xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 68dd5bb1859285b71cb62a10bf107b8ad54064d9)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.109 2023/10/27 19:18:53 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 struct rwlock	*uvmpd_trylockowner(struct vm_page *);
105 void		uvmpd_scan(struct uvm_pmalloc *, struct uvm_constraint_range *);
106 void		uvmpd_scan_inactive(struct uvm_pmalloc *,
107 		    struct uvm_constraint_range *, struct pglist *);
108 void		uvmpd_tune(void);
109 void		uvmpd_drop(struct pglist *);
110 void		uvmpd_dropswap(struct vm_page *);
111 
112 /*
113  * uvm_wait: wait (sleep) for the page daemon to free some pages
114  *
115  * => should be called with all locks released
116  * => should _not_ be called by the page daemon (to avoid deadlock)
117  */
118 
119 void
120 uvm_wait(const char *wmsg)
121 {
122 	uint64_t timo = INFSLP;
123 
124 #ifdef DIAGNOSTIC
125 	if (curproc == &proc0)
126 		panic("%s: cannot sleep for memory during boot", __func__);
127 #endif
128 
129 	/*
130 	 * check for page daemon going to sleep (waiting for itself)
131 	 */
132 	if (curproc == uvm.pagedaemon_proc) {
133 		printf("uvm_wait emergency bufbackoff\n");
134 		if (bufbackoff(NULL, 4) == 0)
135 			return;
136 		/*
137 		 * now we have a problem: the pagedaemon wants to go to
138 		 * sleep until it frees more memory.   but how can it
139 		 * free more memory if it is asleep?  that is a deadlock.
140 		 * we have two options:
141 		 *  [1] panic now
142 		 *  [2] put a timeout on the sleep, thus causing the
143 		 *      pagedaemon to only pause (rather than sleep forever)
144 		 *
145 		 * note that option [2] will only help us if we get lucky
146 		 * and some other process on the system breaks the deadlock
147 		 * by exiting or freeing memory (thus allowing the pagedaemon
148 		 * to continue).  for now we panic if DEBUG is defined,
149 		 * otherwise we hope for the best with option [2] (better
150 		 * yet, this should never happen in the first place!).
151 		 */
152 
153 		printf("pagedaemon: deadlock detected!\n");
154 		timo = MSEC_TO_NSEC(125);	/* set timeout */
155 #if defined(DEBUG)
156 		/* DEBUG: panic so we can debug it */
157 		panic("pagedaemon deadlock");
158 #endif
159 	}
160 
161 	uvm_lock_fpageq();
162 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
163 	msleep_nsec(&uvmexp.free, &uvm.fpageqlock, PVM | PNORELOCK, wmsg, timo);
164 }
165 
166 /*
167  * uvmpd_tune: tune paging parameters
168  *
169  * => called whenever memory is added to (or removed from?) the system
170  * => caller must call with page queues locked
171  */
172 
173 void
174 uvmpd_tune(void)
175 {
176 
177 	uvmexp.freemin = uvmexp.npages / 30;
178 
179 	/* between 16k and 512k */
180 	/* XXX:  what are these values good for? */
181 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
182 #if 0
183 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
184 #endif
185 
186 	/* Make sure there's always a user page free. */
187 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
188 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
189 
190 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
191 	if (uvmexp.freetarg <= uvmexp.freemin)
192 		uvmexp.freetarg = uvmexp.freemin + 1;
193 
194 	/* uvmexp.inactarg: computed in main daemon loop */
195 
196 	uvmexp.wiredmax = uvmexp.npages / 3;
197 }
198 
199 /*
200  * Indicate to the page daemon that a nowait call failed and it should
201  * recover at least some memory in the most restricted region (assumed
202  * to be dma_constraint).
203  */
204 volatile int uvm_nowait_failed;
205 
206 /*
207  * uvm_pageout: the main loop for the pagedaemon
208  */
209 void
210 uvm_pageout(void *arg)
211 {
212 	struct uvm_constraint_range constraint;
213 	struct uvm_pmalloc *pma;
214 	int npages = 0;
215 
216 	/* ensure correct priority and set paging parameters... */
217 	uvm.pagedaemon_proc = curproc;
218 	(void) spl0();
219 	uvm_lock_pageq();
220 	npages = uvmexp.npages;
221 	uvmpd_tune();
222 	uvm_unlock_pageq();
223 
224 	for (;;) {
225 		long size;
226 
227 		uvm_lock_fpageq();
228 		if (!uvm_nowait_failed && TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
229 			msleep_nsec(&uvm.pagedaemon, &uvm.fpageqlock, PVM,
230 			    "pgdaemon", INFSLP);
231 			uvmexp.pdwoke++;
232 		}
233 
234 		if ((pma = TAILQ_FIRST(&uvm.pmr_control.allocs)) != NULL) {
235 			pma->pm_flags |= UVM_PMA_BUSY;
236 			constraint = pma->pm_constraint;
237 		} else {
238 			if (uvm_nowait_failed) {
239 				/*
240 				 * XXX realistically, this is what our
241 				 * nowait callers probably care about
242 				 */
243 				constraint = dma_constraint;
244 				uvm_nowait_failed = 0;
245 			} else
246 				constraint = no_constraint;
247 		}
248 
249 		uvm_unlock_fpageq();
250 
251 		/*
252 		 * now lock page queues and recompute inactive count
253 		 */
254 		uvm_lock_pageq();
255 		if (npages != uvmexp.npages) {	/* check for new pages? */
256 			npages = uvmexp.npages;
257 			uvmpd_tune();
258 		}
259 
260 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
261 		if (uvmexp.inactarg <= uvmexp.freetarg) {
262 			uvmexp.inactarg = uvmexp.freetarg + 1;
263 		}
264 
265 		/* Reclaim pages from the buffer cache if possible. */
266 		size = 0;
267 		if (pma != NULL)
268 			size += pma->pm_size >> PAGE_SHIFT;
269 		if (uvmexp.free - BUFPAGES_DEFICIT < uvmexp.freetarg)
270 			size += uvmexp.freetarg - (uvmexp.free -
271 			    BUFPAGES_DEFICIT);
272 		if (size == 0)
273 			size = 16; /* XXX */
274 		uvm_unlock_pageq();
275 		(void) bufbackoff(&constraint, size * 2);
276 #if NDRM > 0
277 		drmbackoff(size * 2);
278 #endif
279 		uvm_lock_pageq();
280 
281 		/*
282 		 * scan if needed
283 		 */
284 		if (pma != NULL ||
285 		    ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg) ||
286 		    ((uvmexp.inactive + BUFPAGES_INACT) < uvmexp.inactarg)) {
287 			uvmpd_scan(pma, &constraint);
288 		}
289 
290 		/*
291 		 * if there's any free memory to be had,
292 		 * wake up any waiters.
293 		 */
294 		uvm_lock_fpageq();
295 		if (uvmexp.free > uvmexp.reserve_kernel ||
296 		    uvmexp.paging == 0) {
297 			wakeup(&uvmexp.free);
298 		}
299 
300 		if (pma != NULL) {
301 			/*
302 			 * XXX If UVM_PMA_FREED isn't set, no pages
303 			 * were freed.  Should we set UVM_PMA_FAIL in
304 			 * that case?
305 			 */
306 			pma->pm_flags &= ~UVM_PMA_BUSY;
307 			if (pma->pm_flags & UVM_PMA_FREED) {
308 				pma->pm_flags &= ~UVM_PMA_LINKED;
309 				TAILQ_REMOVE(&uvm.pmr_control.allocs, pma,
310 				    pmq);
311 				wakeup(pma);
312 			}
313 		}
314 		uvm_unlock_fpageq();
315 
316 		/*
317 		 * scan done.  unlock page queues (the only lock we are holding)
318 		 */
319 		uvm_unlock_pageq();
320 
321 		sched_pause(yield);
322 	}
323 	/*NOTREACHED*/
324 }
325 
326 
327 /*
328  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
329  */
330 void
331 uvm_aiodone_daemon(void *arg)
332 {
333 	int s, free;
334 	struct buf *bp, *nbp;
335 
336 	uvm.aiodoned_proc = curproc;
337 
338 	for (;;) {
339 		/*
340 		 * Check for done aio structures. If we've got structures to
341 		 * process, do so. Otherwise sleep while avoiding races.
342 		 */
343 		mtx_enter(&uvm.aiodoned_lock);
344 		while ((bp = TAILQ_FIRST(&uvm.aio_done)) == NULL)
345 			msleep_nsec(&uvm.aiodoned, &uvm.aiodoned_lock,
346 			    PVM, "aiodoned", INFSLP);
347 		/* Take the list for ourselves. */
348 		TAILQ_INIT(&uvm.aio_done);
349 		mtx_leave(&uvm.aiodoned_lock);
350 
351 		/* process each i/o that's done. */
352 		free = uvmexp.free;
353 		while (bp != NULL) {
354 			if (bp->b_flags & B_PDAEMON) {
355 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
356 			}
357 			nbp = TAILQ_NEXT(bp, b_freelist);
358 			s = splbio();	/* b_iodone must by called at splbio */
359 			(*bp->b_iodone)(bp);
360 			splx(s);
361 			bp = nbp;
362 
363 			sched_pause(yield);
364 		}
365 		uvm_lock_fpageq();
366 		wakeup(free <= uvmexp.reserve_kernel ? &uvm.pagedaemon :
367 		    &uvmexp.free);
368 		uvm_unlock_fpageq();
369 	}
370 }
371 
372 /*
373  * uvmpd_trylockowner: trylock the page's owner.
374  *
375  * => return the locked rwlock on success.  otherwise, return NULL.
376  */
377 struct rwlock *
378 uvmpd_trylockowner(struct vm_page *pg)
379 {
380 
381 	struct uvm_object *uobj = pg->uobject;
382 	struct rwlock *slock;
383 
384 	if (uobj != NULL) {
385 		slock = uobj->vmobjlock;
386 	} else {
387 		struct vm_anon *anon = pg->uanon;
388 
389 		KASSERT(anon != NULL);
390 		slock = anon->an_lock;
391 	}
392 
393 	if (rw_enter(slock, RW_WRITE|RW_NOSLEEP)) {
394 		return NULL;
395 	}
396 
397 	return slock;
398 }
399 
400 
401 /*
402  * uvmpd_dropswap: free any swap allocated to this page.
403  *
404  * => called with owner locked.
405  */
406 void
407 uvmpd_dropswap(struct vm_page *pg)
408 {
409 	struct vm_anon *anon = pg->uanon;
410 
411 	if ((pg->pg_flags & PQ_ANON) && anon->an_swslot) {
412 		uvm_swap_free(anon->an_swslot, 1);
413 		anon->an_swslot = 0;
414 	} else if (pg->pg_flags & PQ_AOBJ) {
415 		uao_dropswap(pg->uobject, pg->offset >> PAGE_SHIFT);
416 	}
417 }
418 
419 /*
420  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
421  *
422  * => called with page queues locked
423  * => we work on meeting our free target by converting inactive pages
424  *    into free pages.
425  * => we handle the building of swap-backed clusters
426  * => we return TRUE if we are exiting because we met our target
427  */
428 void
429 uvmpd_scan_inactive(struct uvm_pmalloc *pma,
430     struct uvm_constraint_range *constraint, struct pglist *pglst)
431 {
432 	int free, result;
433 	struct vm_page *p, *nextpg;
434 	struct uvm_object *uobj;
435 	struct vm_page *pps[SWCLUSTPAGES], **ppsp;
436 	int npages;
437 	struct vm_page *swpps[SWCLUSTPAGES]; 	/* XXX: see below */
438 	struct rwlock *slock;
439 	int swnpages, swcpages;				/* XXX: see below */
440 	int swslot;
441 	struct vm_anon *anon;
442 	boolean_t swap_backed;
443 	vaddr_t start;
444 	int dirtyreacts;
445 	paddr_t paddr;
446 
447 	/*
448 	 * swslot is non-zero if we are building a swap cluster.  we want
449 	 * to stay in the loop while we have a page to scan or we have
450 	 * a swap-cluster to build.
451 	 */
452 	swslot = 0;
453 	swnpages = swcpages = 0;
454 	dirtyreacts = 0;
455 	p = NULL;
456 
457 	/* Start with the first page on the list that fit in `constraint' */
458 	TAILQ_FOREACH(p, pglst, pageq) {
459 		paddr = atop(VM_PAGE_TO_PHYS(p));
460 		if (paddr >= constraint->ucr_low &&
461 		    paddr < constraint->ucr_high)
462 			break;
463 	}
464 
465 	for (; p != NULL || swslot != 0; p = nextpg) {
466 		/*
467 		 * note that p can be NULL iff we have traversed the whole
468 		 * list and need to do one final swap-backed clustered pageout.
469 		 */
470 		uobj = NULL;
471 		anon = NULL;
472 		if (p) {
473 			/*
474 			 * see if we've met our target
475 			 */
476 			free = uvmexp.free - BUFPAGES_DEFICIT;
477 			if (((pma == NULL || (pma->pm_flags & UVM_PMA_FREED)) &&
478 			    (free + uvmexp.paging >= uvmexp.freetarg << 2)) ||
479 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
480 				if (swslot == 0) {
481 					/* exit now if no swap-i/o pending */
482 					break;
483 				}
484 
485 				/* set p to null to signal final swap i/o */
486 				p = NULL;
487 				nextpg = NULL;
488 			}
489 		}
490 		if (p) {	/* if (we have a new page to consider) */
491 			/*
492 			 * we are below target and have a new page to consider.
493 			 */
494 			uvmexp.pdscans++;
495 			nextpg = TAILQ_NEXT(p, pageq);
496 
497 			anon = p->uanon;
498 			uobj = p->uobject;
499 
500 			/*
501 			 * first we attempt to lock the object that this page
502 			 * belongs to.  if our attempt fails we skip on to
503 			 * the next page (no harm done).  it is important to
504 			 * "try" locking the object as we are locking in the
505 			 * wrong order (pageq -> object) and we don't want to
506 			 * deadlock.
507 			 */
508 			slock = uvmpd_trylockowner(p);
509 			if (slock == NULL) {
510 				continue;
511 			}
512 
513 			/*
514 			 * move referenced pages back to active queue
515 			 * and skip to next page.
516 			 */
517 			if (pmap_is_referenced(p)) {
518 				uvm_pageactivate(p);
519 				rw_exit(slock);
520 				uvmexp.pdreact++;
521 				continue;
522 			}
523 
524 			if (p->pg_flags & PG_BUSY) {
525 				rw_exit(slock);
526 				uvmexp.pdbusy++;
527 				continue;
528 			}
529 
530 			/* does the page belong to an object? */
531 			if (uobj != NULL) {
532 				uvmexp.pdobscan++;
533 			} else {
534 				KASSERT(anon != NULL);
535 				uvmexp.pdanscan++;
536 			}
537 
538 			/*
539 			 * we now have the page queues locked.
540 			 * the page is not busy.   if the page is clean we
541 			 * can free it now and continue.
542 			 */
543 			if (p->pg_flags & PG_CLEAN) {
544 				if (p->pg_flags & PQ_SWAPBACKED) {
545 					/* this page now lives only in swap */
546 					atomic_inc_int(&uvmexp.swpgonly);
547 				}
548 
549 				/* zap all mappings with pmap_page_protect... */
550 				pmap_page_protect(p, PROT_NONE);
551 				uvm_pagefree(p);
552 				uvmexp.pdfreed++;
553 
554 				if (anon) {
555 
556 					/*
557 					 * an anonymous page can only be clean
558 					 * if it has backing store assigned.
559 					 */
560 
561 					KASSERT(anon->an_swslot != 0);
562 
563 					/* remove from object */
564 					anon->an_page = NULL;
565 				}
566 				rw_exit(slock);
567 				continue;
568 			}
569 
570 			/*
571 			 * this page is dirty, skip it if we'll have met our
572 			 * free target when all the current pageouts complete.
573 			 */
574 			if ((pma == NULL || (pma->pm_flags & UVM_PMA_FREED)) &&
575 			    (free + uvmexp.paging > uvmexp.freetarg << 2)) {
576 				rw_exit(slock);
577 				continue;
578 			}
579 
580 			/*
581 			 * this page is dirty, but we can't page it out
582 			 * since all pages in swap are only in swap.
583 			 * reactivate it so that we eventually cycle
584 			 * all pages thru the inactive queue.
585 			 */
586 			if ((p->pg_flags & PQ_SWAPBACKED) && uvm_swapisfull()) {
587 				dirtyreacts++;
588 				uvm_pageactivate(p);
589 				rw_exit(slock);
590 				continue;
591 			}
592 
593 			/*
594 			 * if the page is swap-backed and dirty and swap space
595 			 * is full, free any swap allocated to the page
596 			 * so that other pages can be paged out.
597 			 */
598 			if ((p->pg_flags & PQ_SWAPBACKED) && uvm_swapisfilled())
599 				uvmpd_dropswap(p);
600 
601 			/*
602 			 * the page we are looking at is dirty.   we must
603 			 * clean it before it can be freed.  to do this we
604 			 * first mark the page busy so that no one else will
605 			 * touch the page.   we write protect all the mappings
606 			 * of the page so that no one touches it while it is
607 			 * in I/O.
608 			 */
609 
610 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
611 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
612 			UVM_PAGE_OWN(p, "scan_inactive");
613 			pmap_page_protect(p, PROT_READ);
614 			uvmexp.pgswapout++;
615 
616 			/*
617 			 * for swap-backed pages we need to (re)allocate
618 			 * swap space.
619 			 */
620 			if (swap_backed) {
621 				/* free old swap slot (if any) */
622 				uvmpd_dropswap(p);
623 
624 				/* start new cluster (if necessary) */
625 				if (swslot == 0) {
626 					swnpages = SWCLUSTPAGES;
627 					swslot = uvm_swap_alloc(&swnpages,
628 					    TRUE);
629 					if (swslot == 0) {
630 						/* no swap?  give up! */
631 						atomic_clearbits_int(
632 						    &p->pg_flags,
633 						    PG_BUSY);
634 						UVM_PAGE_OWN(p, NULL);
635 						rw_exit(slock);
636 						continue;
637 					}
638 					swcpages = 0;	/* cluster is empty */
639 				}
640 
641 				/* add block to cluster */
642 				swpps[swcpages] = p;
643 				if (anon)
644 					anon->an_swslot = swslot + swcpages;
645 				else
646 					uao_set_swslot(uobj,
647 					    p->offset >> PAGE_SHIFT,
648 					    swslot + swcpages);
649 				swcpages++;
650 				rw_exit(slock);
651 
652 				/* cluster not full yet? */
653 				if (swcpages < swnpages)
654 					continue;
655 			}
656 		} else {
657 			/* if p == NULL we must be doing a last swap i/o */
658 			swap_backed = TRUE;
659 		}
660 
661 		/*
662 		 * now consider doing the pageout.
663 		 *
664 		 * for swap-backed pages, we do the pageout if we have either
665 		 * filled the cluster (in which case (swnpages == swcpages) or
666 		 * run out of pages (p == NULL).
667 		 *
668 		 * for object pages, we always do the pageout.
669 		 */
670 		if (swap_backed) {
671 			/* starting I/O now... set up for it */
672 			npages = swcpages;
673 			ppsp = swpps;
674 			/* for swap-backed pages only */
675 			start = (vaddr_t) swslot;
676 
677 			/* if this is final pageout we could have a few
678 			 * extra swap blocks */
679 			if (swcpages < swnpages) {
680 				uvm_swap_free(swslot + swcpages,
681 				    (swnpages - swcpages));
682 			}
683 		} else {
684 			/* normal object pageout */
685 			ppsp = pps;
686 			npages = sizeof(pps) / sizeof(struct vm_page *);
687 			/* not looked at because PGO_ALLPAGES is set */
688 			start = 0;
689 		}
690 
691 		/*
692 		 * now do the pageout.
693 		 *
694 		 * for swap_backed pages we have already built the cluster.
695 		 * for !swap_backed pages, uvm_pager_put will call the object's
696 		 * "make put cluster" function to build a cluster on our behalf.
697 		 *
698 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
699 		 * it to free the cluster pages for us on a successful I/O (it
700 		 * always does this for un-successful I/O requests).  this
701 		 * allows us to do clustered pageout without having to deal
702 		 * with cluster pages at this level.
703 		 *
704 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
705 		 *  IN: locked: page queues
706 		 * OUT: locked:
707 		 *     !locked: pageqs
708 		 */
709 
710 		uvmexp.pdpageouts++;
711 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
712 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
713 
714 		/*
715 		 * if we did i/o to swap, zero swslot to indicate that we are
716 		 * no longer building a swap-backed cluster.
717 		 */
718 
719 		if (swap_backed)
720 			swslot = 0;		/* done with this cluster */
721 
722 		/*
723 		 * first, we check for VM_PAGER_PEND which means that the
724 		 * async I/O is in progress and the async I/O done routine
725 		 * will clean up after us.   in this case we move on to the
726 		 * next page.
727 		 *
728 		 * there is a very remote chance that the pending async i/o can
729 		 * finish _before_ we get here.   if that happens, our page "p"
730 		 * may no longer be on the inactive queue.   so we verify this
731 		 * when determining the next page (starting over at the head if
732 		 * we've lost our inactive page).
733 		 */
734 
735 		if (result == VM_PAGER_PEND) {
736 			uvmexp.paging += npages;
737 			uvm_lock_pageq();
738 			uvmexp.pdpending++;
739 			if (p) {
740 				if (p->pg_flags & PQ_INACTIVE)
741 					nextpg = TAILQ_NEXT(p, pageq);
742 				else
743 					nextpg = TAILQ_FIRST(pglst);
744 			} else {
745 				nextpg = NULL;
746 			}
747 			continue;
748 		}
749 
750 		/* clean up "p" if we have one */
751 		if (p) {
752 			/*
753 			 * the I/O request to "p" is done and uvm_pager_put
754 			 * has freed any cluster pages it may have allocated
755 			 * during I/O.  all that is left for us to do is
756 			 * clean up page "p" (which is still PG_BUSY).
757 			 *
758 			 * our result could be one of the following:
759 			 *   VM_PAGER_OK: successful pageout
760 			 *
761 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
762 			 *     to next page
763 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
764 			 *     "reactivate" page to get it out of the way (it
765 			 *     will eventually drift back into the inactive
766 			 *     queue for a retry).
767 			 *   VM_PAGER_UNLOCK: should never see this as it is
768 			 *     only valid for "get" operations
769 			 */
770 
771 			/* relock p's object: page queues not lock yet, so
772 			 * no need for "try" */
773 
774 			/* !swap_backed case: already locked... */
775 			if (swap_backed) {
776 				rw_enter(slock, 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 			rw_exit(slock);
838 
839 			if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
840 				nextpg = TAILQ_FIRST(pglst);	/* reload! */
841 			}
842 		} else {
843 			/*
844 			 * if p is null in this loop, make sure it stays null
845 			 * in the next loop.
846 			 */
847 			nextpg = NULL;
848 
849 			/*
850 			 * lock page queues here just so they're always locked
851 			 * at the end of the loop.
852 			 */
853 			uvm_lock_pageq();
854 		}
855 	}
856 }
857 
858 /*
859  * uvmpd_scan: scan the page queues and attempt to meet our targets.
860  *
861  * => called with pageq's locked
862  */
863 
864 void
865 uvmpd_scan(struct uvm_pmalloc *pma, struct uvm_constraint_range *constraint)
866 {
867 	int free, inactive_shortage, swap_shortage, pages_freed;
868 	struct vm_page *p, *nextpg;
869 	struct rwlock *slock;
870 	paddr_t paddr;
871 
872 	MUTEX_ASSERT_LOCKED(&uvm.pageqlock);
873 
874 	uvmexp.pdrevs++;		/* counter */
875 
876 	/*
877 	 * get current "free" page count
878 	 */
879 	free = uvmexp.free - BUFPAGES_DEFICIT;
880 
881 #ifdef __HAVE_PMAP_COLLECT
882 	/*
883 	 * swap out some processes if we are below our free target.
884 	 * we need to unlock the page queues for this.
885 	 */
886 	if (free < uvmexp.freetarg) {
887 		uvmexp.pdswout++;
888 		uvm_unlock_pageq();
889 		uvm_swapout_threads();
890 		uvm_lock_pageq();
891 	}
892 #endif
893 
894 	/*
895 	 * now we want to work on meeting our targets.   first we work on our
896 	 * free target by converting inactive pages into free pages.  then
897 	 * we work on meeting our inactive target by converting active pages
898 	 * to inactive ones.
899 	 */
900 
901 	pages_freed = uvmexp.pdfreed;
902 	(void) uvmpd_scan_inactive(pma, constraint, &uvm.page_inactive);
903 	pages_freed = uvmexp.pdfreed - pages_freed;
904 
905 	/*
906 	 * we have done the scan to get free pages.   now we work on meeting
907 	 * our inactive target.
908 	 */
909 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive - BUFPAGES_INACT;
910 
911 	/*
912 	 * detect if we're not going to be able to page anything out
913 	 * until we free some swap resources from active pages.
914 	 */
915 	free = uvmexp.free - BUFPAGES_DEFICIT;
916 	swap_shortage = 0;
917 	if (free < uvmexp.freetarg && uvm_swapisfilled() && !uvm_swapisfull() &&
918 	    pages_freed == 0) {
919 		swap_shortage = uvmexp.freetarg - free;
920 	}
921 
922 	for (p = TAILQ_FIRST(&uvm.page_active);
923 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
924 	     p = nextpg) {
925 		nextpg = TAILQ_NEXT(p, pageq);
926 		if (p->pg_flags & PG_BUSY) {
927 			continue;
928 		}
929 
930 		/*
931 		 * skip this page if it doesn't match the constraint.
932 		 */
933 		paddr = atop(VM_PAGE_TO_PHYS(p));
934 		if (paddr < constraint->ucr_low &&
935 		    paddr >= constraint->ucr_high)
936 			continue;
937 
938 		/*
939 		 * lock the page's owner.
940 		 */
941 		slock = uvmpd_trylockowner(p);
942 		if (slock == NULL) {
943 			continue;
944 		}
945 
946 		/*
947 		 * skip this page if it's busy.
948 		 */
949 		if ((p->pg_flags & PG_BUSY) != 0) {
950 			rw_exit(slock);
951 			continue;
952 		}
953 
954 		/*
955 		 * if there's a shortage of swap, free any swap allocated
956 		 * to this page so that other pages can be paged out.
957 		 */
958 		if (swap_shortage > 0) {
959 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
960 				uvm_swap_free(p->uanon->an_swslot, 1);
961 				p->uanon->an_swslot = 0;
962 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
963 				swap_shortage--;
964 			}
965 			if (p->pg_flags & PQ_AOBJ) {
966 				int slot = uao_set_swslot(p->uobject,
967 					p->offset >> PAGE_SHIFT, 0);
968 				if (slot) {
969 					uvm_swap_free(slot, 1);
970 					atomic_clearbits_int(&p->pg_flags,
971 					    PG_CLEAN);
972 					swap_shortage--;
973 				}
974 			}
975 		}
976 
977 		/*
978 		 * deactivate this page if there's a shortage of
979 		 * inactive pages.
980 		 */
981 		if (inactive_shortage > 0) {
982 			pmap_page_protect(p, PROT_NONE);
983 			/* no need to check wire_count as pg is "active" */
984 			uvm_pagedeactivate(p);
985 			uvmexp.pddeact++;
986 			inactive_shortage--;
987 		}
988 
989 		/*
990 		 * we're done with this page.
991 		 */
992 		rw_exit(slock);
993 	}
994 }
995 
996 #ifdef HIBERNATE
997 
998 /*
999  * uvmpd_drop: drop clean pages from list
1000  */
1001 void
1002 uvmpd_drop(struct pglist *pglst)
1003 {
1004 	struct vm_page *p, *nextpg;
1005 
1006 	for (p = TAILQ_FIRST(pglst); p != NULL; p = nextpg) {
1007 		nextpg = TAILQ_NEXT(p, pageq);
1008 
1009 		if (p->pg_flags & PQ_ANON || p->uobject == NULL)
1010 			continue;
1011 
1012 		if (p->pg_flags & PG_BUSY)
1013 			continue;
1014 
1015 		if (p->pg_flags & PG_CLEAN) {
1016 			struct uvm_object * uobj = p->uobject;
1017 
1018 			rw_enter(uobj->vmobjlock, RW_WRITE);
1019 			uvm_lock_pageq();
1020 			/*
1021 			 * we now have the page queues locked.
1022 			 * the page is not busy.   if the page is clean we
1023 			 * can free it now and continue.
1024 			 */
1025 			if (p->pg_flags & PG_CLEAN) {
1026 				if (p->pg_flags & PQ_SWAPBACKED) {
1027 					/* this page now lives only in swap */
1028 					atomic_inc_int(&uvmexp.swpgonly);
1029 				}
1030 
1031 				/* zap all mappings with pmap_page_protect... */
1032 				pmap_page_protect(p, PROT_NONE);
1033 				uvm_pagefree(p);
1034 			}
1035 			uvm_unlock_pageq();
1036 			rw_exit(uobj->vmobjlock);
1037 		}
1038 	}
1039 }
1040 
1041 void
1042 uvmpd_hibernate(void)
1043 {
1044 	uvmpd_drop(&uvm.page_inactive);
1045 	uvmpd_drop(&uvm.page_active);
1046 }
1047 
1048 #endif
1049