->flags &= ~(PG_BUSY|PG_PAGEOUT);
665 					UVM_PAGE_OWN(p, NULL);
666 					uvmexp.paging--;
667 					uvm_pageactivate(p);
668 					simple_unlock(slock);
669 					continue;
670 				}
671 				simple_unlock(slock);
672 			}
673 			swpps[swcpages] = p;
674 			swcpages++;
675 
676 			/*
677 			 * if the cluster isn't full, look for more pages
678 			 * before starting the i/o.
679 			 */
680 
681 			if (swcpages < swnpages) {
682 				continue;
683 			}
684 		}
685 
686 		/*
687 		 * if this is the final pageout we could have a few
688 		 * unused swap blocks.  if so, free them now.
689 		 */
690 
691 		if (swcpages < swnpages) {
692 			uvm_swap_free(swslot + swcpages, (swnpages - swcpages));
693 		}
694 
695 		/*
696 		 * now start the pageout.
697 		 */
698 
699 		uvm_unlock_pageq();
700 		uvmexp.pdpageouts++;
701 		error = uvm_swap_put(swslot, swpps, swcpages, 0);
702 		KASSERT(error == 0);
703 		uvm_lock_pageq();
704 
705 		/*
706 		 * zero swslot to indicate that we are
707 		 * no longer building a swap-backed cluster.
708 		 */
709 
710 		swslot = 0;
711 
712 		/*
713 		 * the pageout is in progress.  bump counters and set up
714 		 * for the next loop.
715 		 */
716 
717 		uvmexp.pdpending++;
718 		if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) {
719 			nextpg = TAILQ_FIRST(pglst);
720 		}
721 	}
722 	return (error);
723 }
724 
725 /*
726  * uvmpd_scan: scan the page queues and attempt to meet our targets.
727  *
728  * => called with pageq's locked
729  */
730 
731 void
732 uvmpd_scan(void)
733 {
734 	int inactive_shortage, swap_shortage, pages_freed;
735 	struct vm_page *p, *nextpg;
736 	struct uvm_object *uobj;
737 	struct vm_anon *anon;
738 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
739 
740 	uvmexp.pdrevs++;
741 	uobj = NULL;
742 	anon = NULL;
743 
744 #ifndef __SWAP_BROKEN
745 
746 	/*
747 	 * swap out some processes if we are below our free target.
748 	 * we need to unlock the page queues for this.
749 	 */
750 
751 	if (uvmexp.free < uvmexp.freetarg && uvmexp.nswapdev != 0) {
752 		uvmexp.pdswout++;
753 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout",
754 		    uvmexp.free, uvmexp.freetarg, 0, 0);
755 		uvm_unlock_pageq();
756 		uvm_swapout_threads();
757 		uvm_lock_pageq();
758 
759 	}
760 #endif
761 
762 	/*
763 	 * now we want to work on meeting our targets.   first we work on our
764 	 * free target by converting inactive pages into free pages.  then
765 	 * we work on meeting our inactive target by converting active pages
766 	 * to inactive ones.
767 	 */
768 
769 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
770 
771 	/*
772 	 * alternate starting queue between swap and object based on the
773 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
774 	 */
775 
776 	pages_freed = uvmexp.pdfreed;
777 	(void) uvmpd_scan_inactive(&uvm.page_inactive);
778 	pages_freed = uvmexp.pdfreed - pages_freed;
779 
780 	/*
781 	 * we have done the scan to get free pages.   now we work on meeting
782 	 * our inactive target.
783 	 */
784 
785 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
786 
787 	/*
788 	 * detect if we're not going to be able to page anything out
789 	 * until we free some swap resources from active pages.
790 	 */
791 
792 	swap_shortage = 0;
793 	if (uvmexp.free < uvmexp.freetarg &&
794 	    uvmexp.swpginuse == uvmexp.swpages &&
795 	    uvmexp.swpgonly < uvmexp.swpages &&
796 	    pages_freed == 0) {
797 		swap_shortage = uvmexp.freetarg - uvmexp.free;
798 	}
799 
800 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
801 		    inactive_shortage, swap_shortage,0,0);
802 	for (p = TAILQ_FIRST(&uvm.page_active);
803 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
804 	     p = nextpg) {
805 		nextpg = TAILQ_NEXT(p, pageq);
806 		if (p->flags & PG_BUSY) {
807 			continue;
808 		}
809 
810 		/*
811 		 * lock the page's owner.
812 		 */
813 		/* is page anon owned or ownerless? */
814 		if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
815 			anon = p->uanon;
816 			KASSERT(anon != NULL);
817 			if (!simple_lock_try(&anon->an_lock)) {
818 				continue;
819 			}
820 
821 			/* take over the page? */
822 			if ((p->pqflags & PQ_ANON) == 0) {
823 				KASSERT(p->loan_count > 0);
824 				p->loan_count--;
825 				p->pqflags |= PQ_ANON;
826 			}
827 		} else {
828 			uobj = p->uobject;
829 			if (!simple_lock_try(&uobj->vmobjlock)) {
830 				continue;
831 			}
832 		}
833 
834 		/*
835 		 * skip this page if it's busy.
836 		 */
837 
838 		if ((p->flags & PG_BUSY) != 0) {
839 			if (p->pqflags & PQ_ANON)
840 				simple_unlock(&anon->an_lock);
841 			else
842 				simple_unlock(&uobj->vmobjlock);
843 			continue;
844 		}
845 
846 		/*
847 		 * if there's a shortage of swap, free any swap allocated
848 		 * to this page so that other pages can be paged out.
849 		 */
850 
851 		if (swap_shortage > 0) {
852 			if ((p->pqflags & PQ_ANON) && anon->an_swslot) {
853 				uvm_swap_free(anon->an_swslot, 1);
854 				anon->an_swslot = 0;
855 				p->flags &= ~PG_CLEAN;
856 				swap_shortage--;
857 			} else if (p->pqflags & PQ_AOBJ) {
858 				int slot = uao_set_swslot(uobj,
859 					p->offset >> PAGE_SHIFT, 0);
860 				if (slot) {
861 					uvm_swap_free(slot, 1);
862 					p->flags &= ~PG_CLEAN;
863 					swap_shortage--;
864 				}
865 			}
866 		}
867 
868 		/*
869 		 * if there's a shortage of inactive pages, deactivate.
870 		 */
871 
872 		if (inactive_shortage > 0) {
873 			/* no need to check wire_count as pg is "active" */
874 			uvm_pagedeactivate(p);
875 			uvmexp.pddeact++;
876 			inactive_shortage--;
877 		}
878 
879 		/*
880 		 * we're done with this page.
881 		 */
882 
883 		if (p->pqflags & PQ_ANON)
884 			simple_unlock(&anon->an_lock);
885 		else
886 			simple_unlock(&uobj->vmobjlock);
887 	}
888 }
889