xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 3a3fbb3f2e2521ab7c4a56b7ff7462ebd9095ec5)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.22 2002/01/02 22:23:25 miod 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 __P((void));
99 static boolean_t	uvmpd_scan_inactive __P((struct pglist *));
100 static void		uvmpd_tune __P((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 			(*bp->b_iodone)(bp);
338 			bp = nbp;
339 		}
340 		if (free <= uvmexp.reserve_kernel) {
341 			s = uvm_lock_fpageq();
342 			wakeup(&uvm.pagedaemon);
343 			uvm_unlock_fpageq(s);
344 		} else {
345 			simple_lock(&uvm.pagedaemon_lock);
346 			wakeup(&uvmexp.free);
347 			simple_unlock(&uvm.pagedaemon_lock);
348 		}
349 	}
350 }
351 
352 
353 
354 /*
355  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
356  *
357  * => called with page queues locked
358  * => we work on meeting our free target by converting inactive pages
359  *    into free pages.
360  * => we handle the building of swap-backed clusters
361  * => we return TRUE if we are exiting because we met our target
362  */
363 
364 static boolean_t
365 uvmpd_scan_inactive(pglst)
366 	struct pglist *pglst;
367 {
368 	boolean_t retval = FALSE;	/* assume we haven't hit target */
369 	int s, free, result;
370 	struct vm_page *p, *nextpg;
371 	struct uvm_object *uobj;
372 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
373 	int npages;
374 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
375 	int swnpages, swcpages;				/* XXX: see below */
376 	int swslot;
377 	struct vm_anon *anon;
378 	boolean_t swap_backed;
379 	vaddr_t start;
380 	int dirtyreacts;
381 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
382 
383 	/*
384 	 * note: we currently keep swap-backed pages on a seperate inactive
385 	 * list from object-backed pages.   however, merging the two lists
386 	 * back together again hasn't been ruled out.   thus, we keep our
387 	 * swap cluster in "swpps" rather than in pps (allows us to mix
388 	 * clustering types in the event of a mixed inactive queue).
389 	 */
390 
391 	/*
392 	 * swslot is non-zero if we are building a swap cluster.  we want
393 	 * to stay in the loop while we have a page to scan or we have
394 	 * a swap-cluster to build.
395 	 */
396 
397 	swslot = 0;
398 	swnpages = swcpages = 0;
399 	free = 0;
400 	dirtyreacts = 0;
401 
402 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
403 
404 		/*
405 		 * note that p can be NULL iff we have traversed the whole
406 		 * list and need to do one final swap-backed clustered pageout.
407 		 */
408 
409 		uobj = NULL;
410 		anon = NULL;
411 
412 		if (p) {
413 
414 			/*
415 			 * update our copy of "free" and see if we've met
416 			 * our target
417 			 */
418 
419 			s = uvm_lock_fpageq();
420 			free = uvmexp.free;
421 			uvm_unlock_fpageq(s);
422 
423 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
424 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
425 				UVMHIST_LOG(pdhist,"  met free target: "
426 					    "exit loop", 0, 0, 0, 0);
427 				retval = TRUE;
428 
429 				if (swslot == 0) {
430 					/* exit now if no swap-i/o pending */
431 					break;
432 				}
433 
434 				/* set p to null to signal final swap i/o */
435 				p = NULL;
436 			}
437 		}
438 
439 		if (p) {	/* if (we have a new page to consider) */
440 
441 			/*
442 			 * we are below target and have a new page to consider.
443 			 */
444 			uvmexp.pdscans++;
445 			nextpg = TAILQ_NEXT(p, pageq);
446 
447 			/*
448 			 * move referenced pages back to active queue and
449 			 * skip to next page (unlikely to happen since
450 			 * inactive pages shouldn't have any valid mappings
451 			 * and we cleared reference before deactivating).
452 			 */
453 
454 			if (pmap_is_referenced(p)) {
455 				uvm_pageactivate(p);
456 				uvmexp.pdreact++;
457 				continue;
458 			}
459 
460 			/*
461 			 * first we attempt to lock the object that this page
462 			 * belongs to.  if our attempt fails we skip on to
463 			 * the next page (no harm done).  it is important to
464 			 * "try" locking the object as we are locking in the
465 			 * wrong order (pageq -> object) and we don't want to
466 			 * deadlock.
467 			 *
468 			 * the only time we expect to see an ownerless page
469 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
470 			 * anon has loaned a page from a uvm_object and the
471 			 * uvm_object has dropped the ownership.  in that
472 			 * case, the anon can "take over" the loaned page
473 			 * and make it its own.
474 			 */
475 
476 			/* is page part of an anon or ownerless ? */
477 			if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
478 				anon = p->uanon;
479 				KASSERT(anon != NULL);
480 				if (!simple_lock_try(&anon->an_lock)) {
481 					/* lock failed, skip this page */
482 					continue;
483 				}
484 
485 				/*
486 				 * if the page is ownerless, claim it in the
487 				 * name of "anon"!
488 				 */
489 
490 				if ((p->pqflags & PQ_ANON) == 0) {
491 					KASSERT(p->loan_count > 0);
492 					p->loan_count--;
493 					p->pqflags |= PQ_ANON;
494 					/* anon now owns it */
495 				}
496 				if (p->flags & PG_BUSY) {
497 					simple_unlock(&anon->an_lock);
498 					uvmexp.pdbusy++;
499 					/* someone else owns page, skip it */
500 					continue;
501 				}
502 				uvmexp.pdanscan++;
503 			} else {
504 				uobj = p->uobject;
505 				KASSERT(uobj != NULL);
506 				if (!simple_lock_try(&uobj->vmobjlock)) {
507 					/* lock failed, skip this page */
508 					continue;
509 				}
510 				if (p->flags & PG_BUSY) {
511 					simple_unlock(&uobj->vmobjlock);
512 					uvmexp.pdbusy++;
513 					/* someone else owns page, skip it */
514 					continue;
515 				}
516 				uvmexp.pdobscan++;
517 			}
518 
519 			/*
520 			 * we now have the object and the page queues locked.
521 			 * the page is not busy.   if the page is clean we
522 			 * can free it now and continue.
523 			 */
524 
525 			if (p->flags & PG_CLEAN) {
526 				if (p->pqflags & PQ_SWAPBACKED) {
527 					/* this page now lives only in swap */
528 					simple_lock(&uvm.swap_data_lock);
529 					uvmexp.swpgonly++;
530 					simple_unlock(&uvm.swap_data_lock);
531 				}
532 
533 				/* zap all mappings with pmap_page_protect... */
534 				pmap_page_protect(p, VM_PROT_NONE);
535 				uvm_pagefree(p);
536 				uvmexp.pdfreed++;
537 
538 				if (anon) {
539 
540 					/*
541 					 * an anonymous page can only be clean
542 					 * if it has backing store assigned.
543 					 */
544 
545 					KASSERT(anon->an_swslot != 0);
546 
547 					/* remove from object */
548 					anon->u.an_page = NULL;
549 					simple_unlock(&anon->an_lock);
550 				} else {
551 					/* pagefree has already removed the
552 					 * page from the object */
553 					simple_unlock(&uobj->vmobjlock);
554 				}
555 				continue;
556 			}
557 
558 			/*
559 			 * this page is dirty, skip it if we'll have met our
560 			 * free target when all the current pageouts complete.
561 			 */
562 
563 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
564 				if (anon) {
565 					simple_unlock(&anon->an_lock);
566 				} else {
567 					simple_unlock(&uobj->vmobjlock);
568 				}
569 				continue;
570 			}
571 
572 			/*
573 			 * this page is dirty, but we can't page it out
574 			 * since all pages in swap are only in swap.
575 			 * reactivate it so that we eventually cycle
576 			 * all pages thru the inactive queue.
577 			 */
578 
579 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
580 			if ((p->pqflags & PQ_SWAPBACKED) &&
581 			    uvmexp.swpgonly == uvmexp.swpages) {
582 				dirtyreacts++;
583 				uvm_pageactivate(p);
584 				if (anon) {
585 					simple_unlock(&anon->an_lock);
586 				} else {
587 					simple_unlock(&uobj->vmobjlock);
588 				}
589 				continue;
590 			}
591 
592 			/*
593 			 * if the page is swap-backed and dirty and swap space
594 			 * is full, free any swap allocated to the page
595 			 * so that other pages can be paged out.
596 			 */
597 
598 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
599 			if ((p->pqflags & PQ_SWAPBACKED) &&
600 			    uvmexp.swpginuse == uvmexp.swpages) {
601 
602 				if ((p->pqflags & PQ_ANON) &&
603 				    p->uanon->an_swslot) {
604 					uvm_swap_free(p->uanon->an_swslot, 1);
605 					p->uanon->an_swslot = 0;
606 				}
607 				if (p->pqflags & PQ_AOBJ) {
608 					uao_dropswap(p->uobject,
609 						     p->offset >> PAGE_SHIFT);
610 				}
611 			}
612 
613 			/*
614 			 * the page we are looking at is dirty.   we must
615 			 * clean it before it can be freed.  to do this we
616 			 * first mark the page busy so that no one else will
617 			 * touch the page.   we write protect all the mappings
618 			 * of the page so that no one touches it while it is
619 			 * in I/O.
620 			 */
621 
622 			swap_backed = ((p->pqflags & PQ_SWAPBACKED) != 0);
623 			p->flags |= PG_BUSY;		/* now we own it */
624 			UVM_PAGE_OWN(p, "scan_inactive");
625 			pmap_page_protect(p, VM_PROT_READ);
626 			uvmexp.pgswapout++;
627 
628 			/*
629 			 * for swap-backed pages we need to (re)allocate
630 			 * swap space.
631 			 */
632 
633 			if (swap_backed) {
634 
635 				/*
636 				 * free old swap slot (if any)
637 				 */
638 
639 				if (anon) {
640 					if (anon->an_swslot) {
641 						uvm_swap_free(anon->an_swslot,
642 						    1);
643 						anon->an_swslot = 0;
644 					}
645 				} else {
646 					uao_dropswap(uobj,
647 						     p->offset >> PAGE_SHIFT);
648 				}
649 
650 				/*
651 				 * start new cluster (if necessary)
652 				 */
653 
654 				if (swslot == 0) {
655 					swnpages = MAXBSIZE >> PAGE_SHIFT;
656 					swslot = uvm_swap_alloc(&swnpages,
657 					    TRUE);
658 					if (swslot == 0) {
659 						/* no swap?  give up! */
660 						p->flags &= ~PG_BUSY;
661 						UVM_PAGE_OWN(p, NULL);
662 						if (anon)
663 							simple_unlock(
664 							    &anon->an_lock);
665 						else
666 							simple_unlock(
667 							    &uobj->vmobjlock);
668 						continue;
669 					}
670 					swcpages = 0;	/* cluster is empty */
671 				}
672 
673 				/*
674 				 * add block to cluster
675 				 */
676 
677 				swpps[swcpages] = p;
678 				if (anon)
679 					anon->an_swslot = swslot + swcpages;
680 				else
681 					uao_set_swslot(uobj,
682 					    p->offset >> PAGE_SHIFT,
683 					    swslot + swcpages);
684 				swcpages++;
685 			}
686 		} else {
687 
688 			/* if p == NULL we must be doing a last swap i/o */
689 			swap_backed = TRUE;
690 		}
691 
692 		/*
693 		 * now consider doing the pageout.
694 		 *
695 		 * for swap-backed pages, we do the pageout if we have either
696 		 * filled the cluster (in which case (swnpages == swcpages) or
697 		 * run out of pages (p == NULL).
698 		 *
699 		 * for object pages, we always do the pageout.
700 		 */
701 
702 		if (swap_backed) {
703 			if (p) {	/* if we just added a page to cluster */
704 				if (anon)
705 					simple_unlock(&anon->an_lock);
706 				else
707 					simple_unlock(&uobj->vmobjlock);
708 
709 				/* cluster not full yet? */
710 				if (swcpages < swnpages)
711 					continue;
712 			}
713 
714 			/* starting I/O now... set up for it */
715 			npages = swcpages;
716 			ppsp = swpps;
717 			/* for swap-backed pages only */
718 			start = (vaddr_t) swslot;
719 
720 			/* if this is final pageout we could have a few
721 			 * extra swap blocks */
722 			if (swcpages < swnpages) {
723 				uvm_swap_free(swslot + swcpages,
724 				    (swnpages - swcpages));
725 			}
726 		} else {
727 			/* normal object pageout */
728 			ppsp = pps;
729 			npages = sizeof(pps) / sizeof(struct vm_page *);
730 			/* not looked at because PGO_ALLPAGES is set */
731 			start = 0;
732 		}
733 
734 		/*
735 		 * now do the pageout.
736 		 *
737 		 * for swap_backed pages we have already built the cluster.
738 		 * for !swap_backed pages, uvm_pager_put will call the object's
739 		 * "make put cluster" function to build a cluster on our behalf.
740 		 *
741 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
742 		 * it to free the cluster pages for us on a successful I/O (it
743 		 * always does this for un-successful I/O requests).  this
744 		 * allows us to do clustered pageout without having to deal
745 		 * with cluster pages at this level.
746 		 *
747 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
748 		 *  IN: locked: uobj (if !swap_backed), page queues
749 		 * OUT: locked: uobj (if !swap_backed && result !=VM_PAGER_PEND)
750 		 *     !locked: pageqs, uobj (if swap_backed || VM_PAGER_PEND)
751 		 *
752 		 * [the bit about VM_PAGER_PEND saves us one lock-unlock pair]
753 		 */
754 
755 		/* locked: uobj (if !swap_backed), page queues */
756 		uvmexp.pdpageouts++;
757 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
758 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
759 		/* locked: uobj (if !swap_backed && result != PEND) */
760 		/* unlocked: pageqs, object (if swap_backed ||result == PEND) */
761 
762 		/*
763 		 * if we did i/o to swap, zero swslot to indicate that we are
764 		 * no longer building a swap-backed cluster.
765 		 */
766 
767 		if (swap_backed)
768 			swslot = 0;		/* done with this cluster */
769 
770 		/*
771 		 * first, we check for VM_PAGER_PEND which means that the
772 		 * async I/O is in progress and the async I/O done routine
773 		 * will clean up after us.   in this case we move on to the
774 		 * next page.
775 		 *
776 		 * there is a very remote chance that the pending async i/o can
777 		 * finish _before_ we get here.   if that happens, our page "p"
778 		 * may no longer be on the inactive queue.   so we verify this
779 		 * when determining the next page (starting over at the head if
780 		 * we've lost our inactive page).
781 		 */
782 
783 		if (result == VM_PAGER_PEND) {
784 			uvmexp.paging += npages;
785 			uvm_lock_pageq();
786 			uvmexp.pdpending++;
787 			if (p) {
788 				if (p->pqflags & PQ_INACTIVE)
789 					nextpg = TAILQ_NEXT(p, pageq);
790 				else
791 					nextpg = TAILQ_FIRST(pglst);
792 			} else {
793 				nextpg = NULL;
794 			}
795 			continue;
796 		}
797 
798 #ifdef UBC
799 		if (result == VM_PAGER_ERROR &&
800 		    curproc == uvm.pagedaemon_proc) {
801 			uvm_lock_pageq();
802 			nextpg = TAILQ_NEXT(p, pageq);
803 			uvm_pageactivate(p);
804 			continue;
805 		}
806 #endif
807 
808 		/*
809 		 * clean up "p" if we have one
810 		 */
811 
812 		if (p) {
813 			/*
814 			 * the I/O request to "p" is done and uvm_pager_put
815 			 * has freed any cluster pages it may have allocated
816 			 * during I/O.  all that is left for us to do is
817 			 * clean up page "p" (which is still PG_BUSY).
818 			 *
819 			 * our result could be one of the following:
820 			 *   VM_PAGER_OK: successful pageout
821 			 *
822 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
823 			 *     to next page
824 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
825 			 *     "reactivate" page to get it out of the way (it
826 			 *     will eventually drift back into the inactive
827 			 *     queue for a retry).
828 			 *   VM_PAGER_UNLOCK: should never see this as it is
829 			 *     only valid for "get" operations
830 			 */
831 
832 			/* relock p's object: page queues not lock yet, so
833 			 * no need for "try" */
834 
835 			/* !swap_backed case: already locked... */
836 			if (swap_backed) {
837 				if (anon)
838 					simple_lock(&anon->an_lock);
839 				else
840 					simple_lock(&uobj->vmobjlock);
841 			}
842 
843 #ifdef DIAGNOSTIC
844 			if (result == VM_PAGER_UNLOCK)
845 				panic("pagedaemon: pageout returned "
846 				    "invalid 'unlock' code");
847 #endif
848 
849 			/* handle PG_WANTED now */
850 			if (p->flags & PG_WANTED)
851 				/* still holding object lock */
852 				wakeup(p);
853 
854 			p->flags &= ~(PG_BUSY|PG_WANTED);
855 			UVM_PAGE_OWN(p, NULL);
856 
857 			/* released during I/O? */
858 			if (p->flags & PG_RELEASED) {
859 				if (anon) {
860 					/* remove page so we can get nextpg */
861 					anon->u.an_page = NULL;
862 
863 					simple_unlock(&anon->an_lock);
864 					uvm_anfree(anon);	/* kills anon */
865 					pmap_page_protect(p, VM_PROT_NONE);
866 					anon = NULL;
867 					uvm_lock_pageq();
868 					nextpg = TAILQ_NEXT(p, pageq);
869 					/* free released page */
870 					uvm_pagefree(p);
871 
872 				} else {
873 
874 					/*
875 					 * pgo_releasepg nukes the page and
876 					 * gets "nextpg" for us.  it returns
877 					 * with the page queues locked (when
878 					 * given nextpg ptr).
879 					 */
880 
881 					if (!uobj->pgops->pgo_releasepg(p,
882 					    &nextpg))
883 						/* uobj died after release */
884 						uobj = NULL;
885 
886 					/*
887 					 * lock page queues here so that they're
888 					 * always locked at the end of the loop.
889 					 */
890 
891 					uvm_lock_pageq();
892 				}
893 			} else {	/* page was not released during I/O */
894 				uvm_lock_pageq();
895 				nextpg = TAILQ_NEXT(p, pageq);
896 				if (result != VM_PAGER_OK) {
897 					/* pageout was a failure... */
898 					if (result != VM_PAGER_AGAIN)
899 						uvm_pageactivate(p);
900 					pmap_clear_reference(p);
901 					/* XXXCDC: if (swap_backed) FREE p's
902 					 * swap block? */
903 				} else {
904 					/* pageout was a success... */
905 					pmap_clear_reference(p);
906 					pmap_clear_modify(p);
907 					p->flags |= PG_CLEAN;
908 				}
909 			}
910 
911 			/*
912 			 * drop object lock (if there is an object left).   do
913 			 * a safety check of nextpg to make sure it is on the
914 			 * inactive queue (it should be since PG_BUSY pages on
915 			 * the inactive queue can't be re-queued [note: not
916 			 * true for active queue]).
917 			 */
918 
919 			if (anon)
920 				simple_unlock(&anon->an_lock);
921 			else if (uobj)
922 				simple_unlock(&uobj->vmobjlock);
923 
924 		} else {
925 
926 			/*
927 			 * if p is null in this loop, make sure it stays null
928 			 * in the next loop.
929 			 */
930 
931 			nextpg = NULL;
932 
933 			/*
934 			 * lock page queues here just so they're always locked
935 			 * at the end of the loop.
936 			 */
937 
938 			uvm_lock_pageq();
939 		}
940 
941 		if (nextpg && (nextpg->pqflags & PQ_INACTIVE) == 0) {
942 			nextpg = TAILQ_FIRST(pglst);	/* reload! */
943 		}
944 	}
945 	return (retval);
946 }
947 
948 /*
949  * uvmpd_scan: scan the page queues and attempt to meet our targets.
950  *
951  * => called with pageq's locked
952  */
953 
954 void
955 uvmpd_scan()
956 {
957 	int s, free, inactive_shortage, swap_shortage, pages_freed;
958 	struct vm_page *p, *nextpg;
959 	struct uvm_object *uobj;
960 	boolean_t got_it;
961 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
962 
963 	uvmexp.pdrevs++;		/* counter */
964 	uobj = NULL;
965 
966 	/*
967 	 * get current "free" page count
968 	 */
969 	s = uvm_lock_fpageq();
970 	free = uvmexp.free;
971 	uvm_unlock_fpageq(s);
972 
973 #ifndef __SWAP_BROKEN
974 	/*
975 	 * swap out some processes if we are below our free target.
976 	 * we need to unlock the page queues for this.
977 	 */
978 	if (free < uvmexp.freetarg) {
979 		uvmexp.pdswout++;
980 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout", free,
981 		    uvmexp.freetarg, 0, 0);
982 		uvm_unlock_pageq();
983 		uvm_swapout_threads();
984 		uvm_lock_pageq();
985 
986 	}
987 #endif
988 
989 	/*
990 	 * now we want to work on meeting our targets.   first we work on our
991 	 * free target by converting inactive pages into free pages.  then
992 	 * we work on meeting our inactive target by converting active pages
993 	 * to inactive ones.
994 	 */
995 
996 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
997 
998 	/*
999 	 * alternate starting queue between swap and object based on the
1000 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
1001 	 */
1002 
1003 	got_it = FALSE;
1004 	pages_freed = uvmexp.pdfreed;
1005 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
1006 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
1007 	if (!got_it)
1008 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
1009 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
1010 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
1011 	pages_freed = uvmexp.pdfreed - pages_freed;
1012 
1013 	/*
1014 	 * we have done the scan to get free pages.   now we work on meeting
1015 	 * our inactive target.
1016 	 */
1017 
1018 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
1019 
1020 	/*
1021 	 * detect if we're not going to be able to page anything out
1022 	 * until we free some swap resources from active pages.
1023 	 */
1024 
1025 	swap_shortage = 0;
1026 	if (uvmexp.free < uvmexp.freetarg &&
1027 	    uvmexp.swpginuse == uvmexp.swpages &&
1028 	    uvmexp.swpgonly < uvmexp.swpages &&
1029 	    pages_freed == 0) {
1030 		swap_shortage = uvmexp.freetarg - uvmexp.free;
1031 	}
1032 
1033 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
1034 		    inactive_shortage, swap_shortage,0,0);
1035 	for (p = TAILQ_FIRST(&uvm.page_active);
1036 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
1037 	     p = nextpg) {
1038 		nextpg = TAILQ_NEXT(p, pageq);
1039 		if (p->flags & PG_BUSY)
1040 			continue;	/* quick check before trying to lock */
1041 
1042 		/*
1043 		 * lock the page's owner.
1044 		 */
1045 		/* is page anon owned or ownerless? */
1046 		if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
1047 			KASSERT(p->uanon != NULL);
1048 			if (!simple_lock_try(&p->uanon->an_lock))
1049 				continue;
1050 
1051 			/* take over the page? */
1052 			if ((p->pqflags & PQ_ANON) == 0) {
1053 				KASSERT(p->loan_count > 0);
1054 				p->loan_count--;
1055 				p->pqflags |= PQ_ANON;
1056 			}
1057 		} else {
1058 			if (!simple_lock_try(&p->uobject->vmobjlock))
1059 				continue;
1060 		}
1061 
1062 		/*
1063 		 * skip this page if it's busy.
1064 		 */
1065 
1066 		if ((p->flags & PG_BUSY) != 0) {
1067 			if (p->pqflags & PQ_ANON)
1068 				simple_unlock(&p->uanon->an_lock);
1069 			else
1070 				simple_unlock(&p->uobject->vmobjlock);
1071 			continue;
1072 		}
1073 
1074 		/*
1075 		 * if there's a shortage of swap, free any swap allocated
1076 		 * to this page so that other pages can be paged out.
1077 		 */
1078 
1079 		if (swap_shortage > 0) {
1080 			if ((p->pqflags & PQ_ANON) && p->uanon->an_swslot) {
1081 				uvm_swap_free(p->uanon->an_swslot, 1);
1082 				p->uanon->an_swslot = 0;
1083 				p->flags &= ~PG_CLEAN;
1084 				swap_shortage--;
1085 			}
1086 			if (p->pqflags & PQ_AOBJ) {
1087 				int slot = uao_set_swslot(p->uobject,
1088 					p->offset >> PAGE_SHIFT, 0);
1089 				if (slot) {
1090 					uvm_swap_free(slot, 1);
1091 					p->flags &= ~PG_CLEAN;
1092 					swap_shortage--;
1093 				}
1094 			}
1095 		}
1096 
1097 		/*
1098 		 * deactivate this page if there's a shortage of
1099 		 * inactive pages.
1100 		 */
1101 
1102 		if (inactive_shortage > 0) {
1103 			pmap_page_protect(p, VM_PROT_NONE);
1104 			/* no need to check wire_count as pg is "active" */
1105 			uvm_pagedeactivate(p);
1106 			uvmexp.pddeact++;
1107 			inactive_shortage--;
1108 		}
1109 		if (p->pqflags & PQ_ANON)
1110 			simple_unlock(&p->uanon->an_lock);
1111 		else
1112 			simple_unlock(&p->uobject->vmobjlock);
1113 	}
1114 }
1115