xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision b7ae68fde0d8ef1c03714e8bbb1ee7c6118ea93b)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.77 2006/09/15 15:51:13 yamt 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 /*
70  * uvm_pdaemon.c: the page daemon
71  */
72 
73 #include <sys/cdefs.h>
74 __KERNEL_RCSID(0, "$NetBSD: uvm_pdaemon.c,v 1.77 2006/09/15 15:51:13 yamt Exp $");
75 
76 #include "opt_uvmhist.h"
77 #include "opt_readahead.h"
78 
79 #include <sys/param.h>
80 #include <sys/proc.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
83 #include <sys/pool.h>
84 #include <sys/buf.h>
85 #include <sys/vnode.h>
86 
87 #include <uvm/uvm.h>
88 #include <uvm/uvm_pdpolicy.h>
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 static void	uvmpd_scan(void);
105 static void	uvmpd_scan_queue(void);
106 static void	uvmpd_tune(void);
107 
108 /*
109  * XXX hack to avoid hangs when large processes fork.
110  */
111 int uvm_extrapages;
112 
113 /*
114  * uvm_wait: wait (sleep) for the page daemon to free some pages
115  *
116  * => should be called with all locks released
117  * => should _not_ be called by the page daemon (to avoid deadlock)
118  */
119 
120 void
121 uvm_wait(const char *wmsg)
122 {
123 	int timo = 0;
124 	int s = splbio();
125 
126 	/*
127 	 * check for page daemon going to sleep (waiting for itself)
128 	 */
129 
130 	if (curproc == uvm.pagedaemon_proc && uvmexp.paging == 0) {
131 		/*
132 		 * now we have a problem: the pagedaemon wants to go to
133 		 * sleep until it frees more memory.   but how can it
134 		 * free more memory if it is asleep?  that is a deadlock.
135 		 * we have two options:
136 		 *  [1] panic now
137 		 *  [2] put a timeout on the sleep, thus causing the
138 		 *      pagedaemon to only pause (rather than sleep forever)
139 		 *
140 		 * note that option [2] will only help us if we get lucky
141 		 * and some other process on the system breaks the deadlock
142 		 * by exiting or freeing memory (thus allowing the pagedaemon
143 		 * to continue).  for now we panic if DEBUG is defined,
144 		 * otherwise we hope for the best with option [2] (better
145 		 * yet, this should never happen in the first place!).
146 		 */
147 
148 		printf("pagedaemon: deadlock detected!\n");
149 		timo = hz >> 3;		/* set timeout */
150 #if defined(DEBUG)
151 		/* DEBUG: panic so we can debug it */
152 		panic("pagedaemon deadlock");
153 #endif
154 	}
155 
156 	simple_lock(&uvm.pagedaemon_lock);
157 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
158 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm.pagedaemon_lock, FALSE, wmsg,
159 	    timo);
160 
161 	splx(s);
162 }
163 
164 /*
165  * uvm_kick_pdaemon: perform checks to determine if we need to
166  * give the pagedaemon a nudge, and do so if necessary.
167  */
168 
169 void
170 uvm_kick_pdaemon(void)
171 {
172 
173 	if (uvmexp.free + uvmexp.paging < uvmexp.freemin ||
174 	    (uvmexp.free + uvmexp.paging < uvmexp.freetarg &&
175 	     uvmpdpol_needsscan_p())) {
176 		wakeup(&uvm.pagedaemon);
177 	}
178 }
179 
180 /*
181  * uvmpd_tune: tune paging parameters
182  *
183  * => called when ever memory is added (or removed?) to the system
184  * => caller must call with page queues locked
185  */
186 
187 static void
188 uvmpd_tune(void)
189 {
190 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
191 
192 	uvmexp.freemin = uvmexp.npages / 20;
193 
194 	/* between 16k and 256k */
195 	/* XXX:  what are these values good for? */
196 	uvmexp.freemin = MAX(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
197 	uvmexp.freemin = MIN(uvmexp.freemin, (256*1024) >> PAGE_SHIFT);
198 
199 	/* Make sure there's always a user page free. */
200 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
201 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
202 
203 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
204 	if (uvmexp.freetarg <= uvmexp.freemin)
205 		uvmexp.freetarg = uvmexp.freemin + 1;
206 
207 	uvmexp.freetarg += uvm_extrapages;
208 	uvm_extrapages = 0;
209 
210 	uvmexp.wiredmax = uvmexp.npages / 3;
211 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
212 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
213 }
214 
215 /*
216  * uvm_pageout: the main loop for the pagedaemon
217  */
218 
219 void
220 uvm_pageout(void *arg)
221 {
222 	int bufcnt, npages = 0;
223 	int extrapages = 0;
224 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
225 
226 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
227 
228 	/*
229 	 * ensure correct priority and set paging parameters...
230 	 */
231 
232 	uvm.pagedaemon_proc = curproc;
233 	uvm_lock_pageq();
234 	npages = uvmexp.npages;
235 	uvmpd_tune();
236 	uvm_unlock_pageq();
237 
238 	/*
239 	 * main loop
240 	 */
241 
242 	for (;;) {
243 		simple_lock(&uvm.pagedaemon_lock);
244 
245 		UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
246 		UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
247 		    &uvm.pagedaemon_lock, FALSE, "pgdaemon", 0);
248 		uvmexp.pdwoke++;
249 		UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
250 
251 		/*
252 		 * now lock page queues and recompute inactive count
253 		 */
254 
255 		uvm_lock_pageq();
256 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
257 			npages = uvmexp.npages;
258 			extrapages = uvm_extrapages;
259 			uvmpd_tune();
260 		}
261 
262 		uvmpdpol_tune();
263 
264 		/*
265 		 * Estimate a hint.  Note that bufmem are returned to
266 		 * system only when entire pool page is empty.
267 		 */
268 		bufcnt = uvmexp.freetarg - uvmexp.free;
269 		if (bufcnt < 0)
270 			bufcnt = 0;
271 
272 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d",
273 		    uvmexp.free, uvmexp.freetarg, 0,0);
274 
275 		/*
276 		 * scan if needed
277 		 */
278 
279 		if (uvmexp.free + uvmexp.paging < uvmexp.freetarg ||
280 		    uvmpdpol_needsscan_p()) {
281 			uvmpd_scan();
282 		}
283 
284 		/*
285 		 * if there's any free memory to be had,
286 		 * wake up any waiters.
287 		 */
288 
289 		if (uvmexp.free > uvmexp.reserve_kernel ||
290 		    uvmexp.paging == 0) {
291 			wakeup(&uvmexp.free);
292 		}
293 
294 		/*
295 		 * scan done.  unlock page queues (the only lock we are holding)
296 		 */
297 
298 		uvm_unlock_pageq();
299 
300 		buf_drain(bufcnt << PAGE_SHIFT);
301 
302 		/*
303 		 * drain pool resources now that we're not holding any locks
304 		 */
305 
306 		pool_drain(0);
307 
308 		/*
309 		 * free any cached u-areas we don't need
310 		 */
311 		uvm_uarea_drain(TRUE);
312 
313 	}
314 	/*NOTREACHED*/
315 }
316 
317 
318 /*
319  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
320  */
321 
322 void
323 uvm_aiodone_daemon(void *arg)
324 {
325 	int s, free;
326 	struct buf *bp, *nbp;
327 	UVMHIST_FUNC("uvm_aiodoned"); UVMHIST_CALLED(pdhist);
328 
329 	for (;;) {
330 
331 		/*
332 		 * carefully attempt to go to sleep (without losing "wakeups"!).
333 		 * we need splbio because we want to make sure the aio_done list
334 		 * is totally empty before we go to sleep.
335 		 */
336 
337 		s = splbio();
338 		simple_lock(&uvm.aiodoned_lock);
339 		if (TAILQ_FIRST(&uvm.aio_done) == NULL) {
340 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
341 			UVM_UNLOCK_AND_WAIT(&uvm.aiodoned,
342 			    &uvm.aiodoned_lock, FALSE, "aiodoned", 0);
343 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
344 
345 			/* relock aiodoned_lock, still at splbio */
346 			simple_lock(&uvm.aiodoned_lock);
347 		}
348 
349 		/*
350 		 * check for done aio structures
351 		 */
352 
353 		bp = TAILQ_FIRST(&uvm.aio_done);
354 		if (bp) {
355 			TAILQ_INIT(&uvm.aio_done);
356 		}
357 
358 		simple_unlock(&uvm.aiodoned_lock);
359 		splx(s);
360 
361 		/*
362 		 * process each i/o that's done.
363 		 */
364 
365 		free = uvmexp.free;
366 		while (bp != NULL) {
367 			nbp = TAILQ_NEXT(bp, b_freelist);
368 			(*bp->b_iodone)(bp);
369 			bp = nbp;
370 		}
371 		if (free <= uvmexp.reserve_kernel) {
372 			s = uvm_lock_fpageq();
373 			wakeup(&uvm.pagedaemon);
374 			uvm_unlock_fpageq(s);
375 		} else {
376 			simple_lock(&uvm.pagedaemon_lock);
377 			wakeup(&uvmexp.free);
378 			simple_unlock(&uvm.pagedaemon_lock);
379 		}
380 	}
381 }
382 
383 /*
384  * uvmpd_trylockowner: trylock the page's owner.
385  *
386  * => called with pageq locked.
387  * => resolve orphaned O->A loaned page.
388  * => return the locked simplelock on success.  otherwise, return NULL.
389  */
390 
391 struct simplelock *
392 uvmpd_trylockowner(struct vm_page *pg)
393 {
394 	struct uvm_object *uobj = pg->uobject;
395 	struct simplelock *slock;
396 
397 	UVM_LOCK_ASSERT_PAGEQ();
398 	if (uobj != NULL) {
399 		slock = &uobj->vmobjlock;
400 	} else {
401 		struct vm_anon *anon = pg->uanon;
402 
403 		KASSERT(anon != NULL);
404 		slock = &anon->an_lock;
405 	}
406 
407 	if (!simple_lock_try(slock)) {
408 		return NULL;
409 	}
410 
411 	if (uobj == NULL) {
412 
413 		/*
414 		 * set PQ_ANON if it isn't set already.
415 		 */
416 
417 		if ((pg->pqflags & PQ_ANON) == 0) {
418 			KASSERT(pg->loan_count > 0);
419 			pg->loan_count--;
420 			pg->pqflags |= PQ_ANON;
421 			/* anon now owns it */
422 		}
423 	}
424 
425 	return slock;
426 }
427 
428 #if defined(VMSWAP)
429 struct swapcluster {
430 	int swc_slot;
431 	int swc_nallocated;
432 	int swc_nused;
433 	struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)];
434 };
435 
436 static void
437 swapcluster_init(struct swapcluster *swc)
438 {
439 
440 	swc->swc_slot = 0;
441 }
442 
443 static int
444 swapcluster_allocslots(struct swapcluster *swc)
445 {
446 	int slot;
447 	int npages;
448 
449 	if (swc->swc_slot != 0) {
450 		return 0;
451 	}
452 
453 	/* Even with strange MAXPHYS, the shift
454 	   implicitly rounds down to a page. */
455 	npages = MAXPHYS >> PAGE_SHIFT;
456 	slot = uvm_swap_alloc(&npages, TRUE);
457 	if (slot == 0) {
458 		return ENOMEM;
459 	}
460 	swc->swc_slot = slot;
461 	swc->swc_nallocated = npages;
462 	swc->swc_nused = 0;
463 
464 	return 0;
465 }
466 
467 static int
468 swapcluster_add(struct swapcluster *swc, struct vm_page *pg)
469 {
470 	int slot;
471 	struct uvm_object *uobj;
472 
473 	KASSERT(swc->swc_slot != 0);
474 	KASSERT(swc->swc_nused < swc->swc_nallocated);
475 	KASSERT((pg->pqflags & PQ_SWAPBACKED) != 0);
476 
477 	slot = swc->swc_slot + swc->swc_nused;
478 	uobj = pg->uobject;
479 	if (uobj == NULL) {
480 		LOCK_ASSERT(simple_lock_held(&pg->uanon->an_lock));
481 		pg->uanon->an_swslot = slot;
482 	} else {
483 		int result;
484 
485 		LOCK_ASSERT(simple_lock_held(&uobj->vmobjlock));
486 		result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot);
487 		if (result == -1) {
488 			return ENOMEM;
489 		}
490 	}
491 	swc->swc_pages[swc->swc_nused] = pg;
492 	swc->swc_nused++;
493 
494 	return 0;
495 }
496 
497 static void
498 swapcluster_flush(struct swapcluster *swc, boolean_t now)
499 {
500 	int slot;
501 	int nused;
502 	int nallocated;
503 	int error;
504 
505 	if (swc->swc_slot == 0) {
506 		return;
507 	}
508 	KASSERT(swc->swc_nused <= swc->swc_nallocated);
509 
510 	slot = swc->swc_slot;
511 	nused = swc->swc_nused;
512 	nallocated = swc->swc_nallocated;
513 
514 	/*
515 	 * if this is the final pageout we could have a few
516 	 * unused swap blocks.  if so, free them now.
517 	 */
518 
519 	if (nused < nallocated) {
520 		if (!now) {
521 			return;
522 		}
523 		uvm_swap_free(slot + nused, nallocated - nused);
524 	}
525 
526 	/*
527 	 * now start the pageout.
528 	 */
529 
530 	uvmexp.pdpageouts++;
531 	error = uvm_swap_put(slot, swc->swc_pages, nused, 0);
532 	KASSERT(error == 0);
533 
534 	/*
535 	 * zero swslot to indicate that we are
536 	 * no longer building a swap-backed cluster.
537 	 */
538 
539 	swc->swc_slot = 0;
540 }
541 
542 /*
543  * uvmpd_dropswap: free any swap allocated to this page.
544  *
545  * => called with owner locked.
546  * => return TRUE if a page had an associated slot.
547  */
548 
549 static boolean_t
550 uvmpd_dropswap(struct vm_page *pg)
551 {
552 	boolean_t result = FALSE;
553 	struct vm_anon *anon = pg->uanon;
554 
555 	if ((pg->pqflags & PQ_ANON) && anon->an_swslot) {
556 		uvm_swap_free(anon->an_swslot, 1);
557 		anon->an_swslot = 0;
558 		pg->flags &= ~PG_CLEAN;
559 		result = TRUE;
560 	} else if (pg->pqflags & PQ_AOBJ) {
561 		int slot = uao_set_swslot(pg->uobject,
562 		    pg->offset >> PAGE_SHIFT, 0);
563 		if (slot) {
564 			uvm_swap_free(slot, 1);
565 			pg->flags &= ~PG_CLEAN;
566 			result = TRUE;
567 		}
568 	}
569 
570 	return result;
571 }
572 
573 /*
574  * uvmpd_trydropswap: try to free any swap allocated to this page.
575  *
576  * => return TRUE if a slot is successfully freed.
577  */
578 
579 boolean_t
580 uvmpd_trydropswap(struct vm_page *pg)
581 {
582 	struct simplelock *slock;
583 	boolean_t result;
584 
585 	if ((pg->flags & PG_BUSY) != 0) {
586 		return FALSE;
587 	}
588 
589 	/*
590 	 * lock the page's owner.
591 	 */
592 
593 	slock = uvmpd_trylockowner(pg);
594 	if (slock == NULL) {
595 		return FALSE;
596 	}
597 
598 	/*
599 	 * skip this page if it's busy.
600 	 */
601 
602 	if ((pg->flags & PG_BUSY) != 0) {
603 		simple_unlock(slock);
604 		return FALSE;
605 	}
606 
607 	result = uvmpd_dropswap(pg);
608 
609 	simple_unlock(slock);
610 
611 	return result;
612 }
613 
614 #endif /* defined(VMSWAP) */
615 
616 /*
617  * uvmpd_scan_queue: scan an replace candidate list for pages
618  * to clean or free.
619  *
620  * => called with page queues locked
621  * => we work on meeting our free target by converting inactive pages
622  *    into free pages.
623  * => we handle the building of swap-backed clusters
624  */
625 
626 static void
627 uvmpd_scan_queue(void)
628 {
629 	struct vm_page *p;
630 	struct uvm_object *uobj;
631 	struct vm_anon *anon;
632 #if defined(VMSWAP)
633 	struct swapcluster swc;
634 #endif /* defined(VMSWAP) */
635 	int dirtyreacts;
636 	struct simplelock *slock;
637 	UVMHIST_FUNC("uvmpd_scan_queue"); UVMHIST_CALLED(pdhist);
638 
639 	/*
640 	 * swslot is non-zero if we are building a swap cluster.  we want
641 	 * to stay in the loop while we have a page to scan or we have
642 	 * a swap-cluster to build.
643 	 */
644 
645 #if defined(VMSWAP)
646 	swapcluster_init(&swc);
647 #endif /* defined(VMSWAP) */
648 
649 	dirtyreacts = 0;
650 	uvmpdpol_scaninit();
651 
652 	while (/* CONSTCOND */ 1) {
653 
654 		/*
655 		 * see if we've met the free target.
656 		 */
657 
658 		if (uvmexp.free + uvmexp.paging >= uvmexp.freetarg << 2 ||
659 		    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
660 			UVMHIST_LOG(pdhist,"  met free target: "
661 				    "exit loop", 0, 0, 0, 0);
662 			break;
663 		}
664 
665 		p = uvmpdpol_selectvictim();
666 		if (p == NULL) {
667 			break;
668 		}
669 		KASSERT(uvmpdpol_pageisqueued_p(p));
670 		KASSERT(p->wire_count == 0);
671 
672 		/*
673 		 * we are below target and have a new page to consider.
674 		 */
675 
676 		anon = p->uanon;
677 		uobj = p->uobject;
678 
679 		/*
680 		 * first we attempt to lock the object that this page
681 		 * belongs to.  if our attempt fails we skip on to
682 		 * the next page (no harm done).  it is important to
683 		 * "try" locking the object as we are locking in the
684 		 * wrong order (pageq -> object) and we don't want to
685 		 * deadlock.
686 		 *
687 		 * the only time we expect to see an ownerless page
688 		 * (i.e. a page with no uobject and !PQ_ANON) is if an
689 		 * anon has loaned a page from a uvm_object and the
690 		 * uvm_object has dropped the ownership.  in that
691 		 * case, the anon can "take over" the loaned page
692 		 * and make it its own.
693 		 */
694 
695 		slock = uvmpd_trylockowner(p);
696 		if (slock == NULL) {
697 			continue;
698 		}
699 		if (p->flags & PG_BUSY) {
700 			simple_unlock(slock);
701 			uvmexp.pdbusy++;
702 			continue;
703 		}
704 
705 		/* does the page belong to an object? */
706 		if (uobj != NULL) {
707 			uvmexp.pdobscan++;
708 		} else {
709 #if defined(VMSWAP)
710 			KASSERT(anon != NULL);
711 			uvmexp.pdanscan++;
712 #else /* defined(VMSWAP) */
713 			panic("%s: anon", __func__);
714 #endif /* defined(VMSWAP) */
715 		}
716 
717 
718 		/*
719 		 * we now have the object and the page queues locked.
720 		 * if the page is not swap-backed, call the object's
721 		 * pager to flush and free the page.
722 		 */
723 
724 #if defined(READAHEAD_STATS)
725 		if ((p->pqflags & PQ_READAHEAD) != 0) {
726 			p->pqflags &= ~PQ_READAHEAD;
727 			uvm_ra_miss.ev_count++;
728 		}
729 #endif /* defined(READAHEAD_STATS) */
730 
731 		if ((p->pqflags & PQ_SWAPBACKED) == 0) {
732 			uvm_unlock_pageq();
733 			(void) (uobj->pgops->pgo_put)(uobj, p->offset,
734 			    p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE);
735 			uvm_lock_pageq();
736 			continue;
737 		}
738 
739 		/*
740 		 * the page is swap-backed.  remove all the permissions
741 		 * from the page so we can sync the modified info
742 		 * without any race conditions.  if the page is clean
743 		 * we can free it now and continue.
744 		 */
745 
746 		pmap_page_protect(p, VM_PROT_NONE);
747 		if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
748 			p->flags &= ~(PG_CLEAN);
749 		}
750 		if (p->flags & PG_CLEAN) {
751 			int slot;
752 			int pageidx;
753 
754 			pageidx = p->offset >> PAGE_SHIFT;
755 			uvm_pagefree(p);
756 			uvmexp.pdfreed++;
757 
758 			/*
759 			 * for anons, we need to remove the page
760 			 * from the anon ourselves.  for aobjs,
761 			 * pagefree did that for us.
762 			 */
763 
764 			if (anon) {
765 				KASSERT(anon->an_swslot != 0);
766 				anon->an_page = NULL;
767 				slot = anon->an_swslot;
768 			} else {
769 				slot = uao_find_swslot(uobj, pageidx);
770 			}
771 			simple_unlock(slock);
772 
773 			if (slot > 0) {
774 				/* this page is now only in swap. */
775 				simple_lock(&uvm.swap_data_lock);
776 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
777 				uvmexp.swpgonly++;
778 				simple_unlock(&uvm.swap_data_lock);
779 			}
780 			continue;
781 		}
782 
783 #if defined(VMSWAP)
784 		/*
785 		 * this page is dirty, skip it if we'll have met our
786 		 * free target when all the current pageouts complete.
787 		 */
788 
789 		if (uvmexp.free + uvmexp.paging > uvmexp.freetarg << 2) {
790 			simple_unlock(slock);
791 			continue;
792 		}
793 
794 		/*
795 		 * free any swap space allocated to the page since
796 		 * we'll have to write it again with its new data.
797 		 */
798 
799 		uvmpd_dropswap(p);
800 
801 		/*
802 		 * if all pages in swap are only in swap,
803 		 * the swap space is full and we can't page out
804 		 * any more swap-backed pages.  reactivate this page
805 		 * so that we eventually cycle all pages through
806 		 * the inactive queue.
807 		 */
808 
809 		if (uvm_swapisfull()) {
810 			dirtyreacts++;
811 			uvm_pageactivate(p);
812 			simple_unlock(slock);
813 			continue;
814 		}
815 
816 		/*
817 		 * start new swap pageout cluster (if necessary).
818 		 */
819 
820 		if (swapcluster_allocslots(&swc)) {
821 			simple_unlock(slock);
822 			dirtyreacts++; /* XXX */
823 			continue;
824 		}
825 
826 		/*
827 		 * at this point, we're definitely going reuse this
828 		 * page.  mark the page busy and delayed-free.
829 		 * we should remove the page from the page queues
830 		 * so we don't ever look at it again.
831 		 * adjust counters and such.
832 		 */
833 
834 		p->flags |= PG_BUSY;
835 		UVM_PAGE_OWN(p, "scan_queue");
836 
837 		p->flags |= PG_PAGEOUT;
838 		uvmexp.paging++;
839 		uvm_pagedequeue(p);
840 
841 		uvmexp.pgswapout++;
842 		uvm_unlock_pageq();
843 
844 		/*
845 		 * add the new page to the cluster.
846 		 */
847 
848 		if (swapcluster_add(&swc, p)) {
849 			p->flags &= ~(PG_BUSY|PG_PAGEOUT);
850 			UVM_PAGE_OWN(p, NULL);
851 			uvm_lock_pageq();
852 			uvmexp.paging--;
853 			dirtyreacts++;
854 			uvm_pageactivate(p);
855 			simple_unlock(slock);
856 			continue;
857 		}
858 		simple_unlock(slock);
859 
860 		swapcluster_flush(&swc, FALSE);
861 		uvm_lock_pageq();
862 
863 		/*
864 		 * the pageout is in progress.  bump counters and set up
865 		 * for the next loop.
866 		 */
867 
868 		uvmexp.pdpending++;
869 
870 #else /* defined(VMSWAP) */
871 		uvm_pageactivate(p);
872 		simple_unlock(slock);
873 #endif /* defined(VMSWAP) */
874 	}
875 
876 #if defined(VMSWAP)
877 	uvm_unlock_pageq();
878 	swapcluster_flush(&swc, TRUE);
879 	uvm_lock_pageq();
880 #endif /* defined(VMSWAP) */
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 static void
890 uvmpd_scan(void)
891 {
892 	int swap_shortage, pages_freed;
893 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
894 
895 	uvmexp.pdrevs++;
896 
897 #ifndef __SWAP_BROKEN
898 
899 	/*
900 	 * swap out some processes if we are below our free target.
901 	 * we need to unlock the page queues for this.
902 	 */
903 
904 	if (uvmexp.free < uvmexp.freetarg && uvmexp.nswapdev != 0) {
905 		uvmexp.pdswout++;
906 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout",
907 		    uvmexp.free, uvmexp.freetarg, 0, 0);
908 		uvm_unlock_pageq();
909 		uvm_swapout_threads();
910 		uvm_lock_pageq();
911 
912 	}
913 #endif
914 
915 	/*
916 	 * now we want to work on meeting our targets.   first we work on our
917 	 * free target by converting inactive pages into free pages.  then
918 	 * we work on meeting our inactive target by converting active pages
919 	 * to inactive ones.
920 	 */
921 
922 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
923 
924 	pages_freed = uvmexp.pdfreed;
925 	uvmpd_scan_queue();
926 	pages_freed = uvmexp.pdfreed - pages_freed;
927 
928 	/*
929 	 * detect if we're not going to be able to page anything out
930 	 * until we free some swap resources from active pages.
931 	 */
932 
933 	swap_shortage = 0;
934 	if (uvmexp.free < uvmexp.freetarg &&
935 	    uvmexp.swpginuse >= uvmexp.swpgavail &&
936 	    !uvm_swapisfull() &&
937 	    pages_freed == 0) {
938 		swap_shortage = uvmexp.freetarg - uvmexp.free;
939 	}
940 
941 	uvmpdpol_balancequeue(swap_shortage);
942 }
943 
944 /*
945  * uvm_reclaimable: decide whether to wait for pagedaemon.
946  *
947  * => return TRUE if it seems to be worth to do uvm_wait.
948  *
949  * XXX should be tunable.
950  * XXX should consider pools, etc?
951  */
952 
953 boolean_t
954 uvm_reclaimable(void)
955 {
956 	int filepages;
957 	int active, inactive;
958 
959 	/*
960 	 * if swap is not full, no problem.
961 	 */
962 
963 	if (!uvm_swapisfull()) {
964 		return TRUE;
965 	}
966 
967 	/*
968 	 * file-backed pages can be reclaimed even when swap is full.
969 	 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim.
970 	 *
971 	 * XXX assume the worst case, ie. all wired pages are file-backed.
972 	 *
973 	 * XXX should consider about other reclaimable memory.
974 	 * XXX ie. pools, traditional buffer cache.
975 	 */
976 
977 	filepages = uvmexp.filepages + uvmexp.execpages - uvmexp.wired;
978 	uvm_estimatepageable(&active, &inactive);
979 	if (filepages >= MIN((active + inactive) >> 4,
980 	    5 * 1024 * 1024 >> PAGE_SHIFT)) {
981 		return TRUE;
982 	}
983 
984 	/*
985 	 * kill the process, fail allocation, etc..
986 	 */
987 
988 	return FALSE;
989 }
990 
991 void
992 uvm_estimatepageable(int *active, int *inactive)
993 {
994 
995 	uvmpdpol_estimatepageable(active, inactive);
996 }
997