xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 76c7fc5f6b13ed0b1508e6b313e88e59977ed78e)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.111 2019/10/01 17:40:22 chs 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 1.111 2019/10/01 17:40:22 chs 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 #include <sys/kthread.h>
83 
84 #include <uvm/uvm.h>
85 #include <uvm/uvm_pdpolicy.h>
86 
87 #ifdef UVMHIST
88 UVMHIST_DEFINE(pdhist);
89 #endif
90 
91 /*
92  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
93  * in a pass thru the inactive list when swap is full.  the value should be
94  * "small"... if it's too large we'll cycle the active pages thru the inactive
95  * queue too quickly to for them to be referenced and avoid being freed.
96  */
97 
98 #define	UVMPD_NUMDIRTYREACTS	16
99 
100 #define	UVMPD_NUMTRYLOCKOWNER	16
101 
102 /*
103  * local prototypes
104  */
105 
106 static void	uvmpd_scan(void);
107 static void	uvmpd_scan_queue(void);
108 static void	uvmpd_tune(void);
109 static void	uvmpd_pool_drain_thread(void *);
110 static void	uvmpd_pool_drain_wakeup(void);
111 
112 static unsigned int uvm_pagedaemon_waiters;
113 
114 /* State for the pool drainer thread */
115 static kmutex_t uvmpd_pool_drain_lock;
116 static kcondvar_t uvmpd_pool_drain_cv;
117 static bool uvmpd_pool_drain_run = false;
118 
119 /*
120  * XXX hack to avoid hangs when large processes fork.
121  */
122 u_int uvm_extrapages;
123 
124 /*
125  * uvm_wait: wait (sleep) for the page daemon to free some pages
126  *
127  * => should be called with all locks released
128  * => should _not_ be called by the page daemon (to avoid deadlock)
129  */
130 
131 void
132 uvm_wait(const char *wmsg)
133 {
134 	int timo = 0;
135 
136 	if (uvm.pagedaemon_lwp == NULL)
137 		panic("out of memory before the pagedaemon thread exists");
138 
139 	mutex_spin_enter(&uvm_fpageqlock);
140 
141 	/*
142 	 * check for page daemon going to sleep (waiting for itself)
143 	 */
144 
145 	if (curlwp == uvm.pagedaemon_lwp && uvmexp.paging == 0) {
146 		/*
147 		 * now we have a problem: the pagedaemon wants to go to
148 		 * sleep until it frees more memory.   but how can it
149 		 * free more memory if it is asleep?  that is a deadlock.
150 		 * we have two options:
151 		 *  [1] panic now
152 		 *  [2] put a timeout on the sleep, thus causing the
153 		 *      pagedaemon to only pause (rather than sleep forever)
154 		 *
155 		 * note that option [2] will only help us if we get lucky
156 		 * and some other process on the system breaks the deadlock
157 		 * by exiting or freeing memory (thus allowing the pagedaemon
158 		 * to continue).  for now we panic if DEBUG is defined,
159 		 * otherwise we hope for the best with option [2] (better
160 		 * yet, this should never happen in the first place!).
161 		 */
162 
163 		printf("pagedaemon: deadlock detected!\n");
164 		timo = hz >> 3;		/* set timeout */
165 #if defined(DEBUG)
166 		/* DEBUG: panic so we can debug it */
167 		panic("pagedaemon deadlock");
168 #endif
169 	}
170 
171 	uvm_pagedaemon_waiters++;
172 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
173 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvm_fpageqlock, false, wmsg, timo);
174 }
175 
176 /*
177  * uvm_kick_pdaemon: perform checks to determine if we need to
178  * give the pagedaemon a nudge, and do so if necessary.
179  *
180  * => called with uvm_fpageqlock held.
181  */
182 
183 void
184 uvm_kick_pdaemon(void)
185 {
186 
187 	KASSERT(mutex_owned(&uvm_fpageqlock));
188 
189 	if (uvmexp.free + uvmexp.paging < uvmexp.freemin ||
190 	    (uvmexp.free + uvmexp.paging < uvmexp.freetarg &&
191 	     uvmpdpol_needsscan_p()) ||
192 	     uvm_km_va_starved_p()) {
193 		wakeup(&uvm.pagedaemon);
194 	}
195 }
196 
197 /*
198  * uvmpd_tune: tune paging parameters
199  *
200  * => called when ever memory is added (or removed?) to the system
201  * => caller must call with page queues locked
202  */
203 
204 static void
205 uvmpd_tune(void)
206 {
207 	int val;
208 
209 	UVMHIST_FUNC("uvmpd_tune"); UVMHIST_CALLED(pdhist);
210 
211 	/*
212 	 * try to keep 0.5% of available RAM free, but limit to between
213 	 * 128k and 1024k per-CPU.  XXX: what are these values good for?
214 	 */
215 	val = uvmexp.npages / 200;
216 	val = MAX(val, (128*1024) >> PAGE_SHIFT);
217 	val = MIN(val, (1024*1024) >> PAGE_SHIFT);
218 	val *= ncpu;
219 
220 	/* Make sure there's always a user page free. */
221 	if (val < uvmexp.reserve_kernel + 1)
222 		val = uvmexp.reserve_kernel + 1;
223 	uvmexp.freemin = val;
224 
225 	/* Calculate free target. */
226 	val = (uvmexp.freemin * 4) / 3;
227 	if (val <= uvmexp.freemin)
228 		val = uvmexp.freemin + 1;
229 	uvmexp.freetarg = val + atomic_swap_uint(&uvm_extrapages, 0);
230 
231 	uvmexp.wiredmax = uvmexp.npages / 3;
232 	UVMHIST_LOG(pdhist, "<- done, freemin=%jd, freetarg=%jd, wiredmax=%jd",
233 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
234 }
235 
236 /*
237  * uvm_pageout: the main loop for the pagedaemon
238  */
239 
240 void
241 uvm_pageout(void *arg)
242 {
243 	int npages = 0;
244 	int extrapages = 0;
245 
246 	UVMHIST_FUNC("uvm_pageout"); UVMHIST_CALLED(pdhist);
247 
248 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
249 
250 	mutex_init(&uvmpd_pool_drain_lock, MUTEX_DEFAULT, IPL_VM);
251 	cv_init(&uvmpd_pool_drain_cv, "pooldrain");
252 
253 	/* Create the pool drainer kernel thread. */
254 	if (kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL,
255 	    uvmpd_pool_drain_thread, NULL, NULL, "pooldrain"))
256 		panic("fork pooldrain");
257 
258 	/*
259 	 * ensure correct priority and set paging parameters...
260 	 */
261 
262 	uvm.pagedaemon_lwp = curlwp;
263 	mutex_enter(&uvm_pageqlock);
264 	npages = uvmexp.npages;
265 	uvmpd_tune();
266 	mutex_exit(&uvm_pageqlock);
267 
268 	/*
269 	 * main loop
270 	 */
271 
272 	for (;;) {
273 		bool needsscan, needsfree, kmem_va_starved;
274 
275 		kmem_va_starved = uvm_km_va_starved_p();
276 
277 		mutex_spin_enter(&uvm_fpageqlock);
278 		if ((uvm_pagedaemon_waiters == 0 || uvmexp.paging > 0) &&
279 		    !kmem_va_starved) {
280 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
281 			UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
282 			    &uvm_fpageqlock, false, "pgdaemon", 0);
283 			uvmexp.pdwoke++;
284 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
285 		} else {
286 			mutex_spin_exit(&uvm_fpageqlock);
287 		}
288 
289 		/*
290 		 * now lock page queues and recompute inactive count
291 		 */
292 
293 		mutex_enter(&uvm_pageqlock);
294 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
295 			npages = uvmexp.npages;
296 			extrapages = uvm_extrapages;
297 			mutex_spin_enter(&uvm_fpageqlock);
298 			uvmpd_tune();
299 			mutex_spin_exit(&uvm_fpageqlock);
300 		}
301 
302 		uvmpdpol_tune();
303 
304 		/*
305 		 * Estimate a hint.  Note that bufmem are returned to
306 		 * system only when entire pool page is empty.
307 		 */
308 		mutex_spin_enter(&uvm_fpageqlock);
309 
310 		UVMHIST_LOG(pdhist,"  free/ftarg=%jd/%jd",
311 		    uvmexp.free, uvmexp.freetarg, 0,0);
312 
313 		needsfree = uvmexp.free + uvmexp.paging < uvmexp.freetarg;
314 		needsscan = needsfree || uvmpdpol_needsscan_p();
315 
316 		/*
317 		 * scan if needed
318 		 */
319 		if (needsscan) {
320 			mutex_spin_exit(&uvm_fpageqlock);
321 			uvmpd_scan();
322 			mutex_spin_enter(&uvm_fpageqlock);
323 		}
324 
325 		/*
326 		 * if there's any free memory to be had,
327 		 * wake up any waiters.
328 		 */
329 		if (uvmexp.free > uvmexp.reserve_kernel ||
330 		    uvmexp.paging == 0) {
331 			wakeup(&uvmexp.free);
332 			uvm_pagedaemon_waiters = 0;
333 		}
334 		mutex_spin_exit(&uvm_fpageqlock);
335 
336 		/*
337 		 * scan done.  unlock page queues (the only lock we are holding)
338 		 */
339 		mutex_exit(&uvm_pageqlock);
340 
341 		/*
342 		 * if we don't need free memory, we're done.
343 		 */
344 
345 		if (!needsfree && !kmem_va_starved)
346 			continue;
347 
348 		/*
349 		 * kick the pool drainer thread.
350 		 */
351 
352 		uvmpd_pool_drain_wakeup();
353 	}
354 	/*NOTREACHED*/
355 }
356 
357 
358 /*
359  * uvm_aiodone_worker: a workqueue callback for the aiodone daemon.
360  */
361 
362 void
363 uvm_aiodone_worker(struct work *wk, void *dummy)
364 {
365 	struct buf *bp = (void *)wk;
366 
367 	KASSERT(&bp->b_work == wk);
368 
369 	/*
370 	 * process an i/o that's done.
371 	 */
372 
373 	(*bp->b_iodone)(bp);
374 }
375 
376 void
377 uvm_pageout_start(int npages)
378 {
379 
380 	mutex_spin_enter(&uvm_fpageqlock);
381 	uvmexp.paging += npages;
382 	mutex_spin_exit(&uvm_fpageqlock);
383 }
384 
385 void
386 uvm_pageout_done(int npages)
387 {
388 
389 	mutex_spin_enter(&uvm_fpageqlock);
390 	KASSERT(uvmexp.paging >= npages);
391 	uvmexp.paging -= npages;
392 
393 	/*
394 	 * wake up either of pagedaemon or LWPs waiting for it.
395 	 */
396 
397 	if (uvmexp.free <= uvmexp.reserve_kernel) {
398 		wakeup(&uvm.pagedaemon);
399 	} else {
400 		wakeup(&uvmexp.free);
401 		uvm_pagedaemon_waiters = 0;
402 	}
403 	mutex_spin_exit(&uvm_fpageqlock);
404 }
405 
406 /*
407  * uvmpd_trylockowner: trylock the page's owner.
408  *
409  * => called with pageq locked.
410  * => resolve orphaned O->A loaned page.
411  * => return the locked mutex on success.  otherwise, return NULL.
412  */
413 
414 kmutex_t *
415 uvmpd_trylockowner(struct vm_page *pg)
416 {
417 	struct uvm_object *uobj = pg->uobject;
418 	kmutex_t *slock;
419 
420 	KASSERT(mutex_owned(&uvm_pageqlock));
421 
422 	if (uobj != NULL) {
423 		slock = uobj->vmobjlock;
424 	} else {
425 		struct vm_anon *anon = pg->uanon;
426 
427 		KASSERT(anon != NULL);
428 		slock = anon->an_lock;
429 	}
430 
431 	if (!mutex_tryenter(slock)) {
432 		return NULL;
433 	}
434 
435 	if (uobj == NULL) {
436 
437 		/*
438 		 * set PQ_ANON if it isn't set already.
439 		 */
440 
441 		if ((pg->pqflags & PQ_ANON) == 0) {
442 			KASSERT(pg->loan_count > 0);
443 			pg->loan_count--;
444 			pg->pqflags |= PQ_ANON;
445 			/* anon now owns it */
446 		}
447 	}
448 
449 	return slock;
450 }
451 
452 #if defined(VMSWAP)
453 struct swapcluster {
454 	int swc_slot;
455 	int swc_nallocated;
456 	int swc_nused;
457 	struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)];
458 };
459 
460 static void
461 swapcluster_init(struct swapcluster *swc)
462 {
463 
464 	swc->swc_slot = 0;
465 	swc->swc_nused = 0;
466 }
467 
468 static int
469 swapcluster_allocslots(struct swapcluster *swc)
470 {
471 	int slot;
472 	int npages;
473 
474 	if (swc->swc_slot != 0) {
475 		return 0;
476 	}
477 
478 	/* Even with strange MAXPHYS, the shift
479 	   implicitly rounds down to a page. */
480 	npages = MAXPHYS >> PAGE_SHIFT;
481 	slot = uvm_swap_alloc(&npages, true);
482 	if (slot == 0) {
483 		return ENOMEM;
484 	}
485 	swc->swc_slot = slot;
486 	swc->swc_nallocated = npages;
487 	swc->swc_nused = 0;
488 
489 	return 0;
490 }
491 
492 static int
493 swapcluster_add(struct swapcluster *swc, struct vm_page *pg)
494 {
495 	int slot;
496 	struct uvm_object *uobj;
497 
498 	KASSERT(swc->swc_slot != 0);
499 	KASSERT(swc->swc_nused < swc->swc_nallocated);
500 	KASSERT((pg->pqflags & PQ_SWAPBACKED) != 0);
501 
502 	slot = swc->swc_slot + swc->swc_nused;
503 	uobj = pg->uobject;
504 	if (uobj == NULL) {
505 		KASSERT(mutex_owned(pg->uanon->an_lock));
506 		pg->uanon->an_swslot = slot;
507 	} else {
508 		int result;
509 
510 		KASSERT(mutex_owned(uobj->vmobjlock));
511 		result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot);
512 		if (result == -1) {
513 			return ENOMEM;
514 		}
515 	}
516 	swc->swc_pages[swc->swc_nused] = pg;
517 	swc->swc_nused++;
518 
519 	return 0;
520 }
521 
522 static void
523 swapcluster_flush(struct swapcluster *swc, bool now)
524 {
525 	int slot;
526 	int nused;
527 	int nallocated;
528 	int error __diagused;
529 
530 	if (swc->swc_slot == 0) {
531 		return;
532 	}
533 	KASSERT(swc->swc_nused <= swc->swc_nallocated);
534 
535 	slot = swc->swc_slot;
536 	nused = swc->swc_nused;
537 	nallocated = swc->swc_nallocated;
538 
539 	/*
540 	 * if this is the final pageout we could have a few
541 	 * unused swap blocks.  if so, free them now.
542 	 */
543 
544 	if (nused < nallocated) {
545 		if (!now) {
546 			return;
547 		}
548 		uvm_swap_free(slot + nused, nallocated - nused);
549 	}
550 
551 	/*
552 	 * now start the pageout.
553 	 */
554 
555 	if (nused > 0) {
556 		uvmexp.pdpageouts++;
557 		uvm_pageout_start(nused);
558 		error = uvm_swap_put(slot, swc->swc_pages, nused, 0);
559 		KASSERT(error == 0 || error == ENOMEM);
560 	}
561 
562 	/*
563 	 * zero swslot to indicate that we are
564 	 * no longer building a swap-backed cluster.
565 	 */
566 
567 	swc->swc_slot = 0;
568 	swc->swc_nused = 0;
569 }
570 
571 static int
572 swapcluster_nused(struct swapcluster *swc)
573 {
574 
575 	return swc->swc_nused;
576 }
577 
578 /*
579  * uvmpd_dropswap: free any swap allocated to this page.
580  *
581  * => called with owner locked.
582  * => return true if a page had an associated slot.
583  */
584 
585 static bool
586 uvmpd_dropswap(struct vm_page *pg)
587 {
588 	bool result = false;
589 	struct vm_anon *anon = pg->uanon;
590 
591 	if ((pg->pqflags & PQ_ANON) && anon->an_swslot) {
592 		uvm_swap_free(anon->an_swslot, 1);
593 		anon->an_swslot = 0;
594 		pg->flags &= ~PG_CLEAN;
595 		result = true;
596 	} else if (pg->pqflags & PQ_AOBJ) {
597 		int slot = uao_set_swslot(pg->uobject,
598 		    pg->offset >> PAGE_SHIFT, 0);
599 		if (slot) {
600 			uvm_swap_free(slot, 1);
601 			pg->flags &= ~PG_CLEAN;
602 			result = true;
603 		}
604 	}
605 
606 	return result;
607 }
608 
609 /*
610  * uvmpd_trydropswap: try to free any swap allocated to this page.
611  *
612  * => return true if a slot is successfully freed.
613  */
614 
615 bool
616 uvmpd_trydropswap(struct vm_page *pg)
617 {
618 	kmutex_t *slock;
619 	bool result;
620 
621 	if ((pg->flags & PG_BUSY) != 0) {
622 		return false;
623 	}
624 
625 	/*
626 	 * lock the page's owner.
627 	 */
628 
629 	slock = uvmpd_trylockowner(pg);
630 	if (slock == NULL) {
631 		return false;
632 	}
633 
634 	/*
635 	 * skip this page if it's busy.
636 	 */
637 
638 	if ((pg->flags & PG_BUSY) != 0) {
639 		mutex_exit(slock);
640 		return false;
641 	}
642 
643 	result = uvmpd_dropswap(pg);
644 
645 	mutex_exit(slock);
646 
647 	return result;
648 }
649 
650 #endif /* defined(VMSWAP) */
651 
652 /*
653  * uvmpd_scan_queue: scan an replace candidate list for pages
654  * to clean or free.
655  *
656  * => called with page queues locked
657  * => we work on meeting our free target by converting inactive pages
658  *    into free pages.
659  * => we handle the building of swap-backed clusters
660  */
661 
662 static void
663 uvmpd_scan_queue(void)
664 {
665 	struct vm_page *p;
666 	struct uvm_object *uobj;
667 	struct vm_anon *anon;
668 #if defined(VMSWAP)
669 	struct swapcluster swc;
670 #endif /* defined(VMSWAP) */
671 	int dirtyreacts;
672 	int lockownerfail;
673 	kmutex_t *slock;
674 	UVMHIST_FUNC("uvmpd_scan_queue"); UVMHIST_CALLED(pdhist);
675 
676 	/*
677 	 * swslot is non-zero if we are building a swap cluster.  we want
678 	 * to stay in the loop while we have a page to scan or we have
679 	 * a swap-cluster to build.
680 	 */
681 
682 #if defined(VMSWAP)
683 	swapcluster_init(&swc);
684 #endif /* defined(VMSWAP) */
685 
686 	dirtyreacts = 0;
687 	lockownerfail = 0;
688 	uvmpdpol_scaninit();
689 
690 	while (/* CONSTCOND */ 1) {
691 
692 		/*
693 		 * see if we've met the free target.
694 		 */
695 
696 		if (uvmexp.free + uvmexp.paging
697 #if defined(VMSWAP)
698 		    + swapcluster_nused(&swc)
699 #endif /* defined(VMSWAP) */
700 		    >= uvmexp.freetarg << 2 ||
701 		    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
702 			UVMHIST_LOG(pdhist,"  met free target: "
703 				    "exit loop", 0, 0, 0, 0);
704 			break;
705 		}
706 
707 		p = uvmpdpol_selectvictim();
708 		if (p == NULL) {
709 			break;
710 		}
711 		KASSERT(uvmpdpol_pageisqueued_p(p));
712 		KASSERT(p->wire_count == 0);
713 
714 		/*
715 		 * we are below target and have a new page to consider.
716 		 */
717 
718 		anon = p->uanon;
719 		uobj = p->uobject;
720 
721 		/*
722 		 * first we attempt to lock the object that this page
723 		 * belongs to.  if our attempt fails we skip on to
724 		 * the next page (no harm done).  it is important to
725 		 * "try" locking the object as we are locking in the
726 		 * wrong order (pageq -> object) and we don't want to
727 		 * deadlock.
728 		 *
729 		 * the only time we expect to see an ownerless page
730 		 * (i.e. a page with no uobject and !PQ_ANON) is if an
731 		 * anon has loaned a page from a uvm_object and the
732 		 * uvm_object has dropped the ownership.  in that
733 		 * case, the anon can "take over" the loaned page
734 		 * and make it its own.
735 		 */
736 
737 		slock = uvmpd_trylockowner(p);
738 		if (slock == NULL) {
739 			/*
740 			 * yield cpu to make a chance for an LWP holding
741 			 * the lock run.  otherwise we can busy-loop too long
742 			 * if the page queue is filled with a lot of pages
743 			 * from few objects.
744 			 */
745 			lockownerfail++;
746 			if (lockownerfail > UVMPD_NUMTRYLOCKOWNER) {
747 				mutex_exit(&uvm_pageqlock);
748 				/* XXX Better than yielding but inadequate. */
749 				kpause("livelock", false, 1, NULL);
750 				mutex_enter(&uvm_pageqlock);
751 				lockownerfail = 0;
752 			}
753 			continue;
754 		}
755 		if (p->flags & PG_BUSY) {
756 			mutex_exit(slock);
757 			uvmexp.pdbusy++;
758 			continue;
759 		}
760 
761 		/* does the page belong to an object? */
762 		if (uobj != NULL) {
763 			uvmexp.pdobscan++;
764 		} else {
765 #if defined(VMSWAP)
766 			KASSERT(anon != NULL);
767 			uvmexp.pdanscan++;
768 #else /* defined(VMSWAP) */
769 			panic("%s: anon", __func__);
770 #endif /* defined(VMSWAP) */
771 		}
772 
773 
774 		/*
775 		 * we now have the object and the page queues locked.
776 		 * if the page is not swap-backed, call the object's
777 		 * pager to flush and free the page.
778 		 */
779 
780 #if defined(READAHEAD_STATS)
781 		if ((p->pqflags & PQ_READAHEAD) != 0) {
782 			p->pqflags &= ~PQ_READAHEAD;
783 			uvm_ra_miss.ev_count++;
784 		}
785 #endif /* defined(READAHEAD_STATS) */
786 
787 		if ((p->pqflags & PQ_SWAPBACKED) == 0) {
788 			KASSERT(uobj != NULL);
789 			mutex_exit(&uvm_pageqlock);
790 			(void) (uobj->pgops->pgo_put)(uobj, p->offset,
791 			    p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE);
792 			mutex_enter(&uvm_pageqlock);
793 			continue;
794 		}
795 
796 		/*
797 		 * the page is swap-backed.  remove all the permissions
798 		 * from the page so we can sync the modified info
799 		 * without any race conditions.  if the page is clean
800 		 * we can free it now and continue.
801 		 */
802 
803 		pmap_page_protect(p, VM_PROT_NONE);
804 		if ((p->flags & PG_CLEAN) && pmap_clear_modify(p)) {
805 			p->flags &= ~(PG_CLEAN);
806 		}
807 		if (p->flags & PG_CLEAN) {
808 			int slot;
809 			int pageidx;
810 
811 			pageidx = p->offset >> PAGE_SHIFT;
812 			uvm_pagefree(p);
813 			uvmexp.pdfreed++;
814 
815 			/*
816 			 * for anons, we need to remove the page
817 			 * from the anon ourselves.  for aobjs,
818 			 * pagefree did that for us.
819 			 */
820 
821 			if (anon) {
822 				KASSERT(anon->an_swslot != 0);
823 				anon->an_page = NULL;
824 				slot = anon->an_swslot;
825 			} else {
826 				slot = uao_find_swslot(uobj, pageidx);
827 			}
828 			mutex_exit(slock);
829 
830 			if (slot > 0) {
831 				/* this page is now only in swap. */
832 				mutex_enter(&uvm_swap_data_lock);
833 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
834 				uvmexp.swpgonly++;
835 				mutex_exit(&uvm_swap_data_lock);
836 			}
837 			continue;
838 		}
839 
840 #if defined(VMSWAP)
841 		/*
842 		 * this page is dirty, skip it if we'll have met our
843 		 * free target when all the current pageouts complete.
844 		 */
845 
846 		if (uvmexp.free + uvmexp.paging > uvmexp.freetarg << 2) {
847 			mutex_exit(slock);
848 			continue;
849 		}
850 
851 		/*
852 		 * free any swap space allocated to the page since
853 		 * we'll have to write it again with its new data.
854 		 */
855 
856 		uvmpd_dropswap(p);
857 
858 		/*
859 		 * start new swap pageout cluster (if necessary).
860 		 *
861 		 * if swap is full reactivate this page so that
862 		 * we eventually cycle all pages through the
863 		 * inactive queue.
864 		 */
865 
866 		if (swapcluster_allocslots(&swc)) {
867 			dirtyreacts++;
868 			uvm_pageactivate(p);
869 			mutex_exit(slock);
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 	 * if still below the minimum target, try unloading kernel
972 	 * modules.
973 	 */
974 
975 	if (uvmexp.free < uvmexp.freemin) {
976 		module_thread_kick();
977 	}
978 }
979 
980 /*
981  * uvm_reclaimable: decide whether to wait for pagedaemon.
982  *
983  * => return true if it seems to be worth to do uvm_wait.
984  *
985  * XXX should be tunable.
986  * XXX should consider pools, etc?
987  */
988 
989 bool
990 uvm_reclaimable(void)
991 {
992 	int filepages;
993 	int active, inactive;
994 
995 	/*
996 	 * if swap is not full, no problem.
997 	 */
998 
999 	if (!uvm_swapisfull()) {
1000 		return true;
1001 	}
1002 
1003 	/*
1004 	 * file-backed pages can be reclaimed even when swap is full.
1005 	 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim.
1006 	 *
1007 	 * XXX assume the worst case, ie. all wired pages are file-backed.
1008 	 *
1009 	 * XXX should consider about other reclaimable memory.
1010 	 * XXX ie. pools, traditional buffer cache.
1011 	 */
1012 
1013 	filepages = uvmexp.filepages + uvmexp.execpages - uvmexp.wired;
1014 	uvm_estimatepageable(&active, &inactive);
1015 	if (filepages >= MIN((active + inactive) >> 4,
1016 	    5 * 1024 * 1024 >> PAGE_SHIFT)) {
1017 		return true;
1018 	}
1019 
1020 	/*
1021 	 * kill the process, fail allocation, etc..
1022 	 */
1023 
1024 	return false;
1025 }
1026 
1027 void
1028 uvm_estimatepageable(int *active, int *inactive)
1029 {
1030 
1031 	uvmpdpol_estimatepageable(active, inactive);
1032 }
1033 
1034 
1035 /*
1036  * Use a separate thread for draining pools.
1037  * This work can't done from the main pagedaemon thread because
1038  * some pool allocators need to take vm_map locks.
1039  */
1040 
1041 static void
1042 uvmpd_pool_drain_thread(void *arg)
1043 {
1044 	int bufcnt;
1045 
1046 	for (;;) {
1047 		mutex_enter(&uvmpd_pool_drain_lock);
1048 		if (!uvmpd_pool_drain_run) {
1049 			cv_wait(&uvmpd_pool_drain_cv, &uvmpd_pool_drain_lock);
1050 		}
1051 		uvmpd_pool_drain_run = false;
1052 		mutex_exit(&uvmpd_pool_drain_lock);
1053 
1054 		/*
1055 		 * kill unused metadata buffers.
1056 		 */
1057 		mutex_spin_enter(&uvm_fpageqlock);
1058 		bufcnt = uvmexp.freetarg - uvmexp.free;
1059 		mutex_spin_exit(&uvm_fpageqlock);
1060 		if (bufcnt < 0)
1061 			bufcnt = 0;
1062 
1063 		mutex_enter(&bufcache_lock);
1064 		buf_drain(bufcnt << PAGE_SHIFT);
1065 		mutex_exit(&bufcache_lock);
1066 
1067 		/*
1068 		 * drain a pool.
1069 		 */
1070 		pool_drain(NULL);
1071 	}
1072 	/*NOTREACHED*/
1073 }
1074 
1075 static void
1076 uvmpd_pool_drain_wakeup(void)
1077 {
1078 
1079 	mutex_enter(&uvmpd_pool_drain_lock);
1080 	uvmpd_pool_drain_run = true;
1081 	cv_signal(&uvmpd_pool_drain_cv);
1082 	mutex_exit(&uvmpd_pool_drain_lock);
1083 }
1084