xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision fff57c5525bbe431aee7bdb3983954f0627a42cb)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.92 2008/02/29 20:35:23 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 1.92 2008/02/29 20:35:23 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 
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 	uvmexp.freemin = uvmexp.npages / 20;
197 
198 	/* between 16k and 256k */
199 	/* XXX:  what are these values good for? */
200 	uvmexp.freemin = MAX(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
201 	uvmexp.freemin = MIN(uvmexp.freemin, (256*1024) >> PAGE_SHIFT);
202 
203 	/* Make sure there's always a user page free. */
204 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
205 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
206 
207 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
208 	if (uvmexp.freetarg <= uvmexp.freemin)
209 		uvmexp.freetarg = uvmexp.freemin + 1;
210 
211 	uvmexp.freetarg += uvm_extrapages;
212 	uvm_extrapages = 0;
213 
214 	uvmexp.wiredmax = uvmexp.npages / 3;
215 	UVMHIST_LOG(pdhist, "<- done, freemin=%d, freetarg=%d, wiredmax=%d",
216 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
217 }
218 
219 /*
220  * uvm_pageout: the main loop for the pagedaemon
221  */
222 
223 void
224 uvm_pageout(void *arg)
225 {
226 	int bufcnt, npages = 0;
227 	int extrapages = 0;
228 	struct pool *pp;
229 	uint64_t where;
230 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
231 
232 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
233 
234 	/*
235 	 * ensure correct priority and set paging parameters...
236 	 */
237 
238 	uvm.pagedaemon_lwp = curlwp;
239 	mutex_enter(&uvm_pageqlock);
240 	npages = uvmexp.npages;
241 	uvmpd_tune();
242 	mutex_exit(&uvm_pageqlock);
243 
244 	/*
245 	 * main loop
246 	 */
247 
248 	for (;;) {
249 		bool needsscan;
250 
251 		mutex_spin_enter(&uvm_fpageqlock);
252 		if (uvm_pagedaemon_waiters == 0 || uvmexp.paging > 0) {
253 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
254 			UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
255 			    &uvm_fpageqlock, false, "pgdaemon", 0);
256 			uvmexp.pdwoke++;
257 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
258 		} else {
259 			mutex_spin_exit(&uvm_fpageqlock);
260 		}
261 
262 		/*
263 		 * now lock page queues and recompute inactive count
264 		 */
265 
266 		mutex_enter(&uvm_pageqlock);
267 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
268 			npages = uvmexp.npages;
269 			extrapages = uvm_extrapages;
270 			mutex_spin_enter(&uvm_fpageqlock);
271 			uvmpd_tune();
272 			mutex_spin_exit(&uvm_fpageqlock);
273 		}
274 
275 		uvmpdpol_tune();
276 
277 		/*
278 		 * Estimate a hint.  Note that bufmem are returned to
279 		 * system only when entire pool page is empty.
280 		 */
281 		mutex_spin_enter(&uvm_fpageqlock);
282 		bufcnt = uvmexp.freetarg - uvmexp.free;
283 		if (bufcnt < 0)
284 			bufcnt = 0;
285 
286 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d",
287 		    uvmexp.free, uvmexp.freetarg, 0,0);
288 
289 		needsscan = uvmexp.free + uvmexp.paging < uvmexp.freetarg ||
290 		    uvmpdpol_needsscan_p();
291 		mutex_spin_exit(&uvm_fpageqlock);
292 
293 		/*
294 		 * scan if needed
295 		 */
296 		if (needsscan)
297 			uvmpd_scan();
298 
299 		/*
300 		 * if there's any free memory to be had,
301 		 * wake up any waiters.
302 		 */
303 
304 		mutex_spin_enter(&uvm_fpageqlock);
305 		if (uvmexp.free > uvmexp.reserve_kernel ||
306 		    uvmexp.paging == 0) {
307 			wakeup(&uvmexp.free);
308 			uvm_pagedaemon_waiters = 0;
309 		}
310 		mutex_spin_exit(&uvm_fpageqlock);
311 
312 		/*
313 		 * scan done.  unlock page queues (the only lock we are holding)
314 		 */
315 		mutex_exit(&uvm_pageqlock);
316 
317 		/*
318 		 * start draining pool resources now that we're not
319 		 * holding any locks.
320 		 */
321 		pool_drain_start(&pp, &where);
322 
323 		/*
324 		 * kill unused metadata buffers.
325 		 */
326 		mutex_enter(&bufcache_lock);
327 		buf_drain(bufcnt << PAGE_SHIFT);
328 		mutex_exit(&bufcache_lock);
329 
330 		/*
331 		 * complete draining the pools.
332 		 */
333 		pool_drain_end(pp, where);
334 	}
335 	/*NOTREACHED*/
336 }
337 
338 
339 /*
340  * uvm_aiodone_worker: a workqueue callback for the aiodone daemon.
341  */
342 
343 void
344 uvm_aiodone_worker(struct work *wk, void *dummy)
345 {
346 	struct buf *bp = (void *)wk;
347 
348 	KASSERT(&bp->b_work == wk);
349 
350 	/*
351 	 * process an i/o that's done.
352 	 */
353 
354 	(*bp->b_iodone)(bp);
355 }
356 
357 void
358 uvm_pageout_start(int npages)
359 {
360 
361 	mutex_spin_enter(&uvm_fpageqlock);
362 	uvmexp.paging += npages;
363 	mutex_spin_exit(&uvm_fpageqlock);
364 }
365 
366 void
367 uvm_pageout_done(int npages)
368 {
369 
370 	mutex_spin_enter(&uvm_fpageqlock);
371 	KASSERT(uvmexp.paging >= npages);
372 	uvmexp.paging -= npages;
373 
374 	/*
375 	 * wake up either of pagedaemon or LWPs waiting for it.
376 	 */
377 
378 	if (uvmexp.free <= uvmexp.reserve_kernel) {
379 		wakeup(&uvm.pagedaemon);
380 	} else {
381 		wakeup(&uvmexp.free);
382 		uvm_pagedaemon_waiters = 0;
383 	}
384 	mutex_spin_exit(&uvm_fpageqlock);
385 }
386 
387 /*
388  * uvmpd_trylockowner: trylock the page's owner.
389  *
390  * => called with pageq locked.
391  * => resolve orphaned O->A loaned page.
392  * => return the locked mutex on success.  otherwise, return NULL.
393  */
394 
395 kmutex_t *
396 uvmpd_trylockowner(struct vm_page *pg)
397 {
398 	struct uvm_object *uobj = pg->uobject;
399 	kmutex_t *slock;
400 
401 	KASSERT(mutex_owned(&uvm_pageqlock));
402 
403 	if (uobj != NULL) {
404 		slock = &uobj->vmobjlock;
405 	} else {
406 		struct vm_anon *anon = pg->uanon;
407 
408 		KASSERT(anon != NULL);
409 		slock = &anon->an_lock;
410 	}
411 
412 	if (!mutex_tryenter(slock)) {
413 		return NULL;
414 	}
415 
416 	if (uobj == NULL) {
417 
418 		/*
419 		 * set PQ_ANON if it isn't set already.
420 		 */
421 
422 		if ((pg->pqflags & PQ_ANON) == 0) {
423 			KASSERT(pg->loan_count > 0);
424 			pg->loan_count--;
425 			pg->pqflags |= PQ_ANON;
426 			/* anon now owns it */
427 		}
428 	}
429 
430 	return slock;
431 }
432 
433 #if defined(VMSWAP)
434 struct swapcluster {
435 	int swc_slot;
436 	int swc_nallocated;
437 	int swc_nused;
438 	struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)];
439 };
440 
441 static void
442 swapcluster_init(struct swapcluster *swc)
443 {
444 
445 	swc->swc_slot = 0;
446 	swc->swc_nused = 0;
447 }
448 
449 static int
450 swapcluster_allocslots(struct swapcluster *swc)
451 {
452 	int slot;
453 	int npages;
454 
455 	if (swc->swc_slot != 0) {
456 		return 0;
457 	}
458 
459 	/* Even with strange MAXPHYS, the shift
460 	   implicitly rounds down to a page. */
461 	npages = MAXPHYS >> PAGE_SHIFT;
462 	slot = uvm_swap_alloc(&npages, true);
463 	if (slot == 0) {
464 		return ENOMEM;
465 	}
466 	swc->swc_slot = slot;
467 	swc->swc_nallocated = npages;
468 	swc->swc_nused = 0;
469 
470 	return 0;
471 }
472 
473 static int
474 swapcluster_add(struct swapcluster *swc, struct vm_page *pg)
475 {
476 	int slot;
477 	struct uvm_object *uobj;
478 
479 	KASSERT(swc->swc_slot != 0);
480 	KASSERT(swc->swc_nused < swc->swc_nallocated);
481 	KASSERT((pg->pqflags & PQ_SWAPBACKED) != 0);
482 
483 	slot = swc->swc_slot + swc->swc_nused;
484 	uobj = pg->uobject;
485 	if (uobj == NULL) {
486 		KASSERT(mutex_owned(&pg->uanon->an_lock));
487 		pg->uanon->an_swslot = slot;
488 	} else {
489 		int result;
490 
491 		KASSERT(mutex_owned(&uobj->vmobjlock));
492 		result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot);
493 		if (result == -1) {
494 			return ENOMEM;
495 		}
496 	}
497 	swc->swc_pages[swc->swc_nused] = pg;
498 	swc->swc_nused++;
499 
500 	return 0;
501 }
502 
503 static void
504 swapcluster_flush(struct swapcluster *swc, bool now)
505 {
506 	int slot;
507 	int nused;
508 	int nallocated;
509 	int error;
510 
511 	if (swc->swc_slot == 0) {
512 		return;
513 	}
514 	KASSERT(swc->swc_nused <= swc->swc_nallocated);
515 
516 	slot = swc->swc_slot;
517 	nused = swc->swc_nused;
518 	nallocated = swc->swc_nallocated;
519 
520 	/*
521 	 * if this is the final pageout we could have a few
522 	 * unused swap blocks.  if so, free them now.
523 	 */
524 
525 	if (nused < nallocated) {
526 		if (!now) {
527 			return;
528 		}
529 		uvm_swap_free(slot + nused, nallocated - nused);
530 	}
531 
532 	/*
533 	 * now start the pageout.
534 	 */
535 
536 	if (nused > 0) {
537 		uvmexp.pdpageouts++;
538 		uvm_pageout_start(nused);
539 		error = uvm_swap_put(slot, swc->swc_pages, nused, 0);
540 		KASSERT(error == 0 || error == ENOMEM);
541 	}
542 
543 	/*
544 	 * zero swslot to indicate that we are
545 	 * no longer building a swap-backed cluster.
546 	 */
547 
548 	swc->swc_slot = 0;
549 	swc->swc_nused = 0;
550 }
551 
552 static int
553 swapcluster_nused(struct swapcluster *swc)
554 {
555 
556 	return swc->swc_nused;
557 }
558 
559 /*
560  * uvmpd_dropswap: free any swap allocated to this page.
561  *
562  * => called with owner locked.
563  * => return true if a page had an associated slot.
564  */
565 
566 static bool
567 uvmpd_dropswap(struct vm_page *pg)
568 {
569 	bool result = false;
570 	struct vm_anon *anon = pg->uanon;
571 
572 	if ((pg->pqflags & PQ_ANON) && anon->an_swslot) {
573 		uvm_swap_free(anon->an_swslot, 1);
574 		anon->an_swslot = 0;
575 		pg->flags &= ~PG_CLEAN;
576 		result = true;
577 	} else if (pg->pqflags & PQ_AOBJ) {
578 		int slot = uao_set_swslot(pg->uobject,
579 		    pg->offset >> PAGE_SHIFT, 0);
580 		if (slot) {
581 			uvm_swap_free(slot, 1);
582 			pg->flags &= ~PG_CLEAN;
583 			result = true;
584 		}
585 	}
586 
587 	return result;
588 }
589 
590 /*
591  * uvmpd_trydropswap: try to free any swap allocated to this page.
592  *
593  * => return true if a slot is successfully freed.
594  */
595 
596 bool
597 uvmpd_trydropswap(struct vm_page *pg)
598 {
599 	kmutex_t *slock;
600 	bool result;
601 
602 	if ((pg->flags & PG_BUSY) != 0) {
603 		return false;
604 	}
605 
606 	/*
607 	 * lock the page's owner.
608 	 */
609 
610 	slock = uvmpd_trylockowner(pg);
611 	if (slock == NULL) {
612 		return false;
613 	}
614 
615 	/*
616 	 * skip this page if it's busy.
617 	 */
618 
619 	if ((pg->flags & PG_BUSY) != 0) {
620 		mutex_exit(slock);
621 		return false;
622 	}
623 
624 	result = uvmpd_dropswap(pg);
625 
626 	mutex_exit(slock);
627 
628 	return result;
629 }
630 
631 #endif /* defined(VMSWAP) */
632 
633 /*
634  * uvmpd_scan_queue: scan an replace candidate list for pages
635  * to clean or free.
636  *
637  * => called with page queues locked
638  * => we work on meeting our free target by converting inactive pages
639  *    into free pages.
640  * => we handle the building of swap-backed clusters
641  */
642 
643 static void
644 uvmpd_scan_queue(void)
645 {
646 	struct vm_page *p;
647 	struct uvm_object *uobj;
648 	struct vm_anon *anon;
649 #if defined(VMSWAP)
650 	struct swapcluster swc;
651 #endif /* defined(VMSWAP) */
652 	int dirtyreacts;
653 	int lockownerfail;
654 	kmutex_t *slock;
655 	UVMHIST_FUNC("uvmpd_scan_queue"); UVMHIST_CALLED(pdhist);
656 
657 	/*
658 	 * swslot is non-zero if we are building a swap cluster.  we want
659 	 * to stay in the loop while we have a page to scan or we have
660 	 * a swap-cluster to build.
661 	 */
662 
663 #if defined(VMSWAP)
664 	swapcluster_init(&swc);
665 #endif /* defined(VMSWAP) */
666 
667 	dirtyreacts = 0;
668 	lockownerfail = 0;
669 	uvmpdpol_scaninit();
670 
671 	while (/* CONSTCOND */ 1) {
672 
673 		/*
674 		 * see if we've met the free target.
675 		 */
676 
677 		if (uvmexp.free + uvmexp.paging
678 #if defined(VMSWAP)
679 		    + swapcluster_nused(&swc)
680 #endif /* defined(VMSWAP) */
681 		    >= uvmexp.freetarg << 2 ||
682 		    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
683 			UVMHIST_LOG(pdhist,"  met free target: "
684 				    "exit loop", 0, 0, 0, 0);
685 			break;
686 		}
687 
688 		p = uvmpdpol_selectvictim();
689 		if (p == NULL) {
690 			break;
691 		}
692 		KASSERT(uvmpdpol_pageisqueued_p(p));
693 		KASSERT(p->wire_count == 0);
694 
695 		/*
696 		 * we are below target and have a new page to consider.
697 		 */
698 
699 		anon = p->uanon;
700 		uobj = p->uobject;
701 
702 		/*
703 		 * first we attempt to lock the object that this page
704 		 * belongs to.  if our attempt fails we skip on to
705 		 * the next page (no harm done).  it is important to
706 		 * "try" locking the object as we are locking in the
707 		 * wrong order (pageq -> object) and we don't want to
708 		 * deadlock.
709 		 *
710 		 * the only time we expect to see an ownerless page
711 		 * (i.e. a page with no uobject and !PQ_ANON) is if an
712 		 * anon has loaned a page from a uvm_object and the
713 		 * uvm_object has dropped the ownership.  in that
714 		 * case, the anon can "take over" the loaned page
715 		 * and make it its own.
716 		 */
717 
718 		slock = uvmpd_trylockowner(p);
719 		if (slock == NULL) {
720 			/*
721 			 * yield cpu to make a chance for an LWP holding
722 			 * the lock run.  otherwise we can busy-loop too long
723 			 * if the page queue is filled with a lot of pages
724 			 * from few objects.
725 			 */
726 			lockownerfail++;
727 			if (lockownerfail > UVMPD_NUMTRYLOCKOWNER) {
728 				mutex_exit(&uvm_pageqlock);
729 				/* XXX Better than yielding but inadequate. */
730 				kpause("livelock", false, 1, NULL);
731 				mutex_enter(&uvm_pageqlock);
732 				lockownerfail = 0;
733 			}
734 			continue;
735 		}
736 		if (p->flags & PG_BUSY) {
737 			mutex_exit(slock);
738 			uvmexp.pdbusy++;
739 			continue;
740 		}
741 
742 		/* does the page belong to an object? */
743 		if (uobj != NULL) {
744 			uvmexp.pdobscan++;
745 		} else {
746 #if defined(VMSWAP)
747 			KASSERT(anon != NULL);
748 			uvmexp.pdanscan++;
749 #else /* defined(VMSWAP) */
750 			panic("%s: anon", __func__);
751 #endif /* defined(VMSWAP) */
752 		}
753 
754 
755 		/*
756 		 * we now have the object and the page queues locked.
757 		 * if the page is not swap-backed, call the object's
758 		 * pager to flush and free the page.
759 		 */
760 
761 #if defined(READAHEAD_STATS)
762 		if ((p->pqflags & PQ_READAHEAD) != 0) {
763 			p->pqflags &= ~PQ_READAHEAD;
764 			uvm_ra_miss.ev_count++;
765 		}
766 #endif /* defined(READAHEAD_STATS) */
767 
768 		if ((p->pqflags & PQ_SWAPBACKED) == 0) {
769 			KASSERT(uobj != NULL);
770 			mutex_exit(&uvm_pageqlock);
771 			(void) (uobj->pgops->pgo_put)(uobj, p->offset,
772 			    p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE);
773 			mutex_enter(&uvm_pageqlock);
774 			continue;
775 		}
776 
777 		/*
778 		 * the page is swap-backed.  remove all the permissions
779 		 * from the page so we can sync the modified info
780 		 * without any race conditions.  if the page is clean
781 		 * we can free it now and continue.
782 		 */
783 
784 		pmap_page_protect(p, VM_PROT_NONE);
785 		if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
786 			p->flags &= ~(PG_CLEAN);
787 		}
788 		if (p->flags & PG_CLEAN) {
789 			int slot;
790 			int pageidx;
791 
792 			pageidx = p->offset >> PAGE_SHIFT;
793 			uvm_pagefree(p);
794 			uvmexp.pdfreed++;
795 
796 			/*
797 			 * for anons, we need to remove the page
798 			 * from the anon ourselves.  for aobjs,
799 			 * pagefree did that for us.
800 			 */
801 
802 			if (anon) {
803 				KASSERT(anon->an_swslot != 0);
804 				anon->an_page = NULL;
805 				slot = anon->an_swslot;
806 			} else {
807 				slot = uao_find_swslot(uobj, pageidx);
808 			}
809 			mutex_exit(slock);
810 
811 			if (slot > 0) {
812 				/* this page is now only in swap. */
813 				mutex_enter(&uvm_swap_data_lock);
814 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
815 				uvmexp.swpgonly++;
816 				mutex_exit(&uvm_swap_data_lock);
817 			}
818 			continue;
819 		}
820 
821 #if defined(VMSWAP)
822 		/*
823 		 * this page is dirty, skip it if we'll have met our
824 		 * free target when all the current pageouts complete.
825 		 */
826 
827 		if (uvmexp.free + uvmexp.paging > uvmexp.freetarg << 2) {
828 			mutex_exit(slock);
829 			continue;
830 		}
831 
832 		/*
833 		 * free any swap space allocated to the page since
834 		 * we'll have to write it again with its new data.
835 		 */
836 
837 		uvmpd_dropswap(p);
838 
839 		/*
840 		 * if all pages in swap are only in swap,
841 		 * the swap space is full and we can't page out
842 		 * any more swap-backed pages.  reactivate this page
843 		 * so that we eventually cycle all pages through
844 		 * the inactive queue.
845 		 */
846 
847 		if (uvm_swapisfull()) {
848 			dirtyreacts++;
849 			uvm_pageactivate(p);
850 			mutex_exit(slock);
851 			continue;
852 		}
853 
854 		/*
855 		 * start new swap pageout cluster (if necessary).
856 		 */
857 
858 		if (swapcluster_allocslots(&swc)) {
859 			mutex_exit(slock);
860 			dirtyreacts++; /* XXX */
861 			continue;
862 		}
863 
864 		/*
865 		 * at this point, we're definitely going reuse this
866 		 * page.  mark the page busy and delayed-free.
867 		 * we should remove the page from the page queues
868 		 * so we don't ever look at it again.
869 		 * adjust counters and such.
870 		 */
871 
872 		p->flags |= PG_BUSY;
873 		UVM_PAGE_OWN(p, "scan_queue");
874 
875 		p->flags |= PG_PAGEOUT;
876 		uvm_pagedequeue(p);
877 
878 		uvmexp.pgswapout++;
879 		mutex_exit(&uvm_pageqlock);
880 
881 		/*
882 		 * add the new page to the cluster.
883 		 */
884 
885 		if (swapcluster_add(&swc, p)) {
886 			p->flags &= ~(PG_BUSY|PG_PAGEOUT);
887 			UVM_PAGE_OWN(p, NULL);
888 			mutex_enter(&uvm_pageqlock);
889 			dirtyreacts++;
890 			uvm_pageactivate(p);
891 			mutex_exit(slock);
892 			continue;
893 		}
894 		mutex_exit(slock);
895 
896 		swapcluster_flush(&swc, false);
897 		mutex_enter(&uvm_pageqlock);
898 
899 		/*
900 		 * the pageout is in progress.  bump counters and set up
901 		 * for the next loop.
902 		 */
903 
904 		uvmexp.pdpending++;
905 
906 #else /* defined(VMSWAP) */
907 		uvm_pageactivate(p);
908 		mutex_exit(slock);
909 #endif /* defined(VMSWAP) */
910 	}
911 
912 #if defined(VMSWAP)
913 	mutex_exit(&uvm_pageqlock);
914 	swapcluster_flush(&swc, true);
915 	mutex_enter(&uvm_pageqlock);
916 #endif /* defined(VMSWAP) */
917 }
918 
919 /*
920  * uvmpd_scan: scan the page queues and attempt to meet our targets.
921  *
922  * => called with pageq's locked
923  */
924 
925 static void
926 uvmpd_scan(void)
927 {
928 	int swap_shortage, pages_freed;
929 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
930 
931 	uvmexp.pdrevs++;
932 
933 #ifndef __SWAP_BROKEN
934 
935 	/*
936 	 * swap out some processes if we are below our free target.
937 	 * we need to unlock the page queues for this.
938 	 */
939 
940 	if (uvmexp.free < uvmexp.freetarg && uvmexp.nswapdev != 0 &&
941 	    uvm.swapout_enabled) {
942 		uvmexp.pdswout++;
943 		UVMHIST_LOG(pdhist,"  free %d < target %d: swapout",
944 		    uvmexp.free, uvmexp.freetarg, 0, 0);
945 		mutex_exit(&uvm_pageqlock);
946 		uvm_swapout_threads();
947 		mutex_enter(&uvm_pageqlock);
948 
949 	}
950 #endif
951 
952 	/*
953 	 * now we want to work on meeting our targets.   first we work on our
954 	 * free target by converting inactive pages into free pages.  then
955 	 * we work on meeting our inactive target by converting active pages
956 	 * to inactive ones.
957 	 */
958 
959 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
960 
961 	pages_freed = uvmexp.pdfreed;
962 	uvmpd_scan_queue();
963 	pages_freed = uvmexp.pdfreed - pages_freed;
964 
965 	/*
966 	 * detect if we're not going to be able to page anything out
967 	 * until we free some swap resources from active pages.
968 	 */
969 
970 	swap_shortage = 0;
971 	if (uvmexp.free < uvmexp.freetarg &&
972 	    uvmexp.swpginuse >= uvmexp.swpgavail &&
973 	    !uvm_swapisfull() &&
974 	    pages_freed == 0) {
975 		swap_shortage = uvmexp.freetarg - uvmexp.free;
976 	}
977 
978 	uvmpdpol_balancequeue(swap_shortage);
979 }
980 
981 /*
982  * uvm_reclaimable: decide whether to wait for pagedaemon.
983  *
984  * => return true if it seems to be worth to do uvm_wait.
985  *
986  * XXX should be tunable.
987  * XXX should consider pools, etc?
988  */
989 
990 bool
991 uvm_reclaimable(void)
992 {
993 	int filepages;
994 	int active, inactive;
995 
996 	/*
997 	 * if swap is not full, no problem.
998 	 */
999 
1000 	if (!uvm_swapisfull()) {
1001 		return true;
1002 	}
1003 
1004 	/*
1005 	 * file-backed pages can be reclaimed even when swap is full.
1006 	 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim.
1007 	 *
1008 	 * XXX assume the worst case, ie. all wired pages are file-backed.
1009 	 *
1010 	 * XXX should consider about other reclaimable memory.
1011 	 * XXX ie. pools, traditional buffer cache.
1012 	 */
1013 
1014 	filepages = uvmexp.filepages + uvmexp.execpages - uvmexp.wired;
1015 	uvm_estimatepageable(&active, &inactive);
1016 	if (filepages >= MIN((active + inactive) >> 4,
1017 	    5 * 1024 * 1024 >> PAGE_SHIFT)) {
1018 		return true;
1019 	}
1020 
1021 	/*
1022 	 * kill the process, fail allocation, etc..
1023 	 */
1024 
1025 	return false;
1026 }
1027 
1028 void
1029 uvm_estimatepageable(int *active, int *inactive)
1030 {
1031 
1032 	uvmpdpol_estimatepageable(active, inactive);
1033 }
1034