xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 062cd249892f31a10349de47b78d800cd8ae2c25)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.106 2012/06/05 22:51:47 jym 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. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)vm_pageout.c        8.5 (Berkeley) 2/14/94
37  * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp
38  *
39  *
40  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
41  * All rights reserved.
42  *
43  * Permission to use, copy, modify and distribute this software and
44  * its documentation is hereby granted, provided that both the copyright
45  * notice and this permission notice appear in all copies of the
46  * software, derivative works or modified versions, and any portions
47  * thereof, and that both notices appear in supporting documentation.
48  *
49  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
50  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
51  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52  *
53  * Carnegie Mellon requests users of this software to return to
54  *
55  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
56  *  School of Computer Science
57  *  Carnegie Mellon University
58  *  Pittsburgh PA 15213-3890
59  *
60  * any improvements or extensions that they make and grant Carnegie the
61  * rights to redistribute these changes.
62  */
63 
64 /*
65  * uvm_pdaemon.c: the page daemon
66  */
67 
68 #include <sys/cdefs.h>
69 __KERNEL_RCSID(0, "$NetBSD: uvm_pdaemon.c,v 1.106 2012/06/05 22:51:47 jym Exp $");
70 
71 #include "opt_uvmhist.h"
72 #include "opt_readahead.h"
73 
74 #include <sys/param.h>
75 #include <sys/proc.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/pool.h>
79 #include <sys/buf.h>
80 #include <sys/module.h>
81 #include <sys/atomic.h>
82 
83 #include <uvm/uvm.h>
84 #include <uvm/uvm_pdpolicy.h>
85 
86 /*
87  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
88  * in a pass thru the inactive list when swap is full.  the value should be
89  * "small"... if it's too large we'll cycle the active pages thru the inactive
90  * queue too quickly to for them to be referenced and avoid being freed.
91  */
92 
93 #define	UVMPD_NUMDIRTYREACTS	16
94 
95 #define	UVMPD_NUMTRYLOCKOWNER	16
96 
97 /*
98  * local prototypes
99  */
100 
101 static void	uvmpd_scan(void);
102 static void	uvmpd_scan_queue(void);
103 static void	uvmpd_tune(void);
104 
105 static unsigned int uvm_pagedaemon_waiters;
106 
107 /*
108  * XXX hack to avoid hangs when large processes fork.
109  */
110 u_int uvm_extrapages;
111 
112 /*
113  * uvm_wait: wait (sleep) for the page daemon to free some pages
114  *
115  * => should be called with all locks released
116  * => should _not_ be called by the page daemon (to avoid deadlock)
117  */
118 
119 void
120 uvm_wait(const char *wmsg)
121 {
122 	int timo = 0;
123 
124 	mutex_spin_enter(&uvm_fpageqlock);
125 
126 	/*
127 	 * check for page daemon going to sleep (waiting for itself)
128 	 */
129 
130 	if (curlwp == uvm.pagedaemon_lwp && 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 	uvm_pagedaemon_waiters++;
157 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
158 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm_fpageqlock, false, wmsg, timo);
159 }
160 
161 /*
162  * uvm_kick_pdaemon: perform checks to determine if we need to
163  * give the pagedaemon a nudge, and do so if necessary.
164  *
165  * => called with uvm_fpageqlock held.
166  */
167 
168 void
169 uvm_kick_pdaemon(void)
170 {
171 
172 	KASSERT(mutex_owned(&uvm_fpageqlock));
173 
174 	if (uvmexp.free + uvmexp.paging < uvmexp.freemin ||
175 	    (uvmexp.free + uvmexp.paging < uvmexp.freetarg &&
176 	     uvmpdpol_needsscan_p()) ||
177 	     uvm_km_va_starved_p()) {
178 		wakeup(&uvm.pagedaemon);
179 	}
180 }
181 
182 /*
183  * uvmpd_tune: tune paging parameters
184  *
185  * => called when ever memory is added (or removed?) to the system
186  * => caller must call with page queues locked
187  */
188 
189 static void
190 uvmpd_tune(void)
191 {
192 	int val;
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 	val = uvmexp.npages / 200;
201 	val = MAX(val, (128*1024) >> PAGE_SHIFT);
202 	val = MIN(val, (1024*1024) >> PAGE_SHIFT);
203 	val *= ncpu;
204 
205 	/* Make sure there's always a user page free. */
206 	if (val < uvmexp.reserve_kernel + 1)
207 		val = uvmexp.reserve_kernel + 1;
208 	uvmexp.freemin = val;
209 
210 	/* Calculate free target. */
211 	val = (uvmexp.freemin * 4) / 3;
212 	if (val <= uvmexp.freemin)
213 		val = uvmexp.freemin + 1;
214 	uvmexp.freetarg = val + atomic_swap_uint(&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 
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, kmem_va_starved;
252 
253 		kmem_va_starved = uvm_km_va_starved_p();
254 
255 		mutex_spin_enter(&uvm_fpageqlock);
256 		if ((uvm_pagedaemon_waiters == 0 || uvmexp.paging > 0) &&
257 		    !kmem_va_starved) {
258 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
259 			UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
260 			    &uvm_fpageqlock, false, "pgdaemon", 0);
261 			uvmexp.pdwoke++;
262 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
263 		} else {
264 			mutex_spin_exit(&uvm_fpageqlock);
265 		}
266 
267 		/*
268 		 * now lock page queues and recompute inactive count
269 		 */
270 
271 		mutex_enter(&uvm_pageqlock);
272 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
273 			npages = uvmexp.npages;
274 			extrapages = uvm_extrapages;
275 			mutex_spin_enter(&uvm_fpageqlock);
276 			uvmpd_tune();
277 			mutex_spin_exit(&uvm_fpageqlock);
278 		}
279 
280 		uvmpdpol_tune();
281 
282 		/*
283 		 * Estimate a hint.  Note that bufmem are returned to
284 		 * system only when entire pool page is empty.
285 		 */
286 		mutex_spin_enter(&uvm_fpageqlock);
287 		bufcnt = uvmexp.freetarg - uvmexp.free;
288 		if (bufcnt < 0)
289 			bufcnt = 0;
290 
291 		UVMHIST_LOG(pdhist,"  free/ftarg=%d/%d",
292 		    uvmexp.free, uvmexp.freetarg, 0,0);
293 
294 		needsfree = uvmexp.free + uvmexp.paging < uvmexp.freetarg;
295 		needsscan = needsfree || uvmpdpol_needsscan_p();
296 
297 		/*
298 		 * scan if needed
299 		 */
300 		if (needsscan) {
301 			mutex_spin_exit(&uvm_fpageqlock);
302 			uvmpd_scan();
303 			mutex_spin_enter(&uvm_fpageqlock);
304 		}
305 
306 		/*
307 		 * if there's any free memory to be had,
308 		 * wake up any waiters.
309 		 */
310 		if (uvmexp.free > uvmexp.reserve_kernel ||
311 		    uvmexp.paging == 0) {
312 			wakeup(&uvmexp.free);
313 			uvm_pagedaemon_waiters = 0;
314 		}
315 		mutex_spin_exit(&uvm_fpageqlock);
316 
317 		/*
318 		 * scan done.  unlock page queues (the only lock we are holding)
319 		 */
320 		mutex_exit(&uvm_pageqlock);
321 
322 		/*
323 		 * if we don't need free memory, we're done.
324 		 */
325 
326 		if (!needsfree && !kmem_va_starved)
327 			continue;
328 
329 		/*
330 		 * kill unused metadata buffers.
331 		 */
332 		mutex_enter(&bufcache_lock);
333 		buf_drain(bufcnt << PAGE_SHIFT);
334 		mutex_exit(&bufcache_lock);
335 
336 		/*
337 		 * drain the pools.
338 		 */
339 		pool_drain(&pp);
340 	}
341 	/*NOTREACHED*/
342 }
343 
344 
345 /*
346  * uvm_aiodone_worker: a workqueue callback for the aiodone daemon.
347  */
348 
349 void
350 uvm_aiodone_worker(struct work *wk, void *dummy)
351 {
352 	struct buf *bp = (void *)wk;
353 
354 	KASSERT(&bp->b_work == wk);
355 
356 	/*
357 	 * process an i/o that's done.
358 	 */
359 
360 	(*bp->b_iodone)(bp);
361 }
362 
363 void
364 uvm_pageout_start(int npages)
365 {
366 
367 	mutex_spin_enter(&uvm_fpageqlock);
368 	uvmexp.paging += npages;
369 	mutex_spin_exit(&uvm_fpageqlock);
370 }
371 
372 void
373 uvm_pageout_done(int npages)
374 {
375 
376 	mutex_spin_enter(&uvm_fpageqlock);
377 	KASSERT(uvmexp.paging >= npages);
378 	uvmexp.paging -= npages;
379 
380 	/*
381 	 * wake up either of pagedaemon or LWPs waiting for it.
382 	 */
383 
384 	if (uvmexp.free <= uvmexp.reserve_kernel) {
385 		wakeup(&uvm.pagedaemon);
386 	} else {
387 		wakeup(&uvmexp.free);
388 		uvm_pagedaemon_waiters = 0;
389 	}
390 	mutex_spin_exit(&uvm_fpageqlock);
391 }
392 
393 /*
394  * uvmpd_trylockowner: trylock the page's owner.
395  *
396  * => called with pageq locked.
397  * => resolve orphaned O->A loaned page.
398  * => return the locked mutex on success.  otherwise, return NULL.
399  */
400 
401 kmutex_t *
402 uvmpd_trylockowner(struct vm_page *pg)
403 {
404 	struct uvm_object *uobj = pg->uobject;
405 	kmutex_t *slock;
406 
407 	KASSERT(mutex_owned(&uvm_pageqlock));
408 
409 	if (uobj != NULL) {
410 		slock = uobj->vmobjlock;
411 	} else {
412 		struct vm_anon *anon = pg->uanon;
413 
414 		KASSERT(anon != NULL);
415 		slock = anon->an_lock;
416 	}
417 
418 	if (!mutex_tryenter(slock)) {
419 		return NULL;
420 	}
421 
422 	if (uobj == NULL) {
423 
424 		/*
425 		 * set PQ_ANON if it isn't set already.
426 		 */
427 
428 		if ((pg->pqflags & PQ_ANON) == 0) {
429 			KASSERT(pg->loan_count > 0);
430 			pg->loan_count--;
431 			pg->pqflags |= PQ_ANON;
432 			/* anon now owns it */
433 		}
434 	}
435 
436 	return slock;
437 }
438 
439 #if defined(VMSWAP)
440 struct swapcluster {
441 	int swc_slot;
442 	int swc_nallocated;
443 	int swc_nused;
444 	struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)];
445 };
446 
447 static void
448 swapcluster_init(struct swapcluster *swc)
449 {
450 
451 	swc->swc_slot = 0;
452 	swc->swc_nused = 0;
453 }
454 
455 static int
456 swapcluster_allocslots(struct swapcluster *swc)
457 {
458 	int slot;
459 	int npages;
460 
461 	if (swc->swc_slot != 0) {
462 		return 0;
463 	}
464 
465 	/* Even with strange MAXPHYS, the shift
466 	   implicitly rounds down to a page. */
467 	npages = MAXPHYS >> PAGE_SHIFT;
468 	slot = uvm_swap_alloc(&npages, true);
469 	if (slot == 0) {
470 		return ENOMEM;
471 	}
472 	swc->swc_slot = slot;
473 	swc->swc_nallocated = npages;
474 	swc->swc_nused = 0;
475 
476 	return 0;
477 }
478 
479 static int
480 swapcluster_add(struct swapcluster *swc, struct vm_page *pg)
481 {
482 	int slot;
483 	struct uvm_object *uobj;
484 
485 	KASSERT(swc->swc_slot != 0);
486 	KASSERT(swc->swc_nused < swc->swc_nallocated);
487 	KASSERT((pg->pqflags & PQ_SWAPBACKED) != 0);
488 
489 	slot = swc->swc_slot + swc->swc_nused;
490 	uobj = pg->uobject;
491 	if (uobj == NULL) {
492 		KASSERT(mutex_owned(pg->uanon->an_lock));
493 		pg->uanon->an_swslot = slot;
494 	} else {
495 		int result;
496 
497 		KASSERT(mutex_owned(uobj->vmobjlock));
498 		result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot);
499 		if (result == -1) {
500 			return ENOMEM;
501 		}
502 	}
503 	swc->swc_pages[swc->swc_nused] = pg;
504 	swc->swc_nused++;
505 
506 	return 0;
507 }
508 
509 static void
510 swapcluster_flush(struct swapcluster *swc, bool now)
511 {
512 	int slot;
513 	int nused;
514 	int nallocated;
515 	int error;
516 
517 	if (swc->swc_slot == 0) {
518 		return;
519 	}
520 	KASSERT(swc->swc_nused <= swc->swc_nallocated);
521 
522 	slot = swc->swc_slot;
523 	nused = swc->swc_nused;
524 	nallocated = swc->swc_nallocated;
525 
526 	/*
527 	 * if this is the final pageout we could have a few
528 	 * unused swap blocks.  if so, free them now.
529 	 */
530 
531 	if (nused < nallocated) {
532 		if (!now) {
533 			return;
534 		}
535 		uvm_swap_free(slot + nused, nallocated - nused);
536 	}
537 
538 	/*
539 	 * now start the pageout.
540 	 */
541 
542 	if (nused > 0) {
543 		uvmexp.pdpageouts++;
544 		uvm_pageout_start(nused);
545 		error = uvm_swap_put(slot, swc->swc_pages, nused, 0);
546 		KASSERT(error == 0 || error == ENOMEM);
547 	}
548 
549 	/*
550 	 * zero swslot to indicate that we are
551 	 * no longer building a swap-backed cluster.
552 	 */
553 
554 	swc->swc_slot = 0;
555 	swc->swc_nused = 0;
556 }
557 
558 static int
559 swapcluster_nused(struct swapcluster *swc)
560 {
561 
562 	return swc->swc_nused;
563 }
564 
565 /*
566  * uvmpd_dropswap: free any swap allocated to this page.
567  *
568  * => called with owner locked.
569  * => return true if a page had an associated slot.
570  */
571 
572 static bool
573 uvmpd_dropswap(struct vm_page *pg)
574 {
575 	bool result = false;
576 	struct vm_anon *anon = pg->uanon;
577 
578 	if ((pg->pqflags & PQ_ANON) && anon->an_swslot) {
579 		uvm_swap_free(anon->an_swslot, 1);
580 		anon->an_swslot = 0;
581 		pg->flags &= ~PG_CLEAN;
582 		result = true;
583 	} else if (pg->pqflags & PQ_AOBJ) {
584 		int slot = uao_set_swslot(pg->uobject,
585 		    pg->offset >> PAGE_SHIFT, 0);
586 		if (slot) {
587 			uvm_swap_free(slot, 1);
588 			pg->flags &= ~PG_CLEAN;
589 			result = true;
590 		}
591 	}
592 
593 	return result;
594 }
595 
596 /*
597  * uvmpd_trydropswap: try to free any swap allocated to this page.
598  *
599  * => return true if a slot is successfully freed.
600  */
601 
602 bool
603 uvmpd_trydropswap(struct vm_page *pg)
604 {
605 	kmutex_t *slock;
606 	bool result;
607 
608 	if ((pg->flags & PG_BUSY) != 0) {
609 		return false;
610 	}
611 
612 	/*
613 	 * lock the page's owner.
614 	 */
615 
616 	slock = uvmpd_trylockowner(pg);
617 	if (slock == NULL) {
618 		return false;
619 	}
620 
621 	/*
622 	 * skip this page if it's busy.
623 	 */
624 
625 	if ((pg->flags & PG_BUSY) != 0) {
626 		mutex_exit(slock);
627 		return false;
628 	}
629 
630 	result = uvmpd_dropswap(pg);
631 
632 	mutex_exit(slock);
633 
634 	return result;
635 }
636 
637 #endif /* defined(VMSWAP) */
638 
639 /*
640  * uvmpd_scan_queue: scan an replace candidate list for pages
641  * to clean or free.
642  *
643  * => called with page queues locked
644  * => we work on meeting our free target by converting inactive pages
645  *    into free pages.
646  * => we handle the building of swap-backed clusters
647  */
648 
649 static void
650 uvmpd_scan_queue(void)
651 {
652 	struct vm_page *p;
653 	struct uvm_object *uobj;
654 	struct vm_anon *anon;
655 #if defined(VMSWAP)
656 	struct swapcluster swc;
657 #endif /* defined(VMSWAP) */
658 	int dirtyreacts;
659 	int lockownerfail;
660 	kmutex_t *slock;
661 	UVMHIST_FUNC("uvmpd_scan_queue"); UVMHIST_CALLED(pdhist);
662 
663 	/*
664 	 * swslot is non-zero if we are building a swap cluster.  we want
665 	 * to stay in the loop while we have a page to scan or we have
666 	 * a swap-cluster to build.
667 	 */
668 
669 #if defined(VMSWAP)
670 	swapcluster_init(&swc);
671 #endif /* defined(VMSWAP) */
672 
673 	dirtyreacts = 0;
674 	lockownerfail = 0;
675 	uvmpdpol_scaninit();
676 
677 	while (/* CONSTCOND */ 1) {
678 
679 		/*
680 		 * see if we've met the free target.
681 		 */
682 
683 		if (uvmexp.free + uvmexp.paging
684 #if defined(VMSWAP)
685 		    + swapcluster_nused(&swc)
686 #endif /* defined(VMSWAP) */
687 		    >= uvmexp.freetarg << 2 ||
688 		    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
689 			UVMHIST_LOG(pdhist,"  met free target: "
690 				    "exit loop", 0, 0, 0, 0);
691 			break;
692 		}
693 
694 		p = uvmpdpol_selectvictim();
695 		if (p == NULL) {
696 			break;
697 		}
698 		KASSERT(uvmpdpol_pageisqueued_p(p));
699 		KASSERT(p->wire_count == 0);
700 
701 		/*
702 		 * we are below target and have a new page to consider.
703 		 */
704 
705 		anon = p->uanon;
706 		uobj = p->uobject;
707 
708 		/*
709 		 * first we attempt to lock the object that this page
710 		 * belongs to.  if our attempt fails we skip on to
711 		 * the next page (no harm done).  it is important to
712 		 * "try" locking the object as we are locking in the
713 		 * wrong order (pageq -> object) and we don't want to
714 		 * deadlock.
715 		 *
716 		 * the only time we expect to see an ownerless page
717 		 * (i.e. a page with no uobject and !PQ_ANON) is if an
718 		 * anon has loaned a page from a uvm_object and the
719 		 * uvm_object has dropped the ownership.  in that
720 		 * case, the anon can "take over" the loaned page
721 		 * and make it its own.
722 		 */
723 
724 		slock = uvmpd_trylockowner(p);
725 		if (slock == NULL) {
726 			/*
727 			 * yield cpu to make a chance for an LWP holding
728 			 * the lock run.  otherwise we can busy-loop too long
729 			 * if the page queue is filled with a lot of pages
730 			 * from few objects.
731 			 */
732 			lockownerfail++;
733 			if (lockownerfail > UVMPD_NUMTRYLOCKOWNER) {
734 				mutex_exit(&uvm_pageqlock);
735 				/* XXX Better than yielding but inadequate. */
736 				kpause("livelock", false, 1, NULL);
737 				mutex_enter(&uvm_pageqlock);
738 				lockownerfail = 0;
739 			}
740 			continue;
741 		}
742 		if (p->flags & PG_BUSY) {
743 			mutex_exit(slock);
744 			uvmexp.pdbusy++;
745 			continue;
746 		}
747 
748 		/* does the page belong to an object? */
749 		if (uobj != NULL) {
750 			uvmexp.pdobscan++;
751 		} else {
752 #if defined(VMSWAP)
753 			KASSERT(anon != NULL);
754 			uvmexp.pdanscan++;
755 #else /* defined(VMSWAP) */
756 			panic("%s: anon", __func__);
757 #endif /* defined(VMSWAP) */
758 		}
759 
760 
761 		/*
762 		 * we now have the object and the page queues locked.
763 		 * if the page is not swap-backed, call the object's
764 		 * pager to flush and free the page.
765 		 */
766 
767 #if defined(READAHEAD_STATS)
768 		if ((p->pqflags & PQ_READAHEAD) != 0) {
769 			p->pqflags &= ~PQ_READAHEAD;
770 			uvm_ra_miss.ev_count++;
771 		}
772 #endif /* defined(READAHEAD_STATS) */
773 
774 		if ((p->pqflags & PQ_SWAPBACKED) == 0) {
775 			KASSERT(uobj != NULL);
776 			mutex_exit(&uvm_pageqlock);
777 			(void) (uobj->pgops->pgo_put)(uobj, p->offset,
778 			    p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE);
779 			mutex_enter(&uvm_pageqlock);
780 			continue;
781 		}
782 
783 		/*
784 		 * the page is swap-backed.  remove all the permissions
785 		 * from the page so we can sync the modified info
786 		 * without any race conditions.  if the page is clean
787 		 * we can free it now and continue.
788 		 */
789 
790 		pmap_page_protect(p, VM_PROT_NONE);
791 		if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
792 			p->flags &= ~(PG_CLEAN);
793 		}
794 		if (p->flags & PG_CLEAN) {
795 			int slot;
796 			int pageidx;
797 
798 			pageidx = p->offset >> PAGE_SHIFT;
799 			uvm_pagefree(p);
800 			uvmexp.pdfreed++;
801 
802 			/*
803 			 * for anons, we need to remove the page
804 			 * from the anon ourselves.  for aobjs,
805 			 * pagefree did that for us.
806 			 */
807 
808 			if (anon) {
809 				KASSERT(anon->an_swslot != 0);
810 				anon->an_page = NULL;
811 				slot = anon->an_swslot;
812 			} else {
813 				slot = uao_find_swslot(uobj, pageidx);
814 			}
815 			mutex_exit(slock);
816 
817 			if (slot > 0) {
818 				/* this page is now only in swap. */
819 				mutex_enter(&uvm_swap_data_lock);
820 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
821 				uvmexp.swpgonly++;
822 				mutex_exit(&uvm_swap_data_lock);
823 			}
824 			continue;
825 		}
826 
827 #if defined(VMSWAP)
828 		/*
829 		 * this page is dirty, skip it if we'll have met our
830 		 * free target when all the current pageouts complete.
831 		 */
832 
833 		if (uvmexp.free + uvmexp.paging > uvmexp.freetarg << 2) {
834 			mutex_exit(slock);
835 			continue;
836 		}
837 
838 		/*
839 		 * free any swap space allocated to the page since
840 		 * we'll have to write it again with its new data.
841 		 */
842 
843 		uvmpd_dropswap(p);
844 
845 		/*
846 		 * start new swap pageout cluster (if necessary).
847 		 *
848 		 * if swap is full reactivate this page so that
849 		 * we eventually cycle all pages through the
850 		 * inactive queue.
851 		 */
852 
853 		if (swapcluster_allocslots(&swc)) {
854 			dirtyreacts++;
855 			uvm_pageactivate(p);
856 			mutex_exit(slock);
857 			continue;
858 		}
859 
860 		/*
861 		 * at this point, we're definitely going reuse this
862 		 * page.  mark the page busy and delayed-free.
863 		 * we should remove the page from the page queues
864 		 * so we don't ever look at it again.
865 		 * adjust counters and such.
866 		 */
867 
868 		p->flags |= PG_BUSY;
869 		UVM_PAGE_OWN(p, "scan_queue");
870 
871 		p->flags |= PG_PAGEOUT;
872 		uvm_pagedequeue(p);
873 
874 		uvmexp.pgswapout++;
875 		mutex_exit(&uvm_pageqlock);
876 
877 		/*
878 		 * add the new page to the cluster.
879 		 */
880 
881 		if (swapcluster_add(&swc, p)) {
882 			p->flags &= ~(PG_BUSY|PG_PAGEOUT);
883 			UVM_PAGE_OWN(p, NULL);
884 			mutex_enter(&uvm_pageqlock);
885 			dirtyreacts++;
886 			uvm_pageactivate(p);
887 			mutex_exit(slock);
888 			continue;
889 		}
890 		mutex_exit(slock);
891 
892 		swapcluster_flush(&swc, false);
893 		mutex_enter(&uvm_pageqlock);
894 
895 		/*
896 		 * the pageout is in progress.  bump counters and set up
897 		 * for the next loop.
898 		 */
899 
900 		uvmexp.pdpending++;
901 
902 #else /* defined(VMSWAP) */
903 		uvm_pageactivate(p);
904 		mutex_exit(slock);
905 #endif /* defined(VMSWAP) */
906 	}
907 
908 #if defined(VMSWAP)
909 	mutex_exit(&uvm_pageqlock);
910 	swapcluster_flush(&swc, true);
911 	mutex_enter(&uvm_pageqlock);
912 #endif /* defined(VMSWAP) */
913 }
914 
915 /*
916  * uvmpd_scan: scan the page queues and attempt to meet our targets.
917  *
918  * => called with pageq's locked
919  */
920 
921 static void
922 uvmpd_scan(void)
923 {
924 	int swap_shortage, pages_freed;
925 	UVMHIST_FUNC("uvmpd_scan"); UVMHIST_CALLED(pdhist);
926 
927 	uvmexp.pdrevs++;
928 
929 	/*
930 	 * work on meeting our targets.   first we work on our free target
931 	 * by converting inactive pages into free pages.  then we work on
932 	 * meeting our inactive target by converting active pages to
933 	 * inactive ones.
934 	 */
935 
936 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
937 
938 	pages_freed = uvmexp.pdfreed;
939 	uvmpd_scan_queue();
940 	pages_freed = uvmexp.pdfreed - pages_freed;
941 
942 	/*
943 	 * detect if we're not going to be able to page anything out
944 	 * until we free some swap resources from active pages.
945 	 */
946 
947 	swap_shortage = 0;
948 	if (uvmexp.free < uvmexp.freetarg &&
949 	    uvmexp.swpginuse >= uvmexp.swpgavail &&
950 	    !uvm_swapisfull() &&
951 	    pages_freed == 0) {
952 		swap_shortage = uvmexp.freetarg - uvmexp.free;
953 	}
954 
955 	uvmpdpol_balancequeue(swap_shortage);
956 
957 	/*
958 	 * if still below the minimum target, try unloading kernel
959 	 * modules.
960 	 */
961 
962 	if (uvmexp.free < uvmexp.freemin) {
963 		module_thread_kick();
964 	}
965 }
966 
967 /*
968  * uvm_reclaimable: decide whether to wait for pagedaemon.
969  *
970  * => return true if it seems to be worth to do uvm_wait.
971  *
972  * XXX should be tunable.
973  * XXX should consider pools, etc?
974  */
975 
976 bool
977 uvm_reclaimable(void)
978 {
979 	int filepages;
980 	int active, inactive;
981 
982 	/*
983 	 * if swap is not full, no problem.
984 	 */
985 
986 	if (!uvm_swapisfull()) {
987 		return true;
988 	}
989 
990 	/*
991 	 * file-backed pages can be reclaimed even when swap is full.
992 	 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim.
993 	 *
994 	 * XXX assume the worst case, ie. all wired pages are file-backed.
995 	 *
996 	 * XXX should consider about other reclaimable memory.
997 	 * XXX ie. pools, traditional buffer cache.
998 	 */
999 
1000 	filepages = uvmexp.filepages + uvmexp.execpages - uvmexp.wired;
1001 	uvm_estimatepageable(&active, &inactive);
1002 	if (filepages >= MIN((active + inactive) >> 4,
1003 	    5 * 1024 * 1024 >> PAGE_SHIFT)) {
1004 		return true;
1005 	}
1006 
1007 	/*
1008 	 * kill the process, fail allocation, etc..
1009 	 */
1010 
1011 	return false;
1012 }
1013 
1014 void
1015 uvm_estimatepageable(int *active, int *inactive)
1016 {
1017 
1018 	uvmpdpol_estimatepageable(active, inactive);
1019 }
1020 
1021