xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 2b0358df1d88d06ef4139321dd05bd5e05d91eaf)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.38 2009/04/06 12:02:52 oga Exp $	*/
2 /*	$NetBSD: uvm_pdaemon.c,v 1.23 2000/08/20 10:24:14 bjh21 Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * Copyright (c) 1991, 1993, The Regents of the University of California.
7  *
8  * All rights reserved.
9  *
10  * This code is derived from software contributed to Berkeley by
11  * The Mach Operating System project at Carnegie-Mellon University.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by Charles D. Cranor,
24  *      Washington University, the University of California, Berkeley and
25  *      its contributors.
26  * 4. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)vm_pageout.c        8.5 (Berkeley) 2/14/94
43  * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp
44  *
45  *
46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47  * All rights reserved.
48  *
49  * Permission to use, copy, modify and distribute this software and
50  * its documentation is hereby granted, provided that both the copyright
51  * notice and this permission notice appear in all copies of the
52  * software, derivative works or modified versions, and any portions
53  * thereof, and that both notices appear in supporting documentation.
54  *
55  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58  *
59  * Carnegie Mellon requests users of this software to return to
60  *
61  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
62  *  School of Computer Science
63  *  Carnegie Mellon University
64  *  Pittsburgh PA 15213-3890
65  *
66  * any improvements or extensions that they make and grant Carnegie the
67  * rights to redistribute these changes.
68  */
69 
70 /*
71  * uvm_pdaemon.c: the page daemon
72  */
73 
74 #include <sys/param.h>
75 #include <sys/proc.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/pool.h>
79 #include <sys/buf.h>
80 #include <sys/vnode.h>
81 #include <sys/mount.h>
82 
83 #include <uvm/uvm.h>
84 
85 /*
86  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
87  * in a pass thru the inactive list when swap is full.  the value should be
88  * "small"... if it's too large we'll cycle the active pages thru the inactive
89  * queue too quickly to for them to be referenced and avoid being freed.
90  */
91 
92 #define UVMPD_NUMDIRTYREACTS 16
93 
94 
95 /*
96  * local prototypes
97  */
98 
99 static void		uvmpd_scan(void);
100 static boolean_t	uvmpd_scan_inactive(struct pglist *);
101 static void		uvmpd_tune(void);
102 
103 /*
104  * uvm_wait: wait (sleep) for the page daemon to free some pages
105  *
106  * => should be called with all locks released
107  * => should _not_ be called by the page daemon (to avoid deadlock)
108  */
109 
110 void
111 uvm_wait(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(void)
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 #if 0
173 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
174 #endif
175 
176 	/* Make sure there's always a user page free. */
177 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
178 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
179 
180 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
181 	if (uvmexp.freetarg <= uvmexp.freemin)
182 		uvmexp.freetarg = uvmexp.freemin + 1;
183 
184 	/* uvmexp.inactarg: computed in main daemon loop */
185 
186 	uvmexp.wiredmax = uvmexp.npages / 3;
187 	UVMHIST_LOG(pdhist, "<- done, freemin=%ld, freetarg=%ld, wiredmax=%ld",
188 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
189 }
190 
191 /*
192  * uvm_pageout: the main loop for the pagedaemon
193  */
194 
195 void
196 uvm_pageout(void *arg)
197 {
198 	int npages = 0;
199 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
200 
201 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
202 
203 	/*
204 	 * ensure correct priority and set paging parameters...
205 	 */
206 
207 	uvm.pagedaemon_proc = curproc;
208 	(void) spl0();
209 	uvm_lock_pageq();
210 	npages = uvmexp.npages;
211 	uvmpd_tune();
212 	uvm_unlock_pageq();
213 
214 	/*
215 	 * main loop
216 	 */
217 
218 	for (;;) {
219 		simple_lock(&uvm.pagedaemon_lock);
220 
221 		UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
222 		UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
223 		    &uvm.pagedaemon_lock, FALSE, "pgdaemon", 0);
224 		uvmexp.pdwoke++;
225 		UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
226 
227 		/*
228 		 * now lock page queues and recompute inactive count
229 		 */
230 
231 		uvm_lock_pageq();
232 		if (npages != uvmexp.npages) {	/* check for new pages? */
233 			npages = uvmexp.npages;
234 			uvmpd_tune();
235 		}
236 
237 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
238 		if (uvmexp.inactarg <= uvmexp.freetarg) {
239 			uvmexp.inactarg = uvmexp.freetarg + 1;
240 		}
241 
242 		UVMHIST_LOG(pdhist,"  free/ftarg=%ld/%ld, inact/itarg=%ld/%ld",
243 		    uvmexp.free, uvmexp.freetarg, uvmexp.inactive,
244 		    uvmexp.inactarg);
245 
246 		/*
247 		 * scan if needed
248 		 */
249 		if ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg ||
250 		    uvmexp.inactive < uvmexp.inactarg) {
251 			uvmpd_scan();
252 		}
253 
254 		/*
255 		 * if there's any free memory to be had,
256 		 * wake up any waiters.
257 		 */
258 
259 		if (uvmexp.free > uvmexp.reserve_kernel ||
260 		    uvmexp.paging == 0) {
261 			wakeup(&uvmexp.free);
262 		}
263 
264 		/*
265 		 * scan done.  unlock page queues (the only lock we are holding)
266 		 */
267 
268 		uvm_unlock_pageq();
269 	}
270 	/*NOTREACHED*/
271 }
272 
273 
274 /*
275  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
276  */
277 
278 void
279 uvm_aiodone_daemon(void *arg)
280 {
281 	int s, free;
282 	struct buf *bp, *nbp;
283 	UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist);
284 
285 	uvm.aiodoned_proc = curproc;
286 
287 	for (;;) {
288 
289 		/*
290 		 * Check for done aio structures. If we've got structures to
291 		 * process, do so. Otherwise sleep while avoiding races.
292 		 */
293 		mtx_enter(&uvm.aiodoned_lock);
294 		while ((bp = TAILQ_FIRST(&uvm.aio_done)) == NULL)
295 			msleep(&uvm.aiodoned, &uvm.aiodoned_lock,
296 			    PVM, "aiodoned", 0);
297 		/* Take the list for ourselves. */
298 		TAILQ_INIT(&uvm.aio_done);
299 		mtx_leave(&uvm.aiodoned_lock);
300 
301 		/*
302 		 * process each i/o that's done.
303 		 */
304 
305 		free = uvmexp.free;
306 		while (bp != NULL) {
307 			if (bp->b_flags & B_PDAEMON) {
308 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
309 			}
310 			nbp = TAILQ_NEXT(bp, b_freelist);
311 			s = splbio();	/* b_iodone must by called at splbio */
312 			(*bp->b_iodone)(bp);
313 			splx(s);
314 			bp = nbp;
315 		}
316 		if (free <= uvmexp.reserve_kernel) {
317 			uvm_lock_fpageq();
318 			wakeup(&uvm.pagedaemon);
319 			uvm_unlock_fpageq();
320 		} else {
321 			simple_lock(&uvm.pagedaemon_lock);
322 			wakeup(&uvmexp.free);
323 			simple_unlock(&uvm.pagedaemon_lock);
324 		}
325 	}
326 }
327 
328 
329 
330 /*
331  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
332  *
333  * => called with page queues locked
334  * => we work on meeting our free target by converting inactive pages
335  *    into free pages.
336  * => we handle the building of swap-backed clusters
337  * => we return TRUE if we are exiting because we met our target
338  */
339 
340 static boolean_t
341 uvmpd_scan_inactive(struct pglist *pglst)
342 {
343 	boolean_t retval = FALSE;	/* assume we haven't hit target */
344 	int free, result;
345 	struct vm_page *p, *nextpg;
346 	struct uvm_object *uobj;
347 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
348 	int npages;
349 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
350 	int swnpages, swcpages;				/* XXX: see below */
351 	int swslot;
352 	struct vm_anon *anon;
353 	boolean_t swap_backed;
354 	vaddr_t start;
355 	int dirtyreacts;
356 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
357 
358 	/*
359 	 * note: we currently keep swap-backed pages on a separate inactive
360 	 * list from object-backed pages.   however, merging the two lists
361 	 * back together again hasn't been ruled out.   thus, we keep our
362 	 * swap cluster in "swpps" rather than in pps (allows us to mix
363 	 * clustering types in the event of a mixed inactive queue).
364 	 */
365 
366 	/*
367 	 * swslot is non-zero if we are building a swap cluster.  we want
368 	 * to stay in the loop while we have a page to scan or we have
369 	 * a swap-cluster to build.
370 	 */
371 
372 	swslot = 0;
373 	swnpages = swcpages = 0;
374 	free = 0;
375 	dirtyreacts = 0;
376 
377 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
378 
379 		/*
380 		 * note that p can be NULL iff we have traversed the whole
381 		 * list and need to do one final swap-backed clustered pageout.
382 		 */
383 
384 		uobj = NULL;
385 		anon = NULL;
386 
387 		if (p) {
388 
389 			/*
390 			 * update our copy of "free" and see if we've met
391 			 * our target
392 			 */
393 
394 			uvm_lock_fpageq();
395 			free = uvmexp.free - BUFPAGES_DEFICIT;
396 			uvm_unlock_fpageq();
397 
398 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
399 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
400 				UVMHIST_LOG(pdhist,"  met free target: "
401 					    "exit loop", 0, 0, 0, 0);
402 				retval = TRUE;
403 
404 				if (swslot == 0) {
405 					/* exit now if no swap-i/o pending */
406 					break;
407 				}
408 
409 				/* set p to null to signal final swap i/o */
410 				p = NULL;
411 			}
412 		}
413 
414 		if (p) {	/* if (we have a new page to consider) */
415 
416 			/*
417 			 * we are below target and have a new page to consider.
418 			 */
419 			uvmexp.pdscans++;
420 			nextpg = TAILQ_NEXT(p, pageq);
421 
422 			/*
423 			 * move referenced pages back to active queue and
424 			 * skip to next page (unlikely to happen since
425 			 * inactive pages shouldn't have any valid mappings
426 			 * and we cleared reference before deactivating).
427 			 */
428 
429 			if (pmap_is_referenced(p)) {
430 				uvm_pageactivate(p);
431 				uvmexp.pdreact++;
432 				continue;
433 			}
434 
435 			/*
436 			 * first we attempt to lock the object that this page
437 			 * belongs to.  if our attempt fails we skip on to
438 			 * the next page (no harm done).  it is important to
439 			 * "try" locking the object as we are locking in the
440 			 * wrong order (pageq -> object) and we don't want to
441 			 * deadlock.
442 			 *
443 			 * the only time we expect to see an ownerless page
444 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
445 			 * anon has loaned a page from a uvm_object and the
446 			 * uvm_object has dropped the ownership.  in that
447 			 * case, the anon can "take over" the loaned page
448 			 * and make it its own.
449 			 */
450 
451 			/* is page part of an anon or ownerless ? */
452 			if ((p->pg_flags & PQ_ANON) || p->uobject == NULL) {
453 				anon = p->uanon;
454 				KASSERT(anon != NULL);
455 				if (!simple_lock_try(&anon->an_lock)) {
456 					/* lock failed, skip this page */
457 					continue;
458 				}
459 
460 				/*
461 				 * if the page is ownerless, claim it in the
462 				 * name of "anon"!
463 				 */
464 
465 				if ((p->pg_flags & PQ_ANON) == 0) {
466 					KASSERT(p->loan_count > 0);
467 					p->loan_count--;
468 					atomic_setbits_int(&p->pg_flags,
469 					    PQ_ANON);
470 					/* anon now owns it */
471 				}
472 				if (p->pg_flags & PG_BUSY) {
473 					simple_unlock(&anon->an_lock);
474 					uvmexp.pdbusy++;
475 					/* someone else owns page, skip it */
476 					continue;
477 				}
478 				uvmexp.pdanscan++;
479 			} else {
480 				uobj = p->uobject;
481 				KASSERT(uobj != NULL);
482 				if (!simple_lock_try(&uobj->vmobjlock)) {
483 					/* lock failed, skip this page */
484 					continue;
485 				}
486 				if (p->pg_flags & PG_BUSY) {
487 					simple_unlock(&uobj->vmobjlock);
488 					uvmexp.pdbusy++;
489 					/* someone else owns page, skip it */
490 					continue;
491 				}
492 				uvmexp.pdobscan++;
493 			}
494 
495 			/*
496 			 * we now have the object and the page queues locked.
497 			 * the page is not busy.   if the page is clean we
498 			 * can free it now and continue.
499 			 */
500 
501 			if (p->pg_flags & PG_CLEAN) {
502 				if (p->pg_flags & PQ_SWAPBACKED) {
503 					/* this page now lives only in swap */
504 					simple_lock(&uvm.swap_data_lock);
505 					uvmexp.swpgonly++;
506 					simple_unlock(&uvm.swap_data_lock);
507 				}
508 
509 				/* zap all mappings with pmap_page_protect... */
510 				pmap_page_protect(p, VM_PROT_NONE);
511 				uvm_pagefree(p);
512 				uvmexp.pdfreed++;
513 
514 				if (anon) {
515 
516 					/*
517 					 * an anonymous page can only be clean
518 					 * if it has backing store assigned.
519 					 */
520 
521 					KASSERT(anon->an_swslot != 0);
522 
523 					/* remove from object */
524 					anon->an_page = NULL;
525 					simple_unlock(&anon->an_lock);
526 				} else {
527 					/* pagefree has already removed the
528 					 * page from the object */
529 					simple_unlock(&uobj->vmobjlock);
530 				}
531 				continue;
532 			}
533 
534 			/*
535 			 * this page is dirty, skip it if we'll have met our
536 			 * free target when all the current pageouts complete.
537 			 */
538 
539 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
540 				if (anon) {
541 					simple_unlock(&anon->an_lock);
542 				} else {
543 					simple_unlock(&uobj->vmobjlock);
544 				}
545 				continue;
546 			}
547 
548 			/*
549 			 * this page is dirty, but we can't page it out
550 			 * since all pages in swap are only in swap.
551 			 * reactivate it so that we eventually cycle
552 			 * all pages thru the inactive queue.
553 			 */
554 
555 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
556 			if ((p->pg_flags & PQ_SWAPBACKED) &&
557 			    uvmexp.swpgonly == uvmexp.swpages) {
558 				dirtyreacts++;
559 				uvm_pageactivate(p);
560 				if (anon) {
561 					simple_unlock(&anon->an_lock);
562 				} else {
563 					simple_unlock(&uobj->vmobjlock);
564 				}
565 				continue;
566 			}
567 
568 			/*
569 			 * if the page is swap-backed and dirty and swap space
570 			 * is full, free any swap allocated to the page
571 			 * so that other pages can be paged out.
572 			 */
573 
574 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
575 			if ((p->pg_flags & PQ_SWAPBACKED) &&
576 			    uvmexp.swpginuse == uvmexp.swpages) {
577 
578 				if ((p->pg_flags & PQ_ANON) &&
579 				    p->uanon->an_swslot) {
580 					uvm_swap_free(p->uanon->an_swslot, 1);
581 					p->uanon->an_swslot = 0;
582 				}
583 				if (p->pg_flags & PQ_AOBJ) {
584 					uao_dropswap(p->uobject,
585 						     p->offset >> PAGE_SHIFT);
586 				}
587 			}
588 
589 			/*
590 			 * the page we are looking at is dirty.   we must
591 			 * clean it before it can be freed.  to do this we
592 			 * first mark the page busy so that no one else will
593 			 * touch the page.   we write protect all the mappings
594 			 * of the page so that no one touches it while it is
595 			 * in I/O.
596 			 */
597 
598 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
599 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
600 			UVM_PAGE_OWN(p, "scan_inactive");
601 			pmap_page_protect(p, VM_PROT_READ);
602 			uvmexp.pgswapout++;
603 
604 			/*
605 			 * for swap-backed pages we need to (re)allocate
606 			 * swap space.
607 			 */
608 
609 			if (swap_backed) {
610 
611 				/*
612 				 * free old swap slot (if any)
613 				 */
614 
615 				if (anon) {
616 					if (anon->an_swslot) {
617 						uvm_swap_free(anon->an_swslot,
618 						    1);
619 						anon->an_swslot = 0;
620 					}
621 				} else {
622 					uao_dropswap(uobj,
623 						     p->offset >> PAGE_SHIFT);
624 				}
625 
626 				/*
627 				 * start new cluster (if necessary)
628 				 */
629 
630 				if (swslot == 0) {
631 					swnpages = MAXBSIZE >> PAGE_SHIFT;
632 					swslot = uvm_swap_alloc(&swnpages,
633 					    TRUE);
634 					if (swslot == 0) {
635 						/* no swap?  give up! */
636 						atomic_clearbits_int(
637 						    &p->pg_flags,
638 						    PG_BUSY);
639 						UVM_PAGE_OWN(p, NULL);
640 						if (anon)
641 							simple_unlock(
642 							    &anon->an_lock);
643 						else
644 							simple_unlock(
645 							    &uobj->vmobjlock);
646 						continue;
647 					}
648 					swcpages = 0;	/* cluster is empty */
649 				}
650 
651 				/*
652 				 * add block to cluster
653 				 */
654 
655 				swpps[swcpages] = p;
656 				if (anon)
657 					anon->an_swslot = swslot + swcpages;
658 				else
659 					uao_set_swslot(uobj,
660 					    p->offset >> PAGE_SHIFT,
661 					    swslot + swcpages);
662 				swcpages++;
663 			}
664 		} else {
665 
666 			/* if p == NULL we must be doing a last swap i/o */
667 			swap_backed = TRUE;
668 		}
669 
670 		/*
671 		 * now consider doing the pageout.
672 		 *
673 		 * for swap-backed pages, we do the pageout if we have either
674 		 * filled the cluster (in which case (swnpages == swcpages) or
675 		 * run out of pages (p == NULL).
676 		 *
677 		 * for object pages, we always do the pageout.
678 		 */
679 
680 		if (swap_backed) {
681 			if (p) {	/* if we just added a page to cluster */
682 				if (anon)
683 					simple_unlock(&anon->an_lock);
684 				else
685 					simple_unlock(&uobj->vmobjlock);
686 
687 				/* cluster not full yet? */
688 				if (swcpages < swnpages)
689 					continue;
690 			}
691 
692 			/* starting I/O now... set up for it */
693 			npages = swcpages;
694 			ppsp = swpps;
695 			/* for swap-backed pages only */
696 			start = (vaddr_t) swslot;
697 
698 			/* if this is final pageout we could have a few
699 			 * extra swap blocks */
700 			if (swcpages < swnpages) {
701 				uvm_swap_free(swslot + swcpages,
702 				    (swnpages - swcpages));
703 			}
704 		} else {
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 		 * now do the pageout.
714 		 *
715 		 * for swap_backed pages we have already built the cluster.
716 		 * for !swap_backed pages, uvm_pager_put will call the object's
717 		 * "make put cluster" function to build a cluster on our behalf.
718 		 *
719 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
720 		 * it to free the cluster pages for us on a successful I/O (it
721 		 * always does this for un-successful I/O requests).  this
722 		 * allows us to do clustered pageout without having to deal
723 		 * with cluster pages at this level.
724 		 *
725 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
726 		 *  IN: locked: uobj (if !swap_backed), page queues
727 		 * OUT: locked: uobj (if !swap_backed && result !=VM_PAGER_PEND)
728 		 *     !locked: pageqs, uobj (if swap_backed || VM_PAGER_PEND)
729 		 *
730 		 * [the bit about VM_PAGER_PEND saves us one lock-unlock pair]
731 		 */
732 
733 		/* locked: uobj (if !swap_backed), page queues */
734 		uvmexp.pdpageouts++;
735 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
736 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
737 		/* locked: uobj (if !swap_backed && result != PEND) */
738 		/* unlocked: pageqs, object (if swap_backed ||result == PEND) */
739 
740 		/*
741 		 * if we did i/o to swap, zero swslot to indicate that we are
742 		 * no longer building a swap-backed cluster.
743 		 */
744 
745 		if (swap_backed)
746 			swslot = 0;		/* done with this cluster */
747 
748 		/*
749 		 * first, we check for VM_PAGER_PEND which means that the
750 		 * async I/O is in progress and the async I/O done routine
751 		 * will clean up after us.   in this case we move on to the
752 		 * next page.
753 		 *
754 		 * there is a very remote chance that the pending async i/o can
755 		 * finish _before_ we get here.   if that happens, our page "p"
756 		 * may no longer be on the inactive queue.   so we verify this
757 		 * when determining the next page (starting over at the head if
758 		 * we've lost our inactive page).
759 		 */
760 
761 		if (result == VM_PAGER_PEND) {
762 			uvmexp.paging += npages;
763 			uvm_lock_pageq();
764 			uvmexp.pdpending++;
765 			if (p) {
766 				if (p->pg_flags & PQ_INACTIVE)
767 					nextpg = TAILQ_NEXT(p, pageq);
768 				else
769 					nextpg = TAILQ_FIRST(pglst);
770 			} else {
771 				nextpg = NULL;
772 			}
773 			continue;
774 		}
775 
776 #ifdef UBC
777 		if (result == VM_PAGER_ERROR &&
778 		    curproc == uvm.pagedaemon_proc) {
779 			uvm_lock_pageq();
780 			nextpg = TAILQ_NEXT(p, pageq);
781 			uvm_pageactivate(p);
782 			continue;
783 		}
784 #endif
785 
786 		/*
787 		 * clean up "p" if we have one
788 		 */
789 
790 		if (p) {
791 			/*
792 			 * the I/O request to "p" is done and uvm_pager_put
793 			 * has freed any cluster pages it may have allocated
794 			 * during I/O.  all that is left for us to do is
795 			 * clean up page "p" (which is still PG_BUSY).
796 			 *
797 			 * our result could be one of the following:
798 			 *   VM_PAGER_OK: successful pageout
799 			 *
800 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
801 			 *     to next page
802 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
803 			 *     "reactivate" page to get it out of the way (it
804 			 *     will eventually drift back into the inactive
805 			 *     queue for a retry).
806 			 *   VM_PAGER_UNLOCK: should never see this as it is
807 			 *     only valid for "get" operations
808 			 */
809 
810 			/* relock p's object: page queues not lock yet, so
811 			 * no need for "try" */
812 
813 			/* !swap_backed case: already locked... */
814 			if (swap_backed) {
815 				if (anon)
816 					simple_lock(&anon->an_lock);
817 				else
818 					simple_lock(&uobj->vmobjlock);
819 			}
820 
821 #ifdef DIAGNOSTIC
822 			if (result == VM_PAGER_UNLOCK)
823 				panic("pagedaemon: pageout returned "
824 				    "invalid 'unlock' code");
825 #endif
826 
827 			/* handle PG_WANTED now */
828 			if (p->pg_flags & PG_WANTED)
829 				/* still holding object lock */
830 				wakeup(p);
831 
832 			atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED);
833 			UVM_PAGE_OWN(p, NULL);
834 
835 			/* released during I/O? */
836 			if (p->pg_flags & PG_RELEASED) {
837 				if (anon) {
838 					/* remove page so we can get nextpg */
839 					anon->an_page = NULL;
840 
841 					simple_unlock(&anon->an_lock);
842 					uvm_anfree(anon);	/* kills anon */
843 					pmap_page_protect(p, VM_PROT_NONE);
844 					anon = NULL;
845 					uvm_lock_pageq();
846 					nextpg = TAILQ_NEXT(p, pageq);
847 					/* free released page */
848 					uvm_pagefree(p);
849 
850 				} else {
851 
852 					/*
853 					 * pgo_releasepg nukes the page and
854 					 * gets "nextpg" for us.  it returns
855 					 * with the page queues locked (when
856 					 * given nextpg ptr).
857 					 */
858 
859 					if (!uobj->pgops->pgo_releasepg(p,
860 					    &nextpg))
861 						/* uobj died after release */
862 						uobj = NULL;
863 
864 					/*
865 					 * lock page queues here so that they're
866 					 * always locked at the end of the loop.
867 					 */
868 
869 					uvm_lock_pageq();
870 				}
871 			} else {	/* page was not released during I/O */
872 				uvm_lock_pageq();
873 				nextpg = TAILQ_NEXT(p, pageq);
874 				if (result != VM_PAGER_OK) {
875 					/* pageout was a failure... */
876 					if (result != VM_PAGER_AGAIN)
877 						uvm_pageactivate(p);
878 					pmap_clear_reference(p);
879 					/* XXXCDC: if (swap_backed) FREE p's
880 					 * swap block? */
881 				} else {
882 					/* pageout was a success... */
883 					pmap_clear_reference(p);
884 					pmap_clear_modify(p);
885 					atomic_setbits_int(&p->pg_flags,
886 					    PG_CLEAN);
887 				}
888 			}
889 
890 			/*
891 			 * drop object lock (if there is an object left).   do
892 			 * a safety check of nextpg to make sure it is on the
893 			 * inactive queue (it should be since PG_BUSY pages on
894 			 * the inactive queue can't be re-queued [note: not
895 			 * true for active queue]).
896 			 */
897 
898 			if (anon)
899 				simple_unlock(&anon->an_lock);
900 			else if (uobj)
901 				simple_unlock(&uobj->vmobjlock);
902 
903 		} else {
904 
905 			/*
906 			 * if p is null in this loop, make sure it stays null
907 			 * in the next loop.
908 			 */
909 
910 			nextpg = NULL;
911 
912 			/*
913 			 * lock page queues here just so they're always locked
914 			 * at the end of the loop.
915 			 */
916 
917 			uvm_lock_pageq();
918 		}
919 
920 		if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
921 			nextpg = TAILQ_FIRST(pglst);	/* reload! */
922 		}
923 	}
924 	return (retval);
925 }
926 
927 /*
928  * uvmpd_scan: scan the page queues and attempt to meet our targets.
929  *
930  * => called with pageq's locked
931  */
932 
933 void
934 uvmpd_scan(void)
935 {
936 	int free, inactive_shortage, swap_shortage, pages_freed;
937 	struct vm_page *p, *nextpg;
938 	struct uvm_object *uobj;
939 	boolean_t got_it;
940 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
941 
942 	uvmexp.pdrevs++;		/* counter */
943 	uobj = NULL;
944 
945 	/*
946 	 * get current "free" page count
947 	 */
948 	uvm_lock_fpageq();
949 	free = uvmexp.free - BUFPAGES_DEFICIT;
950 	uvm_unlock_fpageq();
951 
952 #ifndef __SWAP_BROKEN
953 	/*
954 	 * swap out some processes if we are below our free target.
955 	 * we need to unlock the page queues for this.
956 	 */
957 	if (free < uvmexp.freetarg) {
958 		uvmexp.pdswout++;
959 		UVMHIST_LOG(pdhist,"  free %ld < target %ld: swapout", free,
960 		    uvmexp.freetarg, 0, 0);
961 		uvm_unlock_pageq();
962 		uvm_swapout_threads();
963 		uvm_lock_pageq();
964 	}
965 #endif
966 
967 	/*
968 	 * now we want to work on meeting our targets.   first we work on our
969 	 * free target by converting inactive pages into free pages.  then
970 	 * we work on meeting our inactive target by converting active pages
971 	 * to inactive ones.
972 	 */
973 
974 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
975 
976 	/*
977 	 * alternate starting queue between swap and object based on the
978 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
979 	 */
980 
981 	got_it = FALSE;
982 	pages_freed = uvmexp.pdfreed;	/* XXX - int */
983 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
984 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
985 	if (!got_it)
986 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
987 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
988 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
989 	pages_freed = uvmexp.pdfreed - pages_freed;
990 
991 	/*
992 	 * we have done the scan to get free pages.   now we work on meeting
993 	 * our inactive target.
994 	 */
995 
996 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
997 
998 	/*
999 	 * detect if we're not going to be able to page anything out
1000 	 * until we free some swap resources from active pages.
1001 	 */
1002 
1003 	swap_shortage = 0;
1004 	if (uvmexp.free < uvmexp.freetarg &&
1005 	    uvmexp.swpginuse == uvmexp.swpages &&
1006 	    uvmexp.swpgonly < uvmexp.swpages &&
1007 	    pages_freed == 0) {
1008 		swap_shortage = uvmexp.freetarg - uvmexp.free;
1009 	}
1010 
1011 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%ld swap_shortage=%ld",
1012 		    inactive_shortage, swap_shortage,0,0);
1013 	for (p = TAILQ_FIRST(&uvm.page_active);
1014 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
1015 	     p = nextpg) {
1016 		nextpg = TAILQ_NEXT(p, pageq);
1017 		if (p->pg_flags & PG_BUSY)
1018 			continue;	/* quick check before trying to lock */
1019 
1020 		/*
1021 		 * lock the page's owner.
1022 		 */
1023 		/* is page anon owned or ownerless? */
1024 		if ((p->pg_flags & PQ_ANON) || p->uobject == NULL) {
1025 			KASSERT(p->uanon != NULL);
1026 			if (!simple_lock_try(&p->uanon->an_lock))
1027 				continue;
1028 
1029 			/* take over the page? */
1030 			if ((p->pg_flags & PQ_ANON) == 0) {
1031 				KASSERT(p->loan_count > 0);
1032 				p->loan_count--;
1033 				atomic_setbits_int(&p->pg_flags, PQ_ANON);
1034 			}
1035 		} else {
1036 			if (!simple_lock_try(&p->uobject->vmobjlock))
1037 				continue;
1038 		}
1039 
1040 		/*
1041 		 * skip this page if it's busy.
1042 		 */
1043 
1044 		if ((p->pg_flags & PG_BUSY) != 0) {
1045 			if (p->pg_flags & PQ_ANON)
1046 				simple_unlock(&p->uanon->an_lock);
1047 			else
1048 				simple_unlock(&p->uobject->vmobjlock);
1049 			continue;
1050 		}
1051 
1052 		/*
1053 		 * if there's a shortage of swap, free any swap allocated
1054 		 * to this page so that other pages can be paged out.
1055 		 */
1056 
1057 		if (swap_shortage > 0) {
1058 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
1059 				uvm_swap_free(p->uanon->an_swslot, 1);
1060 				p->uanon->an_swslot = 0;
1061 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
1062 				swap_shortage--;
1063 			}
1064 			if (p->pg_flags & PQ_AOBJ) {
1065 				int slot = uao_set_swslot(p->uobject,
1066 					p->offset >> PAGE_SHIFT, 0);
1067 				if (slot) {
1068 					uvm_swap_free(slot, 1);
1069 					atomic_clearbits_int(&p->pg_flags,
1070 					    PG_CLEAN);
1071 					swap_shortage--;
1072 				}
1073 			}
1074 		}
1075 
1076 		/*
1077 		 * deactivate this page if there's a shortage of
1078 		 * inactive pages.
1079 		 */
1080 
1081 		if (inactive_shortage > 0) {
1082 			pmap_page_protect(p, VM_PROT_NONE);
1083 			/* no need to check wire_count as pg is "active" */
1084 			uvm_pagedeactivate(p);
1085 			uvmexp.pddeact++;
1086 			inactive_shortage--;
1087 		}
1088 		if (p->pg_flags & PQ_ANON)
1089 			simple_unlock(&p->uanon->an_lock);
1090 		else
1091 			simple_unlock(&p->uobject->vmobjlock);
1092 	}
1093 }
1094