xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision b2ea75c1b17e1a9a339660e7ed45cd24946b230e)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.12 2001/08/11 10:57:22 art Exp $	*/
2 /*	$NetBSD: uvm_pdaemon.c,v 1.19 1999/11/04 21:51:42 thorpej 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 
80 #include <vm/vm.h>
81 #include <vm/vm_page.h>
82 #include <vm/vm_kern.h>
83 
84 #include <uvm/uvm.h>
85 
86 /*
87  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedeamon will reactivate
88  * in a pass thru the inactive list when swap is full.  the value should be
89  * "small"... if it's too large we'll cycle the active pages thru the inactive
90  * queue too quickly to for them to be referenced and avoid being freed.
91  */
92 
93 #define UVMPD_NUMDIRTYREACTS 16
94 
95 
96 /*
97  * local prototypes
98  */
99 
100 static void		uvmpd_scan __P((void));
101 static boolean_t	uvmpd_scan_inactive __P((struct pglist *));
102 static void		uvmpd_tune __P((void));
103 
104 
105 /*
106  * uvm_wait: wait (sleep) for the page daemon to free some pages
107  *
108  * => should be called with all locks released
109  * => should _not_ be called by the page daemon (to avoid deadlock)
110  */
111 
112 void
113 uvm_wait(wmsg)
114 	const char *wmsg;
115 {
116 	int timo = 0;
117 	int s = splbio();
118 
119 	/*
120 	 * check for page daemon going to sleep (waiting for itself)
121 	 */
122 
123 	if (curproc == uvm.pagedaemon_proc) {
124 		/*
125 		 * now we have a problem: the pagedaemon wants to go to
126 		 * sleep until it frees more memory.   but how can it
127 		 * free more memory if it is asleep?  that is a deadlock.
128 		 * we have two options:
129 		 *  [1] panic now
130 		 *  [2] put a timeout on the sleep, thus causing the
131 		 *      pagedaemon to only pause (rather than sleep forever)
132 		 *
133 		 * note that option [2] will only help us if we get lucky
134 		 * and some other process on the system breaks the deadlock
135 		 * by exiting or freeing memory (thus allowing the pagedaemon
136 		 * to continue).  for now we panic if DEBUG is defined,
137 		 * otherwise we hope for the best with option [2] (better
138 		 * yet, this should never happen in the first place!).
139 		 */
140 
141 		printf("pagedaemon: deadlock detected!\n");
142 		timo = hz >> 3;		/* set timeout */
143 #if defined(DEBUG)
144 		/* DEBUG: panic so we can debug it */
145 		panic("pagedaemon deadlock");
146 #endif
147 	}
148 
149 	simple_lock(&uvm.pagedaemon_lock);
150 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
151 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, (char *)wmsg,
152 	    timo);
153 
154 	splx(s);
155 }
156 
157 
158 /*
159  * uvmpd_tune: tune paging parameters
160  *
161  * => called when ever memory is added (or removed?) to the system
162  * => caller must call with page queues locked
163  */
164 
165 static void
166 uvmpd_tune()
167 {
168 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
169 
170 	uvmexp.freemin = uvmexp.npages / 30;
171 
172 	/* between 16k and 512k */
173 	/* XXX:  what are these values good for? */
174 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
175 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
176 
177 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
178 	if (uvmexp.freetarg <= uvmexp.freemin)
179 		uvmexp.freetarg = uvmexp.freemin + 1;
180 
181 	/* uvmexp.inactarg: computed in main daemon loop */
182 
183 	uvmexp.wiredmax = uvmexp.npages / 3;
184 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
185 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
186 }
187 
188 /*
189  * uvm_pageout: the main loop for the pagedaemon
190  */
191 
192 void
193 uvm_pageout()
194 {
195 	int npages = 0;
196 	int s;
197 	struct uvm_aiodesc *aio, *nextaio;
198 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
199 
200 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
201 
202 	/*
203 	 * ensure correct priority and set paging parameters...
204 	 */
205 
206 	uvm.pagedaemon_proc = curproc;
207 	(void) spl0();
208 	uvm_lock_pageq();
209 	npages = uvmexp.npages;
210 	uvmpd_tune();
211 	uvm_unlock_pageq();
212 
213 	/*
214 	 * main loop
215 	 */
216 	while (TRUE) {
217 
218 		/*
219 		 * carefully attempt to go to sleep (without losing "wakeups"!).
220 		 * we need splbio because we want to make sure the aio_done list
221 		 * is totally empty before we go to sleep.
222 		 */
223 
224 		s = splbio();
225 		simple_lock(&uvm.pagedaemon_lock);
226 
227 		/*
228 		 * if we've got done aio's, then bypass the sleep
229 		 */
230 
231 		if (uvm.aio_done.tqh_first == NULL) {
232 			UVMHIST_LOG(maphist,"  <<SLEEPING>>",0,0,0,0);
233 			UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
234 			    &uvm.pagedaemon_lock, FALSE, "daemon_slp", 0);
235 			uvmexp.pdwoke++;
236 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
237 
238 			/* relock pagedaemon_lock, still at splbio */
239 			simple_lock(&uvm.pagedaemon_lock);
240 		}
241 
242 		/*
243 		 * check for done aio structures
244 		 */
245 
246 		aio = uvm.aio_done.tqh_first;	/* save current list (if any)*/
247 		if (aio) {
248 			TAILQ_INIT(&uvm.aio_done);	/* zero global list */
249 		}
250 
251 		simple_unlock(&uvm.pagedaemon_lock);	/* unlock */
252 		splx(s);				/* drop splbio */
253 
254 		/*
255 		 * first clear out any pending aios (to free space in case we
256 		 * want to pageout more stuff).
257 		 */
258 
259 		for (/*null*/; aio != NULL ; aio = nextaio) {
260 
261 			uvmexp.paging -= aio->npages;
262 			nextaio = aio->aioq.tqe_next;
263 			aio->aiodone(aio);
264 
265 		}
266 
267 		/* Next, drain pool resources */
268 		pool_drain(0);
269 
270 		/*
271 		 * now lock page queues and recompute inactive count
272 		 */
273 		uvm_lock_pageq();
274 
275 		if (npages != uvmexp.npages) {	/* check for new pages? */
276 			npages = uvmexp.npages;
277 			uvmpd_tune();
278 		}
279 
280 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
281 		if (uvmexp.inactarg <= uvmexp.freetarg)
282 			uvmexp.inactarg = uvmexp.freetarg + 1;
283 
284 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d, inact/itarg=%d/%d",
285 		    uvmexp.free, uvmexp.freetarg, uvmexp.inactive,
286 		    uvmexp.inactarg);
287 
288 		/*
289 		 * scan if needed
290 		 * [XXX: note we are reading uvm.free without locking]
291 		 */
292 		if (uvmexp.free < uvmexp.freetarg ||
293 		    uvmexp.inactive < uvmexp.inactarg)
294 			uvmpd_scan();
295 
296 		/*
297 		 * done scan.  unlock page queues (the only lock we are holding)
298 		 */
299 		uvm_unlock_pageq();
300 
301 		/*
302 		 * done!    restart loop.
303 		 */
304 		if (uvmexp.free > uvmexp.reserve_kernel ||
305 		    uvmexp.paging == 0)
306 			wakeup(&uvmexp.free);
307 	}
308 	/*NOTREACHED*/
309 }
310 
311 /*
312  * uvmpd_scan_inactive: the first loop of uvmpd_scan broken out into
313  * 	its own function for ease of reading.
314  *
315  * => called with page queues locked
316  * => we work on meeting our free target by converting inactive pages
317  *    into free pages.
318  * => we handle the building of swap-backed clusters
319  * => we return TRUE if we are exiting because we met our target
320  */
321 
322 static boolean_t
323 uvmpd_scan_inactive(pglst)
324 	struct pglist *pglst;
325 {
326 	boolean_t retval = FALSE;	/* assume we haven't hit target */
327 	int s, free, result;
328 	struct vm_page *p, *nextpg;
329 	struct uvm_object *uobj;
330 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
331 	int npages;
332 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
333 	int swnpages, swcpages;				/* XXX: see below */
334 	int swslot;
335 	struct vm_anon *anon;
336 	boolean_t swap_backed;
337 	vaddr_t start;
338 	int dirtyreacts;
339 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
340 
341 	/*
342 	 * note: we currently keep swap-backed pages on a separate inactive
343 	 * list from object-backed pages.   however, merging the two lists
344 	 * back together again hasn't been ruled out.   thus, we keep our
345 	 * swap cluster in "swpps" rather than in pps (allows us to mix
346 	 * clustering types in the event of a mixed inactive queue).
347 	 */
348 
349 	/*
350 	 * swslot is non-zero if we are building a swap cluster.  we want
351 	 * to stay in the loop while we have a page to scan or we have
352 	 * a swap-cluster to build.
353 	 */
354 	swslot = 0;
355 	swnpages = swcpages = 0;
356 	free = 0;
357 	dirtyreacts = 0;
358 
359 	for (p = pglst->tqh_first ; p != NULL || swslot != 0 ; p = nextpg) {
360 
361 		/*
362 		 * note that p can be NULL iff we have traversed the whole
363 		 * list and need to do one final swap-backed clustered pageout.
364 		 */
365 		if (p) {
366 			/*
367 			 * update our copy of "free" and see if we've met
368 			 * our target
369 			 */
370 			s = uvm_lock_fpageq();
371 			free = uvmexp.free;
372 			uvm_unlock_fpageq(s);
373 
374 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
375 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
376 				UVMHIST_LOG(pdhist,"  met free target: "
377 				    "exit loop", 0, 0, 0, 0);
378 				retval = TRUE;		/* hit the target! */
379 
380 				if (swslot == 0)
381 					/* exit now if no swap-i/o pending */
382 					break;
383 
384 				/* set p to null to signal final swap i/o */
385 				p = NULL;
386 			}
387 		}
388 
389 		uobj = NULL;	/* be safe and shut gcc up */
390 		anon = NULL;	/* be safe and shut gcc up */
391 
392 		if (p) {	/* if (we have a new page to consider) */
393 			/*
394 			 * we are below target and have a new page to consider.
395 			 */
396 			uvmexp.pdscans++;
397 			nextpg = p->pageq.tqe_next;
398 
399 			/*
400 			 * move referenced pages back to active queue and
401 			 * skip to next page (unlikely to happen since
402 			 * inactive pages shouldn't have any valid mappings
403 			 * and we cleared reference before deactivating).
404 			 */
405 			if (pmap_is_referenced(p)) {
406 				uvm_pageactivate(p);
407 				uvmexp.pdreact++;
408 				continue;
409 			}
410 
411 			/*
412 			 * first we attempt to lock the object that this page
413 			 * belongs to.  if our attempt fails we skip on to
414 			 * the next page (no harm done).  it is important to
415 			 * "try" locking the object as we are locking in the
416 			 * wrong order (pageq -> object) and we don't want to
417 			 * get deadlocked.
418 			 *
419 			 * the only time we exepct to see an ownerless page
420 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
421 			 * anon has loaned a page from a uvm_object and the
422 			 * uvm_object has dropped the ownership.  in that
423 			 * case, the anon can "take over" the loaned page
424 			 * and make it its own.
425 			 */
426 
427 			/* is page part of an anon or ownerless ? */
428 			if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
429 
430 				anon = p->uanon;
431 
432 #ifdef DIAGNOSTIC
433 				/* to be on inactive q, page must be part
434 				 * of _something_ */
435 				if (anon == NULL)
436 					panic("pagedaemon: page with no anon "
437 					    "or object detected - loop 1");
438 #endif
439 
440 				if (!simple_lock_try(&anon->an_lock))
441 					/* lock failed, skip this page */
442 					continue;
443 
444 				/*
445 				 * if the page is ownerless, claim it in the
446 				 * name of "anon"!
447 				 */
448 				if ((p->pqflags & PQ_ANON) == 0) {
449 #ifdef DIAGNOSTIC
450 					if (p->loan_count < 1)
451 						panic("pagedaemon: non-loaned "
452 						    "ownerless page detected -"
453 						    " loop 1");
454 #endif
455 					p->loan_count--;
456 					p->pqflags |= PQ_ANON;      /* anon now owns it */
457 				}
458 
459 				if (p->flags & PG_BUSY) {
460 					simple_unlock(&anon->an_lock);
461 					uvmexp.pdbusy++;
462 					/* someone else owns page, skip it */
463 					continue;
464 				}
465 
466 				uvmexp.pdanscan++;
467 
468 			} else {
469 
470 				uobj = p->uobject;
471 
472 				if (!simple_lock_try(&uobj->vmobjlock))
473 					/* lock failed, skip this page */
474 					continue;
475 
476 				if (p->flags & PG_BUSY) {
477 					simple_unlock(&uobj->vmobjlock);
478 					uvmexp.pdbusy++;
479 					/* someone else owns page, skip it */
480 					continue;
481 				}
482 
483 				uvmexp.pdobscan++;
484 			}
485 
486 			/*
487 			 * we now have the object and the page queues locked.
488 			 * the page is not busy.   if the page is clean we
489 			 * can free it now and continue.
490 			 */
491 
492 			if (p->flags & PG_CLEAN) {
493 				if (p->pqflags & PQ_SWAPBACKED) {
494 					/* this page now lives only in swap */
495 					simple_lock(&uvm.swap_data_lock);
496 					uvmexp.swpgonly++;
497 					simple_unlock(&uvm.swap_data_lock);
498 				}
499 
500 				/* zap all mappings with pmap_page_protect... */
501 				pmap_page_protect(p, VM_PROT_NONE);
502 				uvm_pagefree(p);
503 				uvmexp.pdfreed++;
504 
505 				if (anon) {
506 #ifdef DIAGNOSTIC
507 					/*
508 					 * an anonymous page can only be clean
509 					 * if it has valid backing store.
510 					 */
511 					if (anon->an_swslot == 0)
512 						panic("pagedaemon: clean anon "
513 						 "page without backing store?");
514 #endif
515 					/* remove from object */
516 					anon->u.an_page = NULL;
517 					simple_unlock(&anon->an_lock);
518 				} else {
519 					/* pagefree has already removed the
520 					 * page from the object */
521 					simple_unlock(&uobj->vmobjlock);
522 				}
523 				continue;
524 			}
525 
526 			/*
527 			 * this page is dirty, skip it if we'll have met our
528 			 * free target when all the current pageouts complete.
529 			 */
530 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
531 				if (anon) {
532 					simple_unlock(&anon->an_lock);
533 				} else {
534 					simple_unlock(&uobj->vmobjlock);
535 				}
536 				continue;
537 			}
538 
539 			/*
540 			 * this page is dirty, but we can't page it out
541 			 * since all pages in swap are only in swap.
542 			 * reactivate it so that we eventually cycle
543 			 * all pages thru the inactive queue.
544 			 */
545 #ifdef DIAGNOSTIC
546 			if (uvmexp.swpgonly > uvmexp.swpages) {
547 				panic("uvmexp.swpgonly botch");
548 			}
549 #endif
550 			if ((p->pqflags & PQ_SWAPBACKED) &&
551 			    uvmexp.swpgonly == uvmexp.swpages) {
552 				dirtyreacts++;
553 				uvm_pageactivate(p);
554 				if (anon) {
555 					simple_unlock(&anon->an_lock);
556 				} else {
557 					simple_unlock(&uobj->vmobjlock);
558 				}
559 				continue;
560 			}
561 
562 			/*
563 			 * if the page is swap-backed and dirty and swap space
564 			 * is full, free any swap allocated to the page
565 			 * so that other pages can be paged out.
566 			 */
567 #ifdef DIAGNOSTIC
568 			if (uvmexp.swpginuse > uvmexp.swpages) {
569 				panic("uvmexp.swpginuse botch");
570 			}
571 #endif
572 			if ((p->pqflags & PQ_SWAPBACKED) &&
573 			    uvmexp.swpginuse == uvmexp.swpages) {
574 
575 				if ((p->pqflags & PQ_ANON) &&
576 				    p->uanon->an_swslot) {
577 					uvm_swap_free(p->uanon->an_swslot, 1);
578 					p->uanon->an_swslot = 0;
579 				}
580 				if (p->pqflags & PQ_AOBJ) {
581 					uao_dropswap(p->uobject,
582 						     p->offset >> PAGE_SHIFT);
583 				}
584 			}
585 
586 			/*
587 			 * the page we are looking at is dirty.   we must
588 			 * clean it before it can be freed.  to do this we
589 			 * first mark the page busy so that no one else will
590 			 * touch the page.   we write protect all the mappings
591 			 * of the page so that no one touches it while it is
592 			 * in I/O.
593 			 */
594 
595 			swap_backed = ((p->pqflags & PQ_SWAPBACKED) != 0);
596 			p->flags |= PG_BUSY;		/* now we own it */
597 			UVM_PAGE_OWN(p, "scan_inactive");
598 			pmap_page_protect(p, VM_PROT_READ);
599 			uvmexp.pgswapout++;
600 
601 			/*
602 			 * for swap-backed pages we need to (re)allocate
603 			 * swap space.
604 			 */
605 			if (swap_backed) {
606 
607 				/*
608 				 * free old swap slot (if any)
609 				 */
610 				if (anon) {
611 					if (anon->an_swslot) {
612 						uvm_swap_free(anon->an_swslot,
613 						    1);
614 						anon->an_swslot = 0;
615 					}
616 				} else {
617 					uao_dropswap(uobj,
618 						     p->offset >> PAGE_SHIFT);
619 				}
620 
621 				/*
622 				 * start new cluster (if necessary)
623 				 */
624 				if (swslot == 0) {
625 					/* want this much */
626 					swnpages = MAXBSIZE >> PAGE_SHIFT;
627 
628 					swslot = uvm_swap_alloc(&swnpages,
629 					    TRUE);
630 
631 					if (swslot == 0) {
632 						/* no swap?  give up! */
633 						p->flags &= ~PG_BUSY;
634 						UVM_PAGE_OWN(p, NULL);
635 						if (anon)
636 							simple_unlock(
637 							    &anon->an_lock);
638 						else
639 							simple_unlock(
640 							    &uobj->vmobjlock);
641 						continue;
642 					}
643 					swcpages = 0;	/* cluster is empty */
644 				}
645 
646 				/*
647 				 * add block to cluster
648 				 */
649 				swpps[swcpages] = p;
650 				if (anon)
651 					anon->an_swslot = swslot + swcpages;
652 				else
653 					uao_set_swslot(uobj,
654 					    p->offset >> PAGE_SHIFT,
655 					    swslot + swcpages);
656 				swcpages++;
657 
658 				/* done (swap-backed) */
659 			}
660 
661 			/* end: if (p) ["if we have new page to consider"] */
662 		} else {
663 
664 			/* if p == NULL we must be doing a last swap i/o */
665 			swap_backed = TRUE;
666 		}
667 
668 		/*
669 		 * now consider doing the pageout.
670 		 *
671 		 * for swap-backed pages, we do the pageout if we have either
672 		 * filled the cluster (in which case (swnpages == swcpages) or
673 		 * run out of pages (p == NULL).
674 		 *
675 		 * for object pages, we always do the pageout.
676 		 */
677 		if (swap_backed) {
678 
679 			if (p) {	/* if we just added a page to cluster */
680 				if (anon)
681 					simple_unlock(&anon->an_lock);
682 				else
683 					simple_unlock(&uobj->vmobjlock);
684 
685 				/* cluster not full yet? */
686 				if (swcpages < swnpages)
687 					continue;
688 			}
689 
690 			/* starting I/O now... set up for it */
691 			npages = swcpages;
692 			ppsp = swpps;
693 			/* for swap-backed pages only */
694 			start = (vaddr_t) swslot;
695 
696 			/* if this is final pageout we could have a few
697 			 * extra swap blocks */
698 			if (swcpages < swnpages) {
699 				uvm_swap_free(swslot + swcpages,
700 				    (swnpages - swcpages));
701 			}
702 
703 		} else {
704 
705 			/* normal object pageout */
706 			ppsp = pps;
707 			npages = sizeof(pps) / sizeof(struct vm_page *);
708 			/* not looked at because PGO_ALLPAGES is set */
709 			start = 0;
710 
711 		}
712 
713 		/*
714 		 * now do the pageout.
715 		 *
716 		 * for swap_backed pages we have already built the cluster.
717 		 * for !swap_backed pages, uvm_pager_put will call the object's
718 		 * "make put cluster" function to build a cluster on our behalf.
719 		 *
720 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
721 		 * it to free the cluster pages for us on a successful I/O (it
722 		 * always does this for un-successful I/O requests).  this
723 		 * allows us to do clustered pageout without having to deal
724 		 * with cluster pages at this level.
725 		 *
726 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
727 		 *  IN: locked: uobj (if !swap_backed), page queues
728 		 * OUT: locked: uobj (if !swap_backed && result !=VM_PAGER_PEND)
729 		 *     !locked: pageqs, uobj (if swap_backed || VM_PAGER_PEND)
730 		 *
731 		 * [the bit about VM_PAGER_PEND saves us one lock-unlock pair]
732 		 */
733 
734 		/* locked: uobj (if !swap_backed), page queues */
735 		uvmexp.pdpageouts++;
736 		result = uvm_pager_put((swap_backed) ? NULL : uobj, p,
737 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
738 		/* locked: uobj (if !swap_backed && result != PEND) */
739 		/* unlocked: pageqs, object (if swap_backed ||result == PEND) */
740 
741 		/*
742 		 * if we did i/o to swap, zero swslot to indicate that we are
743 		 * no longer building a swap-backed cluster.
744 		 */
745 
746 		if (swap_backed)
747 			swslot = 0;		/* done with this cluster */
748 
749 		/*
750 		 * first, we check for VM_PAGER_PEND which means that the
751 		 * async I/O is in progress and the async I/O done routine
752 		 * will clean up after us.   in this case we move on to the
753 		 * next page.
754 		 *
755 		 * there is a very remote chance that the pending async i/o can
756 		 * finish _before_ we get here.   if that happens, our page "p"
757 		 * may no longer be on the inactive queue.   so we verify this
758 		 * when determining the next page (starting over at the head if
759 		 * we've lost our inactive page).
760 		 */
761 
762 		if (result == VM_PAGER_PEND) {
763 			uvmexp.paging += npages;
764 			uvm_lock_pageq();		/* relock page queues */
765 			uvmexp.pdpending++;
766 			if (p) {
767 				if (p->pqflags & PQ_INACTIVE)
768 					/* reload! */
769 					nextpg = p->pageq.tqe_next;
770 				else
771 					/* reload! */
772 					nextpg = pglst->tqh_first;
773 				} else {
774 					nextpg = NULL;		/* done list */
775 			}
776 			continue;
777 		}
778 
779 		/*
780 		 * clean up "p" if we have one
781 		 */
782 
783 		if (p) {
784 			/*
785 			 * the I/O request to "p" is done and uvm_pager_put
786 			 * has freed any cluster pages it may have allocated
787 			 * during I/O.  all that is left for us to do is
788 			 * clean up page "p" (which is still PG_BUSY).
789 			 *
790 			 * our result could be one of the following:
791 			 *   VM_PAGER_OK: successful pageout
792 			 *
793 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
794 			 *     to next page
795 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
796 			 *     "reactivate" page to get it out of the way (it
797 			 *     will eventually drift back into the inactive
798 			 *     queue for a retry).
799 			 *   VM_PAGER_UNLOCK: should never see this as it is
800 			 *     only valid for "get" operations
801 			 */
802 
803 			/* relock p's object: page queues not lock yet, so
804 			 * no need for "try" */
805 
806 			/* !swap_backed case: already locked... */
807 			if (swap_backed) {
808 				if (anon)
809 					simple_lock(&anon->an_lock);
810 				else
811 					simple_lock(&uobj->vmobjlock);
812 			}
813 
814 #ifdef DIAGNOSTIC
815 			if (result == VM_PAGER_UNLOCK)
816 				panic("pagedaemon: pageout returned "
817 				    "invalid 'unlock' code");
818 #endif
819 
820 			/* handle PG_WANTED now */
821 			if (p->flags & PG_WANTED)
822 				/* still holding object lock */
823 				wakeup(p);
824 
825 			p->flags &= ~(PG_BUSY|PG_WANTED);
826 			UVM_PAGE_OWN(p, NULL);
827 
828 			/* released during I/O? */
829 			if (p->flags & PG_RELEASED) {
830 				if (anon) {
831 					/* remove page so we can get nextpg */
832 					anon->u.an_page = NULL;
833 
834 					simple_unlock(&anon->an_lock);
835 					uvm_anfree(anon);	/* kills anon */
836 					pmap_page_protect(p, VM_PROT_NONE);
837 					anon = NULL;
838 					uvm_lock_pageq();
839 					nextpg = p->pageq.tqe_next;
840 					/* free released page */
841 					uvm_pagefree(p);
842 
843 				} else {
844 
845 #ifdef DIAGNOSTIC
846 					if (uobj->pgops->pgo_releasepg == NULL)
847 						panic("pagedaemon: no "
848 						   "pgo_releasepg function");
849 #endif
850 
851 					/*
852 					 * pgo_releasepg nukes the page and
853 					 * gets "nextpg" for us.  it returns
854 					 * with the page queues locked (when
855 					 * given nextpg ptr).
856 					 */
857 					if (!uobj->pgops->pgo_releasepg(p,
858 					    &nextpg))
859 						/* uobj died after release */
860 						uobj = NULL;
861 
862 					/*
863 					 * lock page queues here so that they're
864 					 * always locked at the end of the loop.
865 					 */
866 					uvm_lock_pageq();
867 				}
868 
869 			} else {	/* page was not released during I/O */
870 
871 				uvm_lock_pageq();
872 				nextpg = p->pageq.tqe_next;
873 
874 				if (result != VM_PAGER_OK) {
875 
876 					/* pageout was a failure... */
877 					if (result != VM_PAGER_AGAIN)
878 						uvm_pageactivate(p);
879 					pmap_clear_reference(p);
880 					/* XXXCDC: if (swap_backed) FREE p's
881 					 * swap block? */
882 
883 				} else {
884 
885 					/* pageout was a success... */
886 					pmap_clear_reference(p);
887 					pmap_clear_modify(p);
888 					p->flags |= PG_CLEAN;
889 					/* XXX: could free page here, but old
890 					 * pagedaemon does not */
891 
892 				}
893 			}
894 
895 			/*
896 			 * drop object lock (if there is an object left).   do
897 			 * a safety check of nextpg to make sure it is on the
898 			 * inactive queue (it should be since PG_BUSY pages on
899 			 * the inactive queue can't be re-queued [note: not
900 			 * true for active queue]).
901 			 */
902 
903 			if (anon)
904 				simple_unlock(&anon->an_lock);
905 			else if (uobj)
906 				simple_unlock(&uobj->vmobjlock);
907 
908 		} /* if (p) */ else {
909 
910 			/* if p is null in this loop, make sure it stays null
911 			 * in next loop */
912 			nextpg = NULL;
913 
914 			/*
915 			 * lock page queues here just so they're always locked
916 			 * at the end of the loop.
917 			 */
918 			uvm_lock_pageq();
919 		}
920 
921 		if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) {
922 			printf("pagedaemon: invalid nextpg!   reverting to "
923 			    "queue head\n");
924 			nextpg = pglst->tqh_first;	/* reload! */
925 		}
926 
927 	}	/* end of "inactive" 'for' loop */
928 	return (retval);
929 }
930 
931 /*
932  * uvmpd_scan: scan the page queues and attempt to meet our targets.
933  *
934  * => called with pageq's locked
935  */
936 
937 void
938 uvmpd_scan()
939 {
940 	int s, free, inactive_shortage, swap_shortage, pages_freed;
941 	struct vm_page *p, *nextpg;
942 	struct uvm_object *uobj;
943 	boolean_t got_it;
944 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
945 
946 	uvmexp.pdrevs++;		/* counter */
947 
948 #ifdef __GNUC__
949 	uobj = NULL;	/* XXX gcc */
950 #endif
951 	/*
952 	 * get current "free" page count
953 	 */
954 	s = uvm_lock_fpageq();
955 	free = uvmexp.free;
956 	uvm_unlock_fpageq(s);
957 
958 #ifndef __SWAP_BROKEN
959 	/*
960 	 * swap out some processes if we are below our free target.
961 	 * we need to unlock the page queues for this.
962 	 */
963 	if (free < uvmexp.freetarg) {
964 
965 		uvmexp.pdswout++;
966 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout", free,
967 		    uvmexp.freetarg, 0, 0);
968 		uvm_unlock_pageq();
969 		uvm_swapout_threads();
970 		pmap_update();		/* update so we can scan inactive q */
971 		uvm_lock_pageq();
972 
973 	}
974 #endif
975 
976 	/*
977 	 * now we want to work on meeting our targets.   first we work on our
978 	 * free target by converting inactive pages into free pages.  then
979 	 * we work on meeting our inactive target by converting active pages
980 	 * to inactive ones.
981 	 */
982 
983 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
984 
985 	/*
986 	 * do loop #1!   alternate starting queue between swap and object based
987 	 * on the low bit of uvmexp.pdrevs (which we bump by one each call).
988 	 */
989 
990 	got_it = FALSE;
991 	pages_freed = uvmexp.pdfreed;
992 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
993 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
994 	if (!got_it)
995 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
996 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
997 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
998 	pages_freed = uvmexp.pdfreed - pages_freed;
999 
1000 	/*
1001 	 * we have done the scan to get free pages.   now we work on meeting
1002 	 * our inactive target.
1003 	 */
1004 
1005 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
1006 
1007 	/*
1008 	 * detect if we're not going to be able to page anything out
1009 	 * until we free some swap resources from active pages.
1010 	 */
1011 	swap_shortage = 0;
1012 	if (uvmexp.free < uvmexp.freetarg &&
1013 	    uvmexp.swpginuse == uvmexp.swpages &&
1014 	    uvmexp.swpgonly < uvmexp.swpages &&
1015 	    pages_freed == 0) {
1016 		swap_shortage = uvmexp.freetarg - uvmexp.free;
1017 	}
1018 
1019 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
1020 		    inactive_shortage, swap_shortage,0,0);
1021 	for (p = TAILQ_FIRST(&uvm.page_active);
1022 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
1023 	     p = nextpg) {
1024 		nextpg = p->pageq.tqe_next;
1025 		if (p->flags & PG_BUSY)
1026 			continue;	/* quick check before trying to lock */
1027 
1028 		/*
1029 		 * lock the page's owner.
1030 		 */
1031 		/* is page anon owned or ownerless? */
1032 		if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
1033 
1034 #ifdef DIAGNOSTIC
1035 			if (p->uanon == NULL)
1036 				panic("pagedaemon: page with no anon or "
1037 				    "object detected - loop 2");
1038 #endif
1039 			if (!simple_lock_try(&p->uanon->an_lock))
1040 				continue;
1041 
1042 			/* take over the page? */
1043 			if ((p->pqflags & PQ_ANON) == 0) {
1044 #ifdef DIAGNOSTIC
1045 				if (p->loan_count < 1)
1046 					panic("pagedaemon: non-loaned "
1047 					    "ownerless page detected - loop 2");
1048 #endif
1049 				p->loan_count--;
1050 				p->pqflags |= PQ_ANON;
1051 			}
1052 		} else {
1053 			if (!simple_lock_try(&p->uobject->vmobjlock))
1054 				continue;
1055 		}
1056 		/*
1057 		 * skip this page if it's busy.
1058 		 */
1059 		if ((p->flags & PG_BUSY) != 0) {
1060 			if (p->pqflags & PQ_ANON)
1061 				simple_unlock(&p->uanon->an_lock);
1062 			else
1063 				simple_unlock(&p->uobject->vmobjlock);
1064 			continue;
1065 		}
1066 
1067 		/*
1068 		 * if there's a shortage of swap, free any swap allocated
1069 		 * to this page so that other pages can be paged out.
1070 		 */
1071 		if (swap_shortage > 0) {
1072 			if ((p->pqflags & PQ_ANON) && p->uanon->an_swslot) {
1073 				uvm_swap_free(p->uanon->an_swslot, 1);
1074 				p->uanon->an_swslot = 0;
1075 				p->flags &= ~PG_CLEAN;
1076 				swap_shortage--;
1077 			}
1078 			if (p->pqflags & PQ_AOBJ) {
1079 				int slot = uao_set_swslot(p->uobject,
1080 					p->offset >> PAGE_SHIFT, 0);
1081 				if (slot) {
1082 					uvm_swap_free(slot, 1);
1083 					p->flags &= ~PG_CLEAN;
1084 					swap_shortage--;
1085 				}
1086 			}
1087 		}
1088 
1089 		/*
1090 		 * deactivate this page if there's a shortage of
1091 		 * inactive pages.
1092 		 */
1093 		if (inactive_shortage > 0) {
1094 			pmap_page_protect(p, VM_PROT_NONE);
1095 			/* no need to check wire_count as pg is "active" */
1096 			uvm_pagedeactivate(p);
1097 			uvmexp.pddeact++;
1098 			inactive_shortage--;
1099 		}
1100 
1101 		if (p->pqflags & PQ_ANON)
1102 			simple_unlock(&p->uanon->an_lock);
1103 		else
1104 			simple_unlock(&p->uobject->vmobjlock);
1105 	}
1106 }
1107