xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 850e275390052b330d93020bf619a739a3c277ac)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.35 2008/07/02 15:21:33 art Exp $	*/
2 /*	$NetBSD: uvm_pdaemon.c,v 1.23 2000/08/20 10:24:14 bjh21 Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * Copyright (c) 1991, 1993, The Regents of the University of California.
7  *
8  * All rights reserved.
9  *
10  * This code is derived from software contributed to Berkeley by
11  * The Mach Operating System project at Carnegie-Mellon University.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by Charles D. Cranor,
24  *      Washington University, the University of California, Berkeley and
25  *      its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vm_pageout.c        8.5 (Berkeley) 2/14/94
43  * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp
44  *
45  *
46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47  * All rights reserved.
48  *
49  * Permission to use, copy, modify and distribute this software and
50  * its documentation is hereby granted, provided that both the copyright
51  * notice and this permission notice appear in all copies of the
52  * software, derivative works or modified versions, and any portions
53  * thereof, and that both notices appear in supporting documentation.
54  *
55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58  *
59  * Carnegie Mellon requests users of this software to return to
60  *
61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
62  *  School of Computer Science
63  *  Carnegie Mellon University
64  *  Pittsburgh PA 15213-3890
65  *
66  * any improvements or extensions that they make and grant Carnegie the
67  * rights to redistribute these changes.
68  */
69 
70 /*
71  * uvm_pdaemon.c: the page daemon
72  */
73 
74 #include <sys/param.h>
75 #include <sys/proc.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/pool.h>
79 #include <sys/buf.h>
80 #include <sys/vnode.h>
81 #include <sys/mount.h>
82 
83 #include <uvm/uvm.h>
84 
85 /*
86  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
87  * in a pass thru the inactive list when swap is full.  the value should be
88  * "small"... if it's too large we'll cycle the active pages thru the inactive
89  * queue too quickly to for them to be referenced and avoid being freed.
90  */
91 
92 #define UVMPD_NUMDIRTYREACTS 16
93 
94 
95 /*
96  * local prototypes
97  */
98 
99 static void		uvmpd_scan(void);
100 static boolean_t	uvmpd_scan_inactive(struct pglist *);
101 static void		uvmpd_tune(void);
102 
103 /*
104  * uvm_wait: wait (sleep) for the page daemon to free some pages
105  *
106  * => should be called with all locks released
107  * => should _not_ be called by the page daemon (to avoid deadlock)
108  */
109 
110 void
111 uvm_wait(wmsg)
112 	const char *wmsg;
113 {
114 	int timo = 0;
115 	int s = splbio();
116 
117 	/*
118 	 * check for page daemon going to sleep (waiting for itself)
119 	 */
120 
121 	if (curproc == uvm.pagedaemon_proc) {
122 		/*
123 		 * now we have a problem: the pagedaemon wants to go to
124 		 * sleep until it frees more memory.   but how can it
125 		 * free more memory if it is asleep?  that is a deadlock.
126 		 * we have two options:
127 		 *  [1] panic now
128 		 *  [2] put a timeout on the sleep, thus causing the
129 		 *      pagedaemon to only pause (rather than sleep forever)
130 		 *
131 		 * note that option [2] will only help us if we get lucky
132 		 * and some other process on the system breaks the deadlock
133 		 * by exiting or freeing memory (thus allowing the pagedaemon
134 		 * to continue).  for now we panic if DEBUG is defined,
135 		 * otherwise we hope for the best with option [2] (better
136 		 * yet, this should never happen in the first place!).
137 		 */
138 
139 		printf("pagedaemon: deadlock detected!\n");
140 		timo = hz >> 3;		/* set timeout */
141 #if defined(DEBUG)
142 		/* DEBUG: panic so we can debug it */
143 		panic("pagedaemon deadlock");
144 #endif
145 	}
146 
147 	simple_lock(&uvm.pagedaemon_lock);
148 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
149 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, wmsg,
150 	    timo);
151 
152 	splx(s);
153 }
154 
155 
156 /*
157  * uvmpd_tune: tune paging parameters
158  *
159  * => called when ever memory is added (or removed?) to the system
160  * => caller must call with page queues locked
161  */
162 
163 static void
164 uvmpd_tune(void)
165 {
166 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
167 
168 	uvmexp.freemin = uvmexp.npages / 30;
169 
170 	/* between 16k and 512k */
171 	/* XXX:  what are these values good for? */
172 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
173 #if 0
174 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
175 #endif
176 
177 	/* Make sure there's always a user page free. */
178 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
179 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
180 
181 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
182 	if (uvmexp.freetarg <= uvmexp.freemin)
183 		uvmexp.freetarg = uvmexp.freemin + 1;
184 
185 	/* uvmexp.inactarg: computed in main daemon loop */
186 
187 	uvmexp.wiredmax = uvmexp.npages / 3;
188 	UVMHIST_LOG(pdhist, "<- done, freemin=%ld, freetarg=%ld, wiredmax=%ld",
189 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
190 }
191 
192 /*
193  * uvm_pageout: the main loop for the pagedaemon
194  */
195 
196 void
197 uvm_pageout(void *arg)
198 {
199 	int npages = 0;
200 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
201 
202 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
203 
204 	/*
205 	 * ensure correct priority and set paging parameters...
206 	 */
207 
208 	uvm.pagedaemon_proc = curproc;
209 	(void) spl0();
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=%ld/%ld, inact/itarg=%ld/%ld",
244 		    uvmexp.free, uvmexp.freetarg, uvmexp.inactive,
245 		    uvmexp.inactarg);
246 
247 		/*
248 		 * scan if needed
249 		 */
250 		if ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg ||
251 		    uvmexp.inactive < uvmexp.inactarg) {
252 			uvmpd_scan();
253 		}
254 
255 		/*
256 		 * if there's any free memory to be had,
257 		 * wake up any waiters.
258 		 */
259 
260 		if (uvmexp.free > uvmexp.reserve_kernel ||
261 		    uvmexp.paging == 0) {
262 			wakeup(&uvmexp.free);
263 		}
264 
265 		/*
266 		 * scan done.  unlock page queues (the only lock we are holding)
267 		 */
268 
269 		uvm_unlock_pageq();
270 	}
271 	/*NOTREACHED*/
272 }
273 
274 
275 /*
276  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
277  */
278 
279 void
280 uvm_aiodone_daemon(void *arg)
281 {
282 	int s, free;
283 	struct buf *bp, *nbp;
284 	UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist);
285 
286 	for (;;) {
287 
288 		/*
289 		 * carefully attempt to go to sleep (without losing "wakeups"!).
290 		 * we need splbio because we want to make sure the aio_done list
291 		 * is totally empty before we go to sleep.
292 		 */
293 
294 		s = splbio();
295 		simple_lock(&uvm.aiodoned_lock);
296 		if (TAILQ_FIRST(&uvm.aio_done) == NULL) {
297 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
298 			UVM_UNLOCK_AND_WAIT(&uvm.aiodoned,
299 			    &uvm.aiodoned_lock, FALSE, "aiodoned", 0);
300 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
301 
302 			/* relock aiodoned_lock, still at splbio */
303 			simple_lock(&uvm.aiodoned_lock);
304 		}
305 
306 		/*
307 		 * check for done aio structures
308 		 */
309 
310 		bp = TAILQ_FIRST(&uvm.aio_done);
311 		if (bp) {
312 			TAILQ_INIT(&uvm.aio_done);
313 		}
314 
315 		simple_unlock(&uvm.aiodoned_lock);
316 		splx(s);
317 
318 		/*
319 		 * process each i/o that's done.
320 		 */
321 
322 		free = uvmexp.free;
323 		while (bp != NULL) {
324 			if (bp->b_flags & B_PDAEMON) {
325 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
326 			}
327 			nbp = TAILQ_NEXT(bp, b_freelist);
328 			s = splbio();	/* b_iodone must by called at splbio */
329 			(*bp->b_iodone)(bp);
330 			splx(s);
331 			bp = nbp;
332 		}
333 		if (free <= uvmexp.reserve_kernel) {
334 			uvm_lock_fpageq();
335 			wakeup(&uvm.pagedaemon);
336 			uvm_unlock_fpageq();
337 		} else {
338 			simple_lock(&uvm.pagedaemon_lock);
339 			wakeup(&uvmexp.free);
340 			simple_unlock(&uvm.pagedaemon_lock);
341 		}
342 	}
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 static boolean_t
358 uvmpd_scan_inactive(pglst)
359 	struct pglist *pglst;
360 {
361 	boolean_t retval = FALSE;	/* assume we haven't hit target */
362 	int free, result;
363 	struct vm_page *p, *nextpg;
364 	struct uvm_object *uobj;
365 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
366 	int npages;
367 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
368 	int swnpages, swcpages;				/* XXX: see below */
369 	int swslot;
370 	struct vm_anon *anon;
371 	boolean_t swap_backed;
372 	vaddr_t start;
373 	int dirtyreacts;
374 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
375 
376 	/*
377 	 * note: we currently keep swap-backed pages on a separate inactive
378 	 * list from object-backed pages.   however, merging the two lists
379 	 * back together again hasn't been ruled out.   thus, we keep our
380 	 * swap cluster in "swpps" rather than in pps (allows us to mix
381 	 * clustering types in the event of a mixed inactive queue).
382 	 */
383 
384 	/*
385 	 * swslot is non-zero if we are building a swap cluster.  we want
386 	 * to stay in the loop while we have a page to scan or we have
387 	 * a swap-cluster to build.
388 	 */
389 
390 	swslot = 0;
391 	swnpages = swcpages = 0;
392 	free = 0;
393 	dirtyreacts = 0;
394 
395 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
396 
397 		/*
398 		 * note that p can be NULL iff we have traversed the whole
399 		 * list and need to do one final swap-backed clustered pageout.
400 		 */
401 
402 		uobj = NULL;
403 		anon = NULL;
404 
405 		if (p) {
406 
407 			/*
408 			 * update our copy of "free" and see if we've met
409 			 * our target
410 			 */
411 
412 			uvm_lock_fpageq();
413 			free = uvmexp.free - BUFPAGES_DEFICIT;
414 			uvm_unlock_fpageq();
415 
416 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
417 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
418 				UVMHIST_LOG(pdhist,"  met free target: "
419 					    "exit loop", 0, 0, 0, 0);
420 				retval = TRUE;
421 
422 				if (swslot == 0) {
423 					/* exit now if no swap-i/o pending */
424 					break;
425 				}
426 
427 				/* set p to null to signal final swap i/o */
428 				p = NULL;
429 			}
430 		}
431 
432 		if (p) {	/* if (we have a new page to consider) */
433 
434 			/*
435 			 * we are below target and have a new page to consider.
436 			 */
437 			uvmexp.pdscans++;
438 			nextpg = TAILQ_NEXT(p, pageq);
439 
440 			/*
441 			 * move referenced pages back to active queue and
442 			 * skip to next page (unlikely to happen since
443 			 * inactive pages shouldn't have any valid mappings
444 			 * and we cleared reference before deactivating).
445 			 */
446 
447 			if (pmap_is_referenced(p)) {
448 				uvm_pageactivate(p);
449 				uvmexp.pdreact++;
450 				continue;
451 			}
452 
453 			/*
454 			 * first we attempt to lock the object that this page
455 			 * belongs to.  if our attempt fails we skip on to
456 			 * the next page (no harm done).  it is important to
457 			 * "try" locking the object as we are locking in the
458 			 * wrong order (pageq -> object) and we don't want to
459 			 * deadlock.
460 			 *
461 			 * the only time we expect to see an ownerless page
462 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
463 			 * anon has loaned a page from a uvm_object and the
464 			 * uvm_object has dropped the ownership.  in that
465 			 * case, the anon can "take over" the loaned page
466 			 * and make it its own.
467 			 */
468 
469 			/* is page part of an anon or ownerless ? */
470 			if ((p->pg_flags & PQ_ANON) || p->uobject == NULL) {
471 				anon = p->uanon;
472 				KASSERT(anon != NULL);
473 				if (!simple_lock_try(&anon->an_lock)) {
474 					/* lock failed, skip this page */
475 					continue;
476 				}
477 
478 				/*
479 				 * if the page is ownerless, claim it in the
480 				 * name of "anon"!
481 				 */
482 
483 				if ((p->pg_flags & PQ_ANON) == 0) {
484 					KASSERT(p->loan_count > 0);
485 					p->loan_count--;
486 					atomic_setbits_int(&p->pg_flags,
487 					    PQ_ANON);
488 					/* anon now owns it */
489 				}
490 				if (p->pg_flags & PG_BUSY) {
491 					simple_unlock(&anon->an_lock);
492 					uvmexp.pdbusy++;
493 					/* someone else owns page, skip it */
494 					continue;
495 				}
496 				uvmexp.pdanscan++;
497 			} else {
498 				uobj = p->uobject;
499 				KASSERT(uobj != NULL);
500 				if (!simple_lock_try(&uobj->vmobjlock)) {
501 					/* lock failed, skip this page */
502 					continue;
503 				}
504 				if (p->pg_flags & PG_BUSY) {
505 					simple_unlock(&uobj->vmobjlock);
506 					uvmexp.pdbusy++;
507 					/* someone else owns page, skip it */
508 					continue;
509 				}
510 				uvmexp.pdobscan++;
511 			}
512 
513 			/*
514 			 * we now have the object and the page queues locked.
515 			 * the page is not busy.   if the page is clean we
516 			 * can free it now and continue.
517 			 */
518 
519 			if (p->pg_flags & PG_CLEAN) {
520 				if (p->pg_flags & PQ_SWAPBACKED) {
521 					/* this page now lives only in swap */
522 					simple_lock(&uvm.swap_data_lock);
523 					uvmexp.swpgonly++;
524 					simple_unlock(&uvm.swap_data_lock);
525 				}
526 
527 				/* zap all mappings with pmap_page_protect... */
528 				pmap_page_protect(p, VM_PROT_NONE);
529 				uvm_pagefree(p);
530 				uvmexp.pdfreed++;
531 
532 				if (anon) {
533 
534 					/*
535 					 * an anonymous page can only be clean
536 					 * if it has backing store assigned.
537 					 */
538 
539 					KASSERT(anon->an_swslot != 0);
540 
541 					/* remove from object */
542 					anon->an_page = NULL;
543 					simple_unlock(&anon->an_lock);
544 				} else {
545 					/* pagefree has already removed the
546 					 * page from the object */
547 					simple_unlock(&uobj->vmobjlock);
548 				}
549 				continue;
550 			}
551 
552 			/*
553 			 * this page is dirty, skip it if we'll have met our
554 			 * free target when all the current pageouts complete.
555 			 */
556 
557 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
558 				if (anon) {
559 					simple_unlock(&anon->an_lock);
560 				} else {
561 					simple_unlock(&uobj->vmobjlock);
562 				}
563 				continue;
564 			}
565 
566 			/*
567 			 * this page is dirty, but we can't page it out
568 			 * since all pages in swap are only in swap.
569 			 * reactivate it so that we eventually cycle
570 			 * all pages thru the inactive queue.
571 			 */
572 
573 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
574 			if ((p->pg_flags & PQ_SWAPBACKED) &&
575 			    uvmexp.swpgonly == uvmexp.swpages) {
576 				dirtyreacts++;
577 				uvm_pageactivate(p);
578 				if (anon) {
579 					simple_unlock(&anon->an_lock);
580 				} else {
581 					simple_unlock(&uobj->vmobjlock);
582 				}
583 				continue;
584 			}
585 
586 			/*
587 			 * if the page is swap-backed and dirty and swap space
588 			 * is full, free any swap allocated to the page
589 			 * so that other pages can be paged out.
590 			 */
591 
592 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
593 			if ((p->pg_flags & PQ_SWAPBACKED) &&
594 			    uvmexp.swpginuse == uvmexp.swpages) {
595 
596 				if ((p->pg_flags & PQ_ANON) &&
597 				    p->uanon->an_swslot) {
598 					uvm_swap_free(p->uanon->an_swslot, 1);
599 					p->uanon->an_swslot = 0;
600 				}
601 				if (p->pg_flags & PQ_AOBJ) {
602 					uao_dropswap(p->uobject,
603 						     p->offset >> PAGE_SHIFT);
604 				}
605 			}
606 
607 			/*
608 			 * the page we are looking at is dirty.   we must
609 			 * clean it before it can be freed.  to do this we
610 			 * first mark the page busy so that no one else will
611 			 * touch the page.   we write protect all the mappings
612 			 * of the page so that no one touches it while it is
613 			 * in I/O.
614 			 */
615 
616 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
617 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
618 			UVM_PAGE_OWN(p, "scan_inactive");
619 			pmap_page_protect(p, VM_PROT_READ);
620 			uvmexp.pgswapout++;
621 
622 			/*
623 			 * for swap-backed pages we need to (re)allocate
624 			 * swap space.
625 			 */
626 
627 			if (swap_backed) {
628 
629 				/*
630 				 * free old swap slot (if any)
631 				 */
632 
633 				if (anon) {
634 					if (anon->an_swslot) {
635 						uvm_swap_free(anon->an_swslot,
636 						    1);
637 						anon->an_swslot = 0;
638 					}
639 				} else {
640 					uao_dropswap(uobj,
641 						     p->offset >> PAGE_SHIFT);
642 				}
643 
644 				/*
645 				 * start new cluster (if necessary)
646 				 */
647 
648 				if (swslot == 0) {
649 					swnpages = MAXBSIZE >> PAGE_SHIFT;
650 					swslot = uvm_swap_alloc(&swnpages,
651 					    TRUE);
652 					if (swslot == 0) {
653 						/* no swap?  give up! */
654 						atomic_clearbits_int(
655 						    &p->pg_flags,
656 						    PG_BUSY);
657 						UVM_PAGE_OWN(p, NULL);
658 						if (anon)
659 							simple_unlock(
660 							    &anon->an_lock);
661 						else
662 							simple_unlock(
663 							    &uobj->vmobjlock);
664 						continue;
665 					}
666 					swcpages = 0;	/* cluster is empty */
667 				}
668 
669 				/*
670 				 * add block to cluster
671 				 */
672 
673 				swpps[swcpages] = p;
674 				if (anon)
675 					anon->an_swslot = swslot + swcpages;
676 				else
677 					uao_set_swslot(uobj,
678 					    p->offset >> PAGE_SHIFT,
679 					    swslot + swcpages);
680 				swcpages++;
681 			}
682 		} else {
683 
684 			/* if p == NULL we must be doing a last swap i/o */
685 			swap_backed = TRUE;
686 		}
687 
688 		/*
689 		 * now consider doing the pageout.
690 		 *
691 		 * for swap-backed pages, we do the pageout if we have either
692 		 * filled the cluster (in which case (swnpages == swcpages) or
693 		 * run out of pages (p == NULL).
694 		 *
695 		 * for object pages, we always do the pageout.
696 		 */
697 
698 		if (swap_backed) {
699 			if (p) {	/* if we just added a page to cluster */
700 				if (anon)
701 					simple_unlock(&anon->an_lock);
702 				else
703 					simple_unlock(&uobj->vmobjlock);
704 
705 				/* cluster not full yet? */
706 				if (swcpages < swnpages)
707 					continue;
708 			}
709 
710 			/* starting I/O now... set up for it */
711 			npages = swcpages;
712 			ppsp = swpps;
713 			/* for swap-backed pages only */
714 			start = (vaddr_t) swslot;
715 
716 			/* if this is final pageout we could have a few
717 			 * extra swap blocks */
718 			if (swcpages < swnpages) {
719 				uvm_swap_free(swslot + swcpages,
720 				    (swnpages - swcpages));
721 			}
722 		} else {
723 			/* normal object pageout */
724 			ppsp = pps;
725 			npages = sizeof(pps) / sizeof(struct vm_page *);
726 			/* not looked at because PGO_ALLPAGES is set */
727 			start = 0;
728 		}
729 
730 		/*
731 		 * now do the pageout.
732 		 *
733 		 * for swap_backed pages we have already built the cluster.
734 		 * for !swap_backed pages, uvm_pager_put will call the object's
735 		 * "make put cluster" function to build a cluster on our behalf.
736 		 *
737 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
738 		 * it to free the cluster pages for us on a successful I/O (it
739 		 * always does this for un-successful I/O requests).  this
740 		 * allows us to do clustered pageout without having to deal
741 		 * with cluster pages at this level.
742 		 *
743 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
744 		 *  IN: locked: uobj (if !swap_backed), page queues
745 		 * OUT: locked: uobj (if !swap_backed && result !=VM_PAGER_PEND)
746 		 *     !locked: pageqs, uobj (if swap_backed || VM_PAGER_PEND)
747 		 *
748 		 * [the bit about VM_PAGER_PEND saves us one lock-unlock pair]
749 		 */
750 
751 		/* locked: uobj (if !swap_backed), page queues */
752 		uvmexp.pdpageouts++;
753 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
754 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
755 		/* locked: uobj (if !swap_backed && result != PEND) */
756 		/* unlocked: pageqs, object (if swap_backed ||result == PEND) */
757 
758 		/*
759 		 * if we did i/o to swap, zero swslot to indicate that we are
760 		 * no longer building a swap-backed cluster.
761 		 */
762 
763 		if (swap_backed)
764 			swslot = 0;		/* done with this cluster */
765 
766 		/*
767 		 * first, we check for VM_PAGER_PEND which means that the
768 		 * async I/O is in progress and the async I/O done routine
769 		 * will clean up after us.   in this case we move on to the
770 		 * next page.
771 		 *
772 		 * there is a very remote chance that the pending async i/o can
773 		 * finish _before_ we get here.   if that happens, our page "p"
774 		 * may no longer be on the inactive queue.   so we verify this
775 		 * when determining the next page (starting over at the head if
776 		 * we've lost our inactive page).
777 		 */
778 
779 		if (result == VM_PAGER_PEND) {
780 			uvmexp.paging += npages;
781 			uvm_lock_pageq();
782 			uvmexp.pdpending++;
783 			if (p) {
784 				if (p->pg_flags & PQ_INACTIVE)
785 					nextpg = TAILQ_NEXT(p, pageq);
786 				else
787 					nextpg = TAILQ_FIRST(pglst);
788 			} else {
789 				nextpg = NULL;
790 			}
791 			continue;
792 		}
793 
794 #ifdef UBC
795 		if (result == VM_PAGER_ERROR &&
796 		    curproc == uvm.pagedaemon_proc) {
797 			uvm_lock_pageq();
798 			nextpg = TAILQ_NEXT(p, pageq);
799 			uvm_pageactivate(p);
800 			continue;
801 		}
802 #endif
803 
804 		/*
805 		 * clean up "p" if we have one
806 		 */
807 
808 		if (p) {
809 			/*
810 			 * the I/O request to "p" is done and uvm_pager_put
811 			 * has freed any cluster pages it may have allocated
812 			 * during I/O.  all that is left for us to do is
813 			 * clean up page "p" (which is still PG_BUSY).
814 			 *
815 			 * our result could be one of the following:
816 			 *   VM_PAGER_OK: successful pageout
817 			 *
818 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
819 			 *     to next page
820 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
821 			 *     "reactivate" page to get it out of the way (it
822 			 *     will eventually drift back into the inactive
823 			 *     queue for a retry).
824 			 *   VM_PAGER_UNLOCK: should never see this as it is
825 			 *     only valid for "get" operations
826 			 */
827 
828 			/* relock p's object: page queues not lock yet, so
829 			 * no need for "try" */
830 
831 			/* !swap_backed case: already locked... */
832 			if (swap_backed) {
833 				if (anon)
834 					simple_lock(&anon->an_lock);
835 				else
836 					simple_lock(&uobj->vmobjlock);
837 			}
838 
839 #ifdef DIAGNOSTIC
840 			if (result == VM_PAGER_UNLOCK)
841 				panic("pagedaemon: pageout returned "
842 				    "invalid 'unlock' code");
843 #endif
844 
845 			/* handle PG_WANTED now */
846 			if (p->pg_flags & PG_WANTED)
847 				/* still holding object lock */
848 				wakeup(p);
849 
850 			atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED);
851 			UVM_PAGE_OWN(p, NULL);
852 
853 			/* released during I/O? */
854 			if (p->pg_flags & PG_RELEASED) {
855 				if (anon) {
856 					/* remove page so we can get nextpg */
857 					anon->an_page = NULL;
858 
859 					simple_unlock(&anon->an_lock);
860 					uvm_anfree(anon);	/* kills anon */
861 					pmap_page_protect(p, VM_PROT_NONE);
862 					anon = NULL;
863 					uvm_lock_pageq();
864 					nextpg = TAILQ_NEXT(p, pageq);
865 					/* free released page */
866 					uvm_pagefree(p);
867 
868 				} else {
869 
870 					/*
871 					 * pgo_releasepg nukes the page and
872 					 * gets "nextpg" for us.  it returns
873 					 * with the page queues locked (when
874 					 * given nextpg ptr).
875 					 */
876 
877 					if (!uobj->pgops->pgo_releasepg(p,
878 					    &nextpg))
879 						/* uobj died after release */
880 						uobj = NULL;
881 
882 					/*
883 					 * lock page queues here so that they're
884 					 * always locked at the end of the loop.
885 					 */
886 
887 					uvm_lock_pageq();
888 				}
889 			} else {	/* page was not released during I/O */
890 				uvm_lock_pageq();
891 				nextpg = TAILQ_NEXT(p, pageq);
892 				if (result != VM_PAGER_OK) {
893 					/* pageout was a failure... */
894 					if (result != VM_PAGER_AGAIN)
895 						uvm_pageactivate(p);
896 					pmap_clear_reference(p);
897 					/* XXXCDC: if (swap_backed) FREE p's
898 					 * swap block? */
899 				} else {
900 					/* pageout was a success... */
901 					pmap_clear_reference(p);
902 					pmap_clear_modify(p);
903 					atomic_setbits_int(&p->pg_flags,
904 					    PG_CLEAN);
905 				}
906 			}
907 
908 			/*
909 			 * drop object lock (if there is an object left).   do
910 			 * a safety check of nextpg to make sure it is on the
911 			 * inactive queue (it should be since PG_BUSY pages on
912 			 * the inactive queue can't be re-queued [note: not
913 			 * true for active queue]).
914 			 */
915 
916 			if (anon)
917 				simple_unlock(&anon->an_lock);
918 			else if (uobj)
919 				simple_unlock(&uobj->vmobjlock);
920 
921 		} else {
922 
923 			/*
924 			 * if p is null in this loop, make sure it stays null
925 			 * in the next loop.
926 			 */
927 
928 			nextpg = NULL;
929 
930 			/*
931 			 * lock page queues here just so they're always locked
932 			 * at the end of the loop.
933 			 */
934 
935 			uvm_lock_pageq();
936 		}
937 
938 		if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
939 			nextpg = TAILQ_FIRST(pglst);	/* reload! */
940 		}
941 	}
942 	return (retval);
943 }
944 
945 /*
946  * uvmpd_scan: scan the page queues and attempt to meet our targets.
947  *
948  * => called with pageq's locked
949  */
950 
951 void
952 uvmpd_scan(void)
953 {
954 	int free, inactive_shortage, swap_shortage, pages_freed;
955 	struct vm_page *p, *nextpg;
956 	struct uvm_object *uobj;
957 	boolean_t got_it;
958 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
959 
960 	uvmexp.pdrevs++;		/* counter */
961 	uobj = NULL;
962 
963 	/*
964 	 * get current "free" page count
965 	 */
966 	uvm_lock_fpageq();
967 	free = uvmexp.free - BUFPAGES_DEFICIT;
968 	uvm_unlock_fpageq();
969 
970 #ifndef __SWAP_BROKEN
971 	/*
972 	 * swap out some processes if we are below our free target.
973 	 * we need to unlock the page queues for this.
974 	 */
975 	if (free < uvmexp.freetarg) {
976 		uvmexp.pdswout++;
977 		UVMHIST_LOG(pdhist,"  free %ld < target %ld: swapout", free,
978 		    uvmexp.freetarg, 0, 0);
979 		uvm_unlock_pageq();
980 		uvm_swapout_threads();
981 		uvm_lock_pageq();
982 	}
983 #endif
984 
985 	/*
986 	 * now we want to work on meeting our targets.   first we work on our
987 	 * free target by converting inactive pages into free pages.  then
988 	 * we work on meeting our inactive target by converting active pages
989 	 * to inactive ones.
990 	 */
991 
992 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
993 
994 	/*
995 	 * alternate starting queue between swap and object based on the
996 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
997 	 */
998 
999 	got_it = FALSE;
1000 	pages_freed = uvmexp.pdfreed;	/* XXX - int */
1001 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
1002 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
1003 	if (!got_it)
1004 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
1005 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
1006 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
1007 	pages_freed = uvmexp.pdfreed - pages_freed;
1008 
1009 	/*
1010 	 * we have done the scan to get free pages.   now we work on meeting
1011 	 * our inactive target.
1012 	 */
1013 
1014 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
1015 
1016 	/*
1017 	 * detect if we're not going to be able to page anything out
1018 	 * until we free some swap resources from active pages.
1019 	 */
1020 
1021 	swap_shortage = 0;
1022 	if (uvmexp.free < uvmexp.freetarg &&
1023 	    uvmexp.swpginuse == uvmexp.swpages &&
1024 	    uvmexp.swpgonly < uvmexp.swpages &&
1025 	    pages_freed == 0) {
1026 		swap_shortage = uvmexp.freetarg - uvmexp.free;
1027 	}
1028 
1029 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%ld swap_shortage=%ld",
1030 		    inactive_shortage, swap_shortage,0,0);
1031 	for (p = TAILQ_FIRST(&uvm.page_active);
1032 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
1033 	     p = nextpg) {
1034 		nextpg = TAILQ_NEXT(p, pageq);
1035 		if (p->pg_flags & PG_BUSY)
1036 			continue;	/* quick check before trying to lock */
1037 
1038 		/*
1039 		 * lock the page's owner.
1040 		 */
1041 		/* is page anon owned or ownerless? */
1042 		if ((p->pg_flags & PQ_ANON) || p->uobject == NULL) {
1043 			KASSERT(p->uanon != NULL);
1044 			if (!simple_lock_try(&p->uanon->an_lock))
1045 				continue;
1046 
1047 			/* take over the page? */
1048 			if ((p->pg_flags & PQ_ANON) == 0) {
1049 				KASSERT(p->loan_count > 0);
1050 				p->loan_count--;
1051 				atomic_setbits_int(&p->pg_flags, PQ_ANON);
1052 			}
1053 		} else {
1054 			if (!simple_lock_try(&p->uobject->vmobjlock))
1055 				continue;
1056 		}
1057 
1058 		/*
1059 		 * skip this page if it's busy.
1060 		 */
1061 
1062 		if ((p->pg_flags & PG_BUSY) != 0) {
1063 			if (p->pg_flags & PQ_ANON)
1064 				simple_unlock(&p->uanon->an_lock);
1065 			else
1066 				simple_unlock(&p->uobject->vmobjlock);
1067 			continue;
1068 		}
1069 
1070 		/*
1071 		 * if there's a shortage of swap, free any swap allocated
1072 		 * to this page so that other pages can be paged out.
1073 		 */
1074 
1075 		if (swap_shortage > 0) {
1076 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
1077 				uvm_swap_free(p->uanon->an_swslot, 1);
1078 				p->uanon->an_swslot = 0;
1079 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
1080 				swap_shortage--;
1081 			}
1082 			if (p->pg_flags & PQ_AOBJ) {
1083 				int slot = uao_set_swslot(p->uobject,
1084 					p->offset >> PAGE_SHIFT, 0);
1085 				if (slot) {
1086 					uvm_swap_free(slot, 1);
1087 					atomic_clearbits_int(&p->pg_flags,
1088 					    PG_CLEAN);
1089 					swap_shortage--;
1090 				}
1091 			}
1092 		}
1093 
1094 		/*
1095 		 * deactivate this page if there's a shortage of
1096 		 * inactive pages.
1097 		 */
1098 
1099 		if (inactive_shortage > 0) {
1100 			pmap_page_protect(p, VM_PROT_NONE);
1101 			/* no need to check wire_count as pg is "active" */
1102 			uvm_pagedeactivate(p);
1103 			uvmexp.pddeact++;
1104 			inactive_shortage--;
1105 		}
1106 		if (p->pg_flags & PQ_ANON)
1107 			simple_unlock(&p->uanon->an_lock);
1108 		else
1109 			simple_unlock(&p->uobject->vmobjlock);
1110 	}
1111 }
1112