xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 25c4e8bd056e974b28f4a0ffd39d76c190a56013)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.101 2022/06/28 19:31:30 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(struct uvm_pmalloc *);
105 boolean_t	uvmpd_scan_inactive(struct uvm_pmalloc *, 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(pma);
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 uvm_pmalloc *pma, 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[SWCLUSTPAGES], **ppsp;
389 	int npages;
390 	struct vm_page *swpps[SWCLUSTPAGES]; 	/* 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 	p = NULL;
408 
409 	/* Start with the first page on the list that fit in pma's ranges */
410 	if (pma != NULL) {
411 		paddr_t paddr;
412 
413 		TAILQ_FOREACH(p, pglst, pageq) {
414 			paddr = atop(VM_PAGE_TO_PHYS(p));
415 			if (paddr >= pma->pm_constraint.ucr_low &&
416 			    paddr < pma->pm_constraint.ucr_high)
417 				break;
418 		}
419 
420 	}
421 
422 	if (p == NULL) {
423 		p = TAILQ_FIRST(pglst);
424 		pma = NULL;
425 	}
426 
427 	for (; p != NULL || swslot != 0; p = nextpg) {
428 		/*
429 		 * note that p can be NULL iff we have traversed the whole
430 		 * list and need to do one final swap-backed clustered pageout.
431 		 */
432 		uobj = NULL;
433 		anon = NULL;
434 
435 		if (p) {
436 			/*
437 			 * update our copy of "free" and see if we've met
438 			 * our target
439 			 */
440 			free = uvmexp.free - BUFPAGES_DEFICIT;
441 			if (((pma == NULL || (pma->pm_flags & UVM_PMA_FREED)) &&
442 			    (free + uvmexp.paging >= uvmexp.freetarg << 2)) ||
443 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
444 				retval = TRUE;
445 
446 				if (swslot == 0) {
447 					/* exit now if no swap-i/o pending */
448 					break;
449 				}
450 
451 				/* set p to null to signal final swap i/o */
452 				p = NULL;
453 			}
454 		}
455 
456 		if (p) {	/* if (we have a new page to consider) */
457 			/*
458 			 * we are below target and have a new page to consider.
459 			 */
460 			uvmexp.pdscans++;
461 			nextpg = TAILQ_NEXT(p, pageq);
462 
463 			if (p->pg_flags & PQ_ANON) {
464 				anon = p->uanon;
465 				KASSERT(anon != NULL);
466 				if (rw_enter(anon->an_lock,
467 				    RW_WRITE|RW_NOSLEEP)) {
468 					/* lock failed, skip this page */
469 					continue;
470 				}
471 				/*
472 				 * move referenced pages back to active queue
473 				 * and skip to next page.
474 				 */
475 				if (pmap_is_referenced(p)) {
476 					uvm_pageactivate(p);
477 					rw_exit(anon->an_lock);
478 					uvmexp.pdreact++;
479 					continue;
480 				}
481 				if (p->pg_flags & PG_BUSY) {
482 					rw_exit(anon->an_lock);
483 					uvmexp.pdbusy++;
484 					/* someone else owns page, skip it */
485 					continue;
486 				}
487 				uvmexp.pdanscan++;
488 			} else {
489 				uobj = p->uobject;
490 				KASSERT(uobj != NULL);
491 				if (rw_enter(uobj->vmobjlock,
492 				    RW_WRITE|RW_NOSLEEP)) {
493 					/* lock failed, skip this page */
494 					continue;
495 				}
496 				/*
497 				 * move referenced pages back to active queue
498 				 * and skip to next page.
499 				 */
500 				if (pmap_is_referenced(p)) {
501 					uvm_pageactivate(p);
502 					rw_exit(uobj->vmobjlock);
503 					uvmexp.pdreact++;
504 					continue;
505 				}
506 				if (p->pg_flags & PG_BUSY) {
507 					rw_exit(uobj->vmobjlock);
508 					uvmexp.pdbusy++;
509 					/* someone else owns page, skip it */
510 					continue;
511 				}
512 				uvmexp.pdobscan++;
513 			}
514 
515 			/*
516 			 * we now have the page queues locked.
517 			 * the page is not busy.   if the page is clean we
518 			 * can free it now and continue.
519 			 */
520 			if (p->pg_flags & PG_CLEAN) {
521 				if (p->pg_flags & PQ_SWAPBACKED) {
522 					/* this page now lives only in swap */
523 					atomic_inc_int(&uvmexp.swpgonly);
524 				}
525 
526 				/* zap all mappings with pmap_page_protect... */
527 				pmap_page_protect(p, PROT_NONE);
528 				uvm_pagefree(p);
529 				uvmexp.pdfreed++;
530 
531 				if (anon) {
532 
533 					/*
534 					 * an anonymous page can only be clean
535 					 * if it has backing store assigned.
536 					 */
537 
538 					KASSERT(anon->an_swslot != 0);
539 
540 					/* remove from object */
541 					anon->an_page = NULL;
542 					rw_exit(anon->an_lock);
543 				} else {
544 					rw_exit(uobj->vmobjlock);
545 				}
546 				continue;
547 			}
548 
549 			/*
550 			 * this page is dirty, skip it if we'll have met our
551 			 * free target when all the current pageouts complete.
552 			 */
553 			if ((pma == NULL || (pma->pm_flags & UVM_PMA_FREED)) &&
554 			    (free + uvmexp.paging > uvmexp.freetarg << 2)) {
555 				if (anon) {
556 					rw_exit(anon->an_lock);
557 				} else {
558 					rw_exit(uobj->vmobjlock);
559 				}
560 				continue;
561 			}
562 
563 			/*
564 			 * this page is dirty, but we can't page it out
565 			 * since all pages in swap are only in swap.
566 			 * reactivate it so that we eventually cycle
567 			 * all pages thru the inactive queue.
568 			 */
569 			if ((p->pg_flags & PQ_SWAPBACKED) && uvm_swapisfull()) {
570 				dirtyreacts++;
571 				uvm_pageactivate(p);
572 				if (anon) {
573 					rw_exit(anon->an_lock);
574 				} else {
575 					rw_exit(uobj->vmobjlock);
576 				}
577 				continue;
578 			}
579 
580 			/*
581 			 * if the page is swap-backed and dirty and swap space
582 			 * is full, free any swap allocated to the page
583 			 * so that other pages can be paged out.
584 			 */
585 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
586 			if ((p->pg_flags & PQ_SWAPBACKED) &&
587 			    uvmexp.swpginuse == uvmexp.swpages) {
588 
589 				if ((p->pg_flags & PQ_ANON) &&
590 				    p->uanon->an_swslot) {
591 					uvm_swap_free(p->uanon->an_swslot, 1);
592 					p->uanon->an_swslot = 0;
593 				}
594 				if (p->pg_flags & PQ_AOBJ) {
595 					uao_dropswap(p->uobject,
596 						     p->offset >> PAGE_SHIFT);
597 				}
598 			}
599 
600 			/*
601 			 * the page we are looking at is dirty.   we must
602 			 * clean it before it can be freed.  to do this we
603 			 * first mark the page busy so that no one else will
604 			 * touch the page.   we write protect all the mappings
605 			 * of the page so that no one touches it while it is
606 			 * in I/O.
607 			 */
608 
609 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
610 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
611 			UVM_PAGE_OWN(p, "scan_inactive");
612 			pmap_page_protect(p, PROT_READ);
613 			uvmexp.pgswapout++;
614 
615 			/*
616 			 * for swap-backed pages we need to (re)allocate
617 			 * swap space.
618 			 */
619 			if (swap_backed) {
620 				/* free old swap slot (if any) */
621 				if (anon) {
622 					if (anon->an_swslot) {
623 						uvm_swap_free(anon->an_swslot,
624 						    1);
625 						anon->an_swslot = 0;
626 					}
627 				} else {
628 					uao_dropswap(uobj,
629 						     p->offset >> PAGE_SHIFT);
630 				}
631 
632 				/* start new cluster (if necessary) */
633 				if (swslot == 0) {
634 					swnpages = SWCLUSTPAGES;
635 					swslot = uvm_swap_alloc(&swnpages,
636 					    TRUE);
637 					if (swslot == 0) {
638 						/* no swap?  give up! */
639 						atomic_clearbits_int(
640 						    &p->pg_flags,
641 						    PG_BUSY);
642 						UVM_PAGE_OWN(p, NULL);
643 						if (anon)
644 							rw_exit(anon->an_lock);
645 						else
646 							rw_exit(
647 							    uobj->vmobjlock);
648 						continue;
649 					}
650 					swcpages = 0;	/* cluster is empty */
651 				}
652 
653 				/* add block to cluster */
654 				swpps[swcpages] = p;
655 				if (anon)
656 					anon->an_swslot = swslot + swcpages;
657 				else
658 					uao_set_swslot(uobj,
659 					    p->offset >> PAGE_SHIFT,
660 					    swslot + swcpages);
661 				swcpages++;
662 			}
663 		} else {
664 			/* if p == NULL we must be doing a last swap i/o */
665 			swap_backed = TRUE;
666 		}
667 
668 		/*
669 		 * now consider doing the pageout.
670 		 *
671 		 * for swap-backed pages, we do the pageout if we have either
672 		 * filled the cluster (in which case (swnpages == swcpages) or
673 		 * run out of pages (p == NULL).
674 		 *
675 		 * for object pages, we always do the pageout.
676 		 */
677 		if (swap_backed) {
678 			if (p) {	/* if we just added a page to cluster */
679 				if (anon)
680 					rw_exit(anon->an_lock);
681 				else
682 					rw_exit(uobj->vmobjlock);
683 
684 				/* cluster not full yet? */
685 				if (swcpages < swnpages)
686 					continue;
687 			}
688 
689 			/* starting I/O now... set up for it */
690 			npages = swcpages;
691 			ppsp = swpps;
692 			/* for swap-backed pages only */
693 			start = (vaddr_t) swslot;
694 
695 			/* if this is final pageout we could have a few
696 			 * extra swap blocks */
697 			if (swcpages < swnpages) {
698 				uvm_swap_free(swslot + swcpages,
699 				    (swnpages - swcpages));
700 			}
701 		} else {
702 			/* normal object pageout */
703 			ppsp = pps;
704 			npages = sizeof(pps) / sizeof(struct vm_page *);
705 			/* not looked at because PGO_ALLPAGES is set */
706 			start = 0;
707 		}
708 
709 		/*
710 		 * now do the pageout.
711 		 *
712 		 * for swap_backed pages we have already built the cluster.
713 		 * for !swap_backed pages, uvm_pager_put will call the object's
714 		 * "make put cluster" function to build a cluster on our behalf.
715 		 *
716 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
717 		 * it to free the cluster pages for us on a successful I/O (it
718 		 * always does this for un-successful I/O requests).  this
719 		 * allows us to do clustered pageout without having to deal
720 		 * with cluster pages at this level.
721 		 *
722 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
723 		 *  IN: locked: page queues
724 		 * OUT: locked:
725 		 *     !locked: pageqs
726 		 */
727 
728 		uvmexp.pdpageouts++;
729 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
730 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
731 
732 		/*
733 		 * if we did i/o to swap, zero swslot to indicate that we are
734 		 * no longer building a swap-backed cluster.
735 		 */
736 
737 		if (swap_backed)
738 			swslot = 0;		/* done with this cluster */
739 
740 		/*
741 		 * first, we check for VM_PAGER_PEND which means that the
742 		 * async I/O is in progress and the async I/O done routine
743 		 * will clean up after us.   in this case we move on to the
744 		 * next page.
745 		 *
746 		 * there is a very remote chance that the pending async i/o can
747 		 * finish _before_ we get here.   if that happens, our page "p"
748 		 * may no longer be on the inactive queue.   so we verify this
749 		 * when determining the next page (starting over at the head if
750 		 * we've lost our inactive page).
751 		 */
752 
753 		if (result == VM_PAGER_PEND) {
754 			uvmexp.paging += npages;
755 			uvm_lock_pageq();
756 			uvmexp.pdpending++;
757 			if (p) {
758 				if (p->pg_flags & PQ_INACTIVE)
759 					nextpg = TAILQ_NEXT(p, pageq);
760 				else
761 					nextpg = TAILQ_FIRST(pglst);
762 			} else {
763 				nextpg = NULL;
764 			}
765 			continue;
766 		}
767 
768 		/* clean up "p" if we have one */
769 		if (p) {
770 			/*
771 			 * the I/O request to "p" is done and uvm_pager_put
772 			 * has freed any cluster pages it may have allocated
773 			 * during I/O.  all that is left for us to do is
774 			 * clean up page "p" (which is still PG_BUSY).
775 			 *
776 			 * our result could be one of the following:
777 			 *   VM_PAGER_OK: successful pageout
778 			 *
779 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
780 			 *     to next page
781 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
782 			 *     "reactivate" page to get it out of the way (it
783 			 *     will eventually drift back into the inactive
784 			 *     queue for a retry).
785 			 *   VM_PAGER_UNLOCK: should never see this as it is
786 			 *     only valid for "get" operations
787 			 */
788 
789 			/* relock p's object: page queues not lock yet, so
790 			 * no need for "try" */
791 
792 			/* !swap_backed case: already locked... */
793 			if (swap_backed) {
794 				if (anon)
795 					rw_enter(anon->an_lock, RW_WRITE);
796 				else
797 					rw_enter(uobj->vmobjlock, RW_WRITE);
798 			}
799 
800 #ifdef DIAGNOSTIC
801 			if (result == VM_PAGER_UNLOCK)
802 				panic("pagedaemon: pageout returned "
803 				    "invalid 'unlock' code");
804 #endif
805 
806 			/* handle PG_WANTED now */
807 			if (p->pg_flags & PG_WANTED)
808 				wakeup(p);
809 
810 			atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED);
811 			UVM_PAGE_OWN(p, NULL);
812 
813 			/* released during I/O? Can only happen for anons */
814 			if (p->pg_flags & PG_RELEASED) {
815 				KASSERT(anon != NULL);
816 				/*
817 				 * remove page so we can get nextpg,
818 				 * also zero out anon so we don't use
819 				 * it after the free.
820 				 */
821 				anon->an_page = NULL;
822 				p->uanon = NULL;
823 
824 				rw_exit(anon->an_lock);
825 				uvm_anfree(anon);	/* kills anon */
826 				pmap_page_protect(p, PROT_NONE);
827 				anon = NULL;
828 				uvm_lock_pageq();
829 				nextpg = TAILQ_NEXT(p, pageq);
830 				/* free released page */
831 				uvm_pagefree(p);
832 			} else {	/* page was not released during I/O */
833 				uvm_lock_pageq();
834 				nextpg = TAILQ_NEXT(p, pageq);
835 				if (result != VM_PAGER_OK) {
836 					/* pageout was a failure... */
837 					if (result != VM_PAGER_AGAIN)
838 						uvm_pageactivate(p);
839 					pmap_clear_reference(p);
840 					/* XXXCDC: if (swap_backed) FREE p's
841 					 * swap block? */
842 				} else {
843 					/* pageout was a success... */
844 					pmap_clear_reference(p);
845 					pmap_clear_modify(p);
846 					atomic_setbits_int(&p->pg_flags,
847 					    PG_CLEAN);
848 				}
849 			}
850 
851 			/*
852 			 * drop object lock (if there is an object left).   do
853 			 * a safety check of nextpg to make sure it is on the
854 			 * inactive queue (it should be since PG_BUSY pages on
855 			 * the inactive queue can't be re-queued [note: not
856 			 * true for active queue]).
857 			 */
858 			if (anon)
859 				rw_exit(anon->an_lock);
860 			else if (uobj)
861 				rw_exit(uobj->vmobjlock);
862 
863 			if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
864 				nextpg = TAILQ_FIRST(pglst);	/* reload! */
865 			}
866 		} else {
867 			/*
868 			 * if p is null in this loop, make sure it stays null
869 			 * in the next loop.
870 			 */
871 			nextpg = NULL;
872 
873 			/*
874 			 * lock page queues here just so they're always locked
875 			 * at the end of the loop.
876 			 */
877 			uvm_lock_pageq();
878 		}
879 	}
880 	return (retval);
881 }
882 
883 /*
884  * uvmpd_scan: scan the page queues and attempt to meet our targets.
885  *
886  * => called with pageq's locked
887  */
888 
889 void
890 uvmpd_scan(struct uvm_pmalloc *pma)
891 {
892 	int free, inactive_shortage, swap_shortage, pages_freed;
893 	struct vm_page *p, *nextpg;
894 	struct uvm_object *uobj;
895 	struct vm_anon *anon;
896 	struct rwlock *slock;
897 
898 	MUTEX_ASSERT_LOCKED(&uvm.pageqlock);
899 
900 	uvmexp.pdrevs++;		/* counter */
901 	uobj = NULL;
902 
903 	/*
904 	 * get current "free" page count
905 	 */
906 	free = uvmexp.free - BUFPAGES_DEFICIT;
907 
908 #ifndef __SWAP_BROKEN
909 	/*
910 	 * swap out some processes if we are below our free target.
911 	 * we need to unlock the page queues for this.
912 	 */
913 	if (free < uvmexp.freetarg) {
914 		uvmexp.pdswout++;
915 		uvm_unlock_pageq();
916 		uvm_swapout_threads();
917 		uvm_lock_pageq();
918 	}
919 #endif
920 
921 	/*
922 	 * now we want to work on meeting our targets.   first we work on our
923 	 * free target by converting inactive pages into free pages.  then
924 	 * we work on meeting our inactive target by converting active pages
925 	 * to inactive ones.
926 	 */
927 
928 	/*
929 	 * alternate starting queue between swap and object based on the
930 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
931 	 */
932 	pages_freed = uvmexp.pdfreed;
933 	(void) uvmpd_scan_inactive(pma, &uvm.page_inactive);
934 	pages_freed = uvmexp.pdfreed - pages_freed;
935 
936 	/*
937 	 * we have done the scan to get free pages.   now we work on meeting
938 	 * our inactive target.
939 	 */
940 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive - BUFPAGES_INACT;
941 
942 	/*
943 	 * detect if we're not going to be able to page anything out
944 	 * until we free some swap resources from active pages.
945 	 */
946 	free = uvmexp.free - BUFPAGES_DEFICIT;
947 	swap_shortage = 0;
948 	if (free < uvmexp.freetarg &&
949 	    uvmexp.swpginuse == uvmexp.swpages &&
950 	    !uvm_swapisfull() &&
951 	    pages_freed == 0) {
952 		swap_shortage = uvmexp.freetarg - free;
953 	}
954 
955 	for (p = TAILQ_FIRST(&uvm.page_active);
956 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
957 	     p = nextpg) {
958 		nextpg = TAILQ_NEXT(p, pageq);
959 		if (p->pg_flags & PG_BUSY) {
960 			continue;
961 		}
962 
963 		/*
964 		 * lock the page's owner.
965 		 */
966 		if (p->uobject != NULL) {
967 			uobj = p->uobject;
968 			slock = uobj->vmobjlock;
969 			if (rw_enter(slock, RW_WRITE|RW_NOSLEEP)) {
970 				continue;
971 			}
972 		} else {
973 			anon = p->uanon;
974 			KASSERT(p->uanon != NULL);
975 			slock = anon->an_lock;
976 			if (rw_enter(slock, RW_WRITE|RW_NOSLEEP)) {
977 				continue;
978 			}
979 		}
980 
981 		/*
982 		 * skip this page if it's busy.
983 		 */
984 		if ((p->pg_flags & PG_BUSY) != 0) {
985 			rw_exit(slock);
986 			continue;
987 		}
988 
989 		/*
990 		 * if there's a shortage of swap, free any swap allocated
991 		 * to this page so that other pages can be paged out.
992 		 */
993 		if (swap_shortage > 0) {
994 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
995 				uvm_swap_free(p->uanon->an_swslot, 1);
996 				p->uanon->an_swslot = 0;
997 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
998 				swap_shortage--;
999 			}
1000 			if (p->pg_flags & PQ_AOBJ) {
1001 				int slot = uao_set_swslot(p->uobject,
1002 					p->offset >> PAGE_SHIFT, 0);
1003 				if (slot) {
1004 					uvm_swap_free(slot, 1);
1005 					atomic_clearbits_int(&p->pg_flags,
1006 					    PG_CLEAN);
1007 					swap_shortage--;
1008 				}
1009 			}
1010 		}
1011 
1012 		/*
1013 		 * deactivate this page if there's a shortage of
1014 		 * inactive pages.
1015 		 */
1016 		if (inactive_shortage > 0) {
1017 			pmap_page_protect(p, PROT_NONE);
1018 			/* no need to check wire_count as pg is "active" */
1019 			uvm_pagedeactivate(p);
1020 			uvmexp.pddeact++;
1021 			inactive_shortage--;
1022 		}
1023 
1024 		/*
1025 		 * we're done with this page.
1026 		 */
1027 		rw_exit(slock);
1028 	}
1029 }
1030 
1031 #ifdef HIBERNATE
1032 
1033 /*
1034  * uvmpd_drop: drop clean pages from list
1035  */
1036 void
1037 uvmpd_drop(struct pglist *pglst)
1038 {
1039 	struct vm_page *p, *nextpg;
1040 
1041 	for (p = TAILQ_FIRST(pglst); p != NULL; p = nextpg) {
1042 		nextpg = TAILQ_NEXT(p, pageq);
1043 
1044 		if (p->pg_flags & PQ_ANON || p->uobject == NULL)
1045 			continue;
1046 
1047 		if (p->pg_flags & PG_BUSY)
1048 			continue;
1049 
1050 		if (p->pg_flags & PG_CLEAN) {
1051 			struct uvm_object * uobj = p->uobject;
1052 
1053 			rw_enter(uobj->vmobjlock, RW_WRITE);
1054 			uvm_lock_pageq();
1055 			/*
1056 			 * we now have the page queues locked.
1057 			 * the page is not busy.   if the page is clean we
1058 			 * can free it now and continue.
1059 			 */
1060 			if (p->pg_flags & PG_CLEAN) {
1061 				if (p->pg_flags & PQ_SWAPBACKED) {
1062 					/* this page now lives only in swap */
1063 					atomic_inc_int(&uvmexp.swpgonly);
1064 				}
1065 
1066 				/* zap all mappings with pmap_page_protect... */
1067 				pmap_page_protect(p, PROT_NONE);
1068 				uvm_pagefree(p);
1069 			}
1070 			uvm_unlock_pageq();
1071 			rw_exit(uobj->vmobjlock);
1072 		}
1073 	}
1074 }
1075 
1076 void
1077 uvmpd_hibernate(void)
1078 {
1079 	uvmpd_drop(&uvm.page_inactive);
1080 	uvmpd_drop(&uvm.page_active);
1081 }
1082 
1083 #endif
1084