xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 5a38ef86d0b61900239c7913d24a05e7b88a58f0)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.94 2021/12/15 12:53:53 mpi 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. Neither the name of the University nor the names of its contributors
22  *    may be used to endorse or promote products derived from this software
23  *    without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  *
37  *	@(#)vm_pageout.c        8.5 (Berkeley) 2/14/94
38  * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp
39  *
40  *
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  * uvm_pdaemon.c: the page daemon
67  */
68 
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
72 #include <sys/pool.h>
73 #include <sys/proc.h>
74 #include <sys/buf.h>
75 #include <sys/mount.h>
76 #include <sys/atomic.h>
77 
78 #ifdef HIBERNATE
79 #include <sys/hibernate.h>
80 #endif
81 
82 #include <uvm/uvm.h>
83 
84 #include "drm.h"
85 
86 #if NDRM > 0
87 extern void drmbackoff(long);
88 #endif
89 
90 /*
91  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
92  * in a pass thru the inactive list when swap is full.  the value should be
93  * "small"... if it's too large we'll cycle the active pages thru the inactive
94  * queue too quickly to for them to be referenced and avoid being freed.
95  */
96 
97 #define UVMPD_NUMDIRTYREACTS 16
98 
99 
100 /*
101  * local prototypes
102  */
103 
104 void		uvmpd_scan(void);
105 boolean_t	uvmpd_scan_inactive(struct pglist *);
106 void		uvmpd_tune(void);
107 void		uvmpd_drop(struct pglist *);
108 
109 /*
110  * uvm_wait: wait (sleep) for the page daemon to free some pages
111  *
112  * => should be called with all locks released
113  * => should _not_ be called by the page daemon (to avoid deadlock)
114  */
115 
116 void
117 uvm_wait(const char *wmsg)
118 {
119 	uint64_t timo = INFSLP;
120 
121 #ifdef DIAGNOSTIC
122 	if (curproc == &proc0)
123 		panic("%s: cannot sleep for memory during boot", __func__);
124 #endif
125 
126 	/* check for page daemon going to sleep (waiting for itself) */
127 	if (curproc == uvm.pagedaemon_proc) {
128 		printf("uvm_wait emergency bufbackoff\n");
129 		if (bufbackoff(NULL, 4) == 0)
130 			return;
131 		/*
132 		 * now we have a problem: the pagedaemon wants to go to
133 		 * sleep until it frees more memory.   but how can it
134 		 * free more memory if it is asleep?  that is a deadlock.
135 		 * we have two options:
136 		 *  [1] panic now
137 		 *  [2] put a timeout on the sleep, thus causing the
138 		 *      pagedaemon to only pause (rather than sleep forever)
139 		 *
140 		 * note that option [2] will only help us if we get lucky
141 		 * and some other process on the system breaks the deadlock
142 		 * by exiting or freeing memory (thus allowing the pagedaemon
143 		 * to continue).  for now we panic if DEBUG is defined,
144 		 * otherwise we hope for the best with option [2] (better
145 		 * yet, this should never happen in the first place!).
146 		 */
147 
148 		printf("pagedaemon: deadlock detected!\n");
149 		timo = MSEC_TO_NSEC(125);	/* set timeout */
150 #if defined(DEBUG)
151 		/* DEBUG: panic so we can debug it */
152 		panic("pagedaemon deadlock");
153 #endif
154 	}
155 
156 	uvm_lock_fpageq();
157 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
158 	msleep_nsec(&uvmexp.free, &uvm.fpageqlock, PVM | PNORELOCK, wmsg, timo);
159 }
160 
161 /*
162  * uvmpd_tune: tune paging parameters
163  *
164  * => called whenever memory is added to (or removed from?) the system
165  * => caller must call with page queues locked
166  */
167 
168 void
169 uvmpd_tune(void)
170 {
171 
172 	uvmexp.freemin = uvmexp.npages / 30;
173 
174 	/* between 16k and 512k */
175 	/* XXX:  what are these values good for? */
176 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
177 #if 0
178 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
179 #endif
180 
181 	/* Make sure there's always a user page free. */
182 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
183 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
184 
185 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
186 	if (uvmexp.freetarg <= uvmexp.freemin)
187 		uvmexp.freetarg = uvmexp.freemin + 1;
188 
189 	/* uvmexp.inactarg: computed in main daemon loop */
190 
191 	uvmexp.wiredmax = uvmexp.npages / 3;
192 }
193 
194 /*
195  * Indicate to the page daemon that a nowait call failed and it should
196  * recover at least some memory in the most restricted region (assumed
197  * to be dma_constraint).
198  */
199 volatile int uvm_nowait_failed;
200 
201 /*
202  * uvm_pageout: the main loop for the pagedaemon
203  */
204 void
205 uvm_pageout(void *arg)
206 {
207 	struct uvm_constraint_range constraint;
208 	struct uvm_pmalloc *pma;
209 	int npages = 0;
210 
211 	/* ensure correct priority and set paging parameters... */
212 	uvm.pagedaemon_proc = curproc;
213 	(void) spl0();
214 	uvm_lock_pageq();
215 	npages = uvmexp.npages;
216 	uvmpd_tune();
217 	uvm_unlock_pageq();
218 
219 	for (;;) {
220 		long size;
221 
222 		uvm_lock_fpageq();
223 		if (!uvm_nowait_failed && TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
224 			msleep_nsec(&uvm.pagedaemon, &uvm.fpageqlock, PVM,
225 			    "pgdaemon", INFSLP);
226 			uvmexp.pdwoke++;
227 		}
228 
229 		if ((pma = TAILQ_FIRST(&uvm.pmr_control.allocs)) != NULL) {
230 			pma->pm_flags |= UVM_PMA_BUSY;
231 			constraint = pma->pm_constraint;
232 		} else {
233 			if (uvm_nowait_failed) {
234 				/*
235 				 * XXX realisticly, this is what our
236 				 * nowait callers probably care about
237 				 */
238 				constraint = dma_constraint;
239 				uvm_nowait_failed = 0;
240 			} else
241 				constraint = no_constraint;
242 		}
243 
244 		uvm_unlock_fpageq();
245 
246 		/* now lock page queues and recompute inactive count */
247 		uvm_lock_pageq();
248 		if (npages != uvmexp.npages) {	/* check for new pages? */
249 			npages = uvmexp.npages;
250 			uvmpd_tune();
251 		}
252 
253 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
254 		if (uvmexp.inactarg <= uvmexp.freetarg) {
255 			uvmexp.inactarg = uvmexp.freetarg + 1;
256 		}
257 
258 		/* Reclaim pages from the buffer cache if possible. */
259 		size = 0;
260 		if (pma != NULL)
261 			size += pma->pm_size >> PAGE_SHIFT;
262 		if (uvmexp.free - BUFPAGES_DEFICIT < uvmexp.freetarg)
263 			size += uvmexp.freetarg - (uvmexp.free -
264 			    BUFPAGES_DEFICIT);
265 		if (size == 0)
266 			size = 16; /* XXX */
267 		uvm_unlock_pageq();
268 		(void) bufbackoff(&constraint, size * 2);
269 #if NDRM > 0
270 		drmbackoff(size * 2);
271 #endif
272 		uvm_lock_pageq();
273 
274 		/* Scan if needed to meet our targets. */
275 		if (pma != NULL ||
276 		    ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg) ||
277 		    ((uvmexp.inactive + BUFPAGES_INACT) < uvmexp.inactarg)) {
278 			uvmpd_scan();
279 		}
280 
281 		/*
282 		 * if there's any free memory to be had,
283 		 * wake up any waiters.
284 		 */
285 		uvm_lock_fpageq();
286 		if (uvmexp.free > uvmexp.reserve_kernel ||
287 		    uvmexp.paging == 0) {
288 			wakeup(&uvmexp.free);
289 		}
290 
291 		if (pma != NULL) {
292 			/*
293 			 * XXX If UVM_PMA_FREED isn't set, no pages
294 			 * were freed.  Should we set UVM_PMA_FAIL in
295 			 * that case?
296 			 */
297 			pma->pm_flags &= ~UVM_PMA_BUSY;
298 			if (pma->pm_flags & UVM_PMA_FREED) {
299 				pma->pm_flags &= ~UVM_PMA_LINKED;
300 				TAILQ_REMOVE(&uvm.pmr_control.allocs, pma,
301 				    pmq);
302 				wakeup(pma);
303 			}
304 		}
305 		uvm_unlock_fpageq();
306 
307 		/* scan done. unlock page queues (only lock we are holding) */
308 		uvm_unlock_pageq();
309 
310 		sched_pause(yield);
311 	}
312 	/*NOTREACHED*/
313 }
314 
315 
316 /*
317  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
318  */
319 void
320 uvm_aiodone_daemon(void *arg)
321 {
322 	int s, free;
323 	struct buf *bp, *nbp;
324 
325 	uvm.aiodoned_proc = curproc;
326 
327 	for (;;) {
328 		/*
329 		 * Check for done aio structures. If we've got structures to
330 		 * process, do so. Otherwise sleep while avoiding races.
331 		 */
332 		mtx_enter(&uvm.aiodoned_lock);
333 		while ((bp = TAILQ_FIRST(&uvm.aio_done)) == NULL)
334 			msleep_nsec(&uvm.aiodoned, &uvm.aiodoned_lock,
335 			    PVM, "aiodoned", INFSLP);
336 		/* Take the list for ourselves. */
337 		TAILQ_INIT(&uvm.aio_done);
338 		mtx_leave(&uvm.aiodoned_lock);
339 
340 		/* process each i/o that's done. */
341 		free = uvmexp.free;
342 		while (bp != NULL) {
343 			if (bp->b_flags & B_PDAEMON) {
344 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
345 			}
346 			nbp = TAILQ_NEXT(bp, b_freelist);
347 			s = splbio();	/* b_iodone must by called at splbio */
348 			(*bp->b_iodone)(bp);
349 			splx(s);
350 			bp = nbp;
351 
352 			sched_pause(yield);
353 		}
354 		uvm_lock_fpageq();
355 		wakeup(free <= uvmexp.reserve_kernel ? &uvm.pagedaemon :
356 		    &uvmexp.free);
357 		uvm_unlock_fpageq();
358 	}
359 }
360 
361 
362 
363 /*
364  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
365  *
366  * => called with page queues locked
367  * => we work on meeting our free target by converting inactive pages
368  *    into free pages.
369  * => we handle the building of swap-backed clusters
370  * => we return TRUE if we are exiting because we met our target
371  */
372 
373 boolean_t
374 uvmpd_scan_inactive(struct pglist *pglst)
375 {
376 	boolean_t retval = FALSE;	/* assume we haven't hit target */
377 	int free, result;
378 	struct vm_page *p, *nextpg;
379 	struct uvm_object *uobj;
380 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
381 	int npages;
382 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
383 	int swnpages, swcpages;				/* XXX: see below */
384 	int swslot;
385 	struct vm_anon *anon;
386 	boolean_t swap_backed;
387 	vaddr_t start;
388 	int dirtyreacts;
389 
390 	/*
391 	 * note: we currently keep swap-backed pages on a separate inactive
392 	 * list from object-backed pages.   however, merging the two lists
393 	 * back together again hasn't been ruled out.   thus, we keep our
394 	 * swap cluster in "swpps" rather than in pps (allows us to mix
395 	 * clustering types in the event of a mixed inactive queue).
396 	 */
397 	/*
398 	 * swslot is non-zero if we are building a swap cluster.  we want
399 	 * to stay in the loop while we have a page to scan or we have
400 	 * a swap-cluster to build.
401 	 */
402 	swslot = 0;
403 	swnpages = swcpages = 0;
404 	free = 0;
405 	dirtyreacts = 0;
406 
407 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
408 		/*
409 		 * note that p can be NULL iff we have traversed the whole
410 		 * list and need to do one final swap-backed clustered pageout.
411 		 */
412 		uobj = NULL;
413 		anon = NULL;
414 
415 		if (p) {
416 			/*
417 			 * update our copy of "free" and see if we've met
418 			 * our target
419 			 */
420 			free = uvmexp.free - BUFPAGES_DEFICIT;
421 
422 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
423 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
424 				retval = TRUE;
425 
426 				if (swslot == 0) {
427 					/* exit now if no swap-i/o pending */
428 					break;
429 				}
430 
431 				/* set p to null to signal final swap i/o */
432 				p = NULL;
433 			}
434 		}
435 
436 		if (p) {	/* if (we have a new page to consider) */
437 			/*
438 			 * we are below target and have a new page to consider.
439 			 */
440 			uvmexp.pdscans++;
441 			nextpg = TAILQ_NEXT(p, pageq);
442 
443 			if (p->pg_flags & PQ_ANON) {
444 				anon = p->uanon;
445 				KASSERT(anon != NULL);
446 				if (rw_enter(anon->an_lock,
447 				    RW_WRITE|RW_NOSLEEP)) {
448 					/* lock failed, skip this page */
449 					continue;
450 				}
451 				/*
452 				 * move referenced pages back to active queue
453 				 * and skip to next page.
454 				 */
455 				if (pmap_is_referenced(p)) {
456 					uvm_pageactivate(p);
457 					rw_exit(anon->an_lock);
458 					uvmexp.pdreact++;
459 					continue;
460 				}
461 				if (p->pg_flags & PG_BUSY) {
462 					rw_exit(anon->an_lock);
463 					uvmexp.pdbusy++;
464 					/* someone else owns page, skip it */
465 					continue;
466 				}
467 				uvmexp.pdanscan++;
468 			} else {
469 				uobj = p->uobject;
470 				KASSERT(uobj != NULL);
471 				if (rw_enter(uobj->vmobjlock,
472 				    RW_WRITE|RW_NOSLEEP)) {
473 					/* lock failed, skip this page */
474 					continue;
475 				}
476 				/*
477 				 * move referenced pages back to active queue
478 				 * and skip to next page.
479 				 */
480 				if (pmap_is_referenced(p)) {
481 					uvm_pageactivate(p);
482 					rw_exit(uobj->vmobjlock);
483 					uvmexp.pdreact++;
484 					continue;
485 				}
486 				if (p->pg_flags & PG_BUSY) {
487 					rw_exit(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 page queues locked.
497 			 * the page is not busy.   if the page is clean we
498 			 * can free it now and continue.
499 			 */
500 			if (p->pg_flags & PG_CLEAN) {
501 				if (p->pg_flags & PQ_SWAPBACKED) {
502 					/* this page now lives only in swap */
503 					atomic_inc_int(&uvmexp.swpgonly);
504 				}
505 
506 				/* zap all mappings with pmap_page_protect... */
507 				pmap_page_protect(p, PROT_NONE);
508 				uvm_pagefree(p);
509 				uvmexp.pdfreed++;
510 
511 				if (anon) {
512 
513 					/*
514 					 * an anonymous page can only be clean
515 					 * if it has backing store assigned.
516 					 */
517 
518 					KASSERT(anon->an_swslot != 0);
519 
520 					/* remove from object */
521 					anon->an_page = NULL;
522 					rw_exit(anon->an_lock);
523 				} else {
524 					rw_exit(uobj->vmobjlock);
525 				}
526 				continue;
527 			}
528 
529 			/*
530 			 * this page is dirty, skip it if we'll have met our
531 			 * free target when all the current pageouts complete.
532 			 */
533 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
534 				if (anon) {
535 					rw_exit(anon->an_lock);
536 				} else {
537 					rw_exit(uobj->vmobjlock);
538 				}
539 				continue;
540 			}
541 
542 			/*
543 			 * this page is dirty, but we can't page it out
544 			 * since all pages in swap are only in swap.
545 			 * reactivate it so that we eventually cycle
546 			 * all pages thru the inactive queue.
547 			 */
548 			if ((p->pg_flags & PQ_SWAPBACKED) && uvm_swapisfull()) {
549 				dirtyreacts++;
550 				uvm_pageactivate(p);
551 				if (anon) {
552 					rw_exit(anon->an_lock);
553 				} else {
554 					rw_exit(uobj->vmobjlock);
555 				}
556 				continue;
557 			}
558 
559 			/*
560 			 * if the page is swap-backed and dirty and swap space
561 			 * is full, free any swap allocated to the page
562 			 * so that other pages can be paged out.
563 			 */
564 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
565 			if ((p->pg_flags & PQ_SWAPBACKED) &&
566 			    uvmexp.swpginuse == uvmexp.swpages) {
567 
568 				if ((p->pg_flags & PQ_ANON) &&
569 				    p->uanon->an_swslot) {
570 					uvm_swap_free(p->uanon->an_swslot, 1);
571 					p->uanon->an_swslot = 0;
572 				}
573 				if (p->pg_flags & PQ_AOBJ) {
574 					uao_dropswap(p->uobject,
575 						     p->offset >> PAGE_SHIFT);
576 				}
577 			}
578 
579 			/*
580 			 * the page we are looking at is dirty.   we must
581 			 * clean it before it can be freed.  to do this we
582 			 * first mark the page busy so that no one else will
583 			 * touch the page.   we write protect all the mappings
584 			 * of the page so that no one touches it while it is
585 			 * in I/O.
586 			 */
587 
588 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
589 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
590 			UVM_PAGE_OWN(p, "scan_inactive");
591 			pmap_page_protect(p, PROT_READ);
592 			uvmexp.pgswapout++;
593 
594 			/*
595 			 * for swap-backed pages we need to (re)allocate
596 			 * swap space.
597 			 */
598 			if (swap_backed) {
599 				/* free old swap slot (if any) */
600 				if (anon) {
601 					if (anon->an_swslot) {
602 						uvm_swap_free(anon->an_swslot,
603 						    1);
604 						anon->an_swslot = 0;
605 					}
606 				} else {
607 					uao_dropswap(uobj,
608 						     p->offset >> PAGE_SHIFT);
609 				}
610 
611 				/* start new cluster (if necessary) */
612 				if (swslot == 0) {
613 					swnpages = MAXBSIZE >> PAGE_SHIFT;
614 					swslot = uvm_swap_alloc(&swnpages,
615 					    TRUE);
616 					if (swslot == 0) {
617 						/* no swap?  give up! */
618 						atomic_clearbits_int(
619 						    &p->pg_flags,
620 						    PG_BUSY);
621 						UVM_PAGE_OWN(p, NULL);
622 						if (anon)
623 							rw_exit(anon->an_lock);
624 						else
625 							rw_exit(
626 							    uobj->vmobjlock);
627 						continue;
628 					}
629 					swcpages = 0;	/* cluster is empty */
630 				}
631 
632 				/* add block to cluster */
633 				swpps[swcpages] = p;
634 				if (anon)
635 					anon->an_swslot = swslot + swcpages;
636 				else
637 					uao_set_swslot(uobj,
638 					    p->offset >> PAGE_SHIFT,
639 					    swslot + swcpages);
640 				swcpages++;
641 			}
642 		} else {
643 			/* if p == NULL we must be doing a last swap i/o */
644 			swap_backed = TRUE;
645 		}
646 
647 		/*
648 		 * now consider doing the pageout.
649 		 *
650 		 * for swap-backed pages, we do the pageout if we have either
651 		 * filled the cluster (in which case (swnpages == swcpages) or
652 		 * run out of pages (p == NULL).
653 		 *
654 		 * for object pages, we always do the pageout.
655 		 */
656 		if (swap_backed) {
657 			if (p) {	/* if we just added a page to cluster */
658 				if (anon)
659 					rw_exit(anon->an_lock);
660 				else
661 					rw_exit(uobj->vmobjlock);
662 
663 				/* cluster not full yet? */
664 				if (swcpages < swnpages)
665 					continue;
666 			}
667 
668 			/* starting I/O now... set up for it */
669 			npages = swcpages;
670 			ppsp = swpps;
671 			/* for swap-backed pages only */
672 			start = (vaddr_t) swslot;
673 
674 			/* if this is final pageout we could have a few
675 			 * extra swap blocks */
676 			if (swcpages < swnpages) {
677 				uvm_swap_free(swslot + swcpages,
678 				    (swnpages - swcpages));
679 			}
680 		} else {
681 			/* normal object pageout */
682 			ppsp = pps;
683 			npages = sizeof(pps) / sizeof(struct vm_page *);
684 			/* not looked at because PGO_ALLPAGES is set */
685 			start = 0;
686 		}
687 
688 		/*
689 		 * now do the pageout.
690 		 *
691 		 * for swap_backed pages we have already built the cluster.
692 		 * for !swap_backed pages, uvm_pager_put will call the object's
693 		 * "make put cluster" function to build a cluster on our behalf.
694 		 *
695 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
696 		 * it to free the cluster pages for us on a successful I/O (it
697 		 * always does this for un-successful I/O requests).  this
698 		 * allows us to do clustered pageout without having to deal
699 		 * with cluster pages at this level.
700 		 *
701 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
702 		 *  IN: locked: page queues
703 		 * OUT: locked:
704 		 *     !locked: pageqs
705 		 */
706 
707 		uvmexp.pdpageouts++;
708 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
709 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
710 
711 		/*
712 		 * if we did i/o to swap, zero swslot to indicate that we are
713 		 * no longer building a swap-backed cluster.
714 		 */
715 
716 		if (swap_backed)
717 			swslot = 0;		/* done with this cluster */
718 
719 		/*
720 		 * first, we check for VM_PAGER_PEND which means that the
721 		 * async I/O is in progress and the async I/O done routine
722 		 * will clean up after us.   in this case we move on to the
723 		 * next page.
724 		 *
725 		 * there is a very remote chance that the pending async i/o can
726 		 * finish _before_ we get here.   if that happens, our page "p"
727 		 * may no longer be on the inactive queue.   so we verify this
728 		 * when determining the next page (starting over at the head if
729 		 * we've lost our inactive page).
730 		 */
731 
732 		if (result == VM_PAGER_PEND) {
733 			uvmexp.paging += npages;
734 			uvm_lock_pageq();
735 			uvmexp.pdpending++;
736 			if (p) {
737 				if (p->pg_flags & PQ_INACTIVE)
738 					nextpg = TAILQ_NEXT(p, pageq);
739 				else
740 					nextpg = TAILQ_FIRST(pglst);
741 			} else {
742 				nextpg = NULL;
743 			}
744 			continue;
745 		}
746 
747 		/* clean up "p" if we have one */
748 		if (p) {
749 			/*
750 			 * the I/O request to "p" is done and uvm_pager_put
751 			 * has freed any cluster pages it may have allocated
752 			 * during I/O.  all that is left for us to do is
753 			 * clean up page "p" (which is still PG_BUSY).
754 			 *
755 			 * our result could be one of the following:
756 			 *   VM_PAGER_OK: successful pageout
757 			 *
758 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
759 			 *     to next page
760 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
761 			 *     "reactivate" page to get it out of the way (it
762 			 *     will eventually drift back into the inactive
763 			 *     queue for a retry).
764 			 *   VM_PAGER_UNLOCK: should never see this as it is
765 			 *     only valid for "get" operations
766 			 */
767 
768 			/* relock p's object: page queues not lock yet, so
769 			 * no need for "try" */
770 
771 			/* !swap_backed case: already locked... */
772 			if (swap_backed) {
773 				if (anon)
774 					rw_enter(anon->an_lock, RW_WRITE);
775 				else
776 					rw_enter(uobj->vmobjlock, RW_WRITE);
777 			}
778 
779 #ifdef DIAGNOSTIC
780 			if (result == VM_PAGER_UNLOCK)
781 				panic("pagedaemon: pageout returned "
782 				    "invalid 'unlock' code");
783 #endif
784 
785 			/* handle PG_WANTED now */
786 			if (p->pg_flags & PG_WANTED)
787 				wakeup(p);
788 
789 			atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED);
790 			UVM_PAGE_OWN(p, NULL);
791 
792 			/* released during I/O? Can only happen for anons */
793 			if (p->pg_flags & PG_RELEASED) {
794 				KASSERT(anon != NULL);
795 				/*
796 				 * remove page so we can get nextpg,
797 				 * also zero out anon so we don't use
798 				 * it after the free.
799 				 */
800 				anon->an_page = NULL;
801 				p->uanon = NULL;
802 
803 				rw_exit(anon->an_lock);
804 				uvm_anfree(anon);	/* kills anon */
805 				pmap_page_protect(p, PROT_NONE);
806 				anon = NULL;
807 				uvm_lock_pageq();
808 				nextpg = TAILQ_NEXT(p, pageq);
809 				/* free released page */
810 				uvm_pagefree(p);
811 			} else {	/* page was not released during I/O */
812 				uvm_lock_pageq();
813 				nextpg = TAILQ_NEXT(p, pageq);
814 				if (result != VM_PAGER_OK) {
815 					/* pageout was a failure... */
816 					if (result != VM_PAGER_AGAIN)
817 						uvm_pageactivate(p);
818 					pmap_clear_reference(p);
819 					/* XXXCDC: if (swap_backed) FREE p's
820 					 * swap block? */
821 				} else {
822 					/* pageout was a success... */
823 					pmap_clear_reference(p);
824 					pmap_clear_modify(p);
825 					atomic_setbits_int(&p->pg_flags,
826 					    PG_CLEAN);
827 				}
828 			}
829 
830 			/*
831 			 * drop object lock (if there is an object left).   do
832 			 * a safety check of nextpg to make sure it is on the
833 			 * inactive queue (it should be since PG_BUSY pages on
834 			 * the inactive queue can't be re-queued [note: not
835 			 * true for active queue]).
836 			 */
837 			if (anon)
838 				rw_exit(anon->an_lock);
839 			else if (uobj)
840 				rw_exit(uobj->vmobjlock);
841 
842 			if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
843 				nextpg = TAILQ_FIRST(pglst);	/* reload! */
844 			}
845 		} else {
846 			/*
847 			 * if p is null in this loop, make sure it stays null
848 			 * in the next loop.
849 			 */
850 			nextpg = NULL;
851 
852 			/*
853 			 * lock page queues here just so they're always locked
854 			 * at the end of the loop.
855 			 */
856 			uvm_lock_pageq();
857 		}
858 	}
859 	return (retval);
860 }
861 
862 /*
863  * uvmpd_scan: scan the page queues and attempt to meet our targets.
864  *
865  * => called with pageq's locked
866  */
867 
868 void
869 uvmpd_scan(void)
870 {
871 	int free, inactive_shortage, swap_shortage, pages_freed;
872 	struct vm_page *p, *nextpg;
873 	struct uvm_object *uobj;
874 	boolean_t got_it;
875 
876 	MUTEX_ASSERT_LOCKED(&uvm.pageqlock);
877 
878 	uvmexp.pdrevs++;		/* counter */
879 	uobj = NULL;
880 
881 	/*
882 	 * get current "free" page count
883 	 */
884 	free = uvmexp.free - BUFPAGES_DEFICIT;
885 
886 #ifndef __SWAP_BROKEN
887 	/*
888 	 * swap out some processes if we are below our free target.
889 	 * we need to unlock the page queues for this.
890 	 */
891 	if (free < uvmexp.freetarg) {
892 		uvmexp.pdswout++;
893 		uvm_unlock_pageq();
894 		uvm_swapout_threads();
895 		uvm_lock_pageq();
896 	}
897 #endif
898 
899 	/*
900 	 * now we want to work on meeting our targets.   first we work on our
901 	 * free target by converting inactive pages into free pages.  then
902 	 * we work on meeting our inactive target by converting active pages
903 	 * to inactive ones.
904 	 */
905 
906 	/*
907 	 * alternate starting queue between swap and object based on the
908 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
909 	 */
910 	got_it = FALSE;
911 	pages_freed = uvmexp.pdfreed;	/* XXX - int */
912 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
913 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
914 	if (!got_it)
915 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
916 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
917 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
918 	pages_freed = uvmexp.pdfreed - pages_freed;
919 
920 	/*
921 	 * we have done the scan to get free pages.   now we work on meeting
922 	 * our inactive target.
923 	 */
924 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive - BUFPAGES_INACT;
925 
926 	/*
927 	 * detect if we're not going to be able to page anything out
928 	 * until we free some swap resources from active pages.
929 	 */
930 	swap_shortage = 0;
931 	if (uvmexp.free < uvmexp.freetarg &&
932 	    uvmexp.swpginuse == uvmexp.swpages &&
933 	    !uvm_swapisfull() &&
934 	    pages_freed == 0) {
935 		swap_shortage = uvmexp.freetarg - uvmexp.free;
936 	}
937 
938 	for (p = TAILQ_FIRST(&uvm.page_active);
939 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
940 	     p = nextpg) {
941 		nextpg = TAILQ_NEXT(p, pageq);
942 
943 		/* skip this page if it's busy. */
944 		if (p->pg_flags & PG_BUSY)
945 			continue;
946 
947 		if (p->pg_flags & PQ_ANON) {
948 			KASSERT(p->uanon != NULL);
949 			if (rw_enter(p->uanon->an_lock, RW_WRITE|RW_NOSLEEP))
950 				continue;
951 		} else {
952 			KASSERT(p->uobject != NULL);
953 			if (rw_enter(p->uobject->vmobjlock,
954 			    RW_WRITE|RW_NOSLEEP))
955 				continue;
956 		}
957 
958 		/*
959 		 * if there's a shortage of swap, free any swap allocated
960 		 * to this page so that other pages can be paged out.
961 		 */
962 		if (swap_shortage > 0) {
963 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
964 				uvm_swap_free(p->uanon->an_swslot, 1);
965 				p->uanon->an_swslot = 0;
966 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
967 				swap_shortage--;
968 			}
969 			if (p->pg_flags & PQ_AOBJ) {
970 				int slot = uao_set_swslot(p->uobject,
971 					p->offset >> PAGE_SHIFT, 0);
972 				if (slot) {
973 					uvm_swap_free(slot, 1);
974 					atomic_clearbits_int(&p->pg_flags,
975 					    PG_CLEAN);
976 					swap_shortage--;
977 				}
978 			}
979 		}
980 
981 		/*
982 		 * deactivate this page if there's a shortage of
983 		 * inactive pages.
984 		 */
985 		if (inactive_shortage > 0) {
986 			pmap_page_protect(p, PROT_NONE);
987 			/* no need to check wire_count as pg is "active" */
988 			uvm_pagedeactivate(p);
989 			uvmexp.pddeact++;
990 			inactive_shortage--;
991 		}
992 		if (p->pg_flags & PQ_ANON)
993 			rw_exit(p->uanon->an_lock);
994 		else
995 			rw_exit(p->uobject->vmobjlock);
996 	}
997 }
998 
999 #ifdef HIBERNATE
1000 
1001 /*
1002  * uvmpd_drop: drop clean pages from list
1003  */
1004 void
1005 uvmpd_drop(struct pglist *pglst)
1006 {
1007 	struct vm_page *p, *nextpg;
1008 
1009 	for (p = TAILQ_FIRST(pglst); p != NULL; p = nextpg) {
1010 		nextpg = TAILQ_NEXT(p, pageq);
1011 
1012 		if (p->pg_flags & PQ_ANON || p->uobject == NULL)
1013 			continue;
1014 
1015 		if (p->pg_flags & PG_BUSY)
1016 			continue;
1017 
1018 		if (p->pg_flags & PG_CLEAN) {
1019 			struct uvm_object * uobj = p->uobject;
1020 
1021 			rw_enter(uobj->vmobjlock, RW_WRITE);
1022 			uvm_lock_pageq();
1023 			/*
1024 			 * we now have the page queues locked.
1025 			 * the page is not busy.   if the page is clean we
1026 			 * can free it now and continue.
1027 			 */
1028 			if (p->pg_flags & PG_CLEAN) {
1029 				if (p->pg_flags & PQ_SWAPBACKED) {
1030 					/* this page now lives only in swap */
1031 					atomic_inc_int(&uvmexp.swpgonly);
1032 				}
1033 
1034 				/* zap all mappings with pmap_page_protect... */
1035 				pmap_page_protect(p, PROT_NONE);
1036 				uvm_pagefree(p);
1037 			}
1038 			uvm_unlock_pageq();
1039 			rw_exit(uobj->vmobjlock);
1040 		}
1041 	}
1042 }
1043 
1044 void
1045 uvmpd_hibernate(void)
1046 {
1047 	uvmpd_drop(&uvm.page_inactive_swp);
1048 	uvmpd_drop(&uvm.page_inactive_obj);
1049 	uvmpd_drop(&uvm.page_active);
1050 }
1051 
1052 #endif
1053