xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 2218185d8dd7873a1f0c8b5a21ea97ce9a81b966)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.99 2022/05/12 12:49:31 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 	/*
127 	 * check for page daemon going to sleep (waiting for itself)
128 	 */
129 	if (curproc == uvm.pagedaemon_proc) {
130 		printf("uvm_wait emergency bufbackoff\n");
131 		if (bufbackoff(NULL, 4) == 0)
132 			return;
133 		/*
134 		 * now we have a problem: the pagedaemon wants to go to
135 		 * sleep until it frees more memory.   but how can it
136 		 * free more memory if it is asleep?  that is a deadlock.
137 		 * we have two options:
138 		 *  [1] panic now
139 		 *  [2] put a timeout on the sleep, thus causing the
140 		 *      pagedaemon to only pause (rather than sleep forever)
141 		 *
142 		 * note that option [2] will only help us if we get lucky
143 		 * and some other process on the system breaks the deadlock
144 		 * by exiting or freeing memory (thus allowing the pagedaemon
145 		 * to continue).  for now we panic if DEBUG is defined,
146 		 * otherwise we hope for the best with option [2] (better
147 		 * yet, this should never happen in the first place!).
148 		 */
149 
150 		printf("pagedaemon: deadlock detected!\n");
151 		timo = MSEC_TO_NSEC(125);	/* set timeout */
152 #if defined(DEBUG)
153 		/* DEBUG: panic so we can debug it */
154 		panic("pagedaemon deadlock");
155 #endif
156 	}
157 
158 	uvm_lock_fpageq();
159 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
160 	msleep_nsec(&uvmexp.free, &uvm.fpageqlock, PVM | PNORELOCK, wmsg, timo);
161 }
162 
163 /*
164  * uvmpd_tune: tune paging parameters
165  *
166  * => called whenever memory is added to (or removed from?) the system
167  * => caller must call with page queues locked
168  */
169 
170 void
171 uvmpd_tune(void)
172 {
173 
174 	uvmexp.freemin = uvmexp.npages / 30;
175 
176 	/* between 16k and 512k */
177 	/* XXX:  what are these values good for? */
178 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
179 #if 0
180 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
181 #endif
182 
183 	/* Make sure there's always a user page free. */
184 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
185 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
186 
187 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
188 	if (uvmexp.freetarg <= uvmexp.freemin)
189 		uvmexp.freetarg = uvmexp.freemin + 1;
190 
191 	/* uvmexp.inactarg: computed in main daemon loop */
192 
193 	uvmexp.wiredmax = uvmexp.npages / 3;
194 }
195 
196 /*
197  * Indicate to the page daemon that a nowait call failed and it should
198  * recover at least some memory in the most restricted region (assumed
199  * to be dma_constraint).
200  */
201 volatile int uvm_nowait_failed;
202 
203 /*
204  * uvm_pageout: the main loop for the pagedaemon
205  */
206 void
207 uvm_pageout(void *arg)
208 {
209 	struct uvm_constraint_range constraint;
210 	struct uvm_pmalloc *pma;
211 	int npages = 0;
212 
213 	/* ensure correct priority and set paging parameters... */
214 	uvm.pagedaemon_proc = curproc;
215 	(void) spl0();
216 	uvm_lock_pageq();
217 	npages = uvmexp.npages;
218 	uvmpd_tune();
219 	uvm_unlock_pageq();
220 
221 	for (;;) {
222 		long size;
223 
224 		uvm_lock_fpageq();
225 		if (!uvm_nowait_failed && TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
226 			msleep_nsec(&uvm.pagedaemon, &uvm.fpageqlock, PVM,
227 			    "pgdaemon", INFSLP);
228 			uvmexp.pdwoke++;
229 		}
230 
231 		if ((pma = TAILQ_FIRST(&uvm.pmr_control.allocs)) != NULL) {
232 			pma->pm_flags |= UVM_PMA_BUSY;
233 			constraint = pma->pm_constraint;
234 		} else {
235 			if (uvm_nowait_failed) {
236 				/*
237 				 * XXX realisticly, this is what our
238 				 * nowait callers probably care about
239 				 */
240 				constraint = dma_constraint;
241 				uvm_nowait_failed = 0;
242 			} else
243 				constraint = no_constraint;
244 		}
245 
246 		uvm_unlock_fpageq();
247 
248 		/*
249 		 * now lock page queues and recompute inactive count
250 		 */
251 		uvm_lock_pageq();
252 		if (npages != uvmexp.npages) {	/* check for new pages? */
253 			npages = uvmexp.npages;
254 			uvmpd_tune();
255 		}
256 
257 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
258 		if (uvmexp.inactarg <= uvmexp.freetarg) {
259 			uvmexp.inactarg = uvmexp.freetarg + 1;
260 		}
261 
262 		/* Reclaim pages from the buffer cache if possible. */
263 		size = 0;
264 		if (pma != NULL)
265 			size += pma->pm_size >> PAGE_SHIFT;
266 		if (uvmexp.free - BUFPAGES_DEFICIT < uvmexp.freetarg)
267 			size += uvmexp.freetarg - (uvmexp.free -
268 			    BUFPAGES_DEFICIT);
269 		if (size == 0)
270 			size = 16; /* XXX */
271 		uvm_unlock_pageq();
272 		(void) bufbackoff(&constraint, size * 2);
273 #if NDRM > 0
274 		drmbackoff(size * 2);
275 #endif
276 		uvm_lock_pageq();
277 
278 		/*
279 		 * scan if needed
280 		 */
281 		if (pma != NULL ||
282 		    ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg) ||
283 		    ((uvmexp.inactive + BUFPAGES_INACT) < uvmexp.inactarg)) {
284 			uvmpd_scan();
285 		}
286 
287 		/*
288 		 * if there's any free memory to be had,
289 		 * wake up any waiters.
290 		 */
291 		uvm_lock_fpageq();
292 		if (uvmexp.free > uvmexp.reserve_kernel ||
293 		    uvmexp.paging == 0) {
294 			wakeup(&uvmexp.free);
295 		}
296 
297 		if (pma != NULL) {
298 			/*
299 			 * XXX If UVM_PMA_FREED isn't set, no pages
300 			 * were freed.  Should we set UVM_PMA_FAIL in
301 			 * that case?
302 			 */
303 			pma->pm_flags &= ~UVM_PMA_BUSY;
304 			if (pma->pm_flags & UVM_PMA_FREED) {
305 				pma->pm_flags &= ~UVM_PMA_LINKED;
306 				TAILQ_REMOVE(&uvm.pmr_control.allocs, pma,
307 				    pmq);
308 				wakeup(pma);
309 			}
310 		}
311 		uvm_unlock_fpageq();
312 
313 		/*
314 		 * scan done.  unlock page queues (the only lock we are holding)
315 		 */
316 		uvm_unlock_pageq();
317 
318 		sched_pause(yield);
319 	}
320 	/*NOTREACHED*/
321 }
322 
323 
324 /*
325  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
326  */
327 void
328 uvm_aiodone_daemon(void *arg)
329 {
330 	int s, free;
331 	struct buf *bp, *nbp;
332 
333 	uvm.aiodoned_proc = curproc;
334 
335 	for (;;) {
336 		/*
337 		 * Check for done aio structures. If we've got structures to
338 		 * process, do so. Otherwise sleep while avoiding races.
339 		 */
340 		mtx_enter(&uvm.aiodoned_lock);
341 		while ((bp = TAILQ_FIRST(&uvm.aio_done)) == NULL)
342 			msleep_nsec(&uvm.aiodoned, &uvm.aiodoned_lock,
343 			    PVM, "aiodoned", INFSLP);
344 		/* Take the list for ourselves. */
345 		TAILQ_INIT(&uvm.aio_done);
346 		mtx_leave(&uvm.aiodoned_lock);
347 
348 		/* process each i/o that's done. */
349 		free = uvmexp.free;
350 		while (bp != NULL) {
351 			if (bp->b_flags & B_PDAEMON) {
352 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
353 			}
354 			nbp = TAILQ_NEXT(bp, b_freelist);
355 			s = splbio();	/* b_iodone must by called at splbio */
356 			(*bp->b_iodone)(bp);
357 			splx(s);
358 			bp = nbp;
359 
360 			sched_pause(yield);
361 		}
362 		uvm_lock_fpageq();
363 		wakeup(free <= uvmexp.reserve_kernel ? &uvm.pagedaemon :
364 		    &uvmexp.free);
365 		uvm_unlock_fpageq();
366 	}
367 }
368 
369 
370 
371 /*
372  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
373  *
374  * => called with page queues locked
375  * => we work on meeting our free target by converting inactive pages
376  *    into free pages.
377  * => we handle the building of swap-backed clusters
378  * => we return TRUE if we are exiting because we met our target
379  */
380 
381 boolean_t
382 uvmpd_scan_inactive(struct pglist *pglst)
383 {
384 	boolean_t retval = FALSE;	/* assume we haven't hit target */
385 	int free, result;
386 	struct vm_page *p, *nextpg;
387 	struct uvm_object *uobj;
388 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
389 	int npages;
390 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
391 	int swnpages, swcpages;				/* XXX: see below */
392 	int swslot;
393 	struct vm_anon *anon;
394 	boolean_t swap_backed;
395 	vaddr_t start;
396 	int dirtyreacts;
397 
398 	/*
399 	 * swslot is non-zero if we are building a swap cluster.  we want
400 	 * to stay in the loop while we have a page to scan or we have
401 	 * a swap-cluster to build.
402 	 */
403 	swslot = 0;
404 	swnpages = swcpages = 0;
405 	free = 0;
406 	dirtyreacts = 0;
407 
408 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
409 		/*
410 		 * note that p can be NULL iff we have traversed the whole
411 		 * list and need to do one final swap-backed clustered pageout.
412 		 */
413 		uobj = NULL;
414 		anon = NULL;
415 
416 		if (p) {
417 			/*
418 			 * update our copy of "free" and see if we've met
419 			 * our target
420 			 */
421 			free = uvmexp.free - BUFPAGES_DEFICIT;
422 
423 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
424 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
425 				retval = TRUE;
426 
427 				if (swslot == 0) {
428 					/* exit now if no swap-i/o pending */
429 					break;
430 				}
431 
432 				/* set p to null to signal final swap i/o */
433 				p = NULL;
434 			}
435 		}
436 
437 		if (p) {	/* if (we have a new page to consider) */
438 			/*
439 			 * we are below target and have a new page to consider.
440 			 */
441 			uvmexp.pdscans++;
442 			nextpg = TAILQ_NEXT(p, pageq);
443 
444 			if (p->pg_flags & PQ_ANON) {
445 				anon = p->uanon;
446 				KASSERT(anon != NULL);
447 				if (rw_enter(anon->an_lock,
448 				    RW_WRITE|RW_NOSLEEP)) {
449 					/* lock failed, skip this page */
450 					continue;
451 				}
452 				/*
453 				 * move referenced pages back to active queue
454 				 * and skip to next page.
455 				 */
456 				if (pmap_is_referenced(p)) {
457 					uvm_pageactivate(p);
458 					rw_exit(anon->an_lock);
459 					uvmexp.pdreact++;
460 					continue;
461 				}
462 				if (p->pg_flags & PG_BUSY) {
463 					rw_exit(anon->an_lock);
464 					uvmexp.pdbusy++;
465 					/* someone else owns page, skip it */
466 					continue;
467 				}
468 				uvmexp.pdanscan++;
469 			} else {
470 				uobj = p->uobject;
471 				KASSERT(uobj != NULL);
472 				if (rw_enter(uobj->vmobjlock,
473 				    RW_WRITE|RW_NOSLEEP)) {
474 					/* lock failed, skip this page */
475 					continue;
476 				}
477 				/*
478 				 * move referenced pages back to active queue
479 				 * and skip to next page.
480 				 */
481 				if (pmap_is_referenced(p)) {
482 					uvm_pageactivate(p);
483 					rw_exit(uobj->vmobjlock);
484 					uvmexp.pdreact++;
485 					continue;
486 				}
487 				if (p->pg_flags & PG_BUSY) {
488 					rw_exit(uobj->vmobjlock);
489 					uvmexp.pdbusy++;
490 					/* someone else owns page, skip it */
491 					continue;
492 				}
493 				uvmexp.pdobscan++;
494 			}
495 
496 			/*
497 			 * we now have the page queues locked.
498 			 * the page is not busy.   if the page is clean we
499 			 * can free it now and continue.
500 			 */
501 			if (p->pg_flags & PG_CLEAN) {
502 				if (p->pg_flags & PQ_SWAPBACKED) {
503 					/* this page now lives only in swap */
504 					atomic_inc_int(&uvmexp.swpgonly);
505 				}
506 
507 				/* zap all mappings with pmap_page_protect... */
508 				pmap_page_protect(p, PROT_NONE);
509 				uvm_pagefree(p);
510 				uvmexp.pdfreed++;
511 
512 				if (anon) {
513 
514 					/*
515 					 * an anonymous page can only be clean
516 					 * if it has backing store assigned.
517 					 */
518 
519 					KASSERT(anon->an_swslot != 0);
520 
521 					/* remove from object */
522 					anon->an_page = NULL;
523 					rw_exit(anon->an_lock);
524 				} else {
525 					rw_exit(uobj->vmobjlock);
526 				}
527 				continue;
528 			}
529 
530 			/*
531 			 * this page is dirty, skip it if we'll have met our
532 			 * free target when all the current pageouts complete.
533 			 */
534 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
535 				if (anon) {
536 					rw_exit(anon->an_lock);
537 				} else {
538 					rw_exit(uobj->vmobjlock);
539 				}
540 				continue;
541 			}
542 
543 			/*
544 			 * this page is dirty, but we can't page it out
545 			 * since all pages in swap are only in swap.
546 			 * reactivate it so that we eventually cycle
547 			 * all pages thru the inactive queue.
548 			 */
549 			if ((p->pg_flags & PQ_SWAPBACKED) && uvm_swapisfull()) {
550 				dirtyreacts++;
551 				uvm_pageactivate(p);
552 				if (anon) {
553 					rw_exit(anon->an_lock);
554 				} else {
555 					rw_exit(uobj->vmobjlock);
556 				}
557 				continue;
558 			}
559 
560 			/*
561 			 * if the page is swap-backed and dirty and swap space
562 			 * is full, free any swap allocated to the page
563 			 * so that other pages can be paged out.
564 			 */
565 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
566 			if ((p->pg_flags & PQ_SWAPBACKED) &&
567 			    uvmexp.swpginuse == uvmexp.swpages) {
568 
569 				if ((p->pg_flags & PQ_ANON) &&
570 				    p->uanon->an_swslot) {
571 					uvm_swap_free(p->uanon->an_swslot, 1);
572 					p->uanon->an_swslot = 0;
573 				}
574 				if (p->pg_flags & PQ_AOBJ) {
575 					uao_dropswap(p->uobject,
576 						     p->offset >> PAGE_SHIFT);
577 				}
578 			}
579 
580 			/*
581 			 * the page we are looking at is dirty.   we must
582 			 * clean it before it can be freed.  to do this we
583 			 * first mark the page busy so that no one else will
584 			 * touch the page.   we write protect all the mappings
585 			 * of the page so that no one touches it while it is
586 			 * in I/O.
587 			 */
588 
589 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
590 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
591 			UVM_PAGE_OWN(p, "scan_inactive");
592 			pmap_page_protect(p, PROT_READ);
593 			uvmexp.pgswapout++;
594 
595 			/*
596 			 * for swap-backed pages we need to (re)allocate
597 			 * swap space.
598 			 */
599 			if (swap_backed) {
600 				/* free old swap slot (if any) */
601 				if (anon) {
602 					if (anon->an_swslot) {
603 						uvm_swap_free(anon->an_swslot,
604 						    1);
605 						anon->an_swslot = 0;
606 					}
607 				} else {
608 					uao_dropswap(uobj,
609 						     p->offset >> PAGE_SHIFT);
610 				}
611 
612 				/* start new cluster (if necessary) */
613 				if (swslot == 0) {
614 					swnpages = MAXBSIZE >> PAGE_SHIFT;
615 					swslot = uvm_swap_alloc(&swnpages,
616 					    TRUE);
617 					if (swslot == 0) {
618 						/* no swap?  give up! */
619 						atomic_clearbits_int(
620 						    &p->pg_flags,
621 						    PG_BUSY);
622 						UVM_PAGE_OWN(p, NULL);
623 						if (anon)
624 							rw_exit(anon->an_lock);
625 						else
626 							rw_exit(
627 							    uobj->vmobjlock);
628 						continue;
629 					}
630 					swcpages = 0;	/* cluster is empty */
631 				}
632 
633 				/* add block to cluster */
634 				swpps[swcpages] = p;
635 				if (anon)
636 					anon->an_swslot = swslot + swcpages;
637 				else
638 					uao_set_swslot(uobj,
639 					    p->offset >> PAGE_SHIFT,
640 					    swslot + swcpages);
641 				swcpages++;
642 			}
643 		} else {
644 			/* if p == NULL we must be doing a last swap i/o */
645 			swap_backed = TRUE;
646 		}
647 
648 		/*
649 		 * now consider doing the pageout.
650 		 *
651 		 * for swap-backed pages, we do the pageout if we have either
652 		 * filled the cluster (in which case (swnpages == swcpages) or
653 		 * run out of pages (p == NULL).
654 		 *
655 		 * for object pages, we always do the pageout.
656 		 */
657 		if (swap_backed) {
658 			if (p) {	/* if we just added a page to cluster */
659 				if (anon)
660 					rw_exit(anon->an_lock);
661 				else
662 					rw_exit(uobj->vmobjlock);
663 
664 				/* cluster not full yet? */
665 				if (swcpages < swnpages)
666 					continue;
667 			}
668 
669 			/* starting I/O now... set up for it */
670 			npages = swcpages;
671 			ppsp = swpps;
672 			/* for swap-backed pages only */
673 			start = (vaddr_t) swslot;
674 
675 			/* if this is final pageout we could have a few
676 			 * extra swap blocks */
677 			if (swcpages < swnpages) {
678 				uvm_swap_free(swslot + swcpages,
679 				    (swnpages - swcpages));
680 			}
681 		} else {
682 			/* normal object pageout */
683 			ppsp = pps;
684 			npages = sizeof(pps) / sizeof(struct vm_page *);
685 			/* not looked at because PGO_ALLPAGES is set */
686 			start = 0;
687 		}
688 
689 		/*
690 		 * now do the pageout.
691 		 *
692 		 * for swap_backed pages we have already built the cluster.
693 		 * for !swap_backed pages, uvm_pager_put will call the object's
694 		 * "make put cluster" function to build a cluster on our behalf.
695 		 *
696 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
697 		 * it to free the cluster pages for us on a successful I/O (it
698 		 * always does this for un-successful I/O requests).  this
699 		 * allows us to do clustered pageout without having to deal
700 		 * with cluster pages at this level.
701 		 *
702 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
703 		 *  IN: locked: page queues
704 		 * OUT: locked:
705 		 *     !locked: pageqs
706 		 */
707 
708 		uvmexp.pdpageouts++;
709 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
710 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
711 
712 		/*
713 		 * if we did i/o to swap, zero swslot to indicate that we are
714 		 * no longer building a swap-backed cluster.
715 		 */
716 
717 		if (swap_backed)
718 			swslot = 0;		/* done with this cluster */
719 
720 		/*
721 		 * first, we check for VM_PAGER_PEND which means that the
722 		 * async I/O is in progress and the async I/O done routine
723 		 * will clean up after us.   in this case we move on to the
724 		 * next page.
725 		 *
726 		 * there is a very remote chance that the pending async i/o can
727 		 * finish _before_ we get here.   if that happens, our page "p"
728 		 * may no longer be on the inactive queue.   so we verify this
729 		 * when determining the next page (starting over at the head if
730 		 * we've lost our inactive page).
731 		 */
732 
733 		if (result == VM_PAGER_PEND) {
734 			uvmexp.paging += npages;
735 			uvm_lock_pageq();
736 			uvmexp.pdpending++;
737 			if (p) {
738 				if (p->pg_flags & PQ_INACTIVE)
739 					nextpg = TAILQ_NEXT(p, pageq);
740 				else
741 					nextpg = TAILQ_FIRST(pglst);
742 			} else {
743 				nextpg = NULL;
744 			}
745 			continue;
746 		}
747 
748 		/* clean up "p" if we have one */
749 		if (p) {
750 			/*
751 			 * the I/O request to "p" is done and uvm_pager_put
752 			 * has freed any cluster pages it may have allocated
753 			 * during I/O.  all that is left for us to do is
754 			 * clean up page "p" (which is still PG_BUSY).
755 			 *
756 			 * our result could be one of the following:
757 			 *   VM_PAGER_OK: successful pageout
758 			 *
759 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
760 			 *     to next page
761 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
762 			 *     "reactivate" page to get it out of the way (it
763 			 *     will eventually drift back into the inactive
764 			 *     queue for a retry).
765 			 *   VM_PAGER_UNLOCK: should never see this as it is
766 			 *     only valid for "get" operations
767 			 */
768 
769 			/* relock p's object: page queues not lock yet, so
770 			 * no need for "try" */
771 
772 			/* !swap_backed case: already locked... */
773 			if (swap_backed) {
774 				if (anon)
775 					rw_enter(anon->an_lock, RW_WRITE);
776 				else
777 					rw_enter(uobj->vmobjlock, RW_WRITE);
778 			}
779 
780 #ifdef DIAGNOSTIC
781 			if (result == VM_PAGER_UNLOCK)
782 				panic("pagedaemon: pageout returned "
783 				    "invalid 'unlock' code");
784 #endif
785 
786 			/* handle PG_WANTED now */
787 			if (p->pg_flags & PG_WANTED)
788 				wakeup(p);
789 
790 			atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED);
791 			UVM_PAGE_OWN(p, NULL);
792 
793 			/* released during I/O? Can only happen for anons */
794 			if (p->pg_flags & PG_RELEASED) {
795 				KASSERT(anon != NULL);
796 				/*
797 				 * remove page so we can get nextpg,
798 				 * also zero out anon so we don't use
799 				 * it after the free.
800 				 */
801 				anon->an_page = NULL;
802 				p->uanon = NULL;
803 
804 				rw_exit(anon->an_lock);
805 				uvm_anfree(anon);	/* kills anon */
806 				pmap_page_protect(p, PROT_NONE);
807 				anon = NULL;
808 				uvm_lock_pageq();
809 				nextpg = TAILQ_NEXT(p, pageq);
810 				/* free released page */
811 				uvm_pagefree(p);
812 			} else {	/* page was not released during I/O */
813 				uvm_lock_pageq();
814 				nextpg = TAILQ_NEXT(p, pageq);
815 				if (result != VM_PAGER_OK) {
816 					/* pageout was a failure... */
817 					if (result != VM_PAGER_AGAIN)
818 						uvm_pageactivate(p);
819 					pmap_clear_reference(p);
820 					/* XXXCDC: if (swap_backed) FREE p's
821 					 * swap block? */
822 				} else {
823 					/* pageout was a success... */
824 					pmap_clear_reference(p);
825 					pmap_clear_modify(p);
826 					atomic_setbits_int(&p->pg_flags,
827 					    PG_CLEAN);
828 				}
829 			}
830 
831 			/*
832 			 * drop object lock (if there is an object left).   do
833 			 * a safety check of nextpg to make sure it is on the
834 			 * inactive queue (it should be since PG_BUSY pages on
835 			 * the inactive queue can't be re-queued [note: not
836 			 * true for active queue]).
837 			 */
838 			if (anon)
839 				rw_exit(anon->an_lock);
840 			else if (uobj)
841 				rw_exit(uobj->vmobjlock);
842 
843 			if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
844 				nextpg = TAILQ_FIRST(pglst);	/* reload! */
845 			}
846 		} else {
847 			/*
848 			 * if p is null in this loop, make sure it stays null
849 			 * in the next loop.
850 			 */
851 			nextpg = NULL;
852 
853 			/*
854 			 * lock page queues here just so they're always locked
855 			 * at the end of the loop.
856 			 */
857 			uvm_lock_pageq();
858 		}
859 	}
860 	return (retval);
861 }
862 
863 /*
864  * uvmpd_scan: scan the page queues and attempt to meet our targets.
865  *
866  * => called with pageq's locked
867  */
868 
869 void
870 uvmpd_scan(void)
871 {
872 	int free, inactive_shortage, swap_shortage, pages_freed;
873 	struct vm_page *p, *nextpg;
874 	struct uvm_object *uobj;
875 	struct vm_anon *anon;
876 	struct rwlock *slock;
877 
878 	MUTEX_ASSERT_LOCKED(&uvm.pageqlock);
879 
880 	uvmexp.pdrevs++;		/* counter */
881 	uobj = NULL;
882 
883 	/*
884 	 * get current "free" page count
885 	 */
886 	free = uvmexp.free - BUFPAGES_DEFICIT;
887 
888 #ifndef __SWAP_BROKEN
889 	/*
890 	 * swap out some processes if we are below our free target.
891 	 * we need to unlock the page queues for this.
892 	 */
893 	if (free < uvmexp.freetarg) {
894 		uvmexp.pdswout++;
895 		uvm_unlock_pageq();
896 		uvm_swapout_threads();
897 		uvm_lock_pageq();
898 	}
899 #endif
900 
901 	/*
902 	 * now we want to work on meeting our targets.   first we work on our
903 	 * free target by converting inactive pages into free pages.  then
904 	 * we work on meeting our inactive target by converting active pages
905 	 * to inactive ones.
906 	 */
907 
908 	/*
909 	 * alternate starting queue between swap and object based on the
910 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
911 	 */
912 	pages_freed = uvmexp.pdfreed;
913 	(void) uvmpd_scan_inactive(&uvm.page_inactive);
914 	pages_freed = uvmexp.pdfreed - pages_freed;
915 
916 	/*
917 	 * we have done the scan to get free pages.   now we work on meeting
918 	 * our inactive target.
919 	 */
920 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive - BUFPAGES_INACT;
921 
922 	/*
923 	 * detect if we're not going to be able to page anything out
924 	 * until we free some swap resources from active pages.
925 	 */
926 	free = uvmexp.free - BUFPAGES_DEFICIT;
927 	swap_shortage = 0;
928 	if (free < uvmexp.freetarg &&
929 	    uvmexp.swpginuse == uvmexp.swpages &&
930 	    !uvm_swapisfull() &&
931 	    pages_freed == 0) {
932 		swap_shortage = uvmexp.freetarg - free;
933 	}
934 
935 	for (p = TAILQ_FIRST(&uvm.page_active);
936 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
937 	     p = nextpg) {
938 		nextpg = TAILQ_NEXT(p, pageq);
939 		if (p->pg_flags & PG_BUSY) {
940 			continue;
941 		}
942 
943 		/*
944 		 * lock the page's owner.
945 		 */
946 		if (p->uobject != NULL) {
947 			uobj = p->uobject;
948 			slock = uobj->vmobjlock;
949 			if (rw_enter(slock, RW_WRITE|RW_NOSLEEP)) {
950 				continue;
951 			}
952 		} else {
953 			anon = p->uanon;
954 			KASSERT(p->uanon != NULL);
955 			slock = anon->an_lock;
956 			if (rw_enter(slock, RW_WRITE|RW_NOSLEEP)) {
957 				continue;
958 			}
959 		}
960 
961 		/*
962 		 * skip this page if it's busy.
963 		 */
964 		if ((p->pg_flags & PG_BUSY) != 0) {
965 			rw_exit(slock);
966 			continue;
967 		}
968 
969 		/*
970 		 * if there's a shortage of swap, free any swap allocated
971 		 * to this page so that other pages can be paged out.
972 		 */
973 		if (swap_shortage > 0) {
974 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
975 				uvm_swap_free(p->uanon->an_swslot, 1);
976 				p->uanon->an_swslot = 0;
977 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
978 				swap_shortage--;
979 			}
980 			if (p->pg_flags & PQ_AOBJ) {
981 				int slot = uao_set_swslot(p->uobject,
982 					p->offset >> PAGE_SHIFT, 0);
983 				if (slot) {
984 					uvm_swap_free(slot, 1);
985 					atomic_clearbits_int(&p->pg_flags,
986 					    PG_CLEAN);
987 					swap_shortage--;
988 				}
989 			}
990 		}
991 
992 		/*
993 		 * deactivate this page if there's a shortage of
994 		 * inactive pages.
995 		 */
996 		if (inactive_shortage > 0) {
997 			pmap_page_protect(p, PROT_NONE);
998 			/* no need to check wire_count as pg is "active" */
999 			uvm_pagedeactivate(p);
1000 			uvmexp.pddeact++;
1001 			inactive_shortage--;
1002 		}
1003 
1004 		/*
1005 		 * we're done with this page.
1006 		 */
1007 		rw_exit(slock);
1008 	}
1009 }
1010 
1011 #ifdef HIBERNATE
1012 
1013 /*
1014  * uvmpd_drop: drop clean pages from list
1015  */
1016 void
1017 uvmpd_drop(struct pglist *pglst)
1018 {
1019 	struct vm_page *p, *nextpg;
1020 
1021 	for (p = TAILQ_FIRST(pglst); p != NULL; p = nextpg) {
1022 		nextpg = TAILQ_NEXT(p, pageq);
1023 
1024 		if (p->pg_flags & PQ_ANON || p->uobject == NULL)
1025 			continue;
1026 
1027 		if (p->pg_flags & PG_BUSY)
1028 			continue;
1029 
1030 		if (p->pg_flags & PG_CLEAN) {
1031 			struct uvm_object * uobj = p->uobject;
1032 
1033 			rw_enter(uobj->vmobjlock, RW_WRITE);
1034 			uvm_lock_pageq();
1035 			/*
1036 			 * we now have the page queues locked.
1037 			 * the page is not busy.   if the page is clean we
1038 			 * can free it now and continue.
1039 			 */
1040 			if (p->pg_flags & PG_CLEAN) {
1041 				if (p->pg_flags & PQ_SWAPBACKED) {
1042 					/* this page now lives only in swap */
1043 					atomic_inc_int(&uvmexp.swpgonly);
1044 				}
1045 
1046 				/* zap all mappings with pmap_page_protect... */
1047 				pmap_page_protect(p, PROT_NONE);
1048 				uvm_pagefree(p);
1049 			}
1050 			uvm_unlock_pageq();
1051 			rw_exit(uobj->vmobjlock);
1052 		}
1053 	}
1054 }
1055 
1056 void
1057 uvmpd_hibernate(void)
1058 {
1059 	uvmpd_drop(&uvm.page_inactive);
1060 	uvmpd_drop(&uvm.page_active);
1061 }
1062 
1063 #endif
1064