xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.36 2001/06/27 18:52:10 thorpej Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993, The Regents of the University of California.
6  *
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. All advertising materials mentioning features or use of this software
21  *    must display the following acknowledgement:
22  *	This product includes software developed by Charles D. Cranor,
23  *      Washington University, the University of California, Berkeley and
24  *      its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)vm_pageout.c        8.5 (Berkeley) 2/14/94
42  * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp
43  *
44  *
45  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
46  * All rights reserved.
47  *
48  * Permission to use, copy, modify and distribute this software and
49  * its documentation is hereby granted, provided that both the copyright
50  * notice and this permission notice appear in all copies of the
51  * software, derivative works or modified versions, and any portions
52  * thereof, and that both notices appear in supporting documentation.
53  *
54  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
55  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
56  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57  *
58  * Carnegie Mellon requests users of this software to return to
59  *
60  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
61  *  School of Computer Science
62  *  Carnegie Mellon University
63  *  Pittsburgh PA 15213-3890
64  *
65  * any improvements or extensions that they make and grant Carnegie the
66  * rights to redistribute these changes.
67  */
68 
69 #include "opt_uvmhist.h"
70 
71 /*
72  * uvm_pdaemon.c: the page daemon
73  */
74 
75 #include <sys/param.h>
76 #include <sys/proc.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/pool.h>
80 #include <sys/buf.h>
81 #include <sys/vnode.h>
82 
83 #include <uvm/uvm.h>
84 
85 /*
86  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedeamon will reactivate
87  * in a pass thru the inactive list when swap is full.  the value should be
88  * "small"... if it's too large we'll cycle the active pages thru the inactive
89  * queue too quickly to for them to be referenced and avoid being freed.
90  */
91 
92 #define UVMPD_NUMDIRTYREACTS 16
93 
94 
95 /*
96  * local prototypes
97  */
98 
99 static void		uvmpd_scan __P((void));
100 static boolean_t	uvmpd_scan_inactive __P((struct pglist *));
101 static void		uvmpd_tune __P((void));
102 
103 /*
104  * uvm_wait: wait (sleep) for the page daemon to free some pages
105  *
106  * => should be called with all locks released
107  * => should _not_ be called by the page daemon (to avoid deadlock)
108  */
109 
110 void
111 uvm_wait(wmsg)
112 	const char *wmsg;
113 {
114 	int timo = 0;
115 	int s = splbio();
116 
117 	/*
118 	 * check for page daemon going to sleep (waiting for itself)
119 	 */
120 
121 	if (curproc == uvm.pagedaemon_proc) {
122 		/*
123 		 * now we have a problem: the pagedaemon wants to go to
124 		 * sleep until it frees more memory.   but how can it
125 		 * free more memory if it is asleep?  that is a deadlock.
126 		 * we have two options:
127 		 *  [1] panic now
128 		 *  [2] put a timeout on the sleep, thus causing the
129 		 *      pagedaemon to only pause (rather than sleep forever)
130 		 *
131 		 * note that option [2] will only help us if we get lucky
132 		 * and some other process on the system breaks the deadlock
133 		 * by exiting or freeing memory (thus allowing the pagedaemon
134 		 * to continue).  for now we panic if DEBUG is defined,
135 		 * otherwise we hope for the best with option [2] (better
136 		 * yet, this should never happen in the first place!).
137 		 */
138 
139 		printf("pagedaemon: deadlock detected!\n");
140 		timo = hz >> 3;		/* set timeout */
141 #if defined(DEBUG)
142 		/* DEBUG: panic so we can debug it */
143 		panic("pagedaemon deadlock");
144 #endif
145 	}
146 
147 	simple_lock(&uvm.pagedaemon_lock);
148 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
149 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, wmsg,
150 	    timo);
151 
152 	splx(s);
153 }
154 
155 
156 /*
157  * uvmpd_tune: tune paging parameters
158  *
159  * => called when ever memory is added (or removed?) to the system
160  * => caller must call with page queues locked
161  */
162 
163 static void
164 uvmpd_tune()
165 {
166 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
167 
168 	uvmexp.freemin = uvmexp.npages / 20;
169 
170 	/* between 16k and 256k */
171 	/* XXX:  what are these values good for? */
172 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
173 	uvmexp.freemin = min(uvmexp.freemin, (256*1024) >> PAGE_SHIFT);
174 
175 	/* Make sure there's always a user page free. */
176 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
177 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
178 
179 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
180 	if (uvmexp.freetarg <= uvmexp.freemin)
181 		uvmexp.freetarg = uvmexp.freemin + 1;
182 
183 	/* uvmexp.inactarg: computed in main daemon loop */
184 
185 	uvmexp.wiredmax = uvmexp.npages / 3;
186 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
187 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
188 }
189 
190 /*
191  * uvm_pageout: the main loop for the pagedaemon
192  */
193 
194 void
195 uvm_pageout(void *arg)
196 {
197 	int npages = 0;
198 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
199 
200 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
201 
202 	/*
203 	 * ensure correct priority and set paging parameters...
204 	 */
205 
206 	uvm.pagedaemon_proc = curproc;
207 	(void) spl0();
208 	uvm_lock_pageq();
209 	npages = uvmexp.npages;
210 	uvmpd_tune();
211 	uvm_unlock_pageq();
212 
213 	/*
214 	 * main loop
215 	 */
216 
217 	for (;;) {
218 		simple_lock(&uvm.pagedaemon_lock);
219 
220 		UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
221 		UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
222 		    &uvm.pagedaemon_lock, FALSE, "pgdaemon", 0);
223 		uvmexp.pdwoke++;
224 		UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
225 
226 		/* drain pool resources */
227 		pool_drain(0);
228 
229 		/*
230 		 * now lock page queues and recompute inactive count
231 		 */
232 
233 		uvm_lock_pageq();
234 		if (npages != uvmexp.npages) {	/* check for new pages? */
235 			npages = uvmexp.npages;
236 			uvmpd_tune();
237 		}
238 
239 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
240 		if (uvmexp.inactarg <= uvmexp.freetarg) {
241 			uvmexp.inactarg = uvmexp.freetarg + 1;
242 		}
243 
244 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d, inact/itarg=%d/%d",
245 		    uvmexp.free, uvmexp.freetarg, uvmexp.inactive,
246 		    uvmexp.inactarg);
247 
248 		/*
249 		 * scan if needed
250 		 */
251 
252 		if (uvmexp.free + uvmexp.paging < uvmexp.freetarg ||
253 		    uvmexp.inactive < uvmexp.inactarg) {
254 			uvmpd_scan();
255 		}
256 
257 		/*
258 		 * if there's any free memory to be had,
259 		 * wake up any waiters.
260 		 */
261 
262 		if (uvmexp.free > uvmexp.reserve_kernel ||
263 		    uvmexp.paging == 0) {
264 			wakeup(&uvmexp.free);
265 		}
266 
267 		/*
268 		 * scan done.  unlock page queues (the only lock we are holding)
269 		 */
270 
271 		uvm_unlock_pageq();
272 	}
273 	/*NOTREACHED*/
274 }
275 
276 
277 /*
278  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
279  */
280 
281 void
282 uvm_aiodone_daemon(void *arg)
283 {
284 	int s, free;
285 	struct buf *bp, *nbp;
286 	UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist);
287 
288 	for (;;) {
289 
290 		/*
291 		 * carefully attempt to go to sleep (without losing "wakeups"!).
292 		 * we need splbio because we want to make sure the aio_done list
293 		 * is totally empty before we go to sleep.
294 		 */
295 
296 		s = splbio();
297 		simple_lock(&uvm.aiodoned_lock);
298 		if (TAILQ_FIRST(&uvm.aio_done) == NULL) {
299 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
300 			UVM_UNLOCK_AND_WAIT(&uvm.aiodoned,
301 			    &uvm.aiodoned_lock, FALSE, "aiodoned", 0);
302 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
303 
304 			/* relock aiodoned_lock, still at splbio */
305 			simple_lock(&uvm.aiodoned_lock);
306 		}
307 
308 		/*
309 		 * check for done aio structures
310 		 */
311 
312 		bp = TAILQ_FIRST(&uvm.aio_done);
313 		if (bp) {
314 			TAILQ_INIT(&uvm.aio_done);
315 		}
316 
317 		simple_unlock(&uvm.aiodoned_lock);
318 		splx(s);
319 
320 		/*
321 		 * process each i/o that's done.
322 		 */
323 
324 		free = uvmexp.free;
325 		while (bp != NULL) {
326 			if (bp->b_flags & B_PDAEMON) {
327 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
328 			}
329 			nbp = TAILQ_NEXT(bp, b_freelist);
330 			(*bp->b_iodone)(bp);
331 			bp = nbp;
332 		}
333 		if (free <= uvmexp.reserve_kernel) {
334 			s = uvm_lock_fpageq();
335 			wakeup(&uvm.pagedaemon);
336 			uvm_unlock_fpageq(s);
337 		} else {
338 			simple_lock(&uvm.pagedaemon_lock);
339 			wakeup(&uvmexp.free);
340 			simple_unlock(&uvm.pagedaemon_lock);
341 		}
342 	}
343 }
344 
345 
346 
347 /*
348  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
349  *
350  * => called with page queues locked
351  * => we work on meeting our free target by converting inactive pages
352  *    into free pages.
353  * => we handle the building of swap-backed clusters
354  * => we return TRUE if we are exiting because we met our target
355  */
356 
357 static boolean_t
358 uvmpd_scan_inactive(pglst)
359 	struct pglist *pglst;
360 {
361 	boolean_t retval = FALSE;	/* assume we haven't hit target */
362 	int s, free, result;
363 	struct vm_page *p, *nextpg;
364 	struct uvm_object *uobj;
365 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
366 	int npages;
367 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
368 	int swnpages, swcpages;				/* XXX: see below */
369 	int swslot;
370 	struct vm_anon *anon;
371 	boolean_t swap_backed;
372 	vaddr_t start;
373 	int dirtyreacts, t;
374 	UVMHIST_FUNC("uvmpd_scan_inactive"); UVMHIST_CALLED(pdhist);
375 
376 	/*
377 	 * swslot is non-zero if we are building a swap cluster.  we want
378 	 * to stay in the loop while we have a page to scan or we have
379 	 * a swap-cluster to build.
380 	 */
381 
382 	swslot = 0;
383 	swnpages = swcpages = 0;
384 	free = 0;
385 	dirtyreacts = 0;
386 
387 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
388 
389 		/*
390 		 * note that p can be NULL iff we have traversed the whole
391 		 * list and need to do one final swap-backed clustered pageout.
392 		 */
393 
394 		uobj = NULL;
395 		anon = NULL;
396 
397 		if (p) {
398 
399 			/*
400 			 * update our copy of "free" and see if we've met
401 			 * our target
402 			 */
403 
404 			s = uvm_lock_fpageq();
405 			free = uvmexp.free;
406 			uvm_unlock_fpageq(s);
407 
408 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
409 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
410 				UVMHIST_LOG(pdhist,"  met free target: "
411 					    "exit loop", 0, 0, 0, 0);
412 				retval = TRUE;
413 
414 				if (swslot == 0) {
415 					/* exit now if no swap-i/o pending */
416 					break;
417 				}
418 
419 				/* set p to null to signal final swap i/o */
420 				p = NULL;
421 			}
422 		}
423 
424 		if (p) {	/* if (we have a new page to consider) */
425 
426 			/*
427 			 * we are below target and have a new page to consider.
428 			 */
429 			uvmexp.pdscans++;
430 			nextpg = TAILQ_NEXT(p, pageq);
431 
432 			/*
433 			 * move referenced pages back to active queue and
434 			 * skip to next page.
435 			 */
436 
437 			if (pmap_is_referenced(p)) {
438 				uvm_pageactivate(p);
439 				uvmexp.pdreact++;
440 				continue;
441 			}
442 
443 			/*
444 			 * enforce the minimum thresholds on different
445 			 * types of memory usage.  if reusing the current
446 			 * page would reduce that type of usage below its
447 			 * minimum, reactivate the page instead and move
448 			 * on to the next page.
449 			 */
450 
451 			t = uvmexp.active + uvmexp.inactive + uvmexp.free;
452 			if (p->uanon &&
453 			    uvmexp.anonpages <= (t * uvmexp.anonmin) >> 8) {
454 				uvm_pageactivate(p);
455 				uvmexp.pdreanon++;
456 				continue;
457 			}
458 			if (p->uobject && UVM_OBJ_IS_VTEXT(p->uobject) &&
459 			    uvmexp.vtextpages <= (t * uvmexp.vtextmin) >> 8) {
460 				uvm_pageactivate(p);
461 				uvmexp.pdrevtext++;
462 				continue;
463 			}
464 			if (p->uobject && UVM_OBJ_IS_VNODE(p->uobject) &&
465 			    !UVM_OBJ_IS_VTEXT(p->uobject) &&
466 			    uvmexp.vnodepages <= (t * uvmexp.vnodemin) >> 8) {
467 				uvm_pageactivate(p);
468 				uvmexp.pdrevnode++;
469 				continue;
470 			}
471 
472 			/*
473 			 * first we attempt to lock the object that this page
474 			 * belongs to.  if our attempt fails we skip on to
475 			 * the next page (no harm done).  it is important to
476 			 * "try" locking the object as we are locking in the
477 			 * wrong order (pageq -> object) and we don't want to
478 			 * deadlock.
479 			 *
480 			 * the only time we expect to see an ownerless page
481 			 * (i.e. a page with no uobject and !PQ_ANON) is if an
482 			 * anon has loaned a page from a uvm_object and the
483 			 * uvm_object has dropped the ownership.  in that
484 			 * case, the anon can "take over" the loaned page
485 			 * and make it its own.
486 			 */
487 
488 			/* is page part of an anon or ownerless ? */
489 			if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
490 				anon = p->uanon;
491 				KASSERT(anon != NULL);
492 				if (!simple_lock_try(&anon->an_lock)) {
493 					/* lock failed, skip this page */
494 					continue;
495 				}
496 
497 				/*
498 				 * if the page is ownerless, claim it in the
499 				 * name of "anon"!
500 				 */
501 
502 				if ((p->pqflags & PQ_ANON) == 0) {
503 					KASSERT(p->loan_count > 0);
504 					p->loan_count--;
505 					p->pqflags |= PQ_ANON;
506 					/* anon now owns it */
507 				}
508 				if (p->flags & PG_BUSY) {
509 					simple_unlock(&anon->an_lock);
510 					uvmexp.pdbusy++;
511 					/* someone else owns page, skip it */
512 					continue;
513 				}
514 				uvmexp.pdanscan++;
515 			} else {
516 				uobj = p->uobject;
517 				KASSERT(uobj != NULL);
518 				if (!simple_lock_try(&uobj->vmobjlock)) {
519 					/* lock failed, skip this page */
520 					continue;
521 				}
522 				if (p->flags & PG_BUSY) {
523 					simple_unlock(&uobj->vmobjlock);
524 					uvmexp.pdbusy++;
525 					/* someone else owns page, skip it */
526 					continue;
527 				}
528 				uvmexp.pdobscan++;
529 			}
530 
531 			/*
532 			 * we now have the object and the page queues locked.
533 			 * the page is not busy.  remove all the permissions
534 			 * from the page so we can sync the modified info
535 			 * without any race conditions.  if the page is clean
536 			 * we can free it now and continue.
537 			 */
538 
539 			pmap_page_protect(p, VM_PROT_NONE);
540 			if ((p->flags & PG_CLEAN) != 0 && pmap_is_modified(p)) {
541 				p->flags &= ~PG_CLEAN;
542 			}
543 
544 			if (p->flags & PG_CLEAN) {
545 				if (p->pqflags & PQ_SWAPBACKED) {
546 					/* this page now lives only in swap */
547 					simple_lock(&uvm.swap_data_lock);
548 					uvmexp.swpgonly++;
549 					simple_unlock(&uvm.swap_data_lock);
550 				}
551 
552 				uvm_pagefree(p);
553 				uvmexp.pdfreed++;
554 
555 				if (anon) {
556 
557 					/*
558 					 * an anonymous page can only be clean
559 					 * if it has backing store assigned.
560 					 */
561 
562 					KASSERT(anon->an_swslot != 0);
563 
564 					/* remove from object */
565 					anon->u.an_page = NULL;
566 					simple_unlock(&anon->an_lock);
567 				} else {
568 					/* pagefree has already removed the
569 					 * page from the object */
570 					simple_unlock(&uobj->vmobjlock);
571 				}
572 				continue;
573 			}
574 
575 			/*
576 			 * this page is dirty, skip it if we'll have met our
577 			 * free target when all the current pageouts complete.
578 			 */
579 
580 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
581 				if (anon) {
582 					simple_unlock(&anon->an_lock);
583 				} else {
584 					simple_unlock(&uobj->vmobjlock);
585 				}
586 				continue;
587 			}
588 
589 			/*
590 			 * this page is dirty, but we can't page it out
591 			 * since all pages in swap are only in swap.
592 			 * reactivate it so that we eventually cycle
593 			 * all pages thru the inactive queue.
594 			 */
595 
596 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
597 			if ((p->pqflags & PQ_SWAPBACKED) &&
598 			    uvmexp.swpgonly == uvmexp.swpages) {
599 				dirtyreacts++;
600 				uvm_pageactivate(p);
601 				if (anon) {
602 					simple_unlock(&anon->an_lock);
603 				} else {
604 					simple_unlock(&uobj->vmobjlock);
605 				}
606 				continue;
607 			}
608 
609 			/*
610 			 * if the page is swap-backed and dirty and swap space
611 			 * is full, free any swap allocated to the page
612 			 * so that other pages can be paged out.
613 			 */
614 
615 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
616 			if ((p->pqflags & PQ_SWAPBACKED) &&
617 			    uvmexp.swpginuse == uvmexp.swpages) {
618 
619 				if ((p->pqflags & PQ_ANON) &&
620 				    p->uanon->an_swslot) {
621 					uvm_swap_free(p->uanon->an_swslot, 1);
622 					p->uanon->an_swslot = 0;
623 				}
624 				if (p->pqflags & PQ_AOBJ) {
625 					uao_dropswap(p->uobject,
626 						     p->offset >> PAGE_SHIFT);
627 				}
628 			}
629 
630 			/*
631 			 * the page we are looking at is dirty.   we must
632 			 * clean it before it can be freed.  to do this we
633 			 * first mark the page busy so that no one else will
634 			 * touch the page.
635 			 */
636 
637 			swap_backed = ((p->pqflags & PQ_SWAPBACKED) != 0);
638 			p->flags |= PG_BUSY;		/* now we own it */
639 			UVM_PAGE_OWN(p, "scan_inactive");
640 			uvmexp.pgswapout++;
641 
642 			/*
643 			 * for swap-backed pages we need to (re)allocate
644 			 * swap space.
645 			 */
646 
647 			if (swap_backed) {
648 
649 				/*
650 				 * free old swap slot (if any)
651 				 */
652 
653 				if (anon) {
654 					if (anon->an_swslot) {
655 						uvm_swap_free(anon->an_swslot,
656 						    1);
657 						anon->an_swslot = 0;
658 					}
659 				} else {
660 					uao_dropswap(uobj,
661 						     p->offset >> PAGE_SHIFT);
662 				}
663 
664 				/*
665 				 * start new cluster (if necessary)
666 				 */
667 
668 				if (swslot == 0) {
669 					swnpages = MAXBSIZE >> PAGE_SHIFT;
670 					swslot = uvm_swap_alloc(&swnpages,
671 					    TRUE);
672 					if (swslot == 0) {
673 						/* no swap?  give up! */
674 						p->flags &= ~PG_BUSY;
675 						UVM_PAGE_OWN(p, NULL);
676 						if (anon)
677 							simple_unlock(
678 							    &anon->an_lock);
679 						else
680 							simple_unlock(
681 							    &uobj->vmobjlock);
682 						continue;
683 					}
684 					swcpages = 0;	/* cluster is empty */
685 				}
686 
687 				/*
688 				 * add block to cluster
689 				 */
690 
691 				if (anon) {
692 					anon->an_swslot = swslot + swcpages;
693 				} else {
694 					result = uao_set_swslot(uobj,
695 					    p->offset >> PAGE_SHIFT,
696 					    swslot + swcpages);
697 					if (result == -1) {
698 						p->flags &= ~PG_BUSY;
699 						UVM_PAGE_OWN(p, NULL);
700 						simple_unlock(&uobj->vmobjlock);
701 						continue;
702 					}
703 				}
704 				swpps[swcpages] = p;
705 				swcpages++;
706 			}
707 		} else {
708 
709 			/* if p == NULL we must be doing a last swap i/o */
710 			swap_backed = TRUE;
711 		}
712 
713 		/*
714 		 * now consider doing the pageout.
715 		 *
716 		 * for swap-backed pages, we do the pageout if we have either
717 		 * filled the cluster (in which case (swnpages == swcpages) or
718 		 * run out of pages (p == NULL).
719 		 *
720 		 * for object pages, we always do the pageout.
721 		 */
722 
723 		if (swap_backed) {
724 			if (p) {	/* if we just added a page to cluster */
725 				if (anon)
726 					simple_unlock(&anon->an_lock);
727 				else
728 					simple_unlock(&uobj->vmobjlock);
729 
730 				/* cluster not full yet? */
731 				if (swcpages < swnpages)
732 					continue;
733 			}
734 
735 			/* starting I/O now... set up for it */
736 			npages = swcpages;
737 			ppsp = swpps;
738 			/* for swap-backed pages only */
739 			start = (vaddr_t) swslot;
740 
741 			/* if this is final pageout we could have a few
742 			 * extra swap blocks */
743 			if (swcpages < swnpages) {
744 				uvm_swap_free(swslot + swcpages,
745 				    (swnpages - swcpages));
746 			}
747 		} else {
748 			/* normal object pageout */
749 			ppsp = pps;
750 			npages = sizeof(pps) / sizeof(struct vm_page *);
751 			/* not looked at because PGO_ALLPAGES is set */
752 			start = 0;
753 		}
754 
755 		/*
756 		 * now do the pageout.
757 		 *
758 		 * for swap_backed pages we have already built the cluster.
759 		 * for !swap_backed pages, uvm_pager_put will call the object's
760 		 * "make put cluster" function to build a cluster on our behalf.
761 		 *
762 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
763 		 * it to free the cluster pages for us on a successful I/O (it
764 		 * always does this for un-successful I/O requests).  this
765 		 * allows us to do clustered pageout without having to deal
766 		 * with cluster pages at this level.
767 		 *
768 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
769 		 *  IN: locked: uobj (if !swap_backed), page queues
770 		 * OUT:!locked: pageqs, uobj
771 		 */
772 
773 		/* locked: uobj (if !swap_backed), page queues */
774 		uvmexp.pdpageouts++;
775 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
776 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
777 		/* unlocked: pageqs, uobj */
778 
779 		/*
780 		 * if we did i/o to swap, zero swslot to indicate that we are
781 		 * no longer building a swap-backed cluster.
782 		 */
783 
784 		if (swap_backed)
785 			swslot = 0;		/* done with this cluster */
786 
787 		/*
788 		 * if the pageout failed, reactivate the page and continue.
789 		 */
790 
791 		if (result == EIO && curproc == uvm.pagedaemon_proc) {
792 			uvm_lock_pageq();
793 			nextpg = TAILQ_NEXT(p, pageq);
794 			uvm_pageactivate(p);
795 			continue;
796 		}
797 
798 		/*
799 		 * the pageout is in progress.  bump counters and set up
800 		 * for the next loop.
801 		 */
802 
803 		uvm_lock_pageq();
804 		uvmexp.paging += npages;
805 		uvmexp.pdpending++;
806 		if (p) {
807 			if (p->pqflags & PQ_INACTIVE)
808 				nextpg = TAILQ_NEXT(p, pageq);
809 			else
810 				nextpg = TAILQ_FIRST(pglst);
811 		} else {
812 			nextpg = NULL;
813 		}
814 	}
815 	return (retval);
816 }
817 
818 /*
819  * uvmpd_scan: scan the page queues and attempt to meet our targets.
820  *
821  * => called with pageq's locked
822  */
823 
824 void
825 uvmpd_scan()
826 {
827 	int s, free, inactive_shortage, swap_shortage, pages_freed;
828 	struct vm_page *p, *nextpg;
829 	struct uvm_object *uobj;
830 	boolean_t got_it;
831 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
832 
833 	uvmexp.pdrevs++;		/* counter */
834 	uobj = NULL;
835 
836 	/*
837 	 * get current "free" page count
838 	 */
839 	s = uvm_lock_fpageq();
840 	free = uvmexp.free;
841 	uvm_unlock_fpageq(s);
842 
843 #ifndef __SWAP_BROKEN
844 	/*
845 	 * swap out some processes if we are below our free target.
846 	 * we need to unlock the page queues for this.
847 	 */
848 	if (free < uvmexp.freetarg) {
849 		uvmexp.pdswout++;
850 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout", free,
851 		    uvmexp.freetarg, 0, 0);
852 		uvm_unlock_pageq();
853 		uvm_swapout_threads();
854 		uvm_lock_pageq();
855 
856 	}
857 #endif
858 
859 	/*
860 	 * now we want to work on meeting our targets.   first we work on our
861 	 * free target by converting inactive pages into free pages.  then
862 	 * we work on meeting our inactive target by converting active pages
863 	 * to inactive ones.
864 	 */
865 
866 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
867 
868 	/*
869 	 * alternate starting queue between swap and object based on the
870 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
871 	 */
872 
873 	got_it = FALSE;
874 	pages_freed = uvmexp.pdfreed;
875 	(void) uvmpd_scan_inactive(&uvm.page_inactive);
876 	pages_freed = uvmexp.pdfreed - pages_freed;
877 
878 	/*
879 	 * we have done the scan to get free pages.   now we work on meeting
880 	 * our inactive target.
881 	 */
882 
883 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive;
884 
885 	/*
886 	 * detect if we're not going to be able to page anything out
887 	 * until we free some swap resources from active pages.
888 	 */
889 
890 	swap_shortage = 0;
891 	if (uvmexp.free < uvmexp.freetarg &&
892 	    uvmexp.swpginuse == uvmexp.swpages &&
893 	    uvmexp.swpgonly < uvmexp.swpages &&
894 	    pages_freed == 0) {
895 		swap_shortage = uvmexp.freetarg - uvmexp.free;
896 	}
897 
898 	UVMHIST_LOG(pdhist, "  loop 2: inactive_shortage=%d swap_shortage=%d",
899 		    inactive_shortage, swap_shortage,0,0);
900 	for (p = TAILQ_FIRST(&uvm.page_active);
901 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
902 	     p = nextpg) {
903 		nextpg = TAILQ_NEXT(p, pageq);
904 		if (p->flags & PG_BUSY)
905 			continue;	/* quick check before trying to lock */
906 
907 		/*
908 		 * lock the page's owner.
909 		 */
910 		/* is page anon owned or ownerless? */
911 		if ((p->pqflags & PQ_ANON) || p->uobject == NULL) {
912 			KASSERT(p->uanon != NULL);
913 			if (!simple_lock_try(&p->uanon->an_lock))
914 				continue;
915 
916 			/* take over the page? */
917 			if ((p->pqflags & PQ_ANON) == 0) {
918 				KASSERT(p->loan_count > 0);
919 				p->loan_count--;
920 				p->pqflags |= PQ_ANON;
921 			}
922 		} else {
923 			if (!simple_lock_try(&p->uobject->vmobjlock))
924 				continue;
925 		}
926 
927 		/*
928 		 * skip this page if it's busy.
929 		 */
930 
931 		if ((p->flags & PG_BUSY) != 0) {
932 			if (p->pqflags & PQ_ANON)
933 				simple_unlock(&p->uanon->an_lock);
934 			else
935 				simple_unlock(&p->uobject->vmobjlock);
936 			continue;
937 		}
938 
939 		/*
940 		 * if there's a shortage of swap, free any swap allocated
941 		 * to this page so that other pages can be paged out.
942 		 */
943 
944 		if (swap_shortage > 0) {
945 			if ((p->pqflags & PQ_ANON) && p->uanon->an_swslot) {
946 				uvm_swap_free(p->uanon->an_swslot, 1);
947 				p->uanon->an_swslot = 0;
948 				p->flags &= ~PG_CLEAN;
949 				swap_shortage--;
950 			}
951 			if (p->pqflags & PQ_AOBJ) {
952 				int slot = uao_set_swslot(p->uobject,
953 					p->offset >> PAGE_SHIFT, 0);
954 				if (slot) {
955 					uvm_swap_free(slot, 1);
956 					p->flags &= ~PG_CLEAN;
957 					swap_shortage--;
958 				}
959 			}
960 		}
961 
962 		/*
963 		 * If we're short on inactive pages, move this over
964 		 * to the inactive list.  The second hand will sweep
965 		 * it later, and if it has been referenced again, it
966 		 * will be moved back to active.
967 		 */
968 
969 		if (inactive_shortage > 0) {
970 			pmap_clear_reference(p);
971 			/* no need to check wire_count as pg is "active" */
972 			uvm_pagedeactivate(p);
973 			uvmexp.pddeact++;
974 			inactive_shortage--;
975 		}
976 		if (p->pqflags & PQ_ANON)
977 			simple_unlock(&p->uanon->an_lock);
978 		else
979 			simple_unlock(&p->uobject->vmobjlock);
980 	}
981 }
982