xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 96230fab84e26a6435963032070e916a951a8b2e)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.93 2008/09/23 08:55:52 ad 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 1.93 2008/09/23 08:55:52 ad 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 
86 #include <uvm/uvm.h>
87 #include <uvm/uvm_pdpolicy.h>
88 
89 /*
90  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
91  * in a pass thru the inactive list when swap is full.  the value should be
92  * "small"... if it's too large we'll cycle the active pages thru the inactive
93  * queue too quickly to for them to be referenced and avoid being freed.
94  */
95 
96 #define	UVMPD_NUMDIRTYREACTS	16
97 
98 #define	UVMPD_NUMTRYLOCKOWNER	16
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 unsigned int uvm_pagedaemon_waiters;
109 
110 /*
111  * XXX hack to avoid hangs when large processes fork.
112  */
113 int uvm_extrapages;
114 
115 /*
116  * uvm_wait: wait (sleep) for the page daemon to free some pages
117  *
118  * => should be called with all locks released
119  * => should _not_ be called by the page daemon (to avoid deadlock)
120  */
121 
122 void
123 uvm_wait(const char *wmsg)
124 {
125 	int timo = 0;
126 
127 	mutex_spin_enter(&uvm_fpageqlock);
128 
129 	/*
130 	 * check for page daemon going to sleep (waiting for itself)
131 	 */
132 
133 	if (curlwp == uvm.pagedaemon_lwp && uvmexp.paging == 0) {
134 		/*
135 		 * now we have a problem: the pagedaemon wants to go to
136 		 * sleep until it frees more memory.   but how can it
137 		 * free more memory if it is asleep?  that is a deadlock.
138 		 * we have two options:
139 		 *  [1] panic now
140 		 *  [2] put a timeout on the sleep, thus causing the
141 		 *      pagedaemon to only pause (rather than sleep forever)
142 		 *
143 		 * note that option [2] will only help us if we get lucky
144 		 * and some other process on the system breaks the deadlock
145 		 * by exiting or freeing memory (thus allowing the pagedaemon
146 		 * to continue).  for now we panic if DEBUG is defined,
147 		 * otherwise we hope for the best with option [2] (better
148 		 * yet, this should never happen in the first place!).
149 		 */
150 
151 		printf("pagedaemon: deadlock detected!\n");
152 		timo = hz >> 3;		/* set timeout */
153 #if defined(DEBUG)
154 		/* DEBUG: panic so we can debug it */
155 		panic("pagedaemon deadlock");
156 #endif
157 	}
158 
159 	uvm_pagedaemon_waiters++;
160 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
161 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm_fpageqlock, false, wmsg, timo);
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  * => called with uvm_fpageqlock held.
169  */
170 
171 void
172 uvm_kick_pdaemon(void)
173 {
174 
175 	KASSERT(mutex_owned(&uvm_fpageqlock));
176 
177 	if (uvmexp.free + uvmexp.paging < uvmexp.freemin ||
178 	    (uvmexp.free + uvmexp.paging < uvmexp.freetarg &&
179 	     uvmpdpol_needsscan_p())) {
180 		wakeup(&uvm.pagedaemon);
181 	}
182 }
183 
184 /*
185  * uvmpd_tune: tune paging parameters
186  *
187  * => called when ever memory is added (or removed?) to the system
188  * => caller must call with page queues locked
189  */
190 
191 static void
192 uvmpd_tune(void)
193 {
194 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
195 
196 	/*
197 	 * try to keep 0.5% of available RAM free, but limit to between
198 	 * 128k and 1024k per-CPU.  XXX: what are these values good for?
199 	 */
200 	uvmexp.freemin = uvmexp.npages / 200;
201 	uvmexp.freemin = MAX(uvmexp.freemin, (128*1024) >> PAGE_SHIFT);
202 	uvmexp.freemin = MIN(uvmexp.freemin, (1024*1024) >> PAGE_SHIFT);
203 	uvmexp.freemin *= ncpu;
204 
205 	/* Make sure there's always a user page free. */
206 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
207 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
208 
209 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
210 	if (uvmexp.freetarg <= uvmexp.freemin)
211 		uvmexp.freetarg = uvmexp.freemin + 1;
212 
213 	uvmexp.freetarg += uvm_extrapages;
214 	uvm_extrapages = 0;
215 
216 	uvmexp.wiredmax = uvmexp.npages / 3;
217 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
218 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
219 }
220 
221 /*
222  * uvm_pageout: the main loop for the pagedaemon
223  */
224 
225 void
226 uvm_pageout(void *arg)
227 {
228 	int bufcnt, npages = 0;
229 	int extrapages = 0;
230 	struct pool *pp;
231 	uint64_t where;
232 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
233 
234 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
235 
236 	/*
237 	 * ensure correct priority and set paging parameters...
238 	 */
239 
240 	uvm.pagedaemon_lwp = curlwp;
241 	mutex_enter(&uvm_pageqlock);
242 	npages = uvmexp.npages;
243 	uvmpd_tune();
244 	mutex_exit(&uvm_pageqlock);
245 
246 	/*
247 	 * main loop
248 	 */
249 
250 	for (;;) {
251 		bool needsscan, needsfree;
252 
253 		mutex_spin_enter(&uvm_fpageqlock);
254 		if (uvm_pagedaemon_waiters == 0 || uvmexp.paging > 0) {
255 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
256 			UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
257 			    &uvm_fpageqlock, false, "pgdaemon", 0);
258 			uvmexp.pdwoke++;
259 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
260 		} else {
261 			mutex_spin_exit(&uvm_fpageqlock);
262 		}
263 
264 		/*
265 		 * now lock page queues and recompute inactive count
266 		 */
267 
268 		mutex_enter(&uvm_pageqlock);
269 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
270 			npages = uvmexp.npages;
271 			extrapages = uvm_extrapages;
272 			mutex_spin_enter(&uvm_fpageqlock);
273 			uvmpd_tune();
274 			mutex_spin_exit(&uvm_fpageqlock);
275 		}
276 
277 		uvmpdpol_tune();
278 
279 		/*
280 		 * Estimate a hint.  Note that bufmem are returned to
281 		 * system only when entire pool page is empty.
282 		 */
283 		mutex_spin_enter(&uvm_fpageqlock);
284 		bufcnt = uvmexp.freetarg - uvmexp.free;
285 		if (bufcnt < 0)
286 			bufcnt = 0;
287 
288 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d",
289 		    uvmexp.free, uvmexp.freetarg, 0,0);
290 
291 		needsfree = uvmexp.free + uvmexp.paging < uvmexp.freetarg;
292 		needsscan = needsfree || uvmpdpol_needsscan_p();
293 		mutex_spin_exit(&uvm_fpageqlock);
294 
295 		/*
296 		 * scan if needed
297 		 */
298 		if (needsscan)
299 			uvmpd_scan();
300 
301 		/*
302 		 * if there's any free memory to be had,
303 		 * wake up any waiters.
304 		 */
305 
306 		mutex_spin_enter(&uvm_fpageqlock);
307 		if (uvmexp.free > uvmexp.reserve_kernel ||
308 		    uvmexp.paging == 0) {
309 			wakeup(&uvmexp.free);
310 			uvm_pagedaemon_waiters = 0;
311 		}
312 		mutex_spin_exit(&uvm_fpageqlock);
313 
314 		/*
315 		 * scan done.  unlock page queues (the only lock we are holding)
316 		 */
317 		mutex_exit(&uvm_pageqlock);
318 
319 		/*
320 		 * if we don't need free memory, we're done.
321 		 */
322 
323 		if (!needsfree)
324 			continue;
325 
326 		/*
327 		 * start draining pool resources now that we're not
328 		 * holding any locks.
329 		 */
330 		pool_drain_start(&pp, &where);
331 
332 		/*
333 		 * kill unused metadata buffers.
334 		 */
335 		mutex_enter(&bufcache_lock);
336 		buf_drain(bufcnt << PAGE_SHIFT);
337 		mutex_exit(&bufcache_lock);
338 
339 		/*
340 		 * complete draining the pools.
341 		 */
342 		pool_drain_end(pp, where);
343 	}
344 	/*NOTREACHED*/
345 }
346 
347 
348 /*
349  * uvm_aiodone_worker: a workqueue callback for the aiodone daemon.
350  */
351 
352 void
353 uvm_aiodone_worker(struct work *wk, void *dummy)
354 {
355 	struct buf *bp = (void *)wk;
356 
357 	KASSERT(&bp->b_work == wk);
358 
359 	/*
360 	 * process an i/o that's done.
361 	 */
362 
363 	(*bp->b_iodone)(bp);
364 }
365 
366 void
367 uvm_pageout_start(int npages)
368 {
369 
370 	mutex_spin_enter(&uvm_fpageqlock);
371 	uvmexp.paging += npages;
372 	mutex_spin_exit(&uvm_fpageqlock);
373 }
374 
375 void
376 uvm_pageout_done(int npages)
377 {
378 
379 	mutex_spin_enter(&uvm_fpageqlock);
380 	KASSERT(uvmexp.paging >= npages);
381 	uvmexp.paging -= npages;
382 
383 	/*
384 	 * wake up either of pagedaemon or LWPs waiting for it.
385 	 */
386 
387 	if (uvmexp.free <= uvmexp.reserve_kernel) {
388 		wakeup(&uvm.pagedaemon);
389 	} else {
390 		wakeup(&uvmexp.free);
391 		uvm_pagedaemon_waiters = 0;
392 	}
393 	mutex_spin_exit(&uvm_fpageqlock);
394 }
395 
396 /*
397  * uvmpd_trylockowner: trylock the page's owner.
398  *
399  * => called with pageq locked.
400  * => resolve orphaned O->A loaned page.
401  * => return the locked mutex on success.  otherwise, return NULL.
402  */
403 
404 kmutex_t *
405 uvmpd_trylockowner(struct vm_page *pg)
406 {
407 	struct uvm_object *uobj = pg->uobject;
408 	kmutex_t *slock;
409 
410 	KASSERT(mutex_owned(&uvm_pageqlock));
411 
412 	if (uobj != NULL) {
413 		slock = &uobj->vmobjlock;
414 	} else {
415 		struct vm_anon *anon = pg->uanon;
416 
417 		KASSERT(anon != NULL);
418 		slock = &anon->an_lock;
419 	}
420 
421 	if (!mutex_tryenter(slock)) {
422 		return NULL;
423 	}
424 
425 	if (uobj == NULL) {
426 
427 		/*
428 		 * set PQ_ANON if it isn't set already.
429 		 */
430 
431 		if ((pg->pqflags & PQ_ANON) == 0) {
432 			KASSERT(pg->loan_count > 0);
433 			pg->loan_count--;
434 			pg->pqflags |= PQ_ANON;
435 			/* anon now owns it */
436 		}
437 	}
438 
439 	return slock;
440 }
441 
442 #if defined(VMSWAP)
443 struct swapcluster {
444 	int swc_slot;
445 	int swc_nallocated;
446 	int swc_nused;
447 	struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)];
448 };
449 
450 static void
451 swapcluster_init(struct swapcluster *swc)
452 {
453 
454 	swc->swc_slot = 0;
455 	swc->swc_nused = 0;
456 }
457 
458 static int
459 swapcluster_allocslots(struct swapcluster *swc)
460 {
461 	int slot;
462 	int npages;
463 
464 	if (swc->swc_slot != 0) {
465 		return 0;
466 	}
467 
468 	/* Even with strange MAXPHYS, the shift
469 	   implicitly rounds down to a page. */
470 	npages = MAXPHYS >> PAGE_SHIFT;
471 	slot = uvm_swap_alloc(&npages, true);
472 	if (slot == 0) {
473 		return ENOMEM;
474 	}
475 	swc->swc_slot = slot;
476 	swc->swc_nallocated = npages;
477 	swc->swc_nused = 0;
478 
479 	return 0;
480 }
481 
482 static int
483 swapcluster_add(struct swapcluster *swc, struct vm_page *pg)
484 {
485 	int slot;
486 	struct uvm_object *uobj;
487 
488 	KASSERT(swc->swc_slot != 0);
489 	KASSERT(swc->swc_nused < swc->swc_nallocated);
490 	KASSERT((pg->pqflags & PQ_SWAPBACKED) != 0);
491 
492 	slot = swc->swc_slot + swc->swc_nused;
493 	uobj = pg->uobject;
494 	if (uobj == NULL) {
495 		KASSERT(mutex_owned(&pg->uanon->an_lock));
496 		pg->uanon->an_swslot = slot;
497 	} else {
498 		int result;
499 
500 		KASSERT(mutex_owned(&uobj->vmobjlock));
501 		result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot);
502 		if (result == -1) {
503 			return ENOMEM;
504 		}
505 	}
506 	swc->swc_pages[swc->swc_nused] = pg;
507 	swc->swc_nused++;
508 
509 	return 0;
510 }
511 
512 static void
513 swapcluster_flush(struct swapcluster *swc, bool now)
514 {
515 	int slot;
516 	int nused;
517 	int nallocated;
518 	int error;
519 
520 	if (swc->swc_slot == 0) {
521 		return;
522 	}
523 	KASSERT(swc->swc_nused <= swc->swc_nallocated);
524 
525 	slot = swc->swc_slot;
526 	nused = swc->swc_nused;
527 	nallocated = swc->swc_nallocated;
528 
529 	/*
530 	 * if this is the final pageout we could have a few
531 	 * unused swap blocks.  if so, free them now.
532 	 */
533 
534 	if (nused < nallocated) {
535 		if (!now) {
536 			return;
537 		}
538 		uvm_swap_free(slot + nused, nallocated - nused);
539 	}
540 
541 	/*
542 	 * now start the pageout.
543 	 */
544 
545 	if (nused > 0) {
546 		uvmexp.pdpageouts++;
547 		uvm_pageout_start(nused);
548 		error = uvm_swap_put(slot, swc->swc_pages, nused, 0);
549 		KASSERT(error == 0 || error == ENOMEM);
550 	}
551 
552 	/*
553 	 * zero swslot to indicate that we are
554 	 * no longer building a swap-backed cluster.
555 	 */
556 
557 	swc->swc_slot = 0;
558 	swc->swc_nused = 0;
559 }
560 
561 static int
562 swapcluster_nused(struct swapcluster *swc)
563 {
564 
565 	return swc->swc_nused;
566 }
567 
568 /*
569  * uvmpd_dropswap: free any swap allocated to this page.
570  *
571  * => called with owner locked.
572  * => return true if a page had an associated slot.
573  */
574 
575 static bool
576 uvmpd_dropswap(struct vm_page *pg)
577 {
578 	bool result = false;
579 	struct vm_anon *anon = pg->uanon;
580 
581 	if ((pg->pqflags & PQ_ANON) && anon->an_swslot) {
582 		uvm_swap_free(anon->an_swslot, 1);
583 		anon->an_swslot = 0;
584 		pg->flags &= ~PG_CLEAN;
585 		result = true;
586 	} else if (pg->pqflags & PQ_AOBJ) {
587 		int slot = uao_set_swslot(pg->uobject,
588 		    pg->offset >> PAGE_SHIFT, 0);
589 		if (slot) {
590 			uvm_swap_free(slot, 1);
591 			pg->flags &= ~PG_CLEAN;
592 			result = true;
593 		}
594 	}
595 
596 	return result;
597 }
598 
599 /*
600  * uvmpd_trydropswap: try to free any swap allocated to this page.
601  *
602  * => return true if a slot is successfully freed.
603  */
604 
605 bool
606 uvmpd_trydropswap(struct vm_page *pg)
607 {
608 	kmutex_t *slock;
609 	bool result;
610 
611 	if ((pg->flags & PG_BUSY) != 0) {
612 		return false;
613 	}
614 
615 	/*
616 	 * lock the page's owner.
617 	 */
618 
619 	slock = uvmpd_trylockowner(pg);
620 	if (slock == NULL) {
621 		return false;
622 	}
623 
624 	/*
625 	 * skip this page if it's busy.
626 	 */
627 
628 	if ((pg->flags & PG_BUSY) != 0) {
629 		mutex_exit(slock);
630 		return false;
631 	}
632 
633 	result = uvmpd_dropswap(pg);
634 
635 	mutex_exit(slock);
636 
637 	return result;
638 }
639 
640 #endif /* defined(VMSWAP) */
641 
642 /*
643  * uvmpd_scan_queue: scan an replace candidate list for pages
644  * to clean or free.
645  *
646  * => called with page queues locked
647  * => we work on meeting our free target by converting inactive pages
648  *    into free pages.
649  * => we handle the building of swap-backed clusters
650  */
651 
652 static void
653 uvmpd_scan_queue(void)
654 {
655 	struct vm_page *p;
656 	struct uvm_object *uobj;
657 	struct vm_anon *anon;
658 #if defined(VMSWAP)
659 	struct swapcluster swc;
660 #endif /* defined(VMSWAP) */
661 	int dirtyreacts;
662 	int lockownerfail;
663 	kmutex_t *slock;
664 	UVMHIST_FUNC("uvmpd_scan_queue"); UVMHIST_CALLED(pdhist);
665 
666 	/*
667 	 * swslot is non-zero if we are building a swap cluster.  we want
668 	 * to stay in the loop while we have a page to scan or we have
669 	 * a swap-cluster to build.
670 	 */
671 
672 #if defined(VMSWAP)
673 	swapcluster_init(&swc);
674 #endif /* defined(VMSWAP) */
675 
676 	dirtyreacts = 0;
677 	lockownerfail = 0;
678 	uvmpdpol_scaninit();
679 
680 	while (/* CONSTCOND */ 1) {
681 
682 		/*
683 		 * see if we've met the free target.
684 		 */
685 
686 		if (uvmexp.free + uvmexp.paging
687 #if defined(VMSWAP)
688 		    + swapcluster_nused(&swc)
689 #endif /* defined(VMSWAP) */
690 		    >= uvmexp.freetarg << 2 ||
691 		    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
692 			UVMHIST_LOG(pdhist,"  met free target: "
693 				    "exit loop", 0, 0, 0, 0);
694 			break;
695 		}
696 
697 		p = uvmpdpol_selectvictim();
698 		if (p == NULL) {
699 			break;
700 		}
701 		KASSERT(uvmpdpol_pageisqueued_p(p));
702 		KASSERT(p->wire_count == 0);
703 
704 		/*
705 		 * we are below target and have a new page to consider.
706 		 */
707 
708 		anon = p->uanon;
709 		uobj = p->uobject;
710 
711 		/*
712 		 * first we attempt to lock the object that this page
713 		 * belongs to.  if our attempt fails we skip on to
714 		 * the next page (no harm done).  it is important to
715 		 * "try" locking the object as we are locking in the
716 		 * wrong order (pageq -> object) and we don't want to
717 		 * deadlock.
718 		 *
719 		 * the only time we expect to see an ownerless page
720 		 * (i.e. a page with no uobject and !PQ_ANON) is if an
721 		 * anon has loaned a page from a uvm_object and the
722 		 * uvm_object has dropped the ownership.  in that
723 		 * case, the anon can "take over" the loaned page
724 		 * and make it its own.
725 		 */
726 
727 		slock = uvmpd_trylockowner(p);
728 		if (slock == NULL) {
729 			/*
730 			 * yield cpu to make a chance for an LWP holding
731 			 * the lock run.  otherwise we can busy-loop too long
732 			 * if the page queue is filled with a lot of pages
733 			 * from few objects.
734 			 */
735 			lockownerfail++;
736 			if (lockownerfail > UVMPD_NUMTRYLOCKOWNER) {
737 				mutex_exit(&uvm_pageqlock);
738 				/* XXX Better than yielding but inadequate. */
739 				kpause("livelock", false, 1, NULL);
740 				mutex_enter(&uvm_pageqlock);
741 				lockownerfail = 0;
742 			}
743 			continue;
744 		}
745 		if (p->flags & PG_BUSY) {
746 			mutex_exit(slock);
747 			uvmexp.pdbusy++;
748 			continue;
749 		}
750 
751 		/* does the page belong to an object? */
752 		if (uobj != NULL) {
753 			uvmexp.pdobscan++;
754 		} else {
755 #if defined(VMSWAP)
756 			KASSERT(anon != NULL);
757 			uvmexp.pdanscan++;
758 #else /* defined(VMSWAP) */
759 			panic("%s: anon", __func__);
760 #endif /* defined(VMSWAP) */
761 		}
762 
763 
764 		/*
765 		 * we now have the object and the page queues locked.
766 		 * if the page is not swap-backed, call the object's
767 		 * pager to flush and free the page.
768 		 */
769 
770 #if defined(READAHEAD_STATS)
771 		if ((p->pqflags & PQ_READAHEAD) != 0) {
772 			p->pqflags &= ~PQ_READAHEAD;
773 			uvm_ra_miss.ev_count++;
774 		}
775 #endif /* defined(READAHEAD_STATS) */
776 
777 		if ((p->pqflags & PQ_SWAPBACKED) == 0) {
778 			KASSERT(uobj != NULL);
779 			mutex_exit(&uvm_pageqlock);
780 			(void) (uobj->pgops->pgo_put)(uobj, p->offset,
781 			    p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE);
782 			mutex_enter(&uvm_pageqlock);
783 			continue;
784 		}
785 
786 		/*
787 		 * the page is swap-backed.  remove all the permissions
788 		 * from the page so we can sync the modified info
789 		 * without any race conditions.  if the page is clean
790 		 * we can free it now and continue.
791 		 */
792 
793 		pmap_page_protect(p, VM_PROT_NONE);
794 		if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
795 			p->flags &= ~(PG_CLEAN);
796 		}
797 		if (p->flags & PG_CLEAN) {
798 			int slot;
799 			int pageidx;
800 
801 			pageidx = p->offset >> PAGE_SHIFT;
802 			uvm_pagefree(p);
803 			uvmexp.pdfreed++;
804 
805 			/*
806 			 * for anons, we need to remove the page
807 			 * from the anon ourselves.  for aobjs,
808 			 * pagefree did that for us.
809 			 */
810 
811 			if (anon) {
812 				KASSERT(anon->an_swslot != 0);
813 				anon->an_page = NULL;
814 				slot = anon->an_swslot;
815 			} else {
816 				slot = uao_find_swslot(uobj, pageidx);
817 			}
818 			mutex_exit(slock);
819 
820 			if (slot > 0) {
821 				/* this page is now only in swap. */
822 				mutex_enter(&uvm_swap_data_lock);
823 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
824 				uvmexp.swpgonly++;
825 				mutex_exit(&uvm_swap_data_lock);
826 			}
827 			continue;
828 		}
829 
830 #if defined(VMSWAP)
831 		/*
832 		 * this page is dirty, skip it if we'll have met our
833 		 * free target when all the current pageouts complete.
834 		 */
835 
836 		if (uvmexp.free + uvmexp.paging > uvmexp.freetarg << 2) {
837 			mutex_exit(slock);
838 			continue;
839 		}
840 
841 		/*
842 		 * free any swap space allocated to the page since
843 		 * we'll have to write it again with its new data.
844 		 */
845 
846 		uvmpd_dropswap(p);
847 
848 		/*
849 		 * if all pages in swap are only in swap,
850 		 * the swap space is full and we can't page out
851 		 * any more swap-backed pages.  reactivate this page
852 		 * so that we eventually cycle all pages through
853 		 * the inactive queue.
854 		 */
855 
856 		if (uvm_swapisfull()) {
857 			dirtyreacts++;
858 			uvm_pageactivate(p);
859 			mutex_exit(slock);
860 			continue;
861 		}
862 
863 		/*
864 		 * start new swap pageout cluster (if necessary).
865 		 */
866 
867 		if (swapcluster_allocslots(&swc)) {
868 			mutex_exit(slock);
869 			dirtyreacts++; /* XXX */
870 			continue;
871 		}
872 
873 		/*
874 		 * at this point, we're definitely going reuse this
875 		 * page.  mark the page busy and delayed-free.
876 		 * we should remove the page from the page queues
877 		 * so we don't ever look at it again.
878 		 * adjust counters and such.
879 		 */
880 
881 		p->flags |= PG_BUSY;
882 		UVM_PAGE_OWN(p, "scan_queue");
883 
884 		p->flags |= PG_PAGEOUT;
885 		uvm_pagedequeue(p);
886 
887 		uvmexp.pgswapout++;
888 		mutex_exit(&uvm_pageqlock);
889 
890 		/*
891 		 * add the new page to the cluster.
892 		 */
893 
894 		if (swapcluster_add(&swc, p)) {
895 			p->flags &= ~(PG_BUSY|PG_PAGEOUT);
896 			UVM_PAGE_OWN(p, NULL);
897 			mutex_enter(&uvm_pageqlock);
898 			dirtyreacts++;
899 			uvm_pageactivate(p);
900 			mutex_exit(slock);
901 			continue;
902 		}
903 		mutex_exit(slock);
904 
905 		swapcluster_flush(&swc, false);
906 		mutex_enter(&uvm_pageqlock);
907 
908 		/*
909 		 * the pageout is in progress.  bump counters and set up
910 		 * for the next loop.
911 		 */
912 
913 		uvmexp.pdpending++;
914 
915 #else /* defined(VMSWAP) */
916 		uvm_pageactivate(p);
917 		mutex_exit(slock);
918 #endif /* defined(VMSWAP) */
919 	}
920 
921 #if defined(VMSWAP)
922 	mutex_exit(&uvm_pageqlock);
923 	swapcluster_flush(&swc, true);
924 	mutex_enter(&uvm_pageqlock);
925 #endif /* defined(VMSWAP) */
926 }
927 
928 /*
929  * uvmpd_scan: scan the page queues and attempt to meet our targets.
930  *
931  * => called with pageq's locked
932  */
933 
934 static void
935 uvmpd_scan(void)
936 {
937 	int swap_shortage, pages_freed;
938 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
939 
940 	uvmexp.pdrevs++;
941 
942 	/*
943 	 * work on meeting our targets.   first we work on our free target
944 	 * by converting inactive pages into free pages.  then we work on
945 	 * meeting our inactive target by converting active pages to
946 	 * inactive ones.
947 	 */
948 
949 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
950 
951 	pages_freed = uvmexp.pdfreed;
952 	uvmpd_scan_queue();
953 	pages_freed = uvmexp.pdfreed - pages_freed;
954 
955 	/*
956 	 * detect if we're not going to be able to page anything out
957 	 * until we free some swap resources from active pages.
958 	 */
959 
960 	swap_shortage = 0;
961 	if (uvmexp.free < uvmexp.freetarg &&
962 	    uvmexp.swpginuse >= uvmexp.swpgavail &&
963 	    !uvm_swapisfull() &&
964 	    pages_freed == 0) {
965 		swap_shortage = uvmexp.freetarg - uvmexp.free;
966 	}
967 
968 	uvmpdpol_balancequeue(swap_shortage);
969 
970 	/*
971 	 * swap out some processes if we are still below the minimum
972 	 * free target.  we need to unlock the page queues for this.
973 	 */
974 
975 	if (uvmexp.free < uvmexp.freemin && uvmexp.nswapdev != 0 &&
976 	    uvm.swapout_enabled) {
977 		uvmexp.pdswout++;
978 		UVMHIST_LOG(pdhist,"  free %d < min %d: swapout",
979 		    uvmexp.free, uvmexp.freemin, 0, 0);
980 		mutex_exit(&uvm_pageqlock);
981 		uvm_swapout_threads();
982 		mutex_enter(&uvm_pageqlock);
983 
984 	}
985 }
986 
987 /*
988  * uvm_reclaimable: decide whether to wait for pagedaemon.
989  *
990  * => return true if it seems to be worth to do uvm_wait.
991  *
992  * XXX should be tunable.
993  * XXX should consider pools, etc?
994  */
995 
996 bool
997 uvm_reclaimable(void)
998 {
999 	int filepages;
1000 	int active, inactive;
1001 
1002 	/*
1003 	 * if swap is not full, no problem.
1004 	 */
1005 
1006 	if (!uvm_swapisfull()) {
1007 		return true;
1008 	}
1009 
1010 	/*
1011 	 * file-backed pages can be reclaimed even when swap is full.
1012 	 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim.
1013 	 *
1014 	 * XXX assume the worst case, ie. all wired pages are file-backed.
1015 	 *
1016 	 * XXX should consider about other reclaimable memory.
1017 	 * XXX ie. pools, traditional buffer cache.
1018 	 */
1019 
1020 	filepages = uvmexp.filepages + uvmexp.execpages - uvmexp.wired;
1021 	uvm_estimatepageable(&active, &inactive);
1022 	if (filepages >= MIN((active + inactive) >> 4,
1023 	    5 * 1024 * 1024 >> PAGE_SHIFT)) {
1024 		return true;
1025 	}
1026 
1027 	/*
1028 	 * kill the process, fail allocation, etc..
1029 	 */
1030 
1031 	return false;
1032 }
1033 
1034 void
1035 uvm_estimatepageable(int *active, int *inactive)
1036 {
1037 
1038 	uvmpdpol_estimatepageable(active, inactive);
1039 }
1040