xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision b725ae7711052a2233e31a66fefb8a752c388d7a)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.25 2002/05/24 13:10:53 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 
82 #include <uvm/uvm.h>
83 
84 /*
85  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedeamon will reactivate
86  * in a pass thru the inactive list when swap is full.  the value should be
87  * "small"... if it's too large we'll cycle the active pages thru the inactive
88  * queue too quickly to for them to be referenced and avoid being freed.
89  */
90 
91 #define UVMPD_NUMDIRTYREACTS 16
92 
93 
94 /*
95  * local prototypes
96  */
97 
98 static void		uvmpd_scan(void);
99 static boolean_t	uvmpd_scan_inactive(struct pglist *);
100 static void		uvmpd_tune(void);
101 
102 /*
103  * uvm_wait: wait (sleep) for the page daemon to free some pages
104  *
105  * => should be called with all locks released
106  * => should _not_ be called by the page daemon (to avoid deadlock)
107  */
108 
109 void
110 uvm_wait(wmsg)
111 	const char *wmsg;
112 {
113 	int timo = 0;
114 	int s = splbio();
115 
116 	/*
117 	 * check for page daemon going to sleep (waiting for itself)
118 	 */
119 
120 	if (curproc == uvm.pagedaemon_proc) {
121 		/*
122 		 * now we have a problem: the pagedaemon wants to go to
123 		 * sleep until it frees more memory.   but how can it
124 		 * free more memory if it is asleep?  that is a deadlock.
125 		 * we have two options:
126 		 *  [1] panic now
127 		 *  [2] put a timeout on the sleep, thus causing the
128 		 *      pagedaemon to only pause (rather than sleep forever)
129 		 *
130 		 * note that option [2] will only help us if we get lucky
131 		 * and some other process on the system breaks the deadlock
132 		 * by exiting or freeing memory (thus allowing the pagedaemon
133 		 * to continue).  for now we panic if DEBUG is defined,
134 		 * otherwise we hope for the best with option [2] (better
135 		 * yet, this should never happen in the first place!).
136 		 */
137 
138 		printf("pagedaemon: deadlock detected!\n");
139 		timo = hz >> 3;		/* set timeout */
140 #if defined(DEBUG)
141 		/* DEBUG: panic so we can debug it */
142 		panic("pagedaemon deadlock");
143 #endif
144 	}
145 
146 	simple_lock(&uvm.pagedaemon_lock);
147 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
148 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, wmsg,
149 	    timo);
150 
151 	splx(s);
152 }
153 
154 
155 /*
156  * uvmpd_tune: tune paging parameters
157  *
158  * => called when ever memory is added (or removed?) to the system
159  * => caller must call with page queues locked
160  */
161 
162 static void
163 uvmpd_tune()
164 {
165 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
166 
167 	uvmexp.freemin = uvmexp.npages / 30;
168 
169 	/* between 16k and 512k */
170 	/* XXX:  what are these values good for? */
171 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
172 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
173 
174 	/* Make sure there's always a user page free. */
175 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
176 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
177 
178 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
179 	if (uvmexp.freetarg <= uvmexp.freemin)
180 		uvmexp.freetarg = uvmexp.freemin + 1;
181 
182 	/* uvmexp.inactarg: computed in main daemon loop */
183 
184 	uvmexp.wiredmax = uvmexp.npages / 3;
185 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
186 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
187 }
188 
189 /*
190  * uvm_pageout: the main loop for the pagedaemon
191  */
192 
193 void
194 uvm_pageout(void *arg)
195 {
196 	int npages = 0;
197 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
198 
199 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
200 
201 	/*
202 	 * ensure correct priority and set paging parameters...
203 	 */
204 
205 	uvm.pagedaemon_proc = curproc;
206 	(void) spl0();
207 	uvm_lock_pageq();
208 	npages = uvmexp.npages;
209 	uvmpd_tune();
210 	uvm_unlock_pageq();
211 
212 	/*
213 	 * main loop
214 	 */
215 
216 	for (;;) {
217 		simple_lock(&uvm.pagedaemon_lock);
218 
219 		UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
220 		UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
221 		    &uvm.pagedaemon_lock, FALSE, "pgdaemon", 0);
222 		uvmexp.pdwoke++;
223 		UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
224 
225 		/* drain pool resources */
226 		pool_drain(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 #ifdef UBC
252 		if (uvmexp.free + uvmexp.paging < uvmexp.freetarg ||
253 		    uvmexp.inactive < uvmexp.inactarg ||
254 		    uvm_pgcnt_vnode >
255 		    (uvmexp.active + uvmexp.inactive + uvmexp.wired +
256 		     uvmexp.free) * 13 / 16) {
257 #else
258 		if (uvmexp.free < uvmexp.freetarg ||
259 		    uvmexp.inactive < uvmexp.inactarg) {
260 #endif
261 			uvmpd_scan();
262 		}
263 
264 		/*
265 		 * if there's any free memory to be had,
266 		 * wake up any waiters.
267 		 */
268 
269 		if (uvmexp.free > uvmexp.reserve_kernel ||
270 		    uvmexp.paging == 0) {
271 			wakeup(&uvmexp.free);
272 		}
273 
274 		/*
275 		 * scan done.  unlock page queues (the only lock we are holding)
276 		 */
277 
278 		uvm_unlock_pageq();
279 	}
280 	/*NOTREACHED*/
281 }
282 
283 
284 /*
285  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
286  */
287 
288 void
289 uvm_aiodone_daemon(void *arg)
290 {
291 	int s, free;
292 	struct buf *bp, *nbp;
293 	UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist);
294 
295 	for (;;) {
296 
297 		/*
298 		 * carefully attempt to go to sleep (without losing "wakeups"!).
299 		 * we need splbio because we want to make sure the aio_done list
300 		 * is totally empty before we go to sleep.
301 		 */
302 
303 		s = splbio();
304 		simple_lock(&uvm.aiodoned_lock);
305 		if (TAILQ_FIRST(&uvm.aio_done) == NULL) {
306 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
307 			UVM_UNLOCK_AND_WAIT(&uvm.aiodoned,
308 			    &uvm.aiodoned_lock, FALSE, "aiodoned", 0);
309 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
310 
311 			/* relock aiodoned_lock, still at splbio */
312 			simple_lock(&uvm.aiodoned_lock);
313 		}
314 
315 		/*
316 		 * check for done aio structures
317 		 */
318 
319 		bp = TAILQ_FIRST(&uvm.aio_done);
320 		if (bp) {
321 			TAILQ_INIT(&uvm.aio_done);
322 		}
323 
324 		simple_unlock(&uvm.aiodoned_lock);
325 		splx(s);
326 
327 		/*
328 		 * process each i/o that's done.
329 		 */
330 
331 		free = uvmexp.free;
332 		while (bp != NULL) {
333 			if (bp->b_flags & B_PDAEMON) {
334 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
335 			}
336 			nbp = TAILQ_NEXT(bp, b_freelist);
337 			s = splbio();	/* b_iodone must by called at splbio */
338 			(*bp->b_iodone)(bp);
339 			splx(s);
340 			bp = nbp;
341 		}
342 		if (free <= uvmexp.reserve_kernel) {
343 			s = uvm_lock_fpageq();
344 			wakeup(&uvm.pagedaemon);
345 			uvm_unlock_fpageq(s);
346 		} else {
347 			simple_lock(&uvm.pagedaemon_lock);
348 			wakeup(&uvmexp.free);
349 			simple_unlock(&uvm.pagedaemon_lock);
350 		}
351 	}
352 }
353 
354 
355 
356 /*
357  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
358  *
359  * => called with page queues locked
360  * => we work on meeting our free target by converting inactive pages
361  *    into free pages.
362  * => we handle the building of swap-backed clusters
363  * => we return TRUE if we are exiting because we met our target
364  */
365 
366 static boolean_t
367 uvmpd_scan_inactive(pglst)
368 	struct pglist *pglst;
369 {
370 	boolean_t retval = FALSE;	/* assume we haven't hit target */
371 	int s, free, result;
372 	struct vm_page *p, *nextpg;
373 	struct uvm_object *uobj;
374 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
375 	int npages;
376 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
377 	int swnpages, swcpages;				/* XXX: see below */
378 	int swslot;
379 	struct vm_anon *anon;
380 	boolean_t swap_backed;
381 	vaddr_t start;
382 	int dirtyreacts;
383 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
384 
385 	/*
386 	 * note: we currently keep swap-backed pages on a separate inactive
387 	 * list from object-backed pages.   however, merging the two lists
388 	 * back together again hasn't been ruled out.   thus, we keep our
389 	 * swap cluster in "swpps" rather than in pps (allows us to mix
390 	 * clustering types in the event of a mixed inactive queue).
391 	 */
392 
393 	/*
394 	 * swslot is non-zero if we are building a swap cluster.  we want
395 	 * to stay in the loop while we have a page to scan or we have
396 	 * a swap-cluster to build.
397 	 */
398 
399 	swslot = 0;
400 	swnpages = swcpages = 0;
401 	free = 0;
402 	dirtyreacts = 0;
403 
404 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
405 
406 		/*
407 		 * note that p can be NULL iff we have traversed the whole
408 		 * list and need to do one final swap-backed clustered pageout.
409 		 */
410 
411 		uobj = NULL;
412 		anon = NULL;
413 
414 		if (p) {
415 
416 			/*
417 			 * update our copy of "free" and see if we've met
418 			 * our target
419 			 */
420 
421 			s = uvm_lock_fpageq();
422 			free = uvmexp.free;
423 			uvm_unlock_fpageq(s);
424 
425 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
426 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
427 				UVMHIST_LOG(pdhist,"  met free target: "
428 					    "exit loop", 0, 0, 0, 0);
429 				retval = TRUE;
430 
431 				if (swslot == 0) {
432 					/* exit now if no swap-i/o pending */
433 					break;
434 				}
435 
436 				/* set p to null to signal final swap i/o */
437 				p = NULL;
438 			}
439 		}
440 
441 		if (p) {	/* if (we have a new page to consider) */
442 
443 			/*
444 			 * we are below target and have a new page to consider.
445 			 */
446 			uvmexp.pdscans++;
447 			nextpg = TAILQ_NEXT(p, pageq);
448 
449 			/*
450 			 * move referenced pages back to active queue and
451 			 * skip to next page (unlikely to happen since
452 			 * inactive pages shouldn't have any valid mappings
453 			 * and we cleared reference before deactivating).
454 			 */
455 
456 			if (pmap_is_referenced(p)) {
457 				uvm_pageactivate(p);
458 				uvmexp.pdreact++;
459 				continue;
460 			}
461 
462 			/*
463 			 * first we attempt to lock the object that this page
464 			 * belongs to.  if our attempt fails we skip on to
465 			 * the next page (no harm done).  it is important to
466 			 * "try" locking the object as we are locking in the
467 			 * wrong order (pageq -> object) and we don't want to
468 			 * deadlock.
469 			 *
470 			 * the only time we expect to see an ownerless page
471 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
472 			 * anon has loaned a page from a uvm_object and the
473 			 * uvm_object has dropped the ownership.  in that
474 			 * case, the anon can "take over" the loaned page
475 			 * and make it its own.
476 			 */
477 
478 			/* is page part of an anon or ownerless ? */
479 			if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
480 				anon = p->uanon;
481 				KASSERT(anon != NULL);
482 				if (!simple_lock_try(&anon->an_lock)) {
483 					/* lock failed, skip this page */
484 					continue;
485 				}
486 
487 				/*
488 				 * if the page is ownerless, claim it in the
489 				 * name of "anon"!
490 				 */
491 
492 				if ((p->pqflags & PQ_ANON) == 0) {
493 					KASSERT(p->loan_count > 0);
494 					p->loan_count--;
495 					p->pqflags |= PQ_ANON;
496 					/* anon now owns it */
497 				}
498 				if (p->flags & PG_BUSY) {
499 					simple_unlock(&anon->an_lock);
500 					uvmexp.pdbusy++;
501 					/* someone else owns page, skip it */
502 					continue;
503 				}
504 				uvmexp.pdanscan++;
505 			} else {
506 				uobj = p->uobject;
507 				KASSERT(uobj != NULL);
508 				if (!simple_lock_try(&uobj->vmobjlock)) {
509 					/* lock failed, skip this page */
510 					continue;
511 				}
512 				if (p->flags & PG_BUSY) {
513 					simple_unlock(&uobj->vmobjlock);
514 					uvmexp.pdbusy++;
515 					/* someone else owns page, skip it */
516 					continue;
517 				}
518 				uvmexp.pdobscan++;
519 			}
520 
521 			/*
522 			 * we now have the object and the page queues locked.
523 			 * the page is not busy.   if the page is clean we
524 			 * can free it now and continue.
525 			 */
526 
527 			if (p->flags & PG_CLEAN) {
528 				if (p->pqflags & PQ_SWAPBACKED) {
529 					/* this page now lives only in swap */
530 					simple_lock(&uvm.swap_data_lock);
531 					uvmexp.swpgonly++;
532 					simple_unlock(&uvm.swap_data_lock);
533 				}
534 
535 				/* zap all mappings with pmap_page_protect... */
536 				pmap_page_protect(p, VM_PROT_NONE);
537 				uvm_pagefree(p);
538 				uvmexp.pdfreed++;
539 
540 				if (anon) {
541 
542 					/*
543 					 * an anonymous page can only be clean
544 					 * if it has backing store assigned.
545 					 */
546 
547 					KASSERT(anon->an_swslot != 0);
548 
549 					/* remove from object */
550 					anon->u.an_page = NULL;
551 					simple_unlock(&anon->an_lock);
552 				} else {
553 					/* pagefree has already removed the
554 					 * page from the object */
555 					simple_unlock(&uobj->vmobjlock);
556 				}
557 				continue;
558 			}
559 
560 			/*
561 			 * this page is dirty, skip it if we'll have met our
562 			 * free target when all the current pageouts complete.
563 			 */
564 
565 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
566 				if (anon) {
567 					simple_unlock(&anon->an_lock);
568 				} else {
569 					simple_unlock(&uobj->vmobjlock);
570 				}
571 				continue;
572 			}
573 
574 			/*
575 			 * this page is dirty, but we can't page it out
576 			 * since all pages in swap are only in swap.
577 			 * reactivate it so that we eventually cycle
578 			 * all pages thru the inactive queue.
579 			 */
580 
581 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
582 			if ((p->pqflags & PQ_SWAPBACKED) &&
583 			    uvmexp.swpgonly == uvmexp.swpages) {
584 				dirtyreacts++;
585 				uvm_pageactivate(p);
586 				if (anon) {
587 					simple_unlock(&anon->an_lock);
588 				} else {
589 					simple_unlock(&uobj->vmobjlock);
590 				}
591 				continue;
592 			}
593 
594 			/*
595 			 * if the page is swap-backed and dirty and swap space
596 			 * is full, free any swap allocated to the page
597 			 * so that other pages can be paged out.
598 			 */
599 
600 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
601 			if ((p->pqflags & PQ_SWAPBACKED) &&
602 			    uvmexp.swpginuse == uvmexp.swpages) {
603 
604 				if ((p->pqflags & PQ_ANON) &&
605 				    p->uanon->an_swslot) {
606 					uvm_swap_free(p->uanon->an_swslot, 1);
607 					p->uanon->an_swslot = 0;
608 				}
609 				if (p->pqflags & PQ_AOBJ) {
610 					uao_dropswap(p->uobject,
611 						     p->offset >> PAGE_SHIFT);
612 				}
613 			}
614 
615 			/*
616 			 * the page we are looking at is dirty.   we must
617 			 * clean it before it can be freed.  to do this we
618 			 * first mark the page busy so that no one else will
619 			 * touch the page.   we write protect all the mappings
620 			 * of the page so that no one touches it while it is
621 			 * in I/O.
622 			 */
623 
624 			swap_backed = ((p->pqflags & PQ_SWAPBACKED) != 0);
625 			p->flags |= PG_BUSY;		/* now we own it */
626 			UVM_PAGE_OWN(p, "scan_inactive");
627 			pmap_page_protect(p, VM_PROT_READ);
628 			uvmexp.pgswapout++;
629 
630 			/*
631 			 * for swap-backed pages we need to (re)allocate
632 			 * swap space.
633 			 */
634 
635 			if (swap_backed) {
636 
637 				/*
638 				 * free old swap slot (if any)
639 				 */
640 
641 				if (anon) {
642 					if (anon->an_swslot) {
643 						uvm_swap_free(anon->an_swslot,
644 						    1);
645 						anon->an_swslot = 0;
646 					}
647 				} else {
648 					uao_dropswap(uobj,
649 						     p->offset >> PAGE_SHIFT);
650 				}
651 
652 				/*
653 				 * start new cluster (if necessary)
654 				 */
655 
656 				if (swslot == 0) {
657 					swnpages = MAXBSIZE >> PAGE_SHIFT;
658 					swslot = uvm_swap_alloc(&swnpages,
659 					    TRUE);
660 					if (swslot == 0) {
661 						/* no swap?  give up! */
662 						p->flags &= ~PG_BUSY;
663 						UVM_PAGE_OWN(p, NULL);
664 						if (anon)
665 							simple_unlock(
666 							    &anon->an_lock);
667 						else
668 							simple_unlock(
669 							    &uobj->vmobjlock);
670 						continue;
671 					}
672 					swcpages = 0;	/* cluster is empty */
673 				}
674 
675 				/*
676 				 * add block to cluster
677 				 */
678 
679 				swpps[swcpages] = p;
680 				if (anon)
681 					anon->an_swslot = swslot + swcpages;
682 				else
683 					uao_set_swslot(uobj,
684 					    p->offset >> PAGE_SHIFT,
685 					    swslot + swcpages);
686 				swcpages++;
687 			}
688 		} else {
689 
690 			/* if p == NULL we must be doing a last swap i/o */
691 			swap_backed = TRUE;
692 		}
693 
694 		/*
695 		 * now consider doing the pageout.
696 		 *
697 		 * for swap-backed pages, we do the pageout if we have either
698 		 * filled the cluster (in which case (swnpages == swcpages) or
699 		 * run out of pages (p == NULL).
700 		 *
701 		 * for object pages, we always do the pageout.
702 		 */
703 
704 		if (swap_backed) {
705 			if (p) {	/* if we just added a page to cluster */
706 				if (anon)
707 					simple_unlock(&anon->an_lock);
708 				else
709 					simple_unlock(&uobj->vmobjlock);
710 
711 				/* cluster not full yet? */
712 				if (swcpages < swnpages)
713 					continue;
714 			}
715 
716 			/* starting I/O now... set up for it */
717 			npages = swcpages;
718 			ppsp = swpps;
719 			/* for swap-backed pages only */
720 			start = (vaddr_t) swslot;
721 
722 			/* if this is final pageout we could have a few
723 			 * extra swap blocks */
724 			if (swcpages < swnpages) {
725 				uvm_swap_free(swslot + swcpages,
726 				    (swnpages - swcpages));
727 			}
728 		} else {
729 			/* normal object pageout */
730 			ppsp = pps;
731 			npages = sizeof(pps) / sizeof(struct vm_page *);
732 			/* not looked at because PGO_ALLPAGES is set */
733 			start = 0;
734 		}
735 
736 		/*
737 		 * now do the pageout.
738 		 *
739 		 * for swap_backed pages we have already built the cluster.
740 		 * for !swap_backed pages, uvm_pager_put will call the object's
741 		 * "make put cluster" function to build a cluster on our behalf.
742 		 *
743 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
744 		 * it to free the cluster pages for us on a successful I/O (it
745 		 * always does this for un-successful I/O requests).  this
746 		 * allows us to do clustered pageout without having to deal
747 		 * with cluster pages at this level.
748 		 *
749 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
750 		 *  IN: locked: uobj (if !swap_backed), page queues
751 		 * OUT: locked: uobj (if !swap_backed && result !=VM_PAGER_PEND)
752 		 *     !locked: pageqs, uobj (if swap_backed || VM_PAGER_PEND)
753 		 *
754 		 * [the bit about VM_PAGER_PEND saves us one lock-unlock pair]
755 		 */
756 
757 		/* locked: uobj (if !swap_backed), page queues */
758 		uvmexp.pdpageouts++;
759 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
760 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
761 		/* locked: uobj (if !swap_backed && result != PEND) */
762 		/* unlocked: pageqs, object (if swap_backed ||result == PEND) */
763 
764 		/*
765 		 * if we did i/o to swap, zero swslot to indicate that we are
766 		 * no longer building a swap-backed cluster.
767 		 */
768 
769 		if (swap_backed)
770 			swslot = 0;		/* done with this cluster */
771 
772 		/*
773 		 * first, we check for VM_PAGER_PEND which means that the
774 		 * async I/O is in progress and the async I/O done routine
775 		 * will clean up after us.   in this case we move on to the
776 		 * next page.
777 		 *
778 		 * there is a very remote chance that the pending async i/o can
779 		 * finish _before_ we get here.   if that happens, our page "p"
780 		 * may no longer be on the inactive queue.   so we verify this
781 		 * when determining the next page (starting over at the head if
782 		 * we've lost our inactive page).
783 		 */
784 
785 		if (result == VM_PAGER_PEND) {
786 			uvmexp.paging += npages;
787 			uvm_lock_pageq();
788 			uvmexp.pdpending++;
789 			if (p) {
790 				if (p->pqflags & PQ_INACTIVE)
791 					nextpg = TAILQ_NEXT(p, pageq);
792 				else
793 					nextpg = TAILQ_FIRST(pglst);
794 			} else {
795 				nextpg = NULL;
796 			}
797 			continue;
798 		}
799 
800 #ifdef UBC
801 		if (result == VM_PAGER_ERROR &&
802 		    curproc == uvm.pagedaemon_proc) {
803 			uvm_lock_pageq();
804 			nextpg = TAILQ_NEXT(p, pageq);
805 			uvm_pageactivate(p);
806 			continue;
807 		}
808 #endif
809 
810 		/*
811 		 * clean up "p" if we have one
812 		 */
813 
814 		if (p) {
815 			/*
816 			 * the I/O request to "p" is done and uvm_pager_put
817 			 * has freed any cluster pages it may have allocated
818 			 * during I/O.  all that is left for us to do is
819 			 * clean up page "p" (which is still PG_BUSY).
820 			 *
821 			 * our result could be one of the following:
822 			 *   VM_PAGER_OK: successful pageout
823 			 *
824 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
825 			 *     to next page
826 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
827 			 *     "reactivate" page to get it out of the way (it
828 			 *     will eventually drift back into the inactive
829 			 *     queue for a retry).
830 			 *   VM_PAGER_UNLOCK: should never see this as it is
831 			 *     only valid for "get" operations
832 			 */
833 
834 			/* relock p's object: page queues not lock yet, so
835 			 * no need for "try" */
836 
837 			/* !swap_backed case: already locked... */
838 			if (swap_backed) {
839 				if (anon)
840 					simple_lock(&anon->an_lock);
841 				else
842 					simple_lock(&uobj->vmobjlock);
843 			}
844 
845 #ifdef DIAGNOSTIC
846 			if (result == VM_PAGER_UNLOCK)
847 				panic("pagedaemon: pageout returned "
848 				    "invalid 'unlock' code");
849 #endif
850 
851 			/* handle PG_WANTED now */
852 			if (p->flags & PG_WANTED)
853 				/* still holding object lock */
854 				wakeup(p);
855 
856 			p->flags &= ~(PG_BUSY|PG_WANTED);
857 			UVM_PAGE_OWN(p, NULL);
858 
859 			/* released during I/O? */
860 			if (p->flags & PG_RELEASED) {
861 				if (anon) {
862 					/* remove page so we can get nextpg */
863 					anon->u.an_page = NULL;
864 
865 					simple_unlock(&anon->an_lock);
866 					uvm_anfree(anon);	/* kills anon */
867 					pmap_page_protect(p, VM_PROT_NONE);
868 					anon = NULL;
869 					uvm_lock_pageq();
870 					nextpg = TAILQ_NEXT(p, pageq);
871 					/* free released page */
872 					uvm_pagefree(p);
873 
874 				} else {
875 
876 					/*
877 					 * pgo_releasepg nukes the page and
878 					 * gets "nextpg" for us.  it returns
879 					 * with the page queues locked (when
880 					 * given nextpg ptr).
881 					 */
882 
883 					if (!uobj->pgops->pgo_releasepg(p,
884 					    &nextpg))
885 						/* uobj died after release */
886 						uobj = NULL;
887 
888 					/*
889 					 * lock page queues here so that they're
890 					 * always locked at the end of the loop.
891 					 */
892 
893 					uvm_lock_pageq();
894 				}
895 			} else {	/* page was not released during I/O */
896 				uvm_lock_pageq();
897 				nextpg = TAILQ_NEXT(p, pageq);
898 				if (result != VM_PAGER_OK) {
899 					/* pageout was a failure... */
900 					if (result != VM_PAGER_AGAIN)
901 						uvm_pageactivate(p);
902 					pmap_clear_reference(p);
903 					/* XXXCDC: if (swap_backed) FREE p's
904 					 * swap block? */
905 				} else {
906 					/* pageout was a success... */
907 					pmap_clear_reference(p);
908 					pmap_clear_modify(p);
909 					p->flags |= PG_CLEAN;
910 				}
911 			}
912 
913 			/*
914 			 * drop object lock (if there is an object left).   do
915 			 * a safety check of nextpg to make sure it is on the
916 			 * inactive queue (it should be since PG_BUSY pages on
917 			 * the inactive queue can't be re-queued [note: not
918 			 * true for active queue]).
919 			 */
920 
921 			if (anon)
922 				simple_unlock(&anon->an_lock);
923 			else if (uobj)
924 				simple_unlock(&uobj->vmobjlock);
925 
926 		} else {
927 
928 			/*
929 			 * if p is null in this loop, make sure it stays null
930 			 * in the next loop.
931 			 */
932 
933 			nextpg = NULL;
934 
935 			/*
936 			 * lock page queues here just so they're always locked
937 			 * at the end of the loop.
938 			 */
939 
940 			uvm_lock_pageq();
941 		}
942 
943 		if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) {
944 			nextpg = TAILQ_FIRST(pglst);	/* reload! */
945 		}
946 	}
947 	return (retval);
948 }
949 
950 /*
951  * uvmpd_scan: scan the page queues and attempt to meet our targets.
952  *
953  * => called with pageq's locked
954  */
955 
956 void
957 uvmpd_scan()
958 {
959 	int s, free, inactive_shortage, swap_shortage, pages_freed;
960 	struct vm_page *p, *nextpg;
961 	struct uvm_object *uobj;
962 	boolean_t got_it;
963 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
964 
965 	uvmexp.pdrevs++;		/* counter */
966 	uobj = NULL;
967 
968 	/*
969 	 * get current "free" page count
970 	 */
971 	s = uvm_lock_fpageq();
972 	free = uvmexp.free;
973 	uvm_unlock_fpageq(s);
974 
975 #ifndef __SWAP_BROKEN
976 	/*
977 	 * swap out some processes if we are below our free target.
978 	 * we need to unlock the page queues for this.
979 	 */
980 	if (free < uvmexp.freetarg) {
981 		uvmexp.pdswout++;
982 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout", free,
983 		    uvmexp.freetarg, 0, 0);
984 		uvm_unlock_pageq();
985 		uvm_swapout_threads();
986 		uvm_lock_pageq();
987 
988 	}
989 #endif
990 
991 	/*
992 	 * now we want to work on meeting our targets.   first we work on our
993 	 * free target by converting inactive pages into free pages.  then
994 	 * we work on meeting our inactive target by converting active pages
995 	 * to inactive ones.
996 	 */
997 
998 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
999 
1000 	/*
1001 	 * alternate starting queue between swap and object based on the
1002 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
1003 	 */
1004 
1005 	got_it = FALSE;
1006 	pages_freed = uvmexp.pdfreed;
1007 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
1008 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
1009 	if (!got_it)
1010 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
1011 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
1012 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
1013 	pages_freed = uvmexp.pdfreed - pages_freed;
1014 
1015 	/*
1016 	 * we have done the scan to get free pages.   now we work on meeting
1017 	 * our inactive target.
1018 	 */
1019 
1020 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
1021 
1022 	/*
1023 	 * detect if we're not going to be able to page anything out
1024 	 * until we free some swap resources from active pages.
1025 	 */
1026 
1027 	swap_shortage = 0;
1028 	if (uvmexp.free < uvmexp.freetarg &&
1029 	    uvmexp.swpginuse == uvmexp.swpages &&
1030 	    uvmexp.swpgonly < uvmexp.swpages &&
1031 	    pages_freed == 0) {
1032 		swap_shortage = uvmexp.freetarg - uvmexp.free;
1033 	}
1034 
1035 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
1036 		    inactive_shortage, swap_shortage,0,0);
1037 	for (p = TAILQ_FIRST(&uvm.page_active);
1038 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
1039 	     p = nextpg) {
1040 		nextpg = TAILQ_NEXT(p, pageq);
1041 		if (p->flags & PG_BUSY)
1042 			continue;	/* quick check before trying to lock */
1043 
1044 		/*
1045 		 * lock the page's owner.
1046 		 */
1047 		/* is page anon owned or ownerless? */
1048 		if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
1049 			KASSERT(p->uanon != NULL);
1050 			if (!simple_lock_try(&p->uanon->an_lock))
1051 				continue;
1052 
1053 			/* take over the page? */
1054 			if ((p->pqflags & PQ_ANON) == 0) {
1055 				KASSERT(p->loan_count > 0);
1056 				p->loan_count--;
1057 				p->pqflags |= PQ_ANON;
1058 			}
1059 		} else {
1060 			if (!simple_lock_try(&p->uobject->vmobjlock))
1061 				continue;
1062 		}
1063 
1064 		/*
1065 		 * skip this page if it's busy.
1066 		 */
1067 
1068 		if ((p->flags & PG_BUSY) != 0) {
1069 			if (p->pqflags & PQ_ANON)
1070 				simple_unlock(&p->uanon->an_lock);
1071 			else
1072 				simple_unlock(&p->uobject->vmobjlock);
1073 			continue;
1074 		}
1075 
1076 		/*
1077 		 * if there's a shortage of swap, free any swap allocated
1078 		 * to this page so that other pages can be paged out.
1079 		 */
1080 
1081 		if (swap_shortage > 0) {
1082 			if ((p->pqflags & PQ_ANON) && p->uanon->an_swslot) {
1083 				uvm_swap_free(p->uanon->an_swslot, 1);
1084 				p->uanon->an_swslot = 0;
1085 				p->flags &= ~PG_CLEAN;
1086 				swap_shortage--;
1087 			}
1088 			if (p->pqflags & PQ_AOBJ) {
1089 				int slot = uao_set_swslot(p->uobject,
1090 					p->offset >> PAGE_SHIFT, 0);
1091 				if (slot) {
1092 					uvm_swap_free(slot, 1);
1093 					p->flags &= ~PG_CLEAN;
1094 					swap_shortage--;
1095 				}
1096 			}
1097 		}
1098 
1099 		/*
1100 		 * deactivate this page if there's a shortage of
1101 		 * inactive pages.
1102 		 */
1103 
1104 		if (inactive_shortage > 0) {
1105 			pmap_page_protect(p, VM_PROT_NONE);
1106 			/* no need to check wire_count as pg is "active" */
1107 			uvm_pagedeactivate(p);
1108 			uvmexp.pddeact++;
1109 			inactive_shortage--;
1110 		}
1111 		if (p->pqflags & PQ_ANON)
1112 			simple_unlock(&p->uanon->an_lock);
1113 		else
1114 			simple_unlock(&p->uobject->vmobjlock);
1115 	}
1116 }
1117