xref: /openbsd-src/sys/uvm/uvm_pager.c (revision 2b0358df1d88d06ef4139321dd05bd5e05d91eaf)
1 /*	$OpenBSD: uvm_pager.c,v 1.49 2009/04/06 12:02:52 oga Exp $	*/
2 /*	$NetBSD: uvm_pager.c,v 1.36 2000/11/27 18:26:41 chs Exp $	*/
3 
4 /*
5  *
6  * Copyright (c) 1997 Charles D. Cranor and Washington University.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *      This product includes software developed by Charles D. Cranor and
20  *      Washington University.
21  * 4. The name of the author may not be used to endorse or promote products
22  *    derived from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
25  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
26  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
27  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
28  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
29  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
30  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
31  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
32  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
33  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34  *
35  * from: Id: uvm_pager.c,v 1.1.2.23 1998/02/02 20:38:06 chuck Exp
36  */
37 
38 /*
39  * uvm_pager.c: generic functions used to assist the pagers.
40  */
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/proc.h>
45 #include <sys/malloc.h>
46 #include <sys/pool.h>
47 #include <sys/vnode.h>
48 #include <sys/buf.h>
49 
50 #include <uvm/uvm.h>
51 
52 struct pool *uvm_aiobuf_pool;
53 
54 struct uvm_pagerops *uvmpagerops[] = {
55 	&aobj_pager,
56 	&uvm_deviceops,
57 	&uvm_vnodeops,
58 };
59 
60 /*
61  * the pager map: provides KVA for I/O
62  *
63  * Each uvm_pseg has room for MAX_PAGERMAP_SEGS pager io space of
64  * MAXBSIZE bytes.
65  *
66  * The number of uvm_pseg instances is dynamic using an array segs.
67  * At most UVM_PSEG_COUNT instances can exist.
68  *
69  * psegs[0] always exists (so that the pager can always map in pages).
70  * psegs[0] element 0 is always reserved for the pagedaemon.
71  *
72  * Any other pseg is automatically created when no space is available
73  * and automatically destroyed when it is no longer in use.
74  */
75 #define MAX_PAGER_SEGS	16
76 #define PSEG_NUMSEGS	(PAGER_MAP_SIZE / MAX_PAGER_SEGS / MAXBSIZE)
77 struct uvm_pseg {
78 	/* Start of virtual space; 0 if not inited. */
79 	vaddr_t	start;
80 	/* Bitmap of the segments in use in this pseg. */
81 	int	use;
82 };
83 struct	mutex uvm_pseg_lck;
84 struct	uvm_pseg psegs[PSEG_NUMSEGS];
85 
86 #define UVM_PSEG_FULL(pseg)	((pseg)->use == (1 << MAX_PAGER_SEGS) - 1)
87 #define UVM_PSEG_EMPTY(pseg)	((pseg)->use == 0)
88 #define UVM_PSEG_INUSE(pseg,id)	(((pseg)->use & (1 << (id))) != 0)
89 
90 void		uvm_pseg_init(struct uvm_pseg *);
91 void		uvm_pseg_destroy(struct uvm_pseg *);
92 vaddr_t		uvm_pseg_get(int);
93 void		uvm_pseg_release(vaddr_t);
94 
95 struct vm_page	*uvm_pageratop(vaddr_t);
96 
97 /*
98  * uvm_pager_init: init pagers (at boot time)
99  */
100 
101 void
102 uvm_pager_init(void)
103 {
104 	int lcv;
105 
106 	/*
107 	 * init pager map
108 	 */
109 
110 	uvm_pseg_init(&psegs[0]);
111 	mtx_init(&uvm_pseg_lck, IPL_VM);
112 
113 	/*
114 	 * init ASYNC I/O queue
115 	 */
116 
117 	TAILQ_INIT(&uvm.aio_done);
118 
119 	/*
120 	 * call pager init functions
121 	 */
122 	for (lcv = 0 ; lcv < sizeof(uvmpagerops)/sizeof(struct uvm_pagerops *);
123 	    lcv++) {
124 		if (uvmpagerops[lcv]->pgo_init)
125 			uvmpagerops[lcv]->pgo_init();
126 	}
127 }
128 
129 /*
130  * Initialize a uvm_pseg.
131  *
132  * May fail, in which case seg->start == 0.
133  *
134  * Caller locks uvm_pseg_lck.
135  */
136 void
137 uvm_pseg_init(struct uvm_pseg *pseg)
138 {
139 	KASSERT(pseg->start == 0);
140 	KASSERT(pseg->use == 0);
141 	pseg->start = uvm_km_valloc(kernel_map, MAX_PAGER_SEGS * MAXBSIZE);
142 }
143 
144 /*
145  * Destroy a uvm_pseg.
146  *
147  * Never fails.
148  *
149  * Requires that seg != &psegs[0]
150  *
151  * Caller locks uvm_pseg_lck.
152  */
153 void
154 uvm_pseg_destroy(struct uvm_pseg *pseg)
155 {
156 	KASSERT(pseg != &psegs[0]);
157 	KASSERT(pseg->start != 0);
158 	KASSERT(pseg->use == 0);
159 	uvm_km_free(kernel_map, pseg->start, MAX_PAGER_SEGS * MAXBSIZE);
160 	pseg->start = 0;
161 }
162 
163 /*
164  * Acquire a pager map segment.
165  *
166  * Returns a vaddr for paging. 0 on failure.
167  *
168  * Caller does not lock.
169  */
170 vaddr_t
171 uvm_pseg_get(int flags)
172 {
173 	int i;
174 	struct uvm_pseg *pseg;
175 
176 	mtx_enter(&uvm_pseg_lck);
177 
178 pager_seg_restart:
179 	/* Find first pseg that has room. */
180 	for (pseg = &psegs[0]; pseg != &psegs[PSEG_NUMSEGS]; pseg++) {
181 		if (UVM_PSEG_FULL(pseg))
182 			continue;
183 
184 		if (pseg->start == 0) {
185 			/* Need initialization. */
186 			uvm_pseg_init(pseg);
187 			if (pseg->start == 0)
188 				goto pager_seg_fail;
189 		}
190 
191 		/* Keep index 0 reserved for pagedaemon. */
192 		if (pseg == &psegs[0] && curproc != uvm.pagedaemon_proc)
193 			i = 1;
194 		else
195 			i = 0;
196 
197 		for (; i < MAX_PAGER_SEGS; i++) {
198 			if (!UVM_PSEG_INUSE(pseg, i)) {
199 				pseg->use |= 1 << i;
200 				mtx_leave(&uvm_pseg_lck);
201 				return pseg->start + i * MAXBSIZE;
202 			}
203 		}
204 	}
205 
206 pager_seg_fail:
207 	if ((flags & UVMPAGER_MAPIN_WAITOK) != 0) {
208 		msleep(&psegs, &uvm_pseg_lck, PVM, "pagerseg", 0);
209 		goto pager_seg_restart;
210 	}
211 
212 	mtx_leave(&uvm_pseg_lck);
213 	return 0;
214 }
215 
216 /*
217  * Release a pager map segment.
218  *
219  * Caller does not lock.
220  *
221  * Deallocates pseg if it is no longer in use.
222  */
223 void
224 uvm_pseg_release(vaddr_t segaddr)
225 {
226 	int id;
227 	struct uvm_pseg *pseg;
228 
229 	for (pseg = &psegs[0]; pseg != &psegs[PSEG_NUMSEGS]; pseg++) {
230 		if (pseg->start <= segaddr &&
231 		    segaddr < pseg->start + MAX_PAGER_SEGS * MAXBSIZE)
232 			break;
233 	}
234 	KASSERT(pseg != &psegs[PSEG_NUMSEGS]);
235 
236 	id = (segaddr - pseg->start) / MAXBSIZE;
237 	KASSERT(id >= 0 && id < MAX_PAGER_SEGS);
238 
239 	/* test for no remainder */
240 	KDASSERT(segaddr == pseg->start + id * MAXBSIZE);
241 
242 	mtx_enter(&uvm_pseg_lck);
243 
244 	KASSERT(UVM_PSEG_INUSE(pseg, id));
245 
246 	pseg->use &= ~(1 << id);
247 	wakeup(&psegs);
248 
249 	if (pseg != &psegs[0] && UVM_PSEG_EMPTY(pseg))
250 		uvm_pseg_destroy(pseg);
251 
252 	mtx_leave(&uvm_pseg_lck);
253 }
254 
255 /*
256  * uvm_pagermapin: map pages into KVA for I/O that needs mappings
257  *
258  * We basically just km_valloc a blank map entry to reserve the space in the
259  * kernel map and then use pmap_enter() to put the mappings in by hand.
260  */
261 vaddr_t
262 uvm_pagermapin(struct vm_page **pps, int npages, int flags)
263 {
264 	vaddr_t kva, cva;
265 	vm_prot_t prot;
266 	vsize_t size;
267 	struct vm_page *pp;
268 
269 	UVMHIST_FUNC("uvm_pagermapin"); UVMHIST_CALLED(maphist);
270 
271 	UVMHIST_LOG(maphist,"(pps=%p, npages=%ld, flags=%d)",
272 	    pps, npages, flags,0);
273 
274 	prot = VM_PROT_READ;
275 	if (flags & UVMPAGER_MAPIN_READ)
276 		prot |= VM_PROT_WRITE;
277 	size = ptoa(npages);
278 
279 	KASSERT(size <= MAXBSIZE);
280 
281 	kva = uvm_pseg_get(flags);
282 	if (kva == 0) {
283 		UVMHIST_LOG(maphist,"<- NOWAIT failed", 0,0,0,0);
284 		return 0;
285 	}
286 
287 	for (cva = kva ; size != 0 ; size -= PAGE_SIZE, cva += PAGE_SIZE) {
288 		pp = *pps++;
289 		KASSERT(pp);
290 		KASSERT(pp->pg_flags & PG_BUSY);
291 		/* Allow pmap_enter to fail. */
292 		if (pmap_enter(pmap_kernel(), cva, VM_PAGE_TO_PHYS(pp),
293 		    prot, PMAP_WIRED | PMAP_CANFAIL | prot) != 0) {
294 			pmap_remove(pmap_kernel(), kva, cva);
295 			pmap_update(pmap_kernel());
296 			uvm_pseg_release(kva);
297 			UVMHIST_LOG(maphist,"<- pmap_enter failed", 0,0,0,0);
298 			return 0;
299 		}
300 	}
301 	pmap_update(pmap_kernel());
302 	UVMHIST_LOG(maphist, "<- done (KVA=0x%lx)", kva,0,0,0);
303 	return kva;
304 }
305 
306 /*
307  * uvm_pagermapout: remove KVA mapping
308  *
309  * We remove our mappings by hand and then remove the mapping.
310  */
311 void
312 uvm_pagermapout(vaddr_t kva, int npages)
313 {
314 	UVMHIST_FUNC("uvm_pagermapout"); UVMHIST_CALLED(maphist);
315 
316 	UVMHIST_LOG(maphist, " (kva=0x%lx, npages=%ld)", kva, npages,0,0);
317 
318 	pmap_remove(pmap_kernel(), kva, kva + (npages << PAGE_SHIFT));
319 	pmap_update(pmap_kernel());
320 	uvm_pseg_release(kva);
321 
322 	UVMHIST_LOG(maphist,"<- done",0,0,0,0);
323 }
324 
325 /*
326  * uvm_mk_pcluster
327  *
328  * generic "make 'pager put' cluster" function.  a pager can either
329  * [1] set pgo_mk_pcluster to NULL (never cluster), [2] set it to this
330  * generic function, or [3] set it to a pager specific function.
331  *
332  * => caller must lock object _and_ pagequeues (since we need to look
333  *    at active vs. inactive bits, etc.)
334  * => caller must make center page busy and write-protect it
335  * => we mark all cluster pages busy for the caller
336  * => the caller must unbusy all pages (and check wanted/released
337  *    status if it drops the object lock)
338  * => flags:
339  *      PGO_ALLPAGES:  all pages in object are valid targets
340  *      !PGO_ALLPAGES: use "lo" and "hi" to limit range of cluster
341  *      PGO_DOACTCLUST: include active pages in cluster.
342  *        NOTE: the caller should clear PG_CLEANCHK bits if PGO_DOACTCLUST.
343  *              PG_CLEANCHK is only a hint, but clearing will help reduce
344  *		the number of calls we make to the pmap layer.
345  */
346 
347 struct vm_page **
348 uvm_mk_pcluster(struct uvm_object *uobj, struct vm_page **pps, int *npages,
349     struct vm_page *center, int flags, voff_t mlo, voff_t mhi)
350 {
351 	struct vm_page **ppsp, *pclust;
352 	voff_t lo, hi, curoff;
353 	int center_idx, forward, incr;
354 	UVMHIST_FUNC("uvm_mk_pcluster"); UVMHIST_CALLED(maphist);
355 
356 	/*
357 	 * center page should already be busy and write protected.  XXX:
358 	 * suppose page is wired?  if we lock, then a process could
359 	 * fault/block on it.  if we don't lock, a process could write the
360 	 * pages in the middle of an I/O.  (consider an msync()).  let's
361 	 * lock it for now (better to delay than corrupt data?).
362 	 */
363 
364 	/*
365 	 * get cluster boundaries, check sanity, and apply our limits as well.
366 	 */
367 
368 	uobj->pgops->pgo_cluster(uobj, center->offset, &lo, &hi);
369 	if ((flags & PGO_ALLPAGES) == 0) {
370 		if (lo < mlo)
371 			lo = mlo;
372 		if (hi > mhi)
373 			hi = mhi;
374 	}
375 	if ((hi - lo) >> PAGE_SHIFT > *npages) { /* pps too small, bail out! */
376 		pps[0] = center;
377 		*npages = 1;
378 		return(pps);
379 	}
380 
381 	/*
382 	 * now determine the center and attempt to cluster around the
383 	 * edges
384 	 */
385 
386 	center_idx = (center->offset - lo) >> PAGE_SHIFT;
387 	pps[center_idx] = center;	/* plug in the center page */
388 	ppsp = &pps[center_idx];
389 	*npages = 1;
390 
391 	/*
392 	 * attempt to cluster around the left [backward], and then
393 	 * the right side [forward].
394 	 *
395 	 * note that for inactive pages (pages that have been deactivated)
396 	 * there are no valid mappings and PG_CLEAN should be up to date.
397 	 * [i.e. there is no need to query the pmap with pmap_is_modified
398 	 * since there are no mappings].
399 	 */
400 
401 	for (forward  = 0 ; forward <= 1 ; forward++) {
402 		incr = forward ? PAGE_SIZE : -PAGE_SIZE;
403 		curoff = center->offset + incr;
404 		for ( ;(forward == 0 && curoff >= lo) ||
405 		       (forward && curoff < hi);
406 		      curoff += incr) {
407 
408 			pclust = uvm_pagelookup(uobj, curoff); /* lookup page */
409 			if (pclust == NULL) {
410 				break;			/* no page */
411 			}
412 			/* handle active pages */
413 			/* NOTE: inactive pages don't have pmap mappings */
414 			if ((pclust->pg_flags & PQ_INACTIVE) == 0) {
415 				if ((flags & PGO_DOACTCLUST) == 0) {
416 					/* dont want mapped pages at all */
417 					break;
418 				}
419 
420 				/* make sure "clean" bit is sync'd */
421 				if ((pclust->pg_flags & PG_CLEANCHK) == 0) {
422 					if ((pclust->pg_flags & (PG_CLEAN|PG_BUSY))
423 					   == PG_CLEAN &&
424 					   pmap_is_modified(pclust))
425 						atomic_clearbits_int(
426 						    &pclust->pg_flags,
427 						    PG_CLEAN);
428 					/* now checked */
429 					atomic_setbits_int(&pclust->pg_flags,
430 					    PG_CLEANCHK);
431 				}
432 			}
433 
434 			/* is page available for cleaning and does it need it */
435 			if ((pclust->pg_flags & (PG_CLEAN|PG_BUSY)) != 0) {
436 				break;	/* page is already clean or is busy */
437 			}
438 
439 			/* yes!   enroll the page in our array */
440 			atomic_setbits_int(&pclust->pg_flags, PG_BUSY);
441 			UVM_PAGE_OWN(pclust, "uvm_mk_pcluster");
442 
443 			/* XXX: protect wired page?   see above comment. */
444 			pmap_page_protect(pclust, VM_PROT_READ);
445 			if (!forward) {
446 				ppsp--;			/* back up one page */
447 				*ppsp = pclust;
448 			} else {
449 				/* move forward one page */
450 				ppsp[*npages] = pclust;
451 			}
452 			(*npages)++;
453 		}
454 	}
455 
456 	/*
457 	 * done!  return the cluster array to the caller!!!
458 	 */
459 
460 	UVMHIST_LOG(maphist, "<- done",0,0,0,0);
461 	return(ppsp);
462 }
463 
464 /*
465  * uvm_pager_put: high level pageout routine
466  *
467  * we want to pageout page "pg" to backing store, clustering if
468  * possible.
469  *
470  * => page queues must be locked by caller
471  * => if page is not swap-backed, then "uobj" points to the object
472  *	backing it.   this object should be locked by the caller.
473  * => if page is swap-backed, then "uobj" should be NULL.
474  * => "pg" should be PG_BUSY (by caller), and !PG_CLEAN
475  *    for swap-backed memory, "pg" can be NULL if there is no page
476  *    of interest [sometimes the case for the pagedaemon]
477  * => "ppsp_ptr" should point to an array of npages vm_page pointers
478  *	for possible cluster building
479  * => flags (first two for non-swap-backed pages)
480  *	PGO_ALLPAGES: all pages in uobj are valid targets
481  *	PGO_DOACTCLUST: include "PQ_ACTIVE" pages as valid targets
482  *	PGO_SYNCIO: do SYNC I/O (no async)
483  *	PGO_PDFREECLUST: pagedaemon: drop cluster on successful I/O
484  * => start/stop: if (uobj && !PGO_ALLPAGES) limit targets to this range
485  *		  if (!uobj) start is the (daddr64_t) of the starting swapblk
486  * => return state:
487  *	1. we return the VM_PAGER status code of the pageout
488  *	2. we return with the page queues unlocked
489  *	3. if (uobj != NULL) [!swap_backed] we return with
490  *		uobj locked _only_ if PGO_PDFREECLUST is set
491  *		AND result != VM_PAGER_PEND.   in all other cases
492  *		we return with uobj unlocked.   [this is a hack
493  *		that allows the pagedaemon to save one lock/unlock
494  *		pair in the !swap_backed case since we have to
495  *		lock the uobj to drop the cluster anyway]
496  *	4. on errors we always drop the cluster.   thus, if we return
497  *		!PEND, !OK, then the caller only has to worry about
498  *		un-busying the main page (not the cluster pages).
499  *	5. on success, if !PGO_PDFREECLUST, we return the cluster
500  *		with all pages busy (caller must un-busy and check
501  *		wanted/released flags).
502  */
503 
504 int
505 uvm_pager_put(struct uvm_object *uobj, struct vm_page *pg,
506     struct vm_page ***ppsp_ptr, int *npages, int flags,
507     voff_t start, voff_t stop)
508 {
509 	int result;
510 	daddr64_t swblk;
511 	struct vm_page **ppsp = *ppsp_ptr;
512 	UVMHIST_FUNC("uvm_pager_put"); UVMHIST_CALLED(pdhist);
513 
514 	/*
515 	 * note that uobj is null  if we are doing a swap-backed pageout.
516 	 * note that uobj is !null if we are doing normal object pageout.
517 	 * note that the page queues must be locked to cluster.
518 	 */
519 
520 	if (uobj) {	/* if !swap-backed */
521 
522 		/*
523 		 * attempt to build a cluster for pageout using its
524 		 * make-put-cluster function (if it has one).
525 		 */
526 
527 		if (uobj->pgops->pgo_mk_pcluster) {
528 			ppsp = uobj->pgops->pgo_mk_pcluster(uobj, ppsp,
529 			    npages, pg, flags, start, stop);
530 			*ppsp_ptr = ppsp;  /* update caller's pointer */
531 		} else {
532 			ppsp[0] = pg;
533 			*npages = 1;
534 		}
535 
536 		swblk = 0;		/* XXX: keep gcc happy */
537 
538 	} else {
539 
540 		/*
541 		 * for swap-backed pageout, the caller (the pagedaemon) has
542 		 * already built the cluster for us.   the starting swap
543 		 * block we are writing to has been passed in as "start."
544 		 * "pg" could be NULL if there is no page we are especially
545 		 * interested in (in which case the whole cluster gets dropped
546 		 * in the event of an error or a sync "done").
547 		 */
548 		swblk = (daddr64_t) start;
549 		/* ppsp and npages should be ok */
550 	}
551 
552 	/* now that we've clustered we can unlock the page queues */
553 	uvm_unlock_pageq();
554 
555 	/*
556 	 * now attempt the I/O.   if we have a failure and we are
557 	 * clustered, we will drop the cluster and try again.
558 	 */
559 
560 ReTry:
561 	if (uobj) {
562 		/* object is locked */
563 		result = uobj->pgops->pgo_put(uobj, ppsp, *npages, flags);
564 		UVMHIST_LOG(pdhist, "put -> %ld", result, 0,0,0);
565 		/* object is now unlocked */
566 	} else {
567 		/* nothing locked */
568 		/* XXX daddr64_t -> int */
569 		result = uvm_swap_put(swblk, ppsp, *npages, flags);
570 		/* nothing locked */
571 	}
572 
573 	/*
574 	 * we have attempted the I/O.
575 	 *
576 	 * if the I/O was a success then:
577 	 * 	if !PGO_PDFREECLUST, we return the cluster to the
578 	 *		caller (who must un-busy all pages)
579 	 *	else we un-busy cluster pages for the pagedaemon
580 	 *
581 	 * if I/O is pending (async i/o) then we return the pending code.
582 	 * [in this case the async i/o done function must clean up when
583 	 *  i/o is done...]
584 	 */
585 
586 	if (result == VM_PAGER_PEND || result == VM_PAGER_OK) {
587 		if (result == VM_PAGER_OK && (flags & PGO_PDFREECLUST)) {
588 			/*
589 			 * drop cluster and relock object (only if I/O is
590 			 * not pending)
591 			 */
592 			if (uobj)
593 				/* required for dropcluster */
594 				simple_lock(&uobj->vmobjlock);
595 			if (*npages > 1 || pg == NULL)
596 				uvm_pager_dropcluster(uobj, pg, ppsp, npages,
597 				    PGO_PDFREECLUST);
598 			/* if (uobj): object still locked, as per
599 			 * return-state item #3 */
600 		}
601 		return (result);
602 	}
603 
604 	/*
605 	 * a pager error occured (even after dropping the cluster, if there
606 	 * was one).  give up! the caller only has one page ("pg")
607 	 * to worry about.
608 	 */
609 
610 	if (*npages > 1 || pg == NULL) {
611 		if (uobj) {
612 			simple_lock(&uobj->vmobjlock);
613 		}
614 		uvm_pager_dropcluster(uobj, pg, ppsp, npages, PGO_REALLOCSWAP);
615 
616 		/*
617 		 * for failed swap-backed pageouts with a "pg",
618 		 * we need to reset pg's swslot to either:
619 		 * "swblk" (for transient errors, so we can retry),
620 		 * or 0 (for hard errors).
621 		 */
622 
623 		if (uobj == NULL && pg != NULL) {
624 			/* XXX daddr64_t -> int */
625 			int nswblk = (result == VM_PAGER_AGAIN) ? swblk : 0;
626 			if (pg->pg_flags & PQ_ANON) {
627 				simple_lock(&pg->uanon->an_lock);
628 				pg->uanon->an_swslot = nswblk;
629 				simple_unlock(&pg->uanon->an_lock);
630 			} else {
631 				simple_lock(&pg->uobject->vmobjlock);
632 				uao_set_swslot(pg->uobject,
633 					       pg->offset >> PAGE_SHIFT,
634 					       nswblk);
635 				simple_unlock(&pg->uobject->vmobjlock);
636 			}
637 		}
638 		if (result == VM_PAGER_AGAIN) {
639 
640 			/*
641 			 * for transient failures, free all the swslots that
642 			 * we're not going to retry with.
643 			 */
644 
645 			if (uobj == NULL) {
646 				if (pg) {
647 					/* XXX daddr64_t -> int */
648 					uvm_swap_free(swblk + 1, *npages - 1);
649 				} else {
650 					/* XXX daddr64_t -> int */
651 					uvm_swap_free(swblk, *npages);
652 				}
653 			}
654 			if (pg) {
655 				ppsp[0] = pg;
656 				*npages = 1;
657 				goto ReTry;
658 			}
659 		} else if (uobj == NULL) {
660 
661 			/*
662 			 * for hard errors on swap-backed pageouts,
663 			 * mark the swslots as bad.  note that we do not
664 			 * free swslots that we mark bad.
665 			 */
666 
667 			/* XXX daddr64_t -> int */
668 			uvm_swap_markbad(swblk, *npages);
669 		}
670 	}
671 
672 	/*
673 	 * a pager error occurred (even after dropping the cluster, if there
674 	 * was one).    give up!   the caller only has one page ("pg")
675 	 * to worry about.
676 	 */
677 
678 	if (uobj && (flags & PGO_PDFREECLUST) != 0)
679 		simple_lock(&uobj->vmobjlock);
680 	return(result);
681 }
682 
683 /*
684  * uvm_pager_dropcluster: drop a cluster we have built (because we
685  * got an error, or, if PGO_PDFREECLUST we are un-busying the
686  * cluster pages on behalf of the pagedaemon).
687  *
688  * => uobj, if non-null, is a non-swap-backed object that is
689  *	locked by the caller.   we return with this object still
690  *	locked.
691  * => page queues are not locked
692  * => pg is our page of interest (the one we clustered around, can be null)
693  * => ppsp/npages is our current cluster
694  * => flags: PGO_PDFREECLUST: pageout was a success: un-busy cluster
695  *	pages on behalf of the pagedaemon.
696  *           PGO_REALLOCSWAP: drop previously allocated swap slots for
697  *		clustered swap-backed pages (except for "pg" if !NULL)
698  *		"swblk" is the start of swap alloc (e.g. for ppsp[0])
699  *		[only meaningful if swap-backed (uobj == NULL)]
700  */
701 
702 void
703 uvm_pager_dropcluster(struct uvm_object *uobj, struct vm_page *pg,
704     struct vm_page **ppsp, int *npages, int flags)
705 {
706 	int lcv;
707 	boolean_t obj_is_alive;
708 	struct uvm_object *saved_uobj;
709 
710 	/*
711 	 * drop all pages but "pg"
712 	 */
713 
714 	for (lcv = 0 ; lcv < *npages ; lcv++) {
715 
716 		/* skip "pg" or empty slot */
717 		if (ppsp[lcv] == pg || ppsp[lcv] == NULL)
718 			continue;
719 
720 		/*
721 		 * if swap-backed, gain lock on object that owns page.  note
722 		 * that PQ_ANON bit can't change as long as we are holding
723 		 * the PG_BUSY bit (so there is no need to lock the page
724 		 * queues to test it).
725 		 *
726 		 * once we have the lock, dispose of the pointer to swap, if
727 		 * requested
728 		 */
729 		if (!uobj) {
730 			if (ppsp[lcv]->pg_flags & PQ_ANON) {
731 				simple_lock(&ppsp[lcv]->uanon->an_lock);
732 				if (flags & PGO_REALLOCSWAP)
733 					  /* zap swap block */
734 					  ppsp[lcv]->uanon->an_swslot = 0;
735 			} else {
736 				simple_lock(&ppsp[lcv]->uobject->vmobjlock);
737 				if (flags & PGO_REALLOCSWAP)
738 					uao_set_swslot(ppsp[lcv]->uobject,
739 					    ppsp[lcv]->offset >> PAGE_SHIFT, 0);
740 			}
741 		}
742 
743 		/* did someone want the page while we had it busy-locked? */
744 		if (ppsp[lcv]->pg_flags & PG_WANTED) {
745 			/* still holding obj lock */
746 			wakeup(ppsp[lcv]);
747 		}
748 
749 		/* if page was released, release it.  otherwise un-busy it */
750 		if (ppsp[lcv]->pg_flags & PG_RELEASED) {
751 
752 			if (ppsp[lcv]->pg_flags & PQ_ANON) {
753 				/* so that anfree will free */
754 				atomic_clearbits_int(&ppsp[lcv]->pg_flags,
755 				    PG_BUSY);
756 				UVM_PAGE_OWN(ppsp[lcv], NULL);
757 
758 				pmap_page_protect(ppsp[lcv], VM_PROT_NONE);
759 				simple_unlock(&ppsp[lcv]->uanon->an_lock);
760 				/* kills anon and frees pg */
761 				uvm_anfree(ppsp[lcv]->uanon);
762 
763 				continue;
764 			}
765 
766 			/*
767 			 * pgo_releasepg will dump the page for us
768 			 */
769 
770 			saved_uobj = ppsp[lcv]->uobject;
771 			obj_is_alive =
772 			    saved_uobj->pgops->pgo_releasepg(ppsp[lcv], NULL);
773 
774 			/* for normal objects, "pg" is still PG_BUSY by us,
775 			 * so obj can't die */
776 			KASSERT(!uobj || obj_is_alive);
777 
778 			/* only unlock the object if it is still alive...  */
779 			if (obj_is_alive && saved_uobj != uobj)
780 				simple_unlock(&saved_uobj->vmobjlock);
781 
782 			/*
783 			 * XXXCDC: suppose uobj died in the pgo_releasepg?
784 			 * how pass that
785 			 * info up to caller.  we are currently ignoring it...
786 			 */
787 
788 			continue;		/* next page */
789 		} else {
790 			atomic_clearbits_int(&ppsp[lcv]->pg_flags,
791 			    PG_BUSY|PG_WANTED|PG_FAKE);
792 			UVM_PAGE_OWN(ppsp[lcv], NULL);
793 		}
794 
795 		/*
796 		 * if we are operating on behalf of the pagedaemon and we
797 		 * had a successful pageout update the page!
798 		 */
799 		if (flags & PGO_PDFREECLUST) {
800 			pmap_clear_reference(ppsp[lcv]);
801 			pmap_clear_modify(ppsp[lcv]);
802 			atomic_setbits_int(&ppsp[lcv]->pg_flags, PG_CLEAN);
803 		}
804 
805 		/* if anonymous cluster, unlock object and move on */
806 		if (!uobj) {
807 			if (ppsp[lcv]->pg_flags & PQ_ANON)
808 				simple_unlock(&ppsp[lcv]->uanon->an_lock);
809 			else
810 				simple_unlock(&ppsp[lcv]->uobject->vmobjlock);
811 		}
812 	}
813 }
814 
815 #ifdef UBC
816 /*
817  * interrupt-context iodone handler for nested i/o bufs.
818  *
819  * => must be at splbio().
820  */
821 
822 void
823 uvm_aio_biodone1(struct buf *bp)
824 {
825 	struct buf *mbp = bp->b_private;
826 
827 	splassert(IPL_BIO);
828 
829 	KASSERT(mbp != bp);
830 	if (bp->b_flags & B_ERROR) {
831 		mbp->b_flags |= B_ERROR;
832 		mbp->b_error = bp->b_error;
833 	}
834 	mbp->b_resid -= bp->b_bcount;
835 	pool_put(&bufpool, bp);
836 	if (mbp->b_resid == 0) {
837 		biodone(mbp);
838 	}
839 }
840 #endif
841 
842 /*
843  * interrupt-context iodone handler for single-buf i/os
844  * or the top-level buf of a nested-buf i/o.
845  *
846  * => must be at splbio().
847  */
848 
849 void
850 uvm_aio_biodone(struct buf *bp)
851 {
852 	splassert(IPL_BIO);
853 
854 	/* reset b_iodone for when this is a single-buf i/o. */
855 	bp->b_iodone = uvm_aio_aiodone;
856 
857 	mtx_enter(&uvm.aiodoned_lock);	/* locks uvm.aio_done */
858 	TAILQ_INSERT_TAIL(&uvm.aio_done, bp, b_freelist);
859 	wakeup(&uvm.aiodoned);
860 	mtx_leave(&uvm.aiodoned_lock);
861 }
862 
863 /*
864  * uvm_aio_aiodone: do iodone processing for async i/os.
865  * this should be called in thread context, not interrupt context.
866  */
867 
868 void
869 uvm_aio_aiodone(struct buf *bp)
870 {
871 	int npages = bp->b_bufsize >> PAGE_SHIFT;
872 	struct vm_page *pg, *pgs[MAXPHYS >> PAGE_SHIFT];
873 	struct uvm_object *uobj;
874 	int i, error;
875 	boolean_t write, swap;
876 	UVMHIST_FUNC("uvm_aio_aiodone"); UVMHIST_CALLED(pdhist);
877 	UVMHIST_LOG(pdhist, "bp %p", bp, 0,0,0);
878 
879 	KASSERT(npages <= MAXPHYS >> PAGE_SHIFT);
880 	splassert(IPL_BIO);
881 
882 	error = (bp->b_flags & B_ERROR) ? (bp->b_error ? bp->b_error : EIO) : 0;
883 	write = (bp->b_flags & B_READ) == 0;
884 #ifdef UBC
885 	/* XXXUBC B_NOCACHE is for swap pager, should be done differently */
886 	if (write && !(bp->b_flags & B_NOCACHE) && bioops.io_pageiodone) {
887 		(*bioops.io_pageiodone)(bp);
888 	}
889 #endif
890 
891 	uobj = NULL;
892 	for (i = 0; i < npages; i++) {
893 		pgs[i] = uvm_pageratop((vaddr_t)bp->b_data + (i << PAGE_SHIFT));
894 		UVMHIST_LOG(pdhist, "pgs[%ld] = %p", i, pgs[i],0,0);
895 	}
896 	uvm_pagermapout((vaddr_t)bp->b_data, npages);
897 #ifdef UVM_SWAP_ENCRYPT
898 	/*
899 	 * XXX - assumes that we only get ASYNC writes. used to be above.
900 	 */
901 	if (pgs[0]->pg_flags & PQ_ENCRYPT) {
902 		uvm_swap_freepages(pgs, npages);
903 		goto freed;
904 	}
905 #endif /* UVM_SWAP_ENCRYPT */
906 	for (i = 0; i < npages; i++) {
907 		pg = pgs[i];
908 
909 		if (i == 0) {
910 			swap = (pg->pg_flags & PQ_SWAPBACKED) != 0;
911 			if (!swap) {
912 				uobj = pg->uobject;
913 				simple_lock(&uobj->vmobjlock);
914 			}
915 		}
916 		KASSERT(swap || pg->uobject == uobj);
917 		if (swap) {
918 			if (pg->pg_flags & PQ_ANON) {
919 				simple_lock(&pg->uanon->an_lock);
920 			} else {
921 				simple_lock(&pg->uobject->vmobjlock);
922 			}
923 		}
924 
925 		/*
926 		 * if this is a read and we got an error, mark the pages
927 		 * PG_RELEASED so that uvm_page_unbusy() will free them.
928 		 */
929 		if (!write && error) {
930 			atomic_setbits_int(&pg->pg_flags, PG_RELEASED);
931 			continue;
932 		}
933 		KASSERT(!write || (pgs[i]->pg_flags & PG_FAKE) == 0);
934 
935 		/*
936 		 * if this is a read and the page is PG_FAKE,
937 		 * or this was a successful write,
938 		 * mark the page PG_CLEAN and not PG_FAKE.
939 		 */
940 
941 		if ((pgs[i]->pg_flags & PG_FAKE) || (write && error != ENOMEM)) {
942 			pmap_clear_reference(pgs[i]);
943 			pmap_clear_modify(pgs[i]);
944 			atomic_setbits_int(&pgs[i]->pg_flags, PG_CLEAN);
945 			atomic_clearbits_int(&pgs[i]->pg_flags, PG_FAKE);
946 		}
947 		if (swap) {
948 			if (pg->pg_flags & PQ_ANON) {
949 				simple_unlock(&pg->uanon->an_lock);
950 			} else {
951 				simple_unlock(&pg->uobject->vmobjlock);
952 			}
953 		}
954 	}
955 	uvm_page_unbusy(pgs, npages);
956 	if (!swap) {
957 		simple_unlock(&uobj->vmobjlock);
958 	}
959 
960 #ifdef UVM_SWAP_ENCRYPT
961 freed:
962 #endif
963 	if (write && (bp->b_flags & B_AGE) != 0 && bp->b_vp != NULL) {
964 		vwakeup(bp->b_vp);
965 	}
966 	pool_put(&bufpool, bp);
967 }
968 
969 /*
970  * uvm_pageratop: convert KVAs in the pager map back to their page
971  * structures.
972  */
973 struct vm_page *
974 uvm_pageratop(vaddr_t kva)
975 {
976 	struct vm_page *pg;
977 	paddr_t pa;
978 	boolean_t rv;
979 
980 	rv = pmap_extract(pmap_kernel(), kva, &pa);
981 	KASSERT(rv);
982 	pg = PHYS_TO_VM_PAGE(pa);
983 	KASSERT(pg != NULL);
984 	return (pg);
985 }
986