xref: /openbsd-src/sys/uvm/uvm_pager.c (revision 505ee9ea3b177e2387d907a91ca7da069f3f14d8)
1 /*	$OpenBSD: uvm_pager.c,v 1.72 2019/12/08 12:37:45 mpi Exp $	*/
2 /*	$NetBSD: uvm_pager.c,v 1.36 2000/11/27 18:26:41 chs Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  *
28  * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
29  */
30 
31 /*
32  * uvm_pager.c: generic functions used to assist the pagers.
33  */
34 
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/malloc.h>
38 #include <sys/pool.h>
39 #include <sys/buf.h>
40 #include <sys/atomic.h>
41 
42 #include <uvm/uvm.h>
43 
44 struct pool *uvm_aiobuf_pool;
45 
46 struct uvm_pagerops *uvmpagerops[] = {
47 	&aobj_pager,
48 	&uvm_deviceops,
49 	&uvm_vnodeops,
50 };
51 
52 /*
53  * the pager map: provides KVA for I/O
54  *
55  * Each uvm_pseg has room for MAX_PAGERMAP_SEGS pager io space of
56  * MAXBSIZE bytes.
57  *
58  * The number of uvm_pseg instances is dynamic using an array segs.
59  * At most UVM_PSEG_COUNT instances can exist.
60  *
61  * psegs[0] always exists (so that the pager can always map in pages).
62  * psegs[0] element 0 is always reserved for the pagedaemon.
63  *
64  * Any other pseg is automatically created when no space is available
65  * and automatically destroyed when it is no longer in use.
66  */
67 #define MAX_PAGER_SEGS	16
68 #define PSEG_NUMSEGS	(PAGER_MAP_SIZE / MAX_PAGER_SEGS / MAXBSIZE)
69 struct uvm_pseg {
70 	/* Start of virtual space; 0 if not inited. */
71 	vaddr_t	start;
72 	/* Bitmap of the segments in use in this pseg. */
73 	int	use;
74 };
75 struct	mutex uvm_pseg_lck;
76 struct	uvm_pseg psegs[PSEG_NUMSEGS];
77 
78 #define UVM_PSEG_FULL(pseg)	((pseg)->use == (1 << MAX_PAGER_SEGS) - 1)
79 #define UVM_PSEG_EMPTY(pseg)	((pseg)->use == 0)
80 #define UVM_PSEG_INUSE(pseg,id)	(((pseg)->use & (1 << (id))) != 0)
81 
82 void		uvm_pseg_init(struct uvm_pseg *);
83 vaddr_t		uvm_pseg_get(int);
84 void		uvm_pseg_release(vaddr_t);
85 
86 /*
87  * uvm_pager_init: init pagers (at boot time)
88  */
89 void
90 uvm_pager_init(void)
91 {
92 	int lcv;
93 
94 	/* init pager map */
95 	uvm_pseg_init(&psegs[0]);
96 	mtx_init(&uvm_pseg_lck, IPL_VM);
97 
98 	/* init ASYNC I/O queue */
99 	TAILQ_INIT(&uvm.aio_done);
100 
101 	/* call pager init functions */
102 	for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
103 	    lcv++) {
104 		if (uvmpagerops[lcv]->pgo_init)
105 			uvmpagerops[lcv]->pgo_init();
106 	}
107 }
108 
109 /*
110  * Initialize a uvm_pseg.
111  *
112  * May fail, in which case seg->start == 0.
113  *
114  * Caller locks uvm_pseg_lck.
115  */
116 void
117 uvm_pseg_init(struct uvm_pseg *pseg)
118 {
119 	KASSERT(pseg->start == 0);
120 	KASSERT(pseg->use == 0);
121 	pseg->start = uvm_km_valloc_try(kernel_map, MAX_PAGER_SEGS * MAXBSIZE);
122 }
123 
124 /*
125  * Acquire a pager map segment.
126  *
127  * Returns a vaddr for paging. 0 on failure.
128  *
129  * Caller does not lock.
130  */
131 vaddr_t
132 uvm_pseg_get(int flags)
133 {
134 	int i;
135 	struct uvm_pseg *pseg;
136 
137 	mtx_enter(&uvm_pseg_lck);
138 
139 pager_seg_restart:
140 	/* Find first pseg that has room. */
141 	for (pseg = &psegs[0]; pseg != &psegs[PSEG_NUMSEGS]; pseg++) {
142 		if (UVM_PSEG_FULL(pseg))
143 			continue;
144 
145 		if (pseg->start == 0) {
146 			/* Need initialization. */
147 			uvm_pseg_init(pseg);
148 			if (pseg->start == 0)
149 				goto pager_seg_fail;
150 		}
151 
152 		/* Keep index 0 reserved for pagedaemon. */
153 		if (pseg == &psegs[0] && curproc != uvm.pagedaemon_proc)
154 			i = 1;
155 		else
156 			i = 0;
157 
158 		for (; i < MAX_PAGER_SEGS; i++) {
159 			if (!UVM_PSEG_INUSE(pseg, i)) {
160 				pseg->use |= 1 << i;
161 				mtx_leave(&uvm_pseg_lck);
162 				return pseg->start + i * MAXBSIZE;
163 			}
164 		}
165 	}
166 
167 pager_seg_fail:
168 	if ((flags & UVMPAGER_MAPIN_WAITOK) != 0) {
169 		msleep_nsec(&psegs, &uvm_pseg_lck, PVM, "pagerseg", INFSLP);
170 		goto pager_seg_restart;
171 	}
172 
173 	mtx_leave(&uvm_pseg_lck);
174 	return 0;
175 }
176 
177 /*
178  * Release a pager map segment.
179  *
180  * Caller does not lock.
181  *
182  * Deallocates pseg if it is no longer in use.
183  */
184 void
185 uvm_pseg_release(vaddr_t segaddr)
186 {
187 	int id;
188 	struct uvm_pseg *pseg;
189 	vaddr_t va = 0;
190 
191 	for (pseg = &psegs[0]; pseg != &psegs[PSEG_NUMSEGS]; pseg++) {
192 		if (pseg->start <= segaddr &&
193 		    segaddr < pseg->start + MAX_PAGER_SEGS * MAXBSIZE)
194 			break;
195 	}
196 	KASSERT(pseg != &psegs[PSEG_NUMSEGS]);
197 
198 	id = (segaddr - pseg->start) / MAXBSIZE;
199 	KASSERT(id >= 0 && id < MAX_PAGER_SEGS);
200 
201 	/* test for no remainder */
202 	KDASSERT(segaddr == pseg->start + id * MAXBSIZE);
203 
204 	mtx_enter(&uvm_pseg_lck);
205 
206 	KASSERT(UVM_PSEG_INUSE(pseg, id));
207 
208 	pseg->use &= ~(1 << id);
209 	wakeup(&psegs);
210 
211 	if (pseg != &psegs[0] && UVM_PSEG_EMPTY(pseg)) {
212 		va = pseg->start;
213 		pseg->start = 0;
214 	}
215 
216 	mtx_leave(&uvm_pseg_lck);
217 
218 	if (va)
219 		uvm_km_free(kernel_map, va, MAX_PAGER_SEGS * MAXBSIZE);
220 }
221 
222 /*
223  * uvm_pagermapin: map pages into KVA for I/O that needs mappings
224  *
225  * We basically just km_valloc a blank map entry to reserve the space in the
226  * kernel map and then use pmap_enter() to put the mappings in by hand.
227  */
228 vaddr_t
229 uvm_pagermapin(struct vm_page **pps, int npages, int flags)
230 {
231 	vaddr_t kva, cva;
232 	vm_prot_t prot;
233 	vsize_t size;
234 	struct vm_page *pp;
235 
236 	prot = PROT_READ;
237 	if (flags & UVMPAGER_MAPIN_READ)
238 		prot |= PROT_WRITE;
239 	size = ptoa(npages);
240 
241 	KASSERT(size <= MAXBSIZE);
242 
243 	kva = uvm_pseg_get(flags);
244 	if (kva == 0)
245 		return 0;
246 
247 	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
248 		pp = *pps++;
249 		KASSERT(pp);
250 		KASSERT(pp->pg_flags & PG_BUSY);
251 		/* Allow pmap_enter to fail. */
252 		if (pmap_enter(pmap_kernel(), cva, VM_PAGE_TO_PHYS(pp),
253 		    prot, PMAP_WIRED | PMAP_CANFAIL | prot) != 0) {
254 			pmap_remove(pmap_kernel(), kva, cva);
255 			pmap_update(pmap_kernel());
256 			uvm_pseg_release(kva);
257 			return 0;
258 		}
259 	}
260 	pmap_update(pmap_kernel());
261 	return kva;
262 }
263 
264 /*
265  * uvm_pagermapout: remove KVA mapping
266  *
267  * We remove our mappings by hand and then remove the mapping.
268  */
269 void
270 uvm_pagermapout(vaddr_t kva, int npages)
271 {
272 
273 	pmap_remove(pmap_kernel(), kva, kva + ((vsize_t)npages << PAGE_SHIFT));
274 	pmap_update(pmap_kernel());
275 	uvm_pseg_release(kva);
276 
277 }
278 
279 /*
280  * uvm_mk_pcluster
281  *
282  * generic "make 'pager put' cluster" function.  a pager can either
283  * [1] set pgo_mk_pcluster to NULL (never cluster), [2] set it to this
284  * generic function, or [3] set it to a pager specific function.
285  *
286  * => caller must lock object _and_ pagequeues (since we need to look
287  *    at active vs. inactive bits, etc.)
288  * => caller must make center page busy and write-protect it
289  * => we mark all cluster pages busy for the caller
290  * => the caller must unbusy all pages (and check wanted/released
291  *    status if it drops the object lock)
292  * => flags:
293  *      PGO_ALLPAGES:  all pages in object are valid targets
294  *      !PGO_ALLPAGES: use "lo" and "hi" to limit range of cluster
295  *      PGO_DOACTCLUST: include active pages in cluster.
296  *	PGO_FREE: set the PG_RELEASED bits on the cluster so they'll be freed
297  *		in async io (caller must clean on error).
298  *        NOTE: the caller should clear PG_CLEANCHK bits if PGO_DOACTCLUST.
299  *              PG_CLEANCHK is only a hint, but clearing will help reduce
300  *		the number of calls we make to the pmap layer.
301  */
302 
303 struct vm_page **
304 uvm_mk_pcluster(struct uvm_object *uobj, struct vm_page **pps, int *npages,
305     struct vm_page *center, int flags, voff_t mlo, voff_t mhi)
306 {
307 	struct vm_page **ppsp, *pclust;
308 	voff_t lo, hi, curoff;
309 	int center_idx, forward, incr;
310 
311 	/*
312 	 * center page should already be busy and write protected.  XXX:
313 	 * suppose page is wired?  if we lock, then a process could
314 	 * fault/block on it.  if we don't lock, a process could write the
315 	 * pages in the middle of an I/O.  (consider an msync()).  let's
316 	 * lock it for now (better to delay than corrupt data?).
317 	 */
318 	/* get cluster boundaries, check sanity, and apply our limits as well.*/
319 	uobj->pgops->pgo_cluster(uobj, center->offset, &lo, &hi);
320 	if ((flags & PGO_ALLPAGES) == 0) {
321 		if (lo < mlo)
322 			lo = mlo;
323 		if (hi > mhi)
324 			hi = mhi;
325 	}
326 	if ((hi - lo) >> PAGE_SHIFT > *npages) { /* pps too small, bail out! */
327 		pps[0] = center;
328 		*npages = 1;
329 		return(pps);
330 	}
331 
332 	/* now determine the center and attempt to cluster around the edges */
333 	center_idx = (center->offset - lo) >> PAGE_SHIFT;
334 	pps[center_idx] = center;	/* plug in the center page */
335 	ppsp = &pps[center_idx];
336 	*npages = 1;
337 
338 	/*
339 	 * attempt to cluster around the left [backward], and then
340 	 * the right side [forward].
341 	 *
342 	 * note that for inactive pages (pages that have been deactivated)
343 	 * there are no valid mappings and PG_CLEAN should be up to date.
344 	 * [i.e. there is no need to query the pmap with pmap_is_modified
345 	 * since there are no mappings].
346 	 */
347 	for (forward  = 0 ; forward <= 1 ; forward++) {
348 		incr = forward ? PAGE_SIZE : -PAGE_SIZE;
349 		curoff = center->offset + incr;
350 		for ( ;(forward == 0 && curoff >= lo) ||
351 		       (forward && curoff < hi);
352 		      curoff += incr) {
353 
354 			pclust = uvm_pagelookup(uobj, curoff); /* lookup page */
355 			if (pclust == NULL) {
356 				break;			/* no page */
357 			}
358 			/* handle active pages */
359 			/* NOTE: inactive pages don't have pmap mappings */
360 			if ((pclust->pg_flags & PQ_INACTIVE) == 0) {
361 				if ((flags & PGO_DOACTCLUST) == 0) {
362 					/* dont want mapped pages at all */
363 					break;
364 				}
365 
366 				/* make sure "clean" bit is sync'd */
367 				if ((pclust->pg_flags & PG_CLEANCHK) == 0) {
368 					if ((pclust->pg_flags & (PG_CLEAN|PG_BUSY))
369 					   == PG_CLEAN &&
370 					   pmap_is_modified(pclust))
371 						atomic_clearbits_int(
372 						    &pclust->pg_flags,
373 						    PG_CLEAN);
374 					/* now checked */
375 					atomic_setbits_int(&pclust->pg_flags,
376 					    PG_CLEANCHK);
377 				}
378 			}
379 
380 			/* is page available for cleaning and does it need it */
381 			if ((pclust->pg_flags & (PG_CLEAN|PG_BUSY)) != 0) {
382 				break;	/* page is already clean or is busy */
383 			}
384 
385 			/* yes!   enroll the page in our array */
386 			atomic_setbits_int(&pclust->pg_flags, PG_BUSY);
387 			UVM_PAGE_OWN(pclust, "uvm_mk_pcluster");
388 
389 			/*
390 			 * If we want to free after io is done, and we're
391 			 * async, set the released flag
392 			 */
393 			if ((flags & (PGO_FREE|PGO_SYNCIO)) == PGO_FREE)
394 				atomic_setbits_int(&pclust->pg_flags,
395 				    PG_RELEASED);
396 
397 			/* XXX: protect wired page?   see above comment. */
398 			pmap_page_protect(pclust, PROT_READ);
399 			if (!forward) {
400 				ppsp--;			/* back up one page */
401 				*ppsp = pclust;
402 			} else {
403 				/* move forward one page */
404 				ppsp[*npages] = pclust;
405 			}
406 			(*npages)++;
407 		}
408 	}
409 
410 	/*
411 	 * done!  return the cluster array to the caller!!!
412 	 */
413 	return(ppsp);
414 }
415 
416 /*
417  * uvm_pager_put: high level pageout routine
418  *
419  * we want to pageout page "pg" to backing store, clustering if
420  * possible.
421  *
422  * => page queues must be locked by caller
423  * => if page is not swap-backed, then "uobj" points to the object
424  *	backing it.
425  * => if page is swap-backed, then "uobj" should be NULL.
426  * => "pg" should be PG_BUSY (by caller), and !PG_CLEAN
427  *    for swap-backed memory, "pg" can be NULL if there is no page
428  *    of interest [sometimes the case for the pagedaemon]
429  * => "ppsp_ptr" should point to an array of npages vm_page pointers
430  *	for possible cluster building
431  * => flags (first two for non-swap-backed pages)
432  *	PGO_ALLPAGES: all pages in uobj are valid targets
433  *	PGO_DOACTCLUST: include "PQ_ACTIVE" pages as valid targets
434  *	PGO_SYNCIO: do SYNC I/O (no async)
435  *	PGO_PDFREECLUST: pagedaemon: drop cluster on successful I/O
436  *	PGO_FREE: tell the aio daemon to free pages in the async case.
437  * => start/stop: if (uobj && !PGO_ALLPAGES) limit targets to this range
438  *		  if (!uobj) start is the (daddr_t) of the starting swapblk
439  * => return state:
440  *	1. we return the VM_PAGER status code of the pageout
441  *	2. we return with the page queues unlocked
442  *	3. on errors we always drop the cluster.   thus, if we return
443  *		!PEND, !OK, then the caller only has to worry about
444  *		un-busying the main page (not the cluster pages).
445  *	4. on success, if !PGO_PDFREECLUST, we return the cluster
446  *		with all pages busy (caller must un-busy and check
447  *		wanted/released flags).
448  */
449 int
450 uvm_pager_put(struct uvm_object *uobj, struct vm_page *pg,
451     struct vm_page ***ppsp_ptr, int *npages, int flags,
452     voff_t start, voff_t stop)
453 {
454 	int result;
455 	daddr_t swblk;
456 	struct vm_page **ppsp = *ppsp_ptr;
457 
458 	/*
459 	 * note that uobj is null  if we are doing a swap-backed pageout.
460 	 * note that uobj is !null if we are doing normal object pageout.
461 	 * note that the page queues must be locked to cluster.
462 	 */
463 	if (uobj) {	/* if !swap-backed */
464 		/*
465 		 * attempt to build a cluster for pageout using its
466 		 * make-put-cluster function (if it has one).
467 		 */
468 		if (uobj->pgops->pgo_mk_pcluster) {
469 			ppsp = uobj->pgops->pgo_mk_pcluster(uobj, ppsp,
470 			    npages, pg, flags, start, stop);
471 			*ppsp_ptr = ppsp;  /* update caller's pointer */
472 		} else {
473 			ppsp[0] = pg;
474 			*npages = 1;
475 		}
476 
477 		swblk = 0;		/* XXX: keep gcc happy */
478 	} else {
479 		/*
480 		 * for swap-backed pageout, the caller (the pagedaemon) has
481 		 * already built the cluster for us.   the starting swap
482 		 * block we are writing to has been passed in as "start."
483 		 * "pg" could be NULL if there is no page we are especially
484 		 * interested in (in which case the whole cluster gets dropped
485 		 * in the event of an error or a sync "done").
486 		 */
487 		swblk = start;
488 		/* ppsp and npages should be ok */
489 	}
490 
491 	/* now that we've clustered we can unlock the page queues */
492 	uvm_unlock_pageq();
493 
494 	/*
495 	 * now attempt the I/O.   if we have a failure and we are
496 	 * clustered, we will drop the cluster and try again.
497 	 */
498 ReTry:
499 	if (uobj) {
500 		result = uobj->pgops->pgo_put(uobj, ppsp, *npages, flags);
501 	} else {
502 		/* XXX daddr_t -> int */
503 		result = uvm_swap_put(swblk, ppsp, *npages, flags);
504 	}
505 
506 	/*
507 	 * we have attempted the I/O.
508 	 *
509 	 * if the I/O was a success then:
510 	 * 	if !PGO_PDFREECLUST, we return the cluster to the
511 	 *		caller (who must un-busy all pages)
512 	 *	else we un-busy cluster pages for the pagedaemon
513 	 *
514 	 * if I/O is pending (async i/o) then we return the pending code.
515 	 * [in this case the async i/o done function must clean up when
516 	 *  i/o is done...]
517 	 */
518 	if (result == VM_PAGER_PEND || result == VM_PAGER_OK) {
519 		if (result == VM_PAGER_OK && (flags & PGO_PDFREECLUST)) {
520 			/* drop cluster */
521 			if (*npages > 1 || pg == NULL)
522 				uvm_pager_dropcluster(uobj, pg, ppsp, npages,
523 				    PGO_PDFREECLUST);
524 		}
525 		return (result);
526 	}
527 
528 	/*
529 	 * a pager error occured (even after dropping the cluster, if there
530 	 * was one).  give up! the caller only has one page ("pg")
531 	 * to worry about.
532 	 */
533 	if (*npages > 1 || pg == NULL) {
534 		uvm_pager_dropcluster(uobj, pg, ppsp, npages, PGO_REALLOCSWAP);
535 
536 		/*
537 		 * for failed swap-backed pageouts with a "pg",
538 		 * we need to reset pg's swslot to either:
539 		 * "swblk" (for transient errors, so we can retry),
540 		 * or 0 (for hard errors).
541 		 */
542 		if (uobj == NULL && pg != NULL) {
543 			/* XXX daddr_t -> int */
544 			int nswblk = (result == VM_PAGER_AGAIN) ? swblk : 0;
545 			if (pg->pg_flags & PQ_ANON) {
546 				pg->uanon->an_swslot = nswblk;
547 			} else {
548 				uao_set_swslot(pg->uobject,
549 					       pg->offset >> PAGE_SHIFT,
550 					       nswblk);
551 			}
552 		}
553 		if (result == VM_PAGER_AGAIN) {
554 			/*
555 			 * for transient failures, free all the swslots that
556 			 * we're not going to retry with.
557 			 */
558 			if (uobj == NULL) {
559 				if (pg) {
560 					/* XXX daddr_t -> int */
561 					uvm_swap_free(swblk + 1, *npages - 1);
562 				} else {
563 					/* XXX daddr_t -> int */
564 					uvm_swap_free(swblk, *npages);
565 				}
566 			}
567 			if (pg) {
568 				ppsp[0] = pg;
569 				*npages = 1;
570 				goto ReTry;
571 			}
572 		} else if (uobj == NULL) {
573 			/*
574 			 * for hard errors on swap-backed pageouts,
575 			 * mark the swslots as bad.  note that we do not
576 			 * free swslots that we mark bad.
577 			 */
578 			/* XXX daddr_t -> int */
579 			uvm_swap_markbad(swblk, *npages);
580 		}
581 	}
582 
583 	/*
584 	 * a pager error occurred (even after dropping the cluster, if there
585 	 * was one).    give up!   the caller only has one page ("pg")
586 	 * to worry about.
587 	 */
588 
589 	return(result);
590 }
591 
592 /*
593  * uvm_pager_dropcluster: drop a cluster we have built (because we
594  * got an error, or, if PGO_PDFREECLUST we are un-busying the
595  * cluster pages on behalf of the pagedaemon).
596  *
597  * => uobj, if non-null, is a non-swap-backed object
598  * => page queues are not locked
599  * => pg is our page of interest (the one we clustered around, can be null)
600  * => ppsp/npages is our current cluster
601  * => flags: PGO_PDFREECLUST: pageout was a success: un-busy cluster
602  *	pages on behalf of the pagedaemon.
603  *           PGO_REALLOCSWAP: drop previously allocated swap slots for
604  *		clustered swap-backed pages (except for "pg" if !NULL)
605  *		"swblk" is the start of swap alloc (e.g. for ppsp[0])
606  *		[only meaningful if swap-backed (uobj == NULL)]
607  */
608 
609 void
610 uvm_pager_dropcluster(struct uvm_object *uobj, struct vm_page *pg,
611     struct vm_page **ppsp, int *npages, int flags)
612 {
613 	int lcv;
614 
615 	/* drop all pages but "pg" */
616 	for (lcv = 0 ; lcv < *npages ; lcv++) {
617 		/* skip "pg" or empty slot */
618 		if (ppsp[lcv] == pg || ppsp[lcv] == NULL)
619 			continue;
620 
621 		/*
622 		 * Note that PQ_ANON bit can't change as long as we are holding
623 		 * the PG_BUSY bit (so there is no need to lock the page
624 		 * queues to test it).
625 		 */
626 		if (!uobj) {
627 			if (ppsp[lcv]->pg_flags & PQ_ANON) {
628 				if (flags & PGO_REALLOCSWAP)
629 					  /* zap swap block */
630 					  ppsp[lcv]->uanon->an_swslot = 0;
631 			} else {
632 				if (flags & PGO_REALLOCSWAP)
633 					uao_set_swslot(ppsp[lcv]->uobject,
634 					    ppsp[lcv]->offset >> PAGE_SHIFT, 0);
635 			}
636 		}
637 
638 		/* did someone want the page while we had it busy-locked? */
639 		if (ppsp[lcv]->pg_flags & PG_WANTED) {
640 			wakeup(ppsp[lcv]);
641 		}
642 
643 		/* if page was released, release it.  otherwise un-busy it */
644 		if (ppsp[lcv]->pg_flags & PG_RELEASED &&
645 		    ppsp[lcv]->pg_flags & PQ_ANON) {
646 				/* so that anfree will free */
647 				atomic_clearbits_int(&ppsp[lcv]->pg_flags,
648 				    PG_BUSY);
649 				UVM_PAGE_OWN(ppsp[lcv], NULL);
650 
651 				/* kills anon and frees pg */
652 				uvm_anfree(ppsp[lcv]->uanon);
653 
654 				continue;
655 		} else {
656 			/*
657 			 * if we were planning on async io then we would
658 			 * have PG_RELEASED set, clear that with the others.
659 			 */
660 			atomic_clearbits_int(&ppsp[lcv]->pg_flags,
661 			    PG_BUSY|PG_WANTED|PG_FAKE|PG_RELEASED);
662 			UVM_PAGE_OWN(ppsp[lcv], NULL);
663 		}
664 
665 		/*
666 		 * if we are operating on behalf of the pagedaemon and we
667 		 * had a successful pageout update the page!
668 		 */
669 		if (flags & PGO_PDFREECLUST) {
670 			pmap_clear_reference(ppsp[lcv]);
671 			pmap_clear_modify(ppsp[lcv]);
672 			atomic_setbits_int(&ppsp[lcv]->pg_flags, PG_CLEAN);
673 		}
674 	}
675 }
676 
677 /*
678  * interrupt-context iodone handler for single-buf i/os
679  * or the top-level buf of a nested-buf i/o.
680  *
681  * => must be at splbio().
682  */
683 
684 void
685 uvm_aio_biodone(struct buf *bp)
686 {
687 	splassert(IPL_BIO);
688 
689 	/* reset b_iodone for when this is a single-buf i/o. */
690 	bp->b_iodone = uvm_aio_aiodone;
691 
692 	mtx_enter(&uvm.aiodoned_lock);
693 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
694 	wakeup(&uvm.aiodoned);
695 	mtx_leave(&uvm.aiodoned_lock);
696 }
697 
698 /*
699  * uvm_aio_aiodone: do iodone processing for async i/os.
700  * this should be called in thread context, not interrupt context.
701  */
702 void
703 uvm_aio_aiodone(struct buf *bp)
704 {
705 	int npages = bp->b_bufsize >> PAGE_SHIFT;
706 	struct vm_page *pg, *pgs[MAXPHYS >> PAGE_SHIFT];
707 	struct uvm_object *uobj;
708 	int i, error;
709 	boolean_t write, swap;
710 
711 	KASSERT(npages <= MAXPHYS >> PAGE_SHIFT);
712 	splassert(IPL_BIO);
713 
714 	error = (bp->b_flags & B_ERROR) ? (bp->b_error ? bp->b_error : EIO) : 0;
715 	write = (bp->b_flags & B_READ) == 0;
716 
717 	uobj = NULL;
718 	for (i = 0; i < npages; i++)
719 		pgs[i] = uvm_atopg((vaddr_t)bp->b_data +
720 		    ((vsize_t)i << PAGE_SHIFT));
721 	uvm_pagermapout((vaddr_t)bp->b_data, npages);
722 #ifdef UVM_SWAP_ENCRYPT
723 	/*
724 	 * XXX - assumes that we only get ASYNC writes. used to be above.
725 	 */
726 	if (pgs[0]->pg_flags & PQ_ENCRYPT) {
727 		uvm_swap_freepages(pgs, npages);
728 		goto freed;
729 	}
730 #endif /* UVM_SWAP_ENCRYPT */
731 	for (i = 0; i < npages; i++) {
732 		pg = pgs[i];
733 
734 		if (i == 0) {
735 			swap = (pg->pg_flags & PQ_SWAPBACKED) != 0;
736 			if (!swap) {
737 				uobj = pg->uobject;
738 			}
739 		}
740 		KASSERT(swap || pg->uobject == uobj);
741 
742 		/*
743 		 * if this is a read and we got an error, mark the pages
744 		 * PG_RELEASED so that uvm_page_unbusy() will free them.
745 		 */
746 		if (!write && error) {
747 			atomic_setbits_int(&pg->pg_flags, PG_RELEASED);
748 			continue;
749 		}
750 		KASSERT(!write || (pgs[i]->pg_flags & PG_FAKE) == 0);
751 
752 		/*
753 		 * if this is a read and the page is PG_FAKE,
754 		 * or this was a successful write,
755 		 * mark the page PG_CLEAN and not PG_FAKE.
756 		 */
757 		if ((pgs[i]->pg_flags & PG_FAKE) || (write && error != ENOMEM)) {
758 			pmap_clear_reference(pgs[i]);
759 			pmap_clear_modify(pgs[i]);
760 			atomic_setbits_int(&pgs[i]->pg_flags, PG_CLEAN);
761 			atomic_clearbits_int(&pgs[i]->pg_flags, PG_FAKE);
762 		}
763 	}
764 	uvm_page_unbusy(pgs, npages);
765 
766 #ifdef UVM_SWAP_ENCRYPT
767 freed:
768 #endif
769 	pool_put(&bufpool, bp);
770 }
771