xref: /netbsd-src/sys/uvm/uvm_pdaemon.c (revision 3007d967c5768f8d3944c8609318efbc4c6345a4)
1 /*	$NetBSD: uvm_pdaemon.c,v 1.134 2023/09/10 15:01:11 ad Exp $	*/
2 
3 /*
4  * Copyright (c) 1997 Charles D. Cranor and Washington University.
5  * Copyright (c) 1991, 1993, The Regents of the University of California.
6  *
7  * All rights reserved.
8  *
9  * This code is derived from software contributed to Berkeley by
10  * The Mach Operating System project at Carnegie-Mellon University.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. 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 1.134 2023/09/10 15:01:11 ad 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 #ifndef UVMHIST_PDHIST_SIZE
92 #define UVMHIST_PDHIST_SIZE 100
93 #endif
94 static struct kern_history_ent pdhistbuf[UVMHIST_PDHIST_SIZE];
95 UVMHIST_DEFINE(pdhist) = UVMHIST_INITIALIZER(pdhisthist, pdhistbuf);
96 #endif
97 
98 /*
99  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
100  * in a pass thru the inactive list when swap is full.  the value should be
101  * "small"... if it's too large we'll cycle the active pages thru the inactive
102  * queue too quickly to for them to be referenced and avoid being freed.
103  */
104 
105 #define	UVMPD_NUMDIRTYREACTS	16
106 
107 /*
108  * local prototypes
109  */
110 
111 static void	uvmpd_scan(void);
112 static void	uvmpd_scan_queue(void);
113 static void	uvmpd_tune(void);
114 static void	uvmpd_pool_drain_thread(void *);
115 static void	uvmpd_pool_drain_wakeup(void);
116 
117 static unsigned int uvm_pagedaemon_waiters;
118 
119 /* State for the pool drainer thread */
120 static kmutex_t uvmpd_lock __cacheline_aligned;
121 static kcondvar_t uvmpd_pool_drain_cv;
122 static bool uvmpd_pool_drain_run = false;
123 
124 /*
125  * XXX hack to avoid hangs when large processes fork.
126  */
127 u_int uvm_extrapages;
128 
129 /*
130  * uvm_wait: wait (sleep) for the page daemon to free some pages
131  *
132  * => should be called with all locks released
133  * => should _not_ be called by the page daemon (to avoid deadlock)
134  */
135 
136 void
uvm_wait(const char * wmsg)137 uvm_wait(const char *wmsg)
138 {
139 	int timo = 0;
140 
141 	if (uvm.pagedaemon_lwp == NULL)
142 		panic("out of memory before the pagedaemon thread exists");
143 
144 	mutex_spin_enter(&uvmpd_lock);
145 
146 	/*
147 	 * check for page daemon going to sleep (waiting for itself)
148 	 */
149 
150 	if (curlwp == uvm.pagedaemon_lwp && uvmexp.paging == 0) {
151 		/*
152 		 * now we have a problem: the pagedaemon wants to go to
153 		 * sleep until it frees more memory.   but how can it
154 		 * free more memory if it is asleep?  that is a deadlock.
155 		 * we have two options:
156 		 *  [1] panic now
157 		 *  [2] put a timeout on the sleep, thus causing the
158 		 *      pagedaemon to only pause (rather than sleep forever)
159 		 *
160 		 * note that option [2] will only help us if we get lucky
161 		 * and some other process on the system breaks the deadlock
162 		 * by exiting or freeing memory (thus allowing the pagedaemon
163 		 * to continue).  for now we panic if DEBUG is defined,
164 		 * otherwise we hope for the best with option [2] (better
165 		 * yet, this should never happen in the first place!).
166 		 */
167 
168 		printf("pagedaemon: deadlock detected!\n");
169 		timo = hz >> 3;		/* set timeout */
170 #if defined(DEBUG)
171 		/* DEBUG: panic so we can debug it */
172 		panic("pagedaemon deadlock");
173 #endif
174 	}
175 
176 	uvm_pagedaemon_waiters++;
177 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
178 	UVM_UNLOCK_AND_WAIT(&uvmexp.free, &uvmpd_lock, false, wmsg, timo);
179 }
180 
181 /*
182  * uvm_kick_pdaemon: perform checks to determine if we need to
183  * give the pagedaemon a nudge, and do so if necessary.
184  */
185 
186 void
uvm_kick_pdaemon(void)187 uvm_kick_pdaemon(void)
188 {
189 	int fpages = uvm_availmem(false);
190 
191 	if (fpages + uvmexp.paging < uvmexp.freemin ||
192 	    (fpages + uvmexp.paging < uvmexp.freetarg &&
193 	     uvmpdpol_needsscan_p()) ||
194 	     uvm_km_va_starved_p()) {
195 	     	mutex_spin_enter(&uvmpd_lock);
196 		wakeup(&uvm.pagedaemon);
197 	     	mutex_spin_exit(&uvmpd_lock);
198 	}
199 }
200 
201 /*
202  * uvmpd_tune: tune paging parameters
203  *
204  * => called when ever memory is added (or removed?) to the system
205  */
206 
207 static void
uvmpd_tune(void)208 uvmpd_tune(void)
209 {
210 	int val;
211 
212 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
213 
214 	/*
215 	 * try to keep 0.5% of available RAM free, but limit to between
216 	 * 128k and 1024k per-CPU.  XXX: what are these values good for?
217 	 */
218 	val = uvmexp.npages / 200;
219 	val = MAX(val, (128*1024) >> PAGE_SHIFT);
220 	val = MIN(val, (1024*1024) >> PAGE_SHIFT);
221 	val *= ncpu;
222 
223 	/* Make sure there's always a user page free. */
224 	if (val < uvmexp.reserve_kernel + 1)
225 		val = uvmexp.reserve_kernel + 1;
226 	uvmexp.freemin = val;
227 
228 	/* Calculate free target. */
229 	val = (uvmexp.freemin * 4) / 3;
230 	if (val <= uvmexp.freemin)
231 		val = uvmexp.freemin + 1;
232 	uvmexp.freetarg = val + atomic_swap_uint(&uvm_extrapages, 0);
233 
234 	uvmexp.wiredmax = uvmexp.npages / 3;
235 	UVMHIST_LOG(pdhist, "<- done, freemin=%jd, freetarg=%jd, wiredmax=%jd",
236 	      uvmexp.freemin, uvmexp.freetarg, uvmexp.wiredmax, 0);
237 }
238 
239 /*
240  * uvm_pageout: the main loop for the pagedaemon
241  */
242 
243 void
uvm_pageout(void * arg)244 uvm_pageout(void *arg)
245 {
246 	int npages = 0;
247 	int extrapages = 0;
248 	int fpages;
249 
250 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
251 
252 	UVMHIST_LOG(pdhist,"<starting uvm pagedaemon>", 0, 0, 0, 0);
253 
254 	mutex_init(&uvmpd_lock, MUTEX_DEFAULT, IPL_VM);
255 	cv_init(&uvmpd_pool_drain_cv, "pooldrain");
256 
257 	/* Create the pool drainer kernel thread. */
258 	if (kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL,
259 	    uvmpd_pool_drain_thread, NULL, NULL, "pooldrain"))
260 		panic("fork pooldrain");
261 
262 	/*
263 	 * ensure correct priority and set paging parameters...
264 	 */
265 
266 	uvm.pagedaemon_lwp = curlwp;
267 	npages = uvmexp.npages;
268 	uvmpd_tune();
269 
270 	/*
271 	 * main loop
272 	 */
273 
274 	for (;;) {
275 		bool needsscan, needsfree, kmem_va_starved;
276 
277 		kmem_va_starved = uvm_km_va_starved_p();
278 
279 		mutex_spin_enter(&uvmpd_lock);
280 		if ((uvm_pagedaemon_waiters == 0 || uvmexp.paging > 0) &&
281 		    !kmem_va_starved) {
282 			UVMHIST_LOG(pdhist,"  <<SLEEPING>>",0,0,0,0);
283 			UVM_UNLOCK_AND_WAIT(&uvm.pagedaemon,
284 			    &uvmpd_lock, false, "pgdaemon", 0);
285 			uvmexp.pdwoke++;
286 			UVMHIST_LOG(pdhist,"  <<WOKE UP>>",0,0,0,0);
287 		} else {
288 			mutex_spin_exit(&uvmpd_lock);
289 		}
290 
291 		/*
292 		 * now recompute inactive count
293 		 */
294 
295 		if (npages != uvmexp.npages || extrapages != uvm_extrapages) {
296 			npages = uvmexp.npages;
297 			extrapages = uvm_extrapages;
298 			uvmpd_tune();
299 		}
300 
301 		uvmpdpol_tune();
302 
303 		/*
304 		 * Estimate a hint.  Note that bufmem are returned to
305 		 * system only when entire pool page is empty.
306 		 */
307 		fpages = uvm_availmem(false);
308 		UVMHIST_LOG(pdhist,"  free/ftarg=%jd/%jd",
309 		    fpages, uvmexp.freetarg, 0,0);
310 
311 		needsfree = fpages + uvmexp.paging < uvmexp.freetarg;
312 		needsscan = needsfree || uvmpdpol_needsscan_p();
313 
314 		/*
315 		 * scan if needed
316 		 */
317 		if (needsscan) {
318 			uvmpd_scan();
319 		}
320 
321 		/*
322 		 * if there's any free memory to be had,
323 		 * wake up any waiters.
324 		 */
325 		if (uvm_availmem(false) > uvmexp.reserve_kernel ||
326 		    uvmexp.paging == 0) {
327 			mutex_spin_enter(&uvmpd_lock);
328 			wakeup(&uvmexp.free);
329 			uvm_pagedaemon_waiters = 0;
330 			mutex_spin_exit(&uvmpd_lock);
331 		}
332 
333 		/*
334 		 * scan done.  if we don't need free memory, we're done.
335 		 */
336 
337 		if (!needsfree && !kmem_va_starved)
338 			continue;
339 
340 		/*
341 		 * kick the pool drainer thread.
342 		 */
343 
344 		uvmpd_pool_drain_wakeup();
345 	}
346 	/*NOTREACHED*/
347 }
348 
349 void
uvm_pageout_start(int npages)350 uvm_pageout_start(int npages)
351 {
352 
353 	atomic_add_int(&uvmexp.paging, npages);
354 }
355 
356 void
uvm_pageout_done(int npages)357 uvm_pageout_done(int npages)
358 {
359 
360 	KASSERT(atomic_load_relaxed(&uvmexp.paging) >= npages);
361 
362 	if (npages == 0) {
363 		return;
364 	}
365 
366 	atomic_add_int(&uvmexp.paging, -npages);
367 
368 	/*
369 	 * wake up either of pagedaemon or LWPs waiting for it.
370 	 */
371 
372 	mutex_spin_enter(&uvmpd_lock);
373 	if (uvm_availmem(false) <= uvmexp.reserve_kernel) {
374 		wakeup(&uvm.pagedaemon);
375 	} else if (uvm_pagedaemon_waiters != 0) {
376 		wakeup(&uvmexp.free);
377 		uvm_pagedaemon_waiters = 0;
378 	}
379 	mutex_spin_exit(&uvmpd_lock);
380 }
381 
382 static krwlock_t *
uvmpd_page_owner_lock(struct vm_page * pg)383 uvmpd_page_owner_lock(struct vm_page *pg)
384 {
385 	struct uvm_object *uobj = pg->uobject;
386 	struct vm_anon *anon = pg->uanon;
387 	krwlock_t *slock;
388 
389 	KASSERT(mutex_owned(&pg->interlock));
390 
391 #ifdef DEBUG
392 	if (uobj == (void *)0xdeadbeef || anon == (void *)0xdeadbeef) {
393 		return NULL;
394 	}
395 #endif
396 	if (uobj != NULL) {
397 		slock = uobj->vmobjlock;
398 		KASSERTMSG(slock != NULL, "pg %p uobj %p, NULL lock", pg, uobj);
399 	} else if (anon != NULL) {
400 		slock = anon->an_lock;
401 		KASSERTMSG(slock != NULL, "pg %p anon %p, NULL lock", pg, anon);
402 	} else {
403 		slock = NULL;
404 	}
405 	return slock;
406 }
407 
408 /*
409  * uvmpd_trylockowner: trylock the page's owner.
410  *
411  * => called with page interlock held.
412  * => resolve orphaned O->A loaned page.
413  * => return the locked mutex on success.  otherwise, return NULL.
414  */
415 
416 krwlock_t *
uvmpd_trylockowner(struct vm_page * pg)417 uvmpd_trylockowner(struct vm_page *pg)
418 {
419 	krwlock_t *slock, *heldslock = NULL;
420 
421 	KASSERT(mutex_owned(&pg->interlock));
422 
423 	slock = uvmpd_page_owner_lock(pg);
424 	if (slock == NULL) {
425 		/* Page may be in state of flux - ignore. */
426 		mutex_exit(&pg->interlock);
427 		return NULL;
428 	}
429 
430 	if (rw_tryenter(slock, RW_WRITER)) {
431 		goto success;
432 	}
433 
434 	/*
435 	 * The try-lock didn't work, so now do a blocking lock after
436 	 * dropping the page interlock.  Prevent the owner lock from
437 	 * being freed by taking a hold on it first.
438 	 */
439 
440 	rw_obj_hold(slock);
441 	mutex_exit(&pg->interlock);
442 	rw_enter(slock, RW_WRITER);
443 	heldslock = slock;
444 
445 	/*
446 	 * Now we hold some owner lock.  Check if the lock we hold
447 	 * is still the lock for the owner of the page.
448 	 * If it is then return it, otherwise release it and return NULL.
449 	 */
450 
451 	mutex_enter(&pg->interlock);
452 	slock = uvmpd_page_owner_lock(pg);
453 	if (heldslock != slock) {
454 		rw_exit(heldslock);
455 		slock = NULL;
456 	} else {
457 success:
458 		/*
459 		 * Set PG_ANON if it isn't set already.
460 		 */
461 		if (pg->uobject == NULL && (pg->flags & PG_ANON) == 0) {
462 			KASSERT(pg->loan_count > 0);
463 			pg->loan_count--;
464 			pg->flags |= PG_ANON;
465 			/* anon now owns it */
466 		}
467 	}
468 	mutex_exit(&pg->interlock);
469 	if (heldslock != NULL) {
470 		rw_obj_free(heldslock);
471 	}
472 	return slock;
473 }
474 
475 #if defined(VMSWAP)
476 struct swapcluster {
477 	int swc_slot;
478 	int swc_nallocated;
479 	int swc_nused;
480 	struct vm_page *swc_pages[howmany(MAXPHYS, MIN_PAGE_SIZE)];
481 };
482 
483 static void
swapcluster_init(struct swapcluster * swc)484 swapcluster_init(struct swapcluster *swc)
485 {
486 
487 	swc->swc_slot = 0;
488 	swc->swc_nused = 0;
489 }
490 
491 static int
swapcluster_allocslots(struct swapcluster * swc)492 swapcluster_allocslots(struct swapcluster *swc)
493 {
494 	int slot;
495 	int npages;
496 
497 	if (swc->swc_slot != 0) {
498 		return 0;
499 	}
500 
501 	/* Even with strange MAXPHYS, the shift
502 	   implicitly rounds down to a page. */
503 	npages = MAXPHYS >> PAGE_SHIFT;
504 	slot = uvm_swap_alloc(&npages, true);
505 	if (slot == 0) {
506 		return ENOMEM;
507 	}
508 	swc->swc_slot = slot;
509 	swc->swc_nallocated = npages;
510 	swc->swc_nused = 0;
511 
512 	return 0;
513 }
514 
515 static int
swapcluster_add(struct swapcluster * swc,struct vm_page * pg)516 swapcluster_add(struct swapcluster *swc, struct vm_page *pg)
517 {
518 	int slot;
519 	struct uvm_object *uobj;
520 
521 	KASSERT(swc->swc_slot != 0);
522 	KASSERT(swc->swc_nused < swc->swc_nallocated);
523 	KASSERT((pg->flags & PG_SWAPBACKED) != 0);
524 
525 	slot = swc->swc_slot + swc->swc_nused;
526 	uobj = pg->uobject;
527 	if (uobj == NULL) {
528 		KASSERT(rw_write_held(pg->uanon->an_lock));
529 		pg->uanon->an_swslot = slot;
530 	} else {
531 		int result;
532 
533 		KASSERT(rw_write_held(uobj->vmobjlock));
534 		result = uao_set_swslot(uobj, pg->offset >> PAGE_SHIFT, slot);
535 		if (result == -1) {
536 			return ENOMEM;
537 		}
538 	}
539 	swc->swc_pages[swc->swc_nused] = pg;
540 	swc->swc_nused++;
541 
542 	return 0;
543 }
544 
545 static void
swapcluster_flush(struct swapcluster * swc,bool now)546 swapcluster_flush(struct swapcluster *swc, bool now)
547 {
548 	int slot;
549 	int nused;
550 	int nallocated;
551 	int error __diagused;
552 
553 	if (swc->swc_slot == 0) {
554 		return;
555 	}
556 	KASSERT(swc->swc_nused <= swc->swc_nallocated);
557 
558 	slot = swc->swc_slot;
559 	nused = swc->swc_nused;
560 	nallocated = swc->swc_nallocated;
561 
562 	/*
563 	 * if this is the final pageout we could have a few
564 	 * unused swap blocks.  if so, free them now.
565 	 */
566 
567 	if (nused < nallocated) {
568 		if (!now) {
569 			return;
570 		}
571 		uvm_swap_free(slot + nused, nallocated - nused);
572 	}
573 
574 	/*
575 	 * now start the pageout.
576 	 */
577 
578 	if (nused > 0) {
579 		uvmexp.pdpageouts++;
580 		uvm_pageout_start(nused);
581 		error = uvm_swap_put(slot, swc->swc_pages, nused, 0);
582 		KASSERT(error == 0 || error == ENOMEM);
583 	}
584 
585 	/*
586 	 * zero swslot to indicate that we are
587 	 * no longer building a swap-backed cluster.
588 	 */
589 
590 	swc->swc_slot = 0;
591 	swc->swc_nused = 0;
592 }
593 
594 static int
swapcluster_nused(struct swapcluster * swc)595 swapcluster_nused(struct swapcluster *swc)
596 {
597 
598 	return swc->swc_nused;
599 }
600 
601 /*
602  * uvmpd_dropswap: free any swap allocated to this page.
603  *
604  * => called with owner locked.
605  * => return true if a page had an associated slot.
606  */
607 
608 bool
uvmpd_dropswap(struct vm_page * pg)609 uvmpd_dropswap(struct vm_page *pg)
610 {
611 	bool result = false;
612 	struct vm_anon *anon = pg->uanon;
613 
614 	if ((pg->flags & PG_ANON) && anon->an_swslot) {
615 		uvm_swap_free(anon->an_swslot, 1);
616 		anon->an_swslot = 0;
617 		uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
618 		result = true;
619 	} else if (pg->flags & PG_AOBJ) {
620 		int slot = uao_set_swslot(pg->uobject,
621 		    pg->offset >> PAGE_SHIFT, 0);
622 		if (slot) {
623 			uvm_swap_free(slot, 1);
624 			uvm_pagemarkdirty(pg, UVM_PAGE_STATUS_DIRTY);
625 			result = true;
626 		}
627 	}
628 
629 	return result;
630 }
631 
632 #endif /* defined(VMSWAP) */
633 
634 /*
635  * uvmpd_scan_queue: scan an replace candidate list for pages
636  * to clean or free.
637  *
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
uvmpd_scan_queue(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 	krwlock_t *slock;
654 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
655 
656 	/*
657 	 * swslot is non-zero if we are building a swap cluster.  we want
658 	 * to stay in the loop while we have a page to scan or we have
659 	 * a swap-cluster to build.
660 	 */
661 
662 #if defined(VMSWAP)
663 	swapcluster_init(&swc);
664 #endif /* defined(VMSWAP) */
665 
666 	dirtyreacts = 0;
667 	uvmpdpol_scaninit();
668 
669 	while (/* CONSTCOND */ 1) {
670 
671 		/*
672 		 * see if we've met the free target.
673 		 */
674 
675 		if (uvm_availmem(false) + uvmexp.paging
676 #if defined(VMSWAP)
677 		    + swapcluster_nused(&swc)
678 #endif /* defined(VMSWAP) */
679 		    >= uvmexp.freetarg << 2 ||
680 		    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
681 			UVMHIST_LOG(pdhist,"  met free target: "
682 				    "exit loop", 0, 0, 0, 0);
683 			break;
684 		}
685 
686 		/*
687 		 * first we have the pdpolicy select a victim page
688 		 * and attempt to lock the object that the page
689 		 * belongs to.  if our attempt fails we skip on to
690 		 * the next page (no harm done).  it is important to
691 		 * "try" locking the object as we are locking in the
692 		 * wrong order (pageq -> object) and we don't want to
693 		 * deadlock.
694 		 *
695 		 * the only time we expect to see an ownerless page
696 		 * (i.e. a page with no uobject and !PG_ANON) is if an
697 		 * anon has loaned a page from a uvm_object and the
698 		 * uvm_object has dropped the ownership.  in that
699 		 * case, the anon can "take over" the loaned page
700 		 * and make it its own.
701 		 */
702 
703 		p = uvmpdpol_selectvictim(&slock);
704 		if (p == NULL) {
705 			break;
706 		}
707 		KASSERT(uvmpdpol_pageisqueued_p(p));
708 		KASSERT(uvm_page_owner_locked_p(p, true));
709 		KASSERT(p->wire_count == 0);
710 
711 		/*
712 		 * we are below target and have a new page to consider.
713 		 */
714 
715 		anon = p->uanon;
716 		uobj = p->uobject;
717 
718 		if (p->flags & PG_BUSY) {
719 			rw_exit(slock);
720 			uvmexp.pdbusy++;
721 			continue;
722 		}
723 
724 		/* does the page belong to an object? */
725 		if (uobj != NULL) {
726 			uvmexp.pdobscan++;
727 		} else {
728 #if defined(VMSWAP)
729 			KASSERT(anon != NULL);
730 			uvmexp.pdanscan++;
731 #else /* defined(VMSWAP) */
732 			panic("%s: anon", __func__);
733 #endif /* defined(VMSWAP) */
734 		}
735 
736 
737 		/*
738 		 * we now have the object locked.
739 		 * if the page is not swap-backed, call the object's
740 		 * pager to flush and free the page.
741 		 */
742 
743 #if defined(READAHEAD_STATS)
744 		if ((p->flags & PG_READAHEAD) != 0) {
745 			p->flags &= ~PG_READAHEAD;
746 			uvm_ra_miss.ev_count++;
747 		}
748 #endif /* defined(READAHEAD_STATS) */
749 
750 		if ((p->flags & PG_SWAPBACKED) == 0) {
751 			KASSERT(uobj != NULL);
752 			(void) (uobj->pgops->pgo_put)(uobj, p->offset,
753 			    p->offset + PAGE_SIZE, PGO_CLEANIT|PGO_FREE);
754 			continue;
755 		}
756 
757 		/*
758 		 * the page is swap-backed.  remove all the permissions
759 		 * from the page so we can sync the modified info
760 		 * without any race conditions.  if the page is clean
761 		 * we can free it now and continue.
762 		 */
763 
764 		pmap_page_protect(p, VM_PROT_NONE);
765 		if (uvm_pagegetdirty(p) == UVM_PAGE_STATUS_UNKNOWN) {
766 			if (pmap_clear_modify(p)) {
767 				uvm_pagemarkdirty(p, UVM_PAGE_STATUS_DIRTY);
768 			} else {
769 				uvm_pagemarkdirty(p, UVM_PAGE_STATUS_CLEAN);
770 			}
771 		}
772 		if (uvm_pagegetdirty(p) != UVM_PAGE_STATUS_DIRTY) {
773 			int slot;
774 			int pageidx;
775 
776 			pageidx = p->offset >> PAGE_SHIFT;
777 			uvm_pagefree(p);
778 			atomic_inc_uint(&uvmexp.pdfreed);
779 
780 			/*
781 			 * for anons, we need to remove the page
782 			 * from the anon ourselves.  for aobjs,
783 			 * pagefree did that for us.
784 			 */
785 
786 			if (anon) {
787 				KASSERT(anon->an_swslot != 0);
788 				anon->an_page = NULL;
789 				slot = anon->an_swslot;
790 			} else {
791 				slot = uao_find_swslot(uobj, pageidx);
792 			}
793 			if (slot > 0) {
794 				/* this page is now only in swap. */
795 				KASSERT(uvmexp.swpgonly < uvmexp.swpginuse);
796 				atomic_inc_uint(&uvmexp.swpgonly);
797 			}
798 			rw_exit(slock);
799 			continue;
800 		}
801 
802 #if defined(VMSWAP)
803 		/*
804 		 * this page is dirty, skip it if we'll have met our
805 		 * free target when all the current pageouts complete.
806 		 */
807 
808 		if (uvm_availmem(false) + uvmexp.paging >
809 		    uvmexp.freetarg << 2) {
810 			rw_exit(slock);
811 			continue;
812 		}
813 
814 		/*
815 		 * free any swap space allocated to the page since
816 		 * we'll have to write it again with its new data.
817 		 */
818 
819 		uvmpd_dropswap(p);
820 
821 		/*
822 		 * start new swap pageout cluster (if necessary).
823 		 *
824 		 * if swap is full reactivate this page so that
825 		 * we eventually cycle all pages through the
826 		 * inactive queue.
827 		 */
828 
829 		if (swapcluster_allocslots(&swc)) {
830 			dirtyreacts++;
831 			uvm_pagelock(p);
832 			uvm_pageactivate(p);
833 			uvm_pageunlock(p);
834 			rw_exit(slock);
835 			continue;
836 		}
837 
838 		/*
839 		 * at this point, we're definitely going reuse this
840 		 * page.  mark the page busy and delayed-free.
841 		 * we should remove the page from the page queues
842 		 * so we don't ever look at it again.
843 		 * adjust counters and such.
844 		 */
845 
846 		p->flags |= PG_BUSY;
847 		UVM_PAGE_OWN(p, "scan_queue");
848 		p->flags |= PG_PAGEOUT;
849 		uvmexp.pgswapout++;
850 
851 		uvm_pagelock(p);
852 		uvm_pagedequeue(p);
853 		uvm_pageunlock(p);
854 
855 		/*
856 		 * add the new page to the cluster.
857 		 */
858 
859 		if (swapcluster_add(&swc, p)) {
860 			p->flags &= ~(PG_BUSY|PG_PAGEOUT);
861 			UVM_PAGE_OWN(p, NULL);
862 			dirtyreacts++;
863 			uvm_pagelock(p);
864 			uvm_pageactivate(p);
865 			uvm_pageunlock(p);
866 			rw_exit(slock);
867 			continue;
868 		}
869 		rw_exit(slock);
870 
871 		swapcluster_flush(&swc, false);
872 
873 		/*
874 		 * the pageout is in progress.  bump counters and set up
875 		 * for the next loop.
876 		 */
877 
878 		atomic_inc_uint(&uvmexp.pdpending);
879 
880 #else /* defined(VMSWAP) */
881 		uvm_pagelock(p);
882 		uvm_pageactivate(p);
883 		uvm_pageunlock(p);
884 		rw_exit(slock);
885 #endif /* defined(VMSWAP) */
886 	}
887 
888 	uvmpdpol_scanfini();
889 
890 #if defined(VMSWAP)
891 	swapcluster_flush(&swc, true);
892 #endif /* defined(VMSWAP) */
893 }
894 
895 /*
896  * uvmpd_scan: scan the page queues and attempt to meet our targets.
897  */
898 
899 static void
uvmpd_scan(void)900 uvmpd_scan(void)
901 {
902 	int swap_shortage, pages_freed, fpages;
903 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(pdhist);
904 
905 	uvmexp.pdrevs++;
906 
907 	/*
908 	 * work on meeting our targets.   first we work on our free target
909 	 * by converting inactive pages into free pages.  then we work on
910 	 * meeting our inactive target by converting active pages to
911 	 * inactive ones.
912 	 */
913 
914 	UVMHIST_LOG(pdhist, "  starting 'free' loop",0,0,0,0);
915 
916 	pages_freed = uvmexp.pdfreed;
917 	uvmpd_scan_queue();
918 	pages_freed = uvmexp.pdfreed - pages_freed;
919 
920 	/*
921 	 * detect if we're not going to be able to page anything out
922 	 * until we free some swap resources from active pages.
923 	 */
924 
925 	swap_shortage = 0;
926 	fpages = uvm_availmem(false);
927 	if (fpages < uvmexp.freetarg &&
928 	    uvmexp.swpginuse >= uvmexp.swpgavail &&
929 	    !uvm_swapisfull() &&
930 	    pages_freed == 0) {
931 		swap_shortage = uvmexp.freetarg - fpages;
932 	}
933 
934 	uvmpdpol_balancequeue(swap_shortage);
935 
936 	/*
937 	 * if still below the minimum target, try unloading kernel
938 	 * modules.
939 	 */
940 
941 	if (uvm_availmem(false) < uvmexp.freemin) {
942 		module_thread_kick();
943 	}
944 }
945 
946 /*
947  * uvm_reclaimable: decide whether to wait for pagedaemon.
948  *
949  * => return true if it seems to be worth to do uvm_wait.
950  *
951  * XXX should be tunable.
952  * XXX should consider pools, etc?
953  */
954 
955 bool
uvm_reclaimable(void)956 uvm_reclaimable(void)
957 {
958 	int filepages;
959 	int active, inactive;
960 
961 	/*
962 	 * if swap is not full, no problem.
963 	 */
964 
965 	if (!uvm_swapisfull()) {
966 		return true;
967 	}
968 
969 	/*
970 	 * file-backed pages can be reclaimed even when swap is full.
971 	 * if we have more than 1/16 of pageable memory or 5MB, try to reclaim.
972 	 * NB: filepages calculation does not exclude EXECPAGES - intentional.
973 	 *
974 	 * XXX assume the worst case, ie. all wired pages are file-backed.
975 	 *
976 	 * XXX should consider about other reclaimable memory.
977 	 * XXX ie. pools, traditional buffer cache.
978 	 */
979 
980 	cpu_count_sync(false);
981 	filepages = (int)(cpu_count_get(CPU_COUNT_FILECLEAN) +
982 	    cpu_count_get(CPU_COUNT_FILEUNKNOWN) +
983 	    cpu_count_get(CPU_COUNT_FILEDIRTY) - uvmexp.wired);
984 	uvm_estimatepageable(&active, &inactive);
985 	if (filepages >= MIN((active + inactive) >> 4,
986 	    5 * 1024 * 1024 >> PAGE_SHIFT)) {
987 		return true;
988 	}
989 
990 	/*
991 	 * kill the process, fail allocation, etc..
992 	 */
993 
994 	return false;
995 }
996 
997 void
uvm_estimatepageable(int * active,int * inactive)998 uvm_estimatepageable(int *active, int *inactive)
999 {
1000 
1001 	uvmpdpol_estimatepageable(active, inactive);
1002 }
1003 
1004 
1005 /*
1006  * Use a separate thread for draining pools.
1007  * This work can't done from the main pagedaemon thread because
1008  * some pool allocators need to take vm_map locks.
1009  */
1010 
1011 static void
uvmpd_pool_drain_thread(void * arg)1012 uvmpd_pool_drain_thread(void *arg)
1013 {
1014 	struct pool *firstpool, *curpool;
1015 	int bufcnt, lastslept;
1016 	bool cycled;
1017 
1018 	firstpool = NULL;
1019 	cycled = true;
1020 	for (;;) {
1021 		/*
1022 		 * sleep until awoken by the pagedaemon.
1023 		 */
1024 		mutex_enter(&uvmpd_lock);
1025 		if (!uvmpd_pool_drain_run) {
1026 			lastslept = getticks();
1027 			cv_wait(&uvmpd_pool_drain_cv, &uvmpd_lock);
1028 			if (getticks() != lastslept) {
1029 				cycled = false;
1030 				firstpool = NULL;
1031 			}
1032 		}
1033 		uvmpd_pool_drain_run = false;
1034 		mutex_exit(&uvmpd_lock);
1035 
1036 		/*
1037 		 * rate limit draining, otherwise in desperate circumstances
1038 		 * this can totally saturate the system with xcall activity.
1039 		 */
1040 		if (cycled) {
1041 			kpause("uvmpdlmt", false, 1, NULL);
1042 			cycled = false;
1043 			firstpool = NULL;
1044 		}
1045 
1046 		/*
1047 		 * drain and temporarily disable the freelist cache.
1048 		 */
1049 		uvm_pgflcache_pause();
1050 
1051 		/*
1052 		 * kill unused metadata buffers.
1053 		 */
1054 		bufcnt = uvmexp.freetarg - uvm_availmem(false);
1055 		if (bufcnt < 0)
1056 			bufcnt = 0;
1057 
1058 		mutex_enter(&bufcache_lock);
1059 		buf_drain(bufcnt << PAGE_SHIFT);
1060 		mutex_exit(&bufcache_lock);
1061 
1062 		/*
1063 		 * drain a pool, and then re-enable the freelist cache.
1064 		 */
1065 		(void)pool_drain(&curpool);
1066 		KASSERT(curpool != NULL);
1067 		if (firstpool == NULL) {
1068 			firstpool = curpool;
1069 		} else if (firstpool == curpool) {
1070 			cycled = true;
1071 		}
1072 		uvm_pgflcache_resume();
1073 	}
1074 	/*NOTREACHED*/
1075 }
1076 
1077 static void
uvmpd_pool_drain_wakeup(void)1078 uvmpd_pool_drain_wakeup(void)
1079 {
1080 
1081 	mutex_enter(&uvmpd_lock);
1082 	uvmpd_pool_drain_run = true;
1083 	cv_signal(&uvmpd_pool_drain_cv);
1084 	mutex_exit(&uvmpd_lock);
1085 }
1086