xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 481fca6e59249d8ffcf24fef7cfbe7b131bfb080)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.21 2000/06/27 17:29:33 mrg 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 #include "opt_uvmhist.h"
70 
71 /*
72  * uvm_pdaemon.c: the page daemon
73  */
74 
75 #include <sys/param.h>
76 #include <sys/proc.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/pool.h>
80 
81 #include <uvm/uvm.h>
82 
83 /*
84  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedeamon will reactivate
85  * in a pass thru the inactive list when swap is full.  the value should be
86  * "small"... if it's too large we'll cycle the active pages thru the inactive
87  * queue too quickly to for them to be referenced and avoid being freed.
88  */
89 
90 #define UVMPD_NUMDIRTYREACTS 16
91 
92 
93 /*
94  * local prototypes
95  */
96 
97 static void		uvmpd_scan __P((void));
98 static boolean_t	uvmpd_scan_inactive __P((struct pglist *));
99 static void		uvmpd_tune __P((void));
100 
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 / 20;
168 
169 	/* between 16k and 256k */
170 	/* XXX:  what are these values good for? */
171 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
172 	uvmexp.freemin = min(uvmexp.freemin, (256*1024) >> PAGE_SHIFT);
173 
174 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
175 	if (uvmexp.freetarg <= uvmexp.freemin)
176 		uvmexp.freetarg = uvmexp.freemin + 1;
177 
178 	/* uvmexp.inactarg: computed in main daemon loop */
179 
180 	uvmexp.wiredmax = uvmexp.npages / 3;
181 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
182 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
183 }
184 
185 /*
186  * uvm_pageout: the main loop for the pagedaemon
187  */
188 
189 void
190 uvm_pageout()
191 {
192 	int npages = 0;
193 	int s;
194 	struct uvm_aiodesc *aio, *nextaio;
195 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
196 
197 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
198 
199 	/*
200 	 * ensure correct priority and set paging parameters...
201 	 */
202 
203 	uvm.pagedaemon_proc = curproc;
204 	(void) spl0();
205 	uvm_lock_pageq();
206 	npages = uvmexp.npages;
207 	uvmpd_tune();
208 	uvm_unlock_pageq();
209 
210 	/*
211 	 * main loop
212 	 */
213 	while (TRUE) {
214 
215 		/*
216 		 * carefully attempt to go to sleep (without losing "wakeups"!).
217 		 * we need splbio because we want to make sure the aio_done list
218 		 * is totally empty before we go to sleep.
219 		 */
220 
221 		s = splbio();
222 		simple_lock(&uvm.pagedaemon_lock);
223 
224 		/*
225 		 * if we've got done aio's, then bypass the sleep
226 		 */
227 
228 		if (uvm.aio_done.tqh_first == NULL) {
229 			UVMHIST_LOG(maphist,"  <<SLEEPING>>",0,0,0,0);
230 			UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
231 			    &uvm.pagedaemon_lock, FALSE, "daemon_slp", 0);
232 			uvmexp.pdwoke++;
233 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
234 
235 			/* relock pagedaemon_lock, still at splbio */
236 			simple_lock(&uvm.pagedaemon_lock);
237 		}
238 
239 		/*
240 		 * check for done aio structures
241 		 */
242 
243 		aio = uvm.aio_done.tqh_first;	/* save current list (if any)*/
244 		if (aio) {
245 			TAILQ_INIT(&uvm.aio_done);	/* zero global list */
246 		}
247 
248 		simple_unlock(&uvm.pagedaemon_lock);	/* unlock */
249 		splx(s);				/* drop splbio */
250 
251 		/*
252 		 * first clear out any pending aios (to free space in case we
253 		 * want to pageout more stuff).
254 		 */
255 
256 		for (/*null*/; aio != NULL ; aio = nextaio) {
257 
258 			uvmexp.paging -= aio->npages;
259 			nextaio = aio->aioq.tqe_next;
260 			aio->aiodone(aio);
261 
262 		}
263 
264 		/* Next, drain pool resources */
265 		pool_drain(0);
266 
267 		/*
268 		 * now lock page queues and recompute inactive count
269 		 */
270 		uvm_lock_pageq();
271 
272 		if (npages != uvmexp.npages) {	/* check for new pages? */
273 			npages = uvmexp.npages;
274 			uvmpd_tune();
275 		}
276 
277 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
278 		if (uvmexp.inactarg <= uvmexp.freetarg)
279 			uvmexp.inactarg = uvmexp.freetarg + 1;
280 
281 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d, inact/itarg=%d/%d",
282 		    uvmexp.free, uvmexp.freetarg, uvmexp.inactive,
283 		    uvmexp.inactarg);
284 
285 		/*
286 		 * scan if needed
287 		 * [XXX: note we are reading uvm.free without locking]
288 		 */
289 		if (uvmexp.free < uvmexp.freetarg ||
290 		    uvmexp.inactive < uvmexp.inactarg)
291 			uvmpd_scan();
292 
293 		/*
294 		 * done scan.  unlock page queues (the only lock we are holding)
295 		 */
296 		uvm_unlock_pageq();
297 
298 		/*
299 		 * done!    restart loop.
300 		 */
301 		if (uvmexp.free > uvmexp.reserve_kernel ||
302 		    uvmexp.paging == 0)
303 			wakeup(&uvmexp.free);
304 	}
305 	/*NOTREACHED*/
306 }
307 
308 /*
309  * uvmpd_scan_inactive: the first loop of uvmpd_scan broken out into
310  * 	its own function for ease of reading.
311  *
312  * => called with page queues locked
313  * => we work on meeting our free target by converting inactive pages
314  *    into free pages.
315  * => we handle the building of swap-backed clusters
316  * => we return TRUE if we are exiting because we met our target
317  */
318 
319 static boolean_t
320 uvmpd_scan_inactive(pglst)
321 	struct pglist *pglst;
322 {
323 	boolean_t retval = FALSE;	/* assume we haven't hit target */
324 	int s, free, result;
325 	struct vm_page *p, *nextpg;
326 	struct uvm_object *uobj;
327 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
328 	int npages;
329 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
330 	int swnpages, swcpages;				/* XXX: see below */
331 	int swslot;
332 	struct vm_anon *anon;
333 	boolean_t swap_backed;
334 	vaddr_t start;
335 	int dirtyreacts;
336 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
337 
338 	/*
339 	 * note: we currently keep swap-backed pages on a seperate inactive
340 	 * list from object-backed pages.   however, merging the two lists
341 	 * back together again hasn't been ruled out.   thus, we keep our
342 	 * swap cluster in "swpps" rather than in pps (allows us to mix
343 	 * clustering types in the event of a mixed inactive queue).
344 	 */
345 
346 	/*
347 	 * swslot is non-zero if we are building a swap cluster.  we want
348 	 * to stay in the loop while we have a page to scan or we have
349 	 * a swap-cluster to build.
350 	 */
351 	swslot = 0;
352 	swnpages = swcpages = 0;
353 	free = 0;
354 	dirtyreacts = 0;
355 
356 	for (p = pglst->tqh_first ; p != NULL || swslot != 0 ; p = nextpg) {
357 
358 		/*
359 		 * note that p can be NULL iff we have traversed the whole
360 		 * list and need to do one final swap-backed clustered pageout.
361 		 */
362 		if (p) {
363 			/*
364 			 * update our copy of "free" and see if we've met
365 			 * our target
366 			 */
367 			s = uvm_lock_fpageq();
368 			free = uvmexp.free;
369 			uvm_unlock_fpageq(s);
370 
371 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
372 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
373 				UVMHIST_LOG(pdhist,"  met free target: "
374 				    "exit loop", 0, 0, 0, 0);
375 				retval = TRUE;		/* hit the target! */
376 
377 				if (swslot == 0)
378 					/* exit now if no swap-i/o pending */
379 					break;
380 
381 				/* set p to null to signal final swap i/o */
382 				p = NULL;
383 			}
384 		}
385 
386 		uobj = NULL;	/* be safe and shut gcc up */
387 		anon = NULL;	/* be safe and shut gcc up */
388 
389 		if (p) {	/* if (we have a new page to consider) */
390 			/*
391 			 * we are below target and have a new page to consider.
392 			 */
393 			uvmexp.pdscans++;
394 			nextpg = p->pageq.tqe_next;
395 
396 			/*
397 			 * move referenced pages back to active queue and
398 			 * skip to next page (unlikely to happen since
399 			 * inactive pages shouldn't have any valid mappings
400 			 * and we cleared reference before deactivating).
401 			 */
402 			if (pmap_is_referenced(p)) {
403 				uvm_pageactivate(p);
404 				uvmexp.pdreact++;
405 				continue;
406 			}
407 
408 			/*
409 			 * first we attempt to lock the object that this page
410 			 * belongs to.  if our attempt fails we skip on to
411 			 * the next page (no harm done).  it is important to
412 			 * "try" locking the object as we are locking in the
413 			 * wrong order (pageq -> object) and we don't want to
414 			 * get deadlocked.
415 			 *
416 			 * the only time we exepct to see an ownerless page
417 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
418 			 * anon has loaned a page from a uvm_object and the
419 			 * uvm_object has dropped the ownership.  in that
420 			 * case, the anon can "take over" the loaned page
421 			 * and make it its own.
422 			 */
423 
424 			/* is page part of an anon or ownerless ? */
425 			if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
426 
427 				anon = p->uanon;
428 
429 #ifdef DIAGNOSTIC
430 				/* to be on inactive q, page must be part
431 				 * of _something_ */
432 				if (anon == NULL)
433 					panic("pagedaemon: page with no anon "
434 					    "or object detected - loop 1");
435 #endif
436 
437 				if (!simple_lock_try(&anon->an_lock))
438 					/* lock failed, skip this page */
439 					continue;
440 
441 				/*
442 				 * if the page is ownerless, claim it in the
443 				 * name of "anon"!
444 				 */
445 				if ((p->pqflags & PQ_ANON) == 0) {
446 #ifdef DIAGNOSTIC
447 					if (p->loan_count < 1)
448 						panic("pagedaemon: non-loaned "
449 						    "ownerless page detected -"
450 						    " loop 1");
451 #endif
452 					p->loan_count--;
453 					p->pqflags |= PQ_ANON;      /* anon now owns it */
454 				}
455 
456 				if (p->flags & PG_BUSY) {
457 					simple_unlock(&anon->an_lock);
458 					uvmexp.pdbusy++;
459 					/* someone else owns page, skip it */
460 					continue;
461 				}
462 
463 				uvmexp.pdanscan++;
464 
465 			} else {
466 
467 				uobj = p->uobject;
468 
469 				if (!simple_lock_try(&uobj->vmobjlock))
470 					/* lock failed, skip this page */
471 					continue;
472 
473 				if (p->flags & PG_BUSY) {
474 					simple_unlock(&uobj->vmobjlock);
475 					uvmexp.pdbusy++;
476 					/* someone else owns page, skip it */
477 					continue;
478 				}
479 
480 				uvmexp.pdobscan++;
481 			}
482 
483 			/*
484 			 * we now have the object and the page queues locked.
485 			 * the page is not busy.   if the page is clean we
486 			 * can free it now and continue.
487 			 */
488 
489 			if (p->flags & PG_CLEAN) {
490 				if (p->pqflags & PQ_SWAPBACKED) {
491 					/* this page now lives only in swap */
492 					simple_lock(&uvm.swap_data_lock);
493 					uvmexp.swpgonly++;
494 					simple_unlock(&uvm.swap_data_lock);
495 				}
496 
497 				/* zap all mappings with pmap_page_protect... */
498 				pmap_page_protect(p, VM_PROT_NONE);
499 				uvm_pagefree(p);
500 				uvmexp.pdfreed++;
501 
502 				if (anon) {
503 #ifdef DIAGNOSTIC
504 					/*
505 					 * an anonymous page can only be clean
506 					 * if it has valid backing store.
507 					 */
508 					if (anon->an_swslot == 0)
509 						panic("pagedaemon: clean anon "
510 						 "page without backing store?");
511 #endif
512 					/* remove from object */
513 					anon->u.an_page = NULL;
514 					simple_unlock(&anon->an_lock);
515 				} else {
516 					/* pagefree has already removed the
517 					 * page from the object */
518 					simple_unlock(&uobj->vmobjlock);
519 				}
520 				continue;
521 			}
522 
523 			/*
524 			 * this page is dirty, skip it if we'll have met our
525 			 * free target when all the current pageouts complete.
526 			 */
527 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
528 				if (anon) {
529 					simple_unlock(&anon->an_lock);
530 				} else {
531 					simple_unlock(&uobj->vmobjlock);
532 				}
533 				continue;
534 			}
535 
536 			/*
537 			 * this page is dirty, but we can't page it out
538 			 * since all pages in swap are only in swap.
539 			 * reactivate it so that we eventually cycle
540 			 * all pages thru the inactive queue.
541 			 */
542 #ifdef DIAGNOSTIC
543 			if (uvmexp.swpgonly > uvmexp.swpages) {
544 				panic("uvmexp.swpgonly botch");
545 			}
546 #endif
547 			if ((p->pqflags & PQ_SWAPBACKED) &&
548 			    uvmexp.swpgonly == uvmexp.swpages) {
549 				dirtyreacts++;
550 				uvm_pageactivate(p);
551 				if (anon) {
552 					simple_unlock(&anon->an_lock);
553 				} else {
554 					simple_unlock(&uobj->vmobjlock);
555 				}
556 				continue;
557 			}
558 
559 			/*
560 			 * if the page is swap-backed and dirty and swap space
561 			 * is full, free any swap allocated to the page
562 			 * so that other pages can be paged out.
563 			 */
564 #ifdef DIAGNOSTIC
565 			if (uvmexp.swpginuse > uvmexp.swpages) {
566 				panic("uvmexp.swpginuse botch");
567 			}
568 #endif
569 			if ((p->pqflags & PQ_SWAPBACKED) &&
570 			    uvmexp.swpginuse == uvmexp.swpages) {
571 
572 				if ((p->pqflags & PQ_ANON) &&
573 				    p->uanon->an_swslot) {
574 					uvm_swap_free(p->uanon->an_swslot, 1);
575 					p->uanon->an_swslot = 0;
576 				}
577 				if (p->pqflags & PQ_AOBJ) {
578 					uao_dropswap(p->uobject,
579 						     p->offset >> PAGE_SHIFT);
580 				}
581 			}
582 
583 			/*
584 			 * the page we are looking at is dirty.   we must
585 			 * clean it before it can be freed.  to do this we
586 			 * first mark the page busy so that no one else will
587 			 * touch the page.   we write protect all the mappings
588 			 * of the page so that no one touches it while it is
589 			 * in I/O.
590 			 */
591 
592 			swap_backed = ((p->pqflags & PQ_SWAPBACKED) != 0);
593 			p->flags |= PG_BUSY;		/* now we own it */
594 			UVM_PAGE_OWN(p, "scan_inactive");
595 			pmap_page_protect(p, VM_PROT_READ);
596 			uvmexp.pgswapout++;
597 
598 			/*
599 			 * for swap-backed pages we need to (re)allocate
600 			 * swap space.
601 			 */
602 			if (swap_backed) {
603 
604 				/*
605 				 * free old swap slot (if any)
606 				 */
607 				if (anon) {
608 					if (anon->an_swslot) {
609 						uvm_swap_free(anon->an_swslot,
610 						    1);
611 						anon->an_swslot = 0;
612 					}
613 				} else {
614 					uao_dropswap(uobj,
615 						     p->offset >> PAGE_SHIFT);
616 				}
617 
618 				/*
619 				 * start new cluster (if necessary)
620 				 */
621 				if (swslot == 0) {
622 					/* want this much */
623 					swnpages = MAXBSIZE >> PAGE_SHIFT;
624 
625 					swslot = uvm_swap_alloc(&swnpages,
626 					    TRUE);
627 
628 					if (swslot == 0) {
629 						/* no swap?  give up! */
630 						p->flags &= ~PG_BUSY;
631 						UVM_PAGE_OWN(p, NULL);
632 						if (anon)
633 							simple_unlock(
634 							    &anon->an_lock);
635 						else
636 							simple_unlock(
637 							    &uobj->vmobjlock);
638 						continue;
639 					}
640 					swcpages = 0;	/* cluster is empty */
641 				}
642 
643 				/*
644 				 * add block to cluster
645 				 */
646 				swpps[swcpages] = p;
647 				if (anon)
648 					anon->an_swslot = swslot + swcpages;
649 				else
650 					uao_set_swslot(uobj,
651 					    p->offset >> PAGE_SHIFT,
652 					    swslot + swcpages);
653 				swcpages++;
654 
655 				/* done (swap-backed) */
656 			}
657 
658 			/* end: if (p) ["if we have new page to consider"] */
659 		} else {
660 
661 			/* if p == NULL we must be doing a last swap i/o */
662 			swap_backed = TRUE;
663 		}
664 
665 		/*
666 		 * now consider doing the pageout.
667 		 *
668 		 * for swap-backed pages, we do the pageout if we have either
669 		 * filled the cluster (in which case (swnpages == swcpages) or
670 		 * run out of pages (p == NULL).
671 		 *
672 		 * for object pages, we always do the pageout.
673 		 */
674 		if (swap_backed) {
675 
676 			if (p) {	/* if we just added a page to cluster */
677 				if (anon)
678 					simple_unlock(&anon->an_lock);
679 				else
680 					simple_unlock(&uobj->vmobjlock);
681 
682 				/* cluster not full yet? */
683 				if (swcpages < swnpages)
684 					continue;
685 			}
686 
687 			/* starting I/O now... set up for it */
688 			npages = swcpages;
689 			ppsp = swpps;
690 			/* for swap-backed pages only */
691 			start = (vaddr_t) swslot;
692 
693 			/* if this is final pageout we could have a few
694 			 * extra swap blocks */
695 			if (swcpages < swnpages) {
696 				uvm_swap_free(swslot + swcpages,
697 				    (swnpages - swcpages));
698 			}
699 
700 		} else {
701 
702 			/* normal object pageout */
703 			ppsp = pps;
704 			npages = sizeof(pps) / sizeof(struct vm_page *);
705 			/* not looked at because PGO_ALLPAGES is set */
706 			start = 0;
707 
708 		}
709 
710 		/*
711 		 * now do the pageout.
712 		 *
713 		 * for swap_backed pages we have already built the cluster.
714 		 * for !swap_backed pages, uvm_pager_put will call the object's
715 		 * "make put cluster" function to build a cluster on our behalf.
716 		 *
717 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
718 		 * it to free the cluster pages for us on a successful I/O (it
719 		 * always does this for un-successful I/O requests).  this
720 		 * allows us to do clustered pageout without having to deal
721 		 * with cluster pages at this level.
722 		 *
723 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
724 		 *  IN: locked: uobj (if !swap_backed), page queues
725 		 * OUT: locked: uobj (if !swap_backed && result !=VM_PAGER_PEND)
726 		 *     !locked: pageqs, uobj (if swap_backed || VM_PAGER_PEND)
727 		 *
728 		 * [the bit about VM_PAGER_PEND saves us one lock-unlock pair]
729 		 */
730 
731 		/* locked: uobj (if !swap_backed), page queues */
732 		uvmexp.pdpageouts++;
733 		result = uvm_pager_put((swap_backed) ? NULL : uobj, p,
734 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
735 		/* locked: uobj (if !swap_backed && result != PEND) */
736 		/* unlocked: pageqs, object (if swap_backed ||result == PEND) */
737 
738 		/*
739 		 * if we did i/o to swap, zero swslot to indicate that we are
740 		 * no longer building a swap-backed cluster.
741 		 */
742 
743 		if (swap_backed)
744 			swslot = 0;		/* done with this cluster */
745 
746 		/*
747 		 * first, we check for VM_PAGER_PEND which means that the
748 		 * async I/O is in progress and the async I/O done routine
749 		 * will clean up after us.   in this case we move on to the
750 		 * next page.
751 		 *
752 		 * there is a very remote chance that the pending async i/o can
753 		 * finish _before_ we get here.   if that happens, our page "p"
754 		 * may no longer be on the inactive queue.   so we verify this
755 		 * when determining the next page (starting over at the head if
756 		 * we've lost our inactive page).
757 		 */
758 
759 		if (result == VM_PAGER_PEND) {
760 			uvmexp.paging += npages;
761 			uvm_lock_pageq();		/* relock page queues */
762 			uvmexp.pdpending++;
763 			if (p) {
764 				if (p->pqflags & PQ_INACTIVE)
765 					/* reload! */
766 					nextpg = p->pageq.tqe_next;
767 				else
768 					/* reload! */
769 					nextpg = pglst->tqh_first;
770 				} else {
771 					nextpg = NULL;		/* done list */
772 			}
773 			continue;
774 		}
775 
776 		/*
777 		 * clean up "p" if we have one
778 		 */
779 
780 		if (p) {
781 			/*
782 			 * the I/O request to "p" is done and uvm_pager_put
783 			 * has freed any cluster pages it may have allocated
784 			 * during I/O.  all that is left for us to do is
785 			 * clean up page "p" (which is still PG_BUSY).
786 			 *
787 			 * our result could be one of the following:
788 			 *   VM_PAGER_OK: successful pageout
789 			 *
790 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
791 			 *     to next page
792 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
793 			 *     "reactivate" page to get it out of the way (it
794 			 *     will eventually drift back into the inactive
795 			 *     queue for a retry).
796 			 *   VM_PAGER_UNLOCK: should never see this as it is
797 			 *     only valid for "get" operations
798 			 */
799 
800 			/* relock p's object: page queues not lock yet, so
801 			 * no need for "try" */
802 
803 			/* !swap_backed case: already locked... */
804 			if (swap_backed) {
805 				if (anon)
806 					simple_lock(&anon->an_lock);
807 				else
808 					simple_lock(&uobj->vmobjlock);
809 			}
810 
811 #ifdef DIAGNOSTIC
812 			if (result == VM_PAGER_UNLOCK)
813 				panic("pagedaemon: pageout returned "
814 				    "invalid 'unlock' code");
815 #endif
816 
817 			/* handle PG_WANTED now */
818 			if (p->flags & PG_WANTED)
819 				/* still holding object lock */
820 				wakeup(p);
821 
822 			p->flags &= ~(PG_BUSY|PG_WANTED);
823 			UVM_PAGE_OWN(p, NULL);
824 
825 			/* released during I/O? */
826 			if (p->flags & PG_RELEASED) {
827 				if (anon) {
828 					/* remove page so we can get nextpg */
829 					anon->u.an_page = NULL;
830 
831 					simple_unlock(&anon->an_lock);
832 					uvm_anfree(anon);	/* kills anon */
833 					pmap_page_protect(p, VM_PROT_NONE);
834 					anon = NULL;
835 					uvm_lock_pageq();
836 					nextpg = p->pageq.tqe_next;
837 					/* free released page */
838 					uvm_pagefree(p);
839 
840 				} else {
841 
842 #ifdef DIAGNOSTIC
843 					if (uobj->pgops->pgo_releasepg == NULL)
844 						panic("pagedaemon: no "
845 						   "pgo_releasepg function");
846 #endif
847 
848 					/*
849 					 * pgo_releasepg nukes the page and
850 					 * gets "nextpg" for us.  it returns
851 					 * with the page queues locked (when
852 					 * given nextpg ptr).
853 					 */
854 					if (!uobj->pgops->pgo_releasepg(p,
855 					    &nextpg))
856 						/* uobj died after release */
857 						uobj = NULL;
858 
859 					/*
860 					 * lock page queues here so that they're
861 					 * always locked at the end of the loop.
862 					 */
863 					uvm_lock_pageq();
864 				}
865 
866 			} else {	/* page was not released during I/O */
867 
868 				uvm_lock_pageq();
869 				nextpg = p->pageq.tqe_next;
870 
871 				if (result != VM_PAGER_OK) {
872 
873 					/* pageout was a failure... */
874 					if (result != VM_PAGER_AGAIN)
875 						uvm_pageactivate(p);
876 					pmap_clear_reference(p);
877 					/* XXXCDC: if (swap_backed) FREE p's
878 					 * swap block? */
879 
880 				} else {
881 
882 					/* pageout was a success... */
883 					pmap_clear_reference(p);
884 					pmap_clear_modify(p);
885 					p->flags |= PG_CLEAN;
886 					/* XXX: could free page here, but old
887 					 * pagedaemon does not */
888 
889 				}
890 			}
891 
892 			/*
893 			 * drop object lock (if there is an object left).   do
894 			 * a safety check of nextpg to make sure it is on the
895 			 * inactive queue (it should be since PG_BUSY pages on
896 			 * the inactive queue can't be re-queued [note: not
897 			 * true for active queue]).
898 			 */
899 
900 			if (anon)
901 				simple_unlock(&anon->an_lock);
902 			else if (uobj)
903 				simple_unlock(&uobj->vmobjlock);
904 
905 		} /* if (p) */ else {
906 
907 			/* if p is null in this loop, make sure it stays null
908 			 * in next loop */
909 			nextpg = NULL;
910 
911 			/*
912 			 * lock page queues here just so they're always locked
913 			 * at the end of the loop.
914 			 */
915 			uvm_lock_pageq();
916 		}
917 
918 		if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) {
919 			printf("pagedaemon: invalid nextpg!   reverting to "
920 			    "queue head\n");
921 			nextpg = pglst->tqh_first;	/* reload! */
922 		}
923 
924 	}	/* end of "inactive" 'for' loop */
925 	return (retval);
926 }
927 
928 /*
929  * uvmpd_scan: scan the page queues and attempt to meet our targets.
930  *
931  * => called with pageq's locked
932  */
933 
934 void
935 uvmpd_scan()
936 {
937 	int s, free, inactive_shortage, swap_shortage, pages_freed;
938 	struct vm_page *p, *nextpg;
939 	struct uvm_object *uobj;
940 	boolean_t got_it;
941 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
942 
943 	uvmexp.pdrevs++;		/* counter */
944 
945 #ifdef __GNUC__
946 	uobj = NULL;	/* XXX gcc */
947 #endif
948 	/*
949 	 * get current "free" page count
950 	 */
951 	s = uvm_lock_fpageq();
952 	free = uvmexp.free;
953 	uvm_unlock_fpageq(s);
954 
955 #ifndef __SWAP_BROKEN
956 	/*
957 	 * swap out some processes if we are below our free target.
958 	 * we need to unlock the page queues for this.
959 	 */
960 	if (free < uvmexp.freetarg) {
961 
962 		uvmexp.pdswout++;
963 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout", free,
964 		    uvmexp.freetarg, 0, 0);
965 		uvm_unlock_pageq();
966 		uvm_swapout_threads();
967 		pmap_update();		/* update so we can scan inactive q */
968 		uvm_lock_pageq();
969 
970 	}
971 #endif
972 
973 	/*
974 	 * now we want to work on meeting our targets.   first we work on our
975 	 * free target by converting inactive pages into free pages.  then
976 	 * we work on meeting our inactive target by converting active pages
977 	 * to inactive ones.
978 	 */
979 
980 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
981 
982 	/*
983 	 * do loop #1!   alternate starting queue between swap and object based
984 	 * on the low bit of uvmexp.pdrevs (which we bump by one each call).
985 	 */
986 
987 	got_it = FALSE;
988 	pages_freed = uvmexp.pdfreed;
989 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
990 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
991 	if (!got_it)
992 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
993 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
994 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
995 	pages_freed = uvmexp.pdfreed - pages_freed;
996 
997 	/*
998 	 * we have done the scan to get free pages.   now we work on meeting
999 	 * our inactive target.
1000 	 */
1001 
1002 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
1003 
1004 	/*
1005 	 * detect if we're not going to be able to page anything out
1006 	 * until we free some swap resources from active pages.
1007 	 */
1008 	swap_shortage = 0;
1009 	if (uvmexp.free < uvmexp.freetarg &&
1010 	    uvmexp.swpginuse == uvmexp.swpages &&
1011 	    uvmexp.swpgonly < uvmexp.swpages &&
1012 	    pages_freed == 0) {
1013 		swap_shortage = uvmexp.freetarg - uvmexp.free;
1014 	}
1015 
1016 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
1017 		    inactive_shortage, swap_shortage,0,0);
1018 	for (p = TAILQ_FIRST(&uvm.page_active);
1019 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
1020 	     p = nextpg) {
1021 		nextpg = p->pageq.tqe_next;
1022 		if (p->flags & PG_BUSY)
1023 			continue;	/* quick check before trying to lock */
1024 
1025 		/*
1026 		 * lock the page's owner.
1027 		 */
1028 		/* is page anon owned or ownerless? */
1029 		if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
1030 
1031 #ifdef DIAGNOSTIC
1032 			if (p->uanon == NULL)
1033 				panic("pagedaemon: page with no anon or "
1034 				    "object detected - loop 2");
1035 #endif
1036 			if (!simple_lock_try(&p->uanon->an_lock))
1037 				continue;
1038 
1039 			/* take over the page? */
1040 			if ((p->pqflags & PQ_ANON) == 0) {
1041 #ifdef DIAGNOSTIC
1042 				if (p->loan_count < 1)
1043 					panic("pagedaemon: non-loaned "
1044 					    "ownerless page detected - loop 2");
1045 #endif
1046 				p->loan_count--;
1047 				p->pqflags |= PQ_ANON;
1048 			}
1049 		} else {
1050 			if (!simple_lock_try(&p->uobject->vmobjlock))
1051 				continue;
1052 		}
1053 		/*
1054 		 * skip this page if it's busy.
1055 		 */
1056 		if ((p->flags & PG_BUSY) != 0) {
1057 			if (p->pqflags & PQ_ANON)
1058 				simple_unlock(&p->uanon->an_lock);
1059 			else
1060 				simple_unlock(&p->uobject->vmobjlock);
1061 			continue;
1062 		}
1063 
1064 		/*
1065 		 * if there's a shortage of swap, free any swap allocated
1066 		 * to this page so that other pages can be paged out.
1067 		 */
1068 		if (swap_shortage > 0) {
1069 			if ((p->pqflags & PQ_ANON) && p->uanon->an_swslot) {
1070 				uvm_swap_free(p->uanon->an_swslot, 1);
1071 				p->uanon->an_swslot = 0;
1072 				p->flags &= ~PG_CLEAN;
1073 				swap_shortage--;
1074 			}
1075 			if (p->pqflags & PQ_AOBJ) {
1076 				int slot = uao_set_swslot(p->uobject,
1077 					p->offset >> PAGE_SHIFT, 0);
1078 				if (slot) {
1079 					uvm_swap_free(slot, 1);
1080 					p->flags &= ~PG_CLEAN;
1081 					swap_shortage--;
1082 				}
1083 			}
1084 		}
1085 
1086 		/*
1087 		 * deactivate this page if there's a shortage of
1088 		 * inactive pages.
1089 		 */
1090 		if (inactive_shortage > 0) {
1091 			pmap_page_protect(p, VM_PROT_NONE);
1092 			/* no need to check wire_count as pg is "active" */
1093 			uvm_pagedeactivate(p);
1094 			uvmexp.pddeact++;
1095 			inactive_shortage--;
1096 		}
1097 
1098 		if (p->pqflags & PQ_ANON)
1099 			simple_unlock(&p->uanon->an_lock);
1100 		else
1101 			simple_unlock(&p->uobject->vmobjlock);
1102 	}
1103 }
1104