xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 0b7734b3d77bb9b21afec6f4621cae6c805dbd45)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.77 2015/10/08 15:58:38 kettenis 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/buf.h>
74 #include <sys/mount.h>
75 #include <sys/atomic.h>
76 
77 #ifdef HIBERNATE
78 #include <sys/hibernate.h>
79 #endif
80 
81 #include <uvm/uvm.h>
82 
83 /*
84  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
85  * in a pass thru the inactive list when swap is full.  the value should be
86  * "small"... if it's too large we'll cycle the active pages thru the inactive
87  * queue too quickly to for them to be referenced and avoid being freed.
88  */
89 
90 #define UVMPD_NUMDIRTYREACTS 16
91 
92 
93 /*
94  * local prototypes
95  */
96 
97 void		uvmpd_scan(void);
98 boolean_t	uvmpd_scan_inactive(struct pglist *);
99 void		uvmpd_tune(void);
100 void		uvmpd_drop(struct pglist *);
101 
102 /*
103  * uvm_wait: wait (sleep) for the page daemon to free some pages
104  *
105  * => should be called with all locks released
106  * => should _not_ be called by the page daemon (to avoid deadlock)
107  */
108 
109 void
110 uvm_wait(const char *wmsg)
111 {
112 	int	timo = 0;
113 
114 	/* check for page daemon going to sleep (waiting for itself) */
115 	if (curproc == uvm.pagedaemon_proc) {
116 		printf("uvm_wait emergency bufbackoff\n");
117 		if (bufbackoff(NULL, 4) == 0)
118 			return;
119 		/*
120 		 * now we have a problem: the pagedaemon wants to go to
121 		 * sleep until it frees more memory.   but how can it
122 		 * free more memory if it is asleep?  that is a deadlock.
123 		 * we have two options:
124 		 *  [1] panic now
125 		 *  [2] put a timeout on the sleep, thus causing the
126 		 *      pagedaemon to only pause (rather than sleep forever)
127 		 *
128 		 * note that option [2] will only help us if we get lucky
129 		 * and some other process on the system breaks the deadlock
130 		 * by exiting or freeing memory (thus allowing the pagedaemon
131 		 * to continue).  for now we panic if DEBUG is defined,
132 		 * otherwise we hope for the best with option [2] (better
133 		 * yet, this should never happen in the first place!).
134 		 */
135 
136 		printf("pagedaemon: deadlock detected!\n");
137 		timo = hz >> 3;		/* set timeout */
138 #if defined(DEBUG)
139 		/* DEBUG: panic so we can debug it */
140 		panic("pagedaemon deadlock");
141 #endif
142 	}
143 
144 	uvm_lock_fpageq();
145 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
146 	msleep(&uvmexp.free, &uvm.fpageqlock, PVM | PNORELOCK, wmsg, timo);
147 }
148 
149 /*
150  * uvmpd_tune: tune paging parameters
151  *
152  * => called whenever memory is added to (or removed from?) the system
153  * => caller must call with page queues locked
154  */
155 
156 void
157 uvmpd_tune(void)
158 {
159 
160 	uvmexp.freemin = uvmexp.npages / 30;
161 
162 	/* between 16k and 512k */
163 	/* XXX:  what are these values good for? */
164 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
165 #if 0
166 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
167 #endif
168 
169 	/* Make sure there's always a user page free. */
170 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
171 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
172 
173 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
174 	if (uvmexp.freetarg <= uvmexp.freemin)
175 		uvmexp.freetarg = uvmexp.freemin + 1;
176 
177 	/* uvmexp.inactarg: computed in main daemon loop */
178 
179 	uvmexp.wiredmax = uvmexp.npages / 3;
180 }
181 
182 /*
183  * uvm_pageout: the main loop for the pagedaemon
184  */
185 void
186 uvm_pageout(void *arg)
187 {
188 	struct uvm_constraint_range constraint;
189 	struct uvm_pmalloc *pma;
190 	int work_done;
191 	int npages = 0;
192 
193 	/* ensure correct priority and set paging parameters... */
194 	uvm.pagedaemon_proc = curproc;
195 	(void) spl0();
196 	uvm_lock_pageq();
197 	npages = uvmexp.npages;
198 	uvmpd_tune();
199 	uvm_unlock_pageq();
200 
201 	for (;;) {
202 		long size;
203 	  	work_done = 0; /* No work done this iteration. */
204 
205 		uvm_lock_fpageq();
206 
207 		if (TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
208 			msleep(&uvm.pagedaemon, &uvm.fpageqlock, PVM,
209 			    "pgdaemon", 0);
210 			uvmexp.pdwoke++;
211 		}
212 
213 		if ((pma = TAILQ_FIRST(&uvm.pmr_control.allocs)) != NULL) {
214 			pma->pm_flags |= UVM_PMA_BUSY;
215 			constraint = pma->pm_constraint;
216 		} else
217 			constraint = no_constraint;
218 
219 		uvm_unlock_fpageq();
220 
221 		/* now lock page queues and recompute inactive count */
222 		uvm_lock_pageq();
223 		if (npages != uvmexp.npages) {	/* check for new pages? */
224 			npages = uvmexp.npages;
225 			uvmpd_tune();
226 		}
227 
228 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
229 		if (uvmexp.inactarg <= uvmexp.freetarg) {
230 			uvmexp.inactarg = uvmexp.freetarg + 1;
231 		}
232 
233 		/* Reclaim pages from the buffer cache if possible. */
234 		size = 0;
235 		if (pma != NULL)
236 			size += pma->pm_size >> PAGE_SHIFT;
237 		if (uvmexp.free - BUFPAGES_DEFICIT < uvmexp.freetarg)
238 			size += uvmexp.freetarg - (uvmexp.free -
239 			    BUFPAGES_DEFICIT);
240 		uvm_unlock_pageq();
241 		(void) bufbackoff(&constraint, size * 2);
242 		uvm_lock_pageq();
243 
244 		/* Scan if needed to meet our targets. */
245 		if (pma != NULL ||
246 		    ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg) ||
247 		    ((uvmexp.inactive + BUFPAGES_INACT) < uvmexp.inactarg)) {
248 			uvmpd_scan();
249 			work_done = 1; /* XXX we hope... */
250 		}
251 
252 		/*
253 		 * if there's any free memory to be had,
254 		 * wake up any waiters.
255 		 */
256 		uvm_lock_fpageq();
257 		if (uvmexp.free > uvmexp.reserve_kernel ||
258 		    uvmexp.paging == 0) {
259 			wakeup(&uvmexp.free);
260 		}
261 
262 		if (pma != NULL) {
263 			pma->pm_flags &= ~UVM_PMA_BUSY;
264 			if (!work_done)
265 				pma->pm_flags |= UVM_PMA_FAIL;
266 			if (pma->pm_flags & (UVM_PMA_FAIL | UVM_PMA_FREED)) {
267 				pma->pm_flags &= ~UVM_PMA_LINKED;
268 				TAILQ_REMOVE(&uvm.pmr_control.allocs, pma,
269 				    pmq);
270 			}
271 			wakeup(pma);
272 		}
273 		uvm_unlock_fpageq();
274 
275 		/* scan done. unlock page queues (only lock we are holding) */
276 		uvm_unlock_pageq();
277 
278 		sched_pause();
279 	}
280 	/*NOTREACHED*/
281 }
282 
283 
284 /*
285  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
286  */
287 void
288 uvm_aiodone_daemon(void *arg)
289 {
290 	int s, free;
291 	struct buf *bp, *nbp;
292 
293 	uvm.aiodoned_proc = curproc;
294 
295 	for (;;) {
296 		/*
297 		 * Check for done aio structures. If we've got structures to
298 		 * process, do so. Otherwise sleep while avoiding races.
299 		 */
300 		mtx_enter(&uvm.aiodoned_lock);
301 		while ((bp = TAILQ_FIRST(&uvm.aio_done)) == NULL)
302 			msleep(&uvm.aiodoned, &uvm.aiodoned_lock,
303 			    PVM, "aiodoned", 0);
304 		/* Take the list for ourselves. */
305 		TAILQ_INIT(&uvm.aio_done);
306 		mtx_leave(&uvm.aiodoned_lock);
307 
308 		/* process each i/o that's done. */
309 		free = uvmexp.free;
310 		while (bp != NULL) {
311 			if (bp->b_flags & B_PDAEMON) {
312 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
313 			}
314 			nbp = TAILQ_NEXT(bp, b_freelist);
315 			s = splbio();	/* b_iodone must by called at splbio */
316 			(*bp->b_iodone)(bp);
317 			splx(s);
318 			bp = nbp;
319 
320 			sched_pause();
321 		}
322 		uvm_lock_fpageq();
323 		wakeup(free <= uvmexp.reserve_kernel ? &uvm.pagedaemon :
324 		    &uvmexp.free);
325 		uvm_unlock_fpageq();
326 	}
327 }
328 
329 
330 
331 /*
332  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
333  *
334  * => called with page queues locked
335  * => we work on meeting our free target by converting inactive pages
336  *    into free pages.
337  * => we handle the building of swap-backed clusters
338  * => we return TRUE if we are exiting because we met our target
339  */
340 
341 boolean_t
342 uvmpd_scan_inactive(struct pglist *pglst)
343 {
344 	boolean_t retval = FALSE;	/* assume we haven't hit target */
345 	int free, result;
346 	struct vm_page *p, *nextpg;
347 	struct uvm_object *uobj;
348 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
349 	int npages;
350 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
351 	int swnpages, swcpages;				/* XXX: see below */
352 	int swslot;
353 	struct vm_anon *anon;
354 	boolean_t swap_backed;
355 	vaddr_t start;
356 	int dirtyreacts;
357 
358 	/*
359 	 * note: we currently keep swap-backed pages on a separate inactive
360 	 * list from object-backed pages.   however, merging the two lists
361 	 * back together again hasn't been ruled out.   thus, we keep our
362 	 * swap cluster in "swpps" rather than in pps (allows us to mix
363 	 * clustering types in the event of a mixed inactive queue).
364 	 */
365 	/*
366 	 * swslot is non-zero if we are building a swap cluster.  we want
367 	 * to stay in the loop while we have a page to scan or we have
368 	 * a swap-cluster to build.
369 	 */
370 	swslot = 0;
371 	swnpages = swcpages = 0;
372 	free = 0;
373 	dirtyreacts = 0;
374 
375 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
376 		/*
377 		 * note that p can be NULL iff we have traversed the whole
378 		 * list and need to do one final swap-backed clustered pageout.
379 		 */
380 		uobj = NULL;
381 		anon = NULL;
382 
383 		if (p) {
384 			/*
385 			 * update our copy of "free" and see if we've met
386 			 * our target
387 			 */
388 			free = uvmexp.free - BUFPAGES_DEFICIT;
389 
390 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
391 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
392 				retval = TRUE;
393 
394 				if (swslot == 0) {
395 					/* exit now if no swap-i/o pending */
396 					break;
397 				}
398 
399 				/* set p to null to signal final swap i/o */
400 				p = NULL;
401 			}
402 		}
403 
404 		if (p) {	/* if (we have a new page to consider) */
405 			/*
406 			 * we are below target and have a new page to consider.
407 			 */
408 			uvmexp.pdscans++;
409 			nextpg = TAILQ_NEXT(p, pageq);
410 
411 			/*
412 			 * move referenced pages back to active queue and
413 			 * skip to next page (unlikely to happen since
414 			 * inactive pages shouldn't have any valid mappings
415 			 * and we cleared reference before deactivating).
416 			 */
417 
418 			if (pmap_is_referenced(p)) {
419 				uvm_pageactivate(p);
420 				uvmexp.pdreact++;
421 				continue;
422 			}
423 
424 			if (p->pg_flags & PQ_ANON) {
425 				anon = p->uanon;
426 				KASSERT(anon != NULL);
427 				if (p->pg_flags & PG_BUSY) {
428 					uvmexp.pdbusy++;
429 					/* someone else owns page, skip it */
430 					continue;
431 				}
432 				uvmexp.pdanscan++;
433 			} else {
434 				uobj = p->uobject;
435 				KASSERT(uobj != NULL);
436 				if (p->pg_flags & PG_BUSY) {
437 					uvmexp.pdbusy++;
438 					/* someone else owns page, skip it */
439 					continue;
440 				}
441 				uvmexp.pdobscan++;
442 			}
443 
444 			/*
445 			 * we now have the page queues locked.
446 			 * the page is not busy.   if the page is clean we
447 			 * can free it now and continue.
448 			 */
449 			if (p->pg_flags & PG_CLEAN) {
450 				if (p->pg_flags & PQ_SWAPBACKED) {
451 					/* this page now lives only in swap */
452 					uvmexp.swpgonly++;
453 				}
454 
455 				/* zap all mappings with pmap_page_protect... */
456 				pmap_page_protect(p, PROT_NONE);
457 				uvm_pagefree(p);
458 				uvmexp.pdfreed++;
459 
460 				if (anon) {
461 
462 					/*
463 					 * an anonymous page can only be clean
464 					 * if it has backing store assigned.
465 					 */
466 
467 					KASSERT(anon->an_swslot != 0);
468 
469 					/* remove from object */
470 					anon->an_page = NULL;
471 				}
472 				continue;
473 			}
474 
475 			/*
476 			 * this page is dirty, skip it if we'll have met our
477 			 * free target when all the current pageouts complete.
478 			 */
479 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
480 				continue;
481 			}
482 
483 			/*
484 			 * this page is dirty, but we can't page it out
485 			 * since all pages in swap are only in swap.
486 			 * reactivate it so that we eventually cycle
487 			 * all pages thru the inactive queue.
488 			 */
489 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
490 			if ((p->pg_flags & PQ_SWAPBACKED) &&
491 			    uvmexp.swpgonly == uvmexp.swpages) {
492 				dirtyreacts++;
493 				uvm_pageactivate(p);
494 				continue;
495 			}
496 
497 			/*
498 			 * if the page is swap-backed and dirty and swap space
499 			 * is full, free any swap allocated to the page
500 			 * so that other pages can be paged out.
501 			 */
502 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
503 			if ((p->pg_flags & PQ_SWAPBACKED) &&
504 			    uvmexp.swpginuse == uvmexp.swpages) {
505 
506 				if ((p->pg_flags & PQ_ANON) &&
507 				    p->uanon->an_swslot) {
508 					uvm_swap_free(p->uanon->an_swslot, 1);
509 					p->uanon->an_swslot = 0;
510 				}
511 				if (p->pg_flags & PQ_AOBJ) {
512 					uao_dropswap(p->uobject,
513 						     p->offset >> PAGE_SHIFT);
514 				}
515 			}
516 
517 			/*
518 			 * the page we are looking at is dirty.   we must
519 			 * clean it before it can be freed.  to do this we
520 			 * first mark the page busy so that no one else will
521 			 * touch the page.   we write protect all the mappings
522 			 * of the page so that no one touches it while it is
523 			 * in I/O.
524 			 */
525 
526 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
527 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
528 			UVM_PAGE_OWN(p, "scan_inactive");
529 			pmap_page_protect(p, PROT_READ);
530 			uvmexp.pgswapout++;
531 
532 			/*
533 			 * for swap-backed pages we need to (re)allocate
534 			 * swap space.
535 			 */
536 			if (swap_backed) {
537 				/* free old swap slot (if any) */
538 				if (anon) {
539 					if (anon->an_swslot) {
540 						uvm_swap_free(anon->an_swslot,
541 						    1);
542 						anon->an_swslot = 0;
543 					}
544 				} else {
545 					uao_dropswap(uobj,
546 						     p->offset >> PAGE_SHIFT);
547 				}
548 
549 				/* start new cluster (if necessary) */
550 				if (swslot == 0) {
551 					swnpages = MAXBSIZE >> PAGE_SHIFT;
552 					swslot = uvm_swap_alloc(&swnpages,
553 					    TRUE);
554 					if (swslot == 0) {
555 						/* no swap?  give up! */
556 						atomic_clearbits_int(
557 						    &p->pg_flags,
558 						    PG_BUSY);
559 						UVM_PAGE_OWN(p, NULL);
560 						continue;
561 					}
562 					swcpages = 0;	/* cluster is empty */
563 				}
564 
565 				/* add block to cluster */
566 				swpps[swcpages] = p;
567 				if (anon)
568 					anon->an_swslot = swslot + swcpages;
569 				else
570 					uao_set_swslot(uobj,
571 					    p->offset >> PAGE_SHIFT,
572 					    swslot + swcpages);
573 				swcpages++;
574 			}
575 		} else {
576 			/* if p == NULL we must be doing a last swap i/o */
577 			swap_backed = TRUE;
578 		}
579 
580 		/*
581 		 * now consider doing the pageout.
582 		 *
583 		 * for swap-backed pages, we do the pageout if we have either
584 		 * filled the cluster (in which case (swnpages == swcpages) or
585 		 * run out of pages (p == NULL).
586 		 *
587 		 * for object pages, we always do the pageout.
588 		 */
589 		if (swap_backed) {
590 			if (p) {	/* if we just added a page to cluster */
591 				/* cluster not full yet? */
592 				if (swcpages < swnpages)
593 					continue;
594 			}
595 
596 			/* starting I/O now... set up for it */
597 			npages = swcpages;
598 			ppsp = swpps;
599 			/* for swap-backed pages only */
600 			start = (vaddr_t) swslot;
601 
602 			/* if this is final pageout we could have a few
603 			 * extra swap blocks */
604 			if (swcpages < swnpages) {
605 				uvm_swap_free(swslot + swcpages,
606 				    (swnpages - swcpages));
607 			}
608 		} else {
609 			/* normal object pageout */
610 			ppsp = pps;
611 			npages = sizeof(pps) / sizeof(struct vm_page *);
612 			/* not looked at because PGO_ALLPAGES is set */
613 			start = 0;
614 		}
615 
616 		/*
617 		 * now do the pageout.
618 		 *
619 		 * for swap_backed pages we have already built the cluster.
620 		 * for !swap_backed pages, uvm_pager_put will call the object's
621 		 * "make put cluster" function to build a cluster on our behalf.
622 		 *
623 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
624 		 * it to free the cluster pages for us on a successful I/O (it
625 		 * always does this for un-successful I/O requests).  this
626 		 * allows us to do clustered pageout without having to deal
627 		 * with cluster pages at this level.
628 		 *
629 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
630 		 *  IN: locked: page queues
631 		 * OUT: locked:
632 		 *     !locked: pageqs
633 		 */
634 
635 		uvmexp.pdpageouts++;
636 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
637 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
638 
639 		/*
640 		 * if we did i/o to swap, zero swslot to indicate that we are
641 		 * no longer building a swap-backed cluster.
642 		 */
643 
644 		if (swap_backed)
645 			swslot = 0;		/* done with this cluster */
646 
647 		/*
648 		 * first, we check for VM_PAGER_PEND which means that the
649 		 * async I/O is in progress and the async I/O done routine
650 		 * will clean up after us.   in this case we move on to the
651 		 * next page.
652 		 *
653 		 * there is a very remote chance that the pending async i/o can
654 		 * finish _before_ we get here.   if that happens, our page "p"
655 		 * may no longer be on the inactive queue.   so we verify this
656 		 * when determining the next page (starting over at the head if
657 		 * we've lost our inactive page).
658 		 */
659 
660 		if (result == VM_PAGER_PEND) {
661 			uvmexp.paging += npages;
662 			uvm_lock_pageq();
663 			uvmexp.pdpending++;
664 			if (p) {
665 				if (p->pg_flags & PQ_INACTIVE)
666 					nextpg = TAILQ_NEXT(p, pageq);
667 				else
668 					nextpg = TAILQ_FIRST(pglst);
669 			} else {
670 				nextpg = NULL;
671 			}
672 			continue;
673 		}
674 
675 		/* clean up "p" if we have one */
676 		if (p) {
677 			/*
678 			 * the I/O request to "p" is done and uvm_pager_put
679 			 * has freed any cluster pages it may have allocated
680 			 * during I/O.  all that is left for us to do is
681 			 * clean up page "p" (which is still PG_BUSY).
682 			 *
683 			 * our result could be one of the following:
684 			 *   VM_PAGER_OK: successful pageout
685 			 *
686 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
687 			 *     to next page
688 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
689 			 *     "reactivate" page to get it out of the way (it
690 			 *     will eventually drift back into the inactive
691 			 *     queue for a retry).
692 			 *   VM_PAGER_UNLOCK: should never see this as it is
693 			 *     only valid for "get" operations
694 			 */
695 
696 			/* relock p's object: page queues not lock yet, so
697 			 * no need for "try" */
698 
699 #ifdef DIAGNOSTIC
700 			if (result == VM_PAGER_UNLOCK)
701 				panic("pagedaemon: pageout returned "
702 				    "invalid 'unlock' code");
703 #endif
704 
705 			/* handle PG_WANTED now */
706 			if (p->pg_flags & PG_WANTED)
707 				wakeup(p);
708 
709 			atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED);
710 			UVM_PAGE_OWN(p, NULL);
711 
712 			/* released during I/O? Can only happen for anons */
713 			if (p->pg_flags & PG_RELEASED) {
714 				KASSERT(anon != NULL);
715 				/*
716 				 * remove page so we can get nextpg,
717 				 * also zero out anon so we don't use
718 				 * it after the free.
719 				 */
720 				anon->an_page = NULL;
721 				p->uanon = NULL;
722 
723 				uvm_anfree(anon);	/* kills anon */
724 				pmap_page_protect(p, PROT_NONE);
725 				anon = NULL;
726 				uvm_lock_pageq();
727 				nextpg = TAILQ_NEXT(p, pageq);
728 				/* free released page */
729 				uvm_pagefree(p);
730 			} else {	/* page was not released during I/O */
731 				uvm_lock_pageq();
732 				nextpg = TAILQ_NEXT(p, pageq);
733 				if (result != VM_PAGER_OK) {
734 					/* pageout was a failure... */
735 					if (result != VM_PAGER_AGAIN)
736 						uvm_pageactivate(p);
737 					pmap_clear_reference(p);
738 					/* XXXCDC: if (swap_backed) FREE p's
739 					 * swap block? */
740 				} else {
741 					/* pageout was a success... */
742 					pmap_clear_reference(p);
743 					pmap_clear_modify(p);
744 					atomic_setbits_int(&p->pg_flags,
745 					    PG_CLEAN);
746 				}
747 			}
748 
749 			/*
750 			 * drop object lock (if there is an object left).   do
751 			 * a safety check of nextpg to make sure it is on the
752 			 * inactive queue (it should be since PG_BUSY pages on
753 			 * the inactive queue can't be re-queued [note: not
754 			 * true for active queue]).
755 			 */
756 
757 			if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
758 				nextpg = TAILQ_FIRST(pglst);	/* reload! */
759 			}
760 		} else {
761 			/*
762 			 * if p is null in this loop, make sure it stays null
763 			 * in the next loop.
764 			 */
765 			nextpg = NULL;
766 
767 			/*
768 			 * lock page queues here just so they're always locked
769 			 * at the end of the loop.
770 			 */
771 			uvm_lock_pageq();
772 		}
773 	}
774 	return (retval);
775 }
776 
777 /*
778  * uvmpd_scan: scan the page queues and attempt to meet our targets.
779  *
780  * => called with pageq's locked
781  */
782 
783 void
784 uvmpd_scan(void)
785 {
786 	int free, inactive_shortage, swap_shortage, pages_freed;
787 	struct vm_page *p, *nextpg;
788 	struct uvm_object *uobj;
789 	boolean_t got_it;
790 
791 	uvmexp.pdrevs++;		/* counter */
792 	uobj = NULL;
793 
794 	/*
795 	 * get current "free" page count
796 	 */
797 	free = uvmexp.free - BUFPAGES_DEFICIT;
798 
799 #ifndef __SWAP_BROKEN
800 	/*
801 	 * swap out some processes if we are below our free target.
802 	 * we need to unlock the page queues for this.
803 	 */
804 	if (free < uvmexp.freetarg) {
805 		uvmexp.pdswout++;
806 		uvm_unlock_pageq();
807 		uvm_swapout_threads();
808 		uvm_lock_pageq();
809 	}
810 #endif
811 
812 	/*
813 	 * now we want to work on meeting our targets.   first we work on our
814 	 * free target by converting inactive pages into free pages.  then
815 	 * we work on meeting our inactive target by converting active pages
816 	 * to inactive ones.
817 	 */
818 
819 	/*
820 	 * alternate starting queue between swap and object based on the
821 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
822 	 */
823 	got_it = FALSE;
824 	pages_freed = uvmexp.pdfreed;	/* XXX - int */
825 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
826 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
827 	if (!got_it)
828 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
829 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
830 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
831 	pages_freed = uvmexp.pdfreed - pages_freed;
832 
833 	/*
834 	 * we have done the scan to get free pages.   now we work on meeting
835 	 * our inactive target.
836 	 */
837 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive - BUFPAGES_INACT;
838 
839 	/*
840 	 * detect if we're not going to be able to page anything out
841 	 * until we free some swap resources from active pages.
842 	 */
843 	swap_shortage = 0;
844 	if (uvmexp.free < uvmexp.freetarg &&
845 	    uvmexp.swpginuse == uvmexp.swpages &&
846 	    uvmexp.swpgonly < uvmexp.swpages &&
847 	    pages_freed == 0) {
848 		swap_shortage = uvmexp.freetarg - uvmexp.free;
849 	}
850 
851 	for (p = TAILQ_FIRST(&uvm.page_active);
852 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
853 	     p = nextpg) {
854 		nextpg = TAILQ_NEXT(p, pageq);
855 
856 		/* skip this page if it's busy. */
857 		if (p->pg_flags & PG_BUSY)
858 			continue;
859 
860 		if (p->pg_flags & PQ_ANON)
861 			KASSERT(p->uanon != NULL);
862 		else
863 			KASSERT(p->uobject != NULL);
864 
865 		/*
866 		 * if there's a shortage of swap, free any swap allocated
867 		 * to this page so that other pages can be paged out.
868 		 */
869 		if (swap_shortage > 0) {
870 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
871 				uvm_swap_free(p->uanon->an_swslot, 1);
872 				p->uanon->an_swslot = 0;
873 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
874 				swap_shortage--;
875 			}
876 			if (p->pg_flags & PQ_AOBJ) {
877 				int slot = uao_set_swslot(p->uobject,
878 					p->offset >> PAGE_SHIFT, 0);
879 				if (slot) {
880 					uvm_swap_free(slot, 1);
881 					atomic_clearbits_int(&p->pg_flags,
882 					    PG_CLEAN);
883 					swap_shortage--;
884 				}
885 			}
886 		}
887 
888 		/*
889 		 * deactivate this page if there's a shortage of
890 		 * inactive pages.
891 		 */
892 		if (inactive_shortage > 0) {
893 			pmap_page_protect(p, PROT_NONE);
894 			/* no need to check wire_count as pg is "active" */
895 			uvm_pagedeactivate(p);
896 			uvmexp.pddeact++;
897 			inactive_shortage--;
898 		}
899 	}
900 }
901 
902 #ifdef HIBERNATE
903 
904 /*
905  * uvmpd_drop: drop clean pages from list
906  */
907 void
908 uvmpd_drop(struct pglist *pglst)
909 {
910 	struct vm_page *p, *nextpg;
911 
912 	for (p = TAILQ_FIRST(pglst); p != NULL; p = nextpg) {
913 		nextpg = TAILQ_NEXT(p, pageq);
914 
915 		if (p->pg_flags & PQ_ANON || p->uobject == NULL)
916 			continue;
917 
918 		if (p->pg_flags & PG_BUSY)
919 			continue;
920 
921 		if (p->pg_flags & PG_CLEAN) {
922 			/*
923 			 * we now have the page queues locked.
924 			 * the page is not busy.   if the page is clean we
925 			 * can free it now and continue.
926 			 */
927 			if (p->pg_flags & PG_CLEAN) {
928 				if (p->pg_flags & PQ_SWAPBACKED) {
929 					/* this page now lives only in swap */
930 					uvmexp.swpgonly++;
931 				}
932 
933 				/* zap all mappings with pmap_page_protect... */
934 				pmap_page_protect(p, PROT_NONE);
935 				uvm_pagefree(p);
936 			}
937 		}
938 	}
939 }
940 
941 void
942 uvmpd_hibernate(void)
943 {
944 	uvm_lock_pageq();
945 
946 	uvmpd_drop(&uvm.page_inactive_swp);
947 	uvmpd_drop(&uvm.page_inactive_obj);
948 	uvmpd_drop(&uvm.page_active);
949 
950 	uvm_unlock_pageq();
951 }
952 
953 #endif
954