xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 181254a7b1bdde6873432bffef2d2decc4b5c22f)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.130 2020/07/09 05:57:15 skrll 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 1.130 2020/07/09 05:57:15 skrll Exp $");
70 
71 #include "opt_uvmhist.h"
72 #include "opt_readahead.h"
73 
74 #define	__RWLOCK_PRIVATE
75 
76 #include <sys/param.h>
77 #include <sys/proc.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/pool.h>
81 #include <sys/buf.h>
82 #include <sys/module.h>
83 #include <sys/atomic.h>
84 #include <sys/kthread.h>
85 
86 #include <uvm/uvm.h>
87 #include <uvm/uvm_pdpolicy.h>
88 #include <uvm/uvm_pgflcache.h>
89 
90 #ifdef UVMHIST
91 UVMHIST_DEFINE(pdhist);
92 #endif
93 
94 /*
95  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
96  * in a pass thru the inactive list when swap is full.  the value should be
97  * "small"... if it's too large we'll cycle the active pages thru the inactive
98  * queue too quickly to for them to be referenced and avoid being freed.
99  */
100 
101 #define	UVMPD_NUMDIRTYREACTS	16
102 
103 #define	UVMPD_NUMTRYLOCKOWNER	128
104 
105 /*
106  * local prototypes
107  */
108 
109 static void	uvmpd_scan(void);
110 static void	uvmpd_scan_queue(void);
111 static void	uvmpd_tune(void);
112 static void	uvmpd_pool_drain_thread(void *);
113 static void	uvmpd_pool_drain_wakeup(void);
114 
115 static unsigned int uvm_pagedaemon_waiters;
116 
117 /* State for the pool drainer thread */
118 static kmutex_t uvmpd_lock __cacheline_aligned;
119 static kcondvar_t uvmpd_pool_drain_cv;
120 static bool uvmpd_pool_drain_run = false;
121 
122 /*
123  * XXX hack to avoid hangs when large processes fork.
124  */
125 u_int uvm_extrapages;
126 
127 /*
128  * uvm_wait: wait (sleep) for the page daemon to free some pages
129  *
130  * => should be called with all locks released
131  * => should _not_ be called by the page daemon (to avoid deadlock)
132  */
133 
134 void
135 uvm_wait(const char *wmsg)
136 {
137 	int timo = 0;
138 
139 	if (uvm.pagedaemon_lwp == NULL)
140 		panic("out of memory before the pagedaemon thread exists");
141 
142 	mutex_spin_enter(&uvmpd_lock);
143 
144 	/*
145 	 * check for page daemon going to sleep (waiting for itself)
146 	 */
147 
148 	if (curlwp == uvm.pagedaemon_lwp && uvmexp.paging == 0) {
149 		/*
150 		 * now we have a problem: the pagedaemon wants to go to
151 		 * sleep until it frees more memory.   but how can it
152 		 * free more memory if it is asleep?  that is a deadlock.
153 		 * we have two options:
154 		 *  [1] panic now
155 		 *  [2] put a timeout on the sleep, thus causing the
156 		 *      pagedaemon to only pause (rather than sleep forever)
157 		 *
158 		 * note that option [2] will only help us if we get lucky
159 		 * and some other process on the system breaks the deadlock
160 		 * by exiting or freeing memory (thus allowing the pagedaemon
161 		 * to continue).  for now we panic if DEBUG is defined,
162 		 * otherwise we hope for the best with option [2] (better
163 		 * yet, this should never happen in the first place!).
164 		 */
165 
166 		printf("pagedaemon: deadlock detected!\n");
167 		timo = hz >> 3;		/* set timeout */
168 #if defined(DEBUG)
169 		/* DEBUG: panic so we can debug it */
170 		panic("pagedaemon deadlock");
171 #endif
172 	}
173 
174 	uvm_pagedaemon_waiters++;
175 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
176 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvmpd_lock, false, wmsg, timo);
177 }
178 
179 /*
180  * uvm_kick_pdaemon: perform checks to determine if we need to
181  * give the pagedaemon a nudge, and do so if necessary.
182  */
183 
184 void
185 uvm_kick_pdaemon(void)
186 {
187 	int fpages = uvm_availmem(false);
188 
189 	if (fpages + uvmexp.paging < uvmexp.freemin ||
190 	    (fpages + uvmexp.paging < uvmexp.freetarg &&
191 	     uvmpdpol_needsscan_p()) ||
192 	     uvm_km_va_starved_p()) {
193 	     	mutex_spin_enter(&uvmpd_lock);
194 		wakeup(&uvm.pagedaemon);
195 	     	mutex_spin_exit(&uvmpd_lock);
196 	}
197 }
198 
199 /*
200  * uvmpd_tune: tune paging parameters
201  *
202  * => called when ever memory is added (or removed?) to the system
203  */
204 
205 static void
206 uvmpd_tune(void)
207 {
208 	int val;
209 
210 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
211 
212 	/*
213 	 * try to keep 0.5% of available RAM free, but limit to between
214 	 * 128k and 1024k per-CPU.  XXX: what are these values good for?
215 	 */
216 	val = uvmexp.npages / 200;
217 	val = MAX(val, (128*1024) >> PAGE_SHIFT);
218 	val = MIN(val, (1024*1024) >> PAGE_SHIFT);
219 	val *= ncpu;
220 
221 	/* Make sure there's always a user page free. */
222 	if (val < uvmexp.reserve_kernel + 1)
223 		val = uvmexp.reserve_kernel + 1;
224 	uvmexp.freemin = val;
225 
226 	/* Calculate free target. */
227 	val = (uvmexp.freemin * 4) / 3;
228 	if (val <= uvmexp.freemin)
229 		val = uvmexp.freemin + 1;
230 	uvmexp.freetarg = val + atomic_swap_uint(&uvm_extrapages, 0);
231 
232 	uvmexp.wiredmax = uvmexp.npages / 3;
233 	UVMHIST_LOG(pdhist, "<- done, freemin=%jd, freetarg=%jd, wiredmax=%jd",
234 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
235 }
236 
237 /*
238  * uvm_pageout: the main loop for the pagedaemon
239  */
240 
241 void
242 uvm_pageout(void *arg)
243 {
244 	int npages = 0;
245 	int extrapages = 0;
246 	int fpages;
247 
248 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
249 
250 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
251 
252 	mutex_init(&uvmpd_lock, MUTEX_DEFAULT, IPL_VM);
253 	cv_init(&uvmpd_pool_drain_cv, "pooldrain");
254 
255 	/* Create the pool drainer kernel thread. */
256 	if (kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL,
257 	    uvmpd_pool_drain_thread, NULL, NULL, "pooldrain"))
258 		panic("fork pooldrain");
259 
260 	/*
261 	 * ensure correct priority and set paging parameters...
262 	 */
263 
264 	uvm.pagedaemon_lwp = curlwp;
265 	npages = uvmexp.npages;
266 	uvmpd_tune();
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(&uvmpd_lock);
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 			    &uvmpd_lock, false, "pgdaemon", 0);
283 			uvmexp.pdwoke++;
284 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
285 		} else {
286 			mutex_spin_exit(&uvmpd_lock);
287 		}
288 
289 		/*
290 		 * now recompute inactive count
291 		 */
292 
293 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
294 			npages = uvmexp.npages;
295 			extrapages = uvm_extrapages;
296 			uvmpd_tune();
297 		}
298 
299 		uvmpdpol_tune();
300 
301 		/*
302 		 * Estimate a hint.  Note that bufmem are returned to
303 		 * system only when entire pool page is empty.
304 		 */
305 		fpages = uvm_availmem(false);
306 		UVMHIST_LOG(pdhist,"  free/ftarg=%jd/%jd",
307 		    fpages, uvmexp.freetarg, 0,0);
308 
309 		needsfree = fpages + uvmexp.paging < uvmexp.freetarg;
310 		needsscan = needsfree || uvmpdpol_needsscan_p();
311 
312 		/*
313 		 * scan if needed
314 		 */
315 		if (needsscan) {
316 			uvmpd_scan();
317 		}
318 
319 		/*
320 		 * if there's any free memory to be had,
321 		 * wake up any waiters.
322 		 */
323 		if (uvm_availmem(false) > uvmexp.reserve_kernel ||
324 		    uvmexp.paging == 0) {
325 			mutex_spin_enter(&uvmpd_lock);
326 			wakeup(&uvmexp.free);
327 			uvm_pagedaemon_waiters = 0;
328 			mutex_spin_exit(&uvmpd_lock);
329 		}
330 
331 		/*
332 		 * scan done.  if we don't need free memory, we're done.
333 		 */
334 
335 		if (!needsfree && !kmem_va_starved)
336 			continue;
337 
338 		/*
339 		 * kick the pool drainer thread.
340 		 */
341 
342 		uvmpd_pool_drain_wakeup();
343 	}
344 	/*NOTREACHED*/
345 }
346 
347 void
348 uvm_pageout_start(int npages)
349 {
350 
351 	atomic_add_int(&uvmexp.paging, npages);
352 }
353 
354 void
355 uvm_pageout_done(int npages)
356 {
357 
358 	KASSERT(atomic_load_relaxed(&uvmexp.paging) >= npages);
359 
360 	if (npages == 0) {
361 		return;
362 	}
363 
364 	atomic_add_int(&uvmexp.paging, -npages);
365 
366 	/*
367 	 * wake up either of pagedaemon or LWPs waiting for it.
368 	 */
369 
370 	mutex_spin_enter(&uvmpd_lock);
371 	if (uvm_availmem(false) <= uvmexp.reserve_kernel) {
372 		wakeup(&uvm.pagedaemon);
373 	} else if (uvm_pagedaemon_waiters != 0) {
374 		wakeup(&uvmexp.free);
375 		uvm_pagedaemon_waiters = 0;
376 	}
377 	mutex_spin_exit(&uvmpd_lock);
378 }
379 
380 /*
381  * uvmpd_trylockowner: trylock the page's owner.
382  *
383  * => called with page interlock held.
384  * => resolve orphaned O->A loaned page.
385  * => return the locked mutex on success.  otherwise, return NULL.
386  */
387 
388 krwlock_t *
389 uvmpd_trylockowner(struct vm_page *pg)
390 {
391 	struct uvm_object *uobj = pg->uobject;
392 	struct vm_anon *anon = pg->uanon;
393 	int tries, count;
394 	bool running;
395 	krwlock_t *slock;
396 
397 	KASSERT(mutex_owned(&pg->interlock));
398 
399 	if (uobj != NULL) {
400 		slock = uobj->vmobjlock;
401 		KASSERTMSG(slock != NULL, "pg %p uobj %p, NULL lock", pg, uobj);
402 	} else if (anon != NULL) {
403 		slock = anon->an_lock;
404 		KASSERTMSG(slock != NULL, "pg %p anon %p, NULL lock", pg, anon);
405 	} else {
406 		/* Page may be in state of flux - ignore. */
407 		mutex_exit(&pg->interlock);
408 		return NULL;
409 	}
410 
411 	/*
412 	 * Now try to lock the objects.  We'll try hard, but don't really
413 	 * plan on spending more than a millisecond or so here.
414 	 */
415 	tries = (curlwp == uvm.pagedaemon_lwp ? UVMPD_NUMTRYLOCKOWNER : 1);
416 	for (;;) {
417 		if (rw_tryenter(slock, RW_WRITER)) {
418 			if (uobj == NULL) {
419 				/*
420 				 * set PG_ANON if it isn't set already.
421 				 */
422 				if ((pg->flags & PG_ANON) == 0) {
423 					KASSERT(pg->loan_count > 0);
424 					pg->loan_count--;
425 					pg->flags |= PG_ANON;
426 					/* anon now owns it */
427 				}
428 			}
429 			mutex_exit(&pg->interlock);
430 			return slock;
431 		}
432 		running = rw_owner_running(slock);
433 		if (!running || --tries <= 0) {
434 			break;
435 		}
436 		count = SPINLOCK_BACKOFF_MAX;
437 		SPINLOCK_BACKOFF(count);
438 	}
439 
440 	/*
441 	 * We didn't get the lock; chances are the very next page on the
442 	 * queue also has the same lock, so if the lock owner is not running
443 	 * take a breather and allow them to make progress.  There could be
444 	 * only 1 CPU in the system, or the pagedaemon could have preempted
445 	 * the owner in kernel, or any number of other things could be going
446 	 * on.
447 	 */
448 	mutex_exit(&pg->interlock);
449 	if (curlwp == uvm.pagedaemon_lwp) {
450 		if (!running) {
451 			(void)kpause("pdpglock", false, 1, NULL);
452 		}
453 		uvmexp.pdbusy++;
454 	}
455 	return NULL;
456 }
457 
458 #if defined(VMSWAP)
459 struct swapcluster {
460 	int swc_slot;
461 	int swc_nallocated;
462 	int swc_nused;
463 	struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)];
464 };
465 
466 static void
467 swapcluster_init(struct swapcluster *swc)
468 {
469 
470 	swc->swc_slot = 0;
471 	swc->swc_nused = 0;
472 }
473 
474 static int
475 swapcluster_allocslots(struct swapcluster *swc)
476 {
477 	int slot;
478 	int npages;
479 
480 	if (swc->swc_slot != 0) {
481 		return 0;
482 	}
483 
484 	/* Even with strange MAXPHYS, the shift
485 	   implicitly rounds down to a page. */
486 	npages = MAXPHYS >> PAGE_SHIFT;
487 	slot = uvm_swap_alloc(&npages, true);
488 	if (slot == 0) {
489 		return ENOMEM;
490 	}
491 	swc->swc_slot = slot;
492 	swc->swc_nallocated = npages;
493 	swc->swc_nused = 0;
494 
495 	return 0;
496 }
497 
498 static int
499 swapcluster_add(struct swapcluster *swc, struct vm_page *pg)
500 {
501 	int slot;
502 	struct uvm_object *uobj;
503 
504 	KASSERT(swc->swc_slot != 0);
505 	KASSERT(swc->swc_nused < swc->swc_nallocated);
506 	KASSERT((pg->flags & PG_SWAPBACKED) != 0);
507 
508 	slot = swc->swc_slot + swc->swc_nused;
509 	uobj = pg->uobject;
510 	if (uobj == NULL) {
511 		KASSERT(rw_write_held(pg->uanon->an_lock));
512 		pg->uanon->an_swslot = slot;
513 	} else {
514 		int result;
515 
516 		KASSERT(rw_write_held(uobj->vmobjlock));
517 		result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot);
518 		if (result == -1) {
519 			return ENOMEM;
520 		}
521 	}
522 	swc->swc_pages[swc->swc_nused] = pg;
523 	swc->swc_nused++;
524 
525 	return 0;
526 }
527 
528 static void
529 swapcluster_flush(struct swapcluster *swc, bool now)
530 {
531 	int slot;
532 	int nused;
533 	int nallocated;
534 	int error __diagused;
535 
536 	if (swc->swc_slot == 0) {
537 		return;
538 	}
539 	KASSERT(swc->swc_nused <= swc->swc_nallocated);
540 
541 	slot = swc->swc_slot;
542 	nused = swc->swc_nused;
543 	nallocated = swc->swc_nallocated;
544 
545 	/*
546 	 * if this is the final pageout we could have a few
547 	 * unused swap blocks.  if so, free them now.
548 	 */
549 
550 	if (nused < nallocated) {
551 		if (!now) {
552 			return;
553 		}
554 		uvm_swap_free(slot + nused, nallocated - nused);
555 	}
556 
557 	/*
558 	 * now start the pageout.
559 	 */
560 
561 	if (nused > 0) {
562 		uvmexp.pdpageouts++;
563 		uvm_pageout_start(nused);
564 		error = uvm_swap_put(slot, swc->swc_pages, nused, 0);
565 		KASSERT(error == 0 || error == ENOMEM);
566 	}
567 
568 	/*
569 	 * zero swslot to indicate that we are
570 	 * no longer building a swap-backed cluster.
571 	 */
572 
573 	swc->swc_slot = 0;
574 	swc->swc_nused = 0;
575 }
576 
577 static int
578 swapcluster_nused(struct swapcluster *swc)
579 {
580 
581 	return swc->swc_nused;
582 }
583 
584 /*
585  * uvmpd_dropswap: free any swap allocated to this page.
586  *
587  * => called with owner locked.
588  * => return true if a page had an associated slot.
589  */
590 
591 bool
592 uvmpd_dropswap(struct vm_page *pg)
593 {
594 	bool result = false;
595 	struct vm_anon *anon = pg->uanon;
596 
597 	if ((pg->flags & PG_ANON) && anon->an_swslot) {
598 		uvm_swap_free(anon->an_swslot, 1);
599 		anon->an_swslot = 0;
600 		uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
601 		result = true;
602 	} else if (pg->flags & PG_AOBJ) {
603 		int slot = uao_set_swslot(pg->uobject,
604 		    pg->offset >> PAGE_SHIFT, 0);
605 		if (slot) {
606 			uvm_swap_free(slot, 1);
607 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
608 			result = true;
609 		}
610 	}
611 
612 	return result;
613 }
614 
615 #endif /* defined(VMSWAP) */
616 
617 /*
618  * uvmpd_scan_queue: scan an replace candidate list for pages
619  * to clean or free.
620  *
621  * => we work on meeting our free target by converting inactive pages
622  *    into free pages.
623  * => we handle the building of swap-backed clusters
624  */
625 
626 static void
627 uvmpd_scan_queue(void)
628 {
629 	struct vm_page *p;
630 	struct uvm_object *uobj;
631 	struct vm_anon *anon;
632 #if defined(VMSWAP)
633 	struct swapcluster swc;
634 #endif /* defined(VMSWAP) */
635 	int dirtyreacts;
636 	krwlock_t *slock;
637 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
638 
639 	/*
640 	 * swslot is non-zero if we are building a swap cluster.  we want
641 	 * to stay in the loop while we have a page to scan or we have
642 	 * a swap-cluster to build.
643 	 */
644 
645 #if defined(VMSWAP)
646 	swapcluster_init(&swc);
647 #endif /* defined(VMSWAP) */
648 
649 	dirtyreacts = 0;
650 	uvmpdpol_scaninit();
651 
652 	while (/* CONSTCOND */ 1) {
653 
654 		/*
655 		 * see if we've met the free target.
656 		 */
657 
658 		if (uvm_availmem(false) + uvmexp.paging
659 #if defined(VMSWAP)
660 		    + swapcluster_nused(&swc)
661 #endif /* defined(VMSWAP) */
662 		    >= uvmexp.freetarg << 2 ||
663 		    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
664 			UVMHIST_LOG(pdhist,"  met free target: "
665 				    "exit loop", 0, 0, 0, 0);
666 			break;
667 		}
668 
669 		/*
670 		 * first we have the pdpolicy select a victim page
671 		 * and attempt to lock the object that the page
672 		 * belongs to.  if our attempt fails we skip on to
673 		 * the next page (no harm done).  it is important to
674 		 * "try" locking the object as we are locking in the
675 		 * wrong order (pageq -> object) and we don't want to
676 		 * deadlock.
677 		 *
678 		 * the only time we expect to see an ownerless page
679 		 * (i.e. a page with no uobject and !PG_ANON) is if an
680 		 * anon has loaned a page from a uvm_object and the
681 		 * uvm_object has dropped the ownership.  in that
682 		 * case, the anon can "take over" the loaned page
683 		 * and make it its own.
684 		 */
685 
686 		p = uvmpdpol_selectvictim(&slock);
687 		if (p == NULL) {
688 			break;
689 		}
690 		KASSERT(uvmpdpol_pageisqueued_p(p));
691 		KASSERT(uvm_page_owner_locked_p(p, true));
692 		KASSERT(p->wire_count == 0);
693 
694 		/*
695 		 * we are below target and have a new page to consider.
696 		 */
697 
698 		anon = p->uanon;
699 		uobj = p->uobject;
700 
701 		if (p->flags & PG_BUSY) {
702 			rw_exit(slock);
703 			uvmexp.pdbusy++;
704 			continue;
705 		}
706 
707 		/* does the page belong to an object? */
708 		if (uobj != NULL) {
709 			uvmexp.pdobscan++;
710 		} else {
711 #if defined(VMSWAP)
712 			KASSERT(anon != NULL);
713 			uvmexp.pdanscan++;
714 #else /* defined(VMSWAP) */
715 			panic("%s: anon", __func__);
716 #endif /* defined(VMSWAP) */
717 		}
718 
719 
720 		/*
721 		 * we now have the object locked.
722 		 * if the page is not swap-backed, call the object's
723 		 * pager to flush and free the page.
724 		 */
725 
726 #if defined(READAHEAD_STATS)
727 		if ((p->flags & PG_READAHEAD) != 0) {
728 			p->flags &= ~PG_READAHEAD;
729 			uvm_ra_miss.ev_count++;
730 		}
731 #endif /* defined(READAHEAD_STATS) */
732 
733 		if ((p->flags & PG_SWAPBACKED) == 0) {
734 			KASSERT(uobj != NULL);
735 			(void) (uobj->pgops->pgo_put)(uobj, p->offset,
736 			    p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE);
737 			continue;
738 		}
739 
740 		/*
741 		 * the page is swap-backed.  remove all the permissions
742 		 * from the page so we can sync the modified info
743 		 * without any race conditions.  if the page is clean
744 		 * we can free it now and continue.
745 		 */
746 
747 		pmap_page_protect(p, VM_PROT_NONE);
748 		if (uvm_pagegetdirty(p) == UVM_PAGE_STATUS_UNKNOWN) {
749 			if (pmap_clear_modify(p)) {
750 				uvm_pagemarkdirty(p, UVM_PAGE_STATUS_DIRTY);
751 			} else {
752 				uvm_pagemarkdirty(p, UVM_PAGE_STATUS_CLEAN);
753 			}
754 		}
755 		if (uvm_pagegetdirty(p) != UVM_PAGE_STATUS_DIRTY) {
756 			int slot;
757 			int pageidx;
758 
759 			pageidx = p->offset >> PAGE_SHIFT;
760 			uvm_pagefree(p);
761 			atomic_inc_uint(&uvmexp.pdfreed);
762 
763 			/*
764 			 * for anons, we need to remove the page
765 			 * from the anon ourselves.  for aobjs,
766 			 * pagefree did that for us.
767 			 */
768 
769 			if (anon) {
770 				KASSERT(anon->an_swslot != 0);
771 				anon->an_page = NULL;
772 				slot = anon->an_swslot;
773 			} else {
774 				slot = uao_find_swslot(uobj, pageidx);
775 			}
776 			if (slot > 0) {
777 				/* this page is now only in swap. */
778 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
779 				atomic_inc_uint(&uvmexp.swpgonly);
780 			}
781 			rw_exit(slock);
782 			continue;
783 		}
784 
785 #if defined(VMSWAP)
786 		/*
787 		 * this page is dirty, skip it if we'll have met our
788 		 * free target when all the current pageouts complete.
789 		 */
790 
791 		if (uvm_availmem(false) + uvmexp.paging >
792 		    uvmexp.freetarg << 2) {
793 			rw_exit(slock);
794 			continue;
795 		}
796 
797 		/*
798 		 * free any swap space allocated to the page since
799 		 * we'll have to write it again with its new data.
800 		 */
801 
802 		uvmpd_dropswap(p);
803 
804 		/*
805 		 * start new swap pageout cluster (if necessary).
806 		 *
807 		 * if swap is full reactivate this page so that
808 		 * we eventually cycle all pages through the
809 		 * inactive queue.
810 		 */
811 
812 		if (swapcluster_allocslots(&swc)) {
813 			dirtyreacts++;
814 			uvm_pagelock(p);
815 			uvm_pageactivate(p);
816 			uvm_pageunlock(p);
817 			rw_exit(slock);
818 			continue;
819 		}
820 
821 		/*
822 		 * at this point, we're definitely going reuse this
823 		 * page.  mark the page busy and delayed-free.
824 		 * we should remove the page from the page queues
825 		 * so we don't ever look at it again.
826 		 * adjust counters and such.
827 		 */
828 
829 		p->flags |= PG_BUSY;
830 		UVM_PAGE_OWN(p, "scan_queue");
831 		p->flags |= PG_PAGEOUT;
832 		uvmexp.pgswapout++;
833 
834 		uvm_pagelock(p);
835 		uvm_pagedequeue(p);
836 		uvm_pageunlock(p);
837 
838 		/*
839 		 * add the new page to the cluster.
840 		 */
841 
842 		if (swapcluster_add(&swc, p)) {
843 			p->flags &= ~(PG_BUSY|PG_PAGEOUT);
844 			UVM_PAGE_OWN(p, NULL);
845 			dirtyreacts++;
846 			uvm_pagelock(p);
847 			uvm_pageactivate(p);
848 			uvm_pageunlock(p);
849 			rw_exit(slock);
850 			continue;
851 		}
852 		rw_exit(slock);
853 
854 		swapcluster_flush(&swc, false);
855 
856 		/*
857 		 * the pageout is in progress.  bump counters and set up
858 		 * for the next loop.
859 		 */
860 
861 		atomic_inc_uint(&uvmexp.pdpending);
862 
863 #else /* defined(VMSWAP) */
864 		uvm_pagelock(p);
865 		uvm_pageactivate(p);
866 		uvm_pageunlock(p);
867 		rw_exit(slock);
868 #endif /* defined(VMSWAP) */
869 	}
870 
871 	uvmpdpol_scanfini();
872 
873 #if defined(VMSWAP)
874 	swapcluster_flush(&swc, true);
875 #endif /* defined(VMSWAP) */
876 }
877 
878 /*
879  * uvmpd_scan: scan the page queues and attempt to meet our targets.
880  */
881 
882 static void
883 uvmpd_scan(void)
884 {
885 	int swap_shortage, pages_freed, fpages;
886 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
887 
888 	uvmexp.pdrevs++;
889 
890 	/*
891 	 * work on meeting our targets.   first we work on our free target
892 	 * by converting inactive pages into free pages.  then we work on
893 	 * meeting our inactive target by converting active pages to
894 	 * inactive ones.
895 	 */
896 
897 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
898 
899 	pages_freed = uvmexp.pdfreed;
900 	uvmpd_scan_queue();
901 	pages_freed = uvmexp.pdfreed - pages_freed;
902 
903 	/*
904 	 * detect if we're not going to be able to page anything out
905 	 * until we free some swap resources from active pages.
906 	 */
907 
908 	swap_shortage = 0;
909 	fpages = uvm_availmem(false);
910 	if (fpages < uvmexp.freetarg &&
911 	    uvmexp.swpginuse >= uvmexp.swpgavail &&
912 	    !uvm_swapisfull() &&
913 	    pages_freed == 0) {
914 		swap_shortage = uvmexp.freetarg - fpages;
915 	}
916 
917 	uvmpdpol_balancequeue(swap_shortage);
918 
919 	/*
920 	 * if still below the minimum target, try unloading kernel
921 	 * modules.
922 	 */
923 
924 	if (uvm_availmem(false) < uvmexp.freemin) {
925 		module_thread_kick();
926 	}
927 }
928 
929 /*
930  * uvm_reclaimable: decide whether to wait for pagedaemon.
931  *
932  * => return true if it seems to be worth to do uvm_wait.
933  *
934  * XXX should be tunable.
935  * XXX should consider pools, etc?
936  */
937 
938 bool
939 uvm_reclaimable(void)
940 {
941 	int filepages;
942 	int active, inactive;
943 
944 	/*
945 	 * if swap is not full, no problem.
946 	 */
947 
948 	if (!uvm_swapisfull()) {
949 		return true;
950 	}
951 
952 	/*
953 	 * file-backed pages can be reclaimed even when swap is full.
954 	 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim.
955 	 * NB: filepages calculation does not exclude EXECPAGES - intentional.
956 	 *
957 	 * XXX assume the worst case, ie. all wired pages are file-backed.
958 	 *
959 	 * XXX should consider about other reclaimable memory.
960 	 * XXX ie. pools, traditional buffer cache.
961 	 */
962 
963 	cpu_count_sync(false);
964 	filepages = (int)(cpu_count_get(CPU_COUNT_FILECLEAN) +
965 	    cpu_count_get(CPU_COUNT_FILEUNKNOWN) +
966 	    cpu_count_get(CPU_COUNT_FILEDIRTY) - uvmexp.wired);
967 	uvm_estimatepageable(&active, &inactive);
968 	if (filepages >= MIN((active + inactive) >> 4,
969 	    5 * 1024 * 1024 >> PAGE_SHIFT)) {
970 		return true;
971 	}
972 
973 	/*
974 	 * kill the process, fail allocation, etc..
975 	 */
976 
977 	return false;
978 }
979 
980 void
981 uvm_estimatepageable(int *active, int *inactive)
982 {
983 
984 	uvmpdpol_estimatepageable(active, inactive);
985 }
986 
987 
988 /*
989  * Use a separate thread for draining pools.
990  * This work can't done from the main pagedaemon thread because
991  * some pool allocators need to take vm_map locks.
992  */
993 
994 static void
995 uvmpd_pool_drain_thread(void *arg)
996 {
997 	struct pool *firstpool, *curpool;
998 	int bufcnt, lastslept;
999 	bool cycled;
1000 
1001 	firstpool = NULL;
1002 	cycled = true;
1003 	for (;;) {
1004 		/*
1005 		 * sleep until awoken by the pagedaemon.
1006 		 */
1007 		mutex_enter(&uvmpd_lock);
1008 		if (!uvmpd_pool_drain_run) {
1009 			lastslept = getticks();
1010 			cv_wait(&uvmpd_pool_drain_cv, &uvmpd_lock);
1011 			if (getticks() != lastslept) {
1012 				cycled = false;
1013 				firstpool = NULL;
1014 			}
1015 		}
1016 		uvmpd_pool_drain_run = false;
1017 		mutex_exit(&uvmpd_lock);
1018 
1019 		/*
1020 		 * rate limit draining, otherwise in desperate circumstances
1021 		 * this can totally saturate the system with xcall activity.
1022 		 */
1023 		if (cycled) {
1024 			kpause("uvmpdlmt", false, 1, NULL);
1025 			cycled = false;
1026 			firstpool = NULL;
1027 		}
1028 
1029 		/*
1030 		 * drain and temporarily disable the freelist cache.
1031 		 */
1032 		uvm_pgflcache_pause();
1033 
1034 		/*
1035 		 * kill unused metadata buffers.
1036 		 */
1037 		bufcnt = uvmexp.freetarg - uvm_availmem(false);
1038 		if (bufcnt < 0)
1039 			bufcnt = 0;
1040 
1041 		mutex_enter(&bufcache_lock);
1042 		buf_drain(bufcnt << PAGE_SHIFT);
1043 		mutex_exit(&bufcache_lock);
1044 
1045 		/*
1046 		 * drain a pool, and then re-enable the freelist cache.
1047 		 */
1048 		(void)pool_drain(&curpool);
1049 		KASSERT(curpool != NULL);
1050 		if (firstpool == NULL) {
1051 			firstpool = curpool;
1052 		} else if (firstpool == curpool) {
1053 			cycled = true;
1054 		}
1055 		uvm_pgflcache_resume();
1056 	}
1057 	/*NOTREACHED*/
1058 }
1059 
1060 static void
1061 uvmpd_pool_drain_wakeup(void)
1062 {
1063 
1064 	mutex_enter(&uvmpd_lock);
1065 	uvmpd_pool_drain_run = true;
1066 	cv_signal(&uvmpd_pool_drain_cv);
1067 	mutex_exit(&uvmpd_lock);
1068 }
1069