xref: /freebsd/sys/fs/fuse/fuse_io.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 2007-2009 Google Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions are
9  * met:
10  *
11  * * Redistributions of source code must retain the above copyright
12  *   notice, this list of conditions and the following disclaimer.
13  * * Redistributions in binary form must reproduce the above
14  *   copyright notice, this list of conditions and the following disclaimer
15  *   in the documentation and/or other materials provided with the
16  *   distribution.
17  * * Neither the name of Google Inc. nor the names of its
18  *   contributors may be used to endorse or promote products derived from
19  *   this software without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  *
33  * Copyright (C) 2005 Csaba Henk.
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  *
45  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND
46  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
47  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
48  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
49  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
50  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
51  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
52  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
53  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
54  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
55  * SUCH DAMAGE.
56  */
57 
58 #include <sys/cdefs.h>
59 __FBSDID("$FreeBSD$");
60 
61 #include <sys/types.h>
62 #include <sys/module.h>
63 #include <sys/systm.h>
64 #include <sys/errno.h>
65 #include <sys/param.h>
66 #include <sys/kernel.h>
67 #include <sys/conf.h>
68 #include <sys/uio.h>
69 #include <sys/malloc.h>
70 #include <sys/queue.h>
71 #include <sys/lock.h>
72 #include <sys/sx.h>
73 #include <sys/mutex.h>
74 #include <sys/rwlock.h>
75 #include <sys/proc.h>
76 #include <sys/mount.h>
77 #include <sys/vnode.h>
78 #include <sys/stat.h>
79 #include <sys/unistd.h>
80 #include <sys/filedesc.h>
81 #include <sys/file.h>
82 #include <sys/fcntl.h>
83 #include <sys/bio.h>
84 #include <sys/buf.h>
85 #include <sys/sysctl.h>
86 
87 #include <vm/vm.h>
88 #include <vm/vm_extern.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_object.h>
93 
94 #include "fuse.h"
95 #include "fuse_file.h"
96 #include "fuse_node.h"
97 #include "fuse_internal.h"
98 #include "fuse_ipc.h"
99 #include "fuse_io.h"
100 
101 #define FUSE_DEBUG_MODULE IO
102 #include "fuse_debug.h"
103 
104 
105 static int
106 fuse_read_directbackend(struct vnode *vp, struct uio *uio,
107     struct ucred *cred, struct fuse_filehandle *fufh);
108 static int
109 fuse_read_biobackend(struct vnode *vp, struct uio *uio,
110     struct ucred *cred, struct fuse_filehandle *fufh);
111 static int
112 fuse_write_directbackend(struct vnode *vp, struct uio *uio,
113     struct ucred *cred, struct fuse_filehandle *fufh, int ioflag);
114 static int
115 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
116     struct ucred *cred, struct fuse_filehandle *fufh, int ioflag);
117 
118 int
119 fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag,
120     struct ucred *cred)
121 {
122 	struct fuse_filehandle *fufh;
123 	int err, directio;
124 
125 	MPASS(vp->v_type == VREG || vp->v_type == VDIR);
126 
127 	err = fuse_filehandle_getrw(vp,
128 	    (uio->uio_rw == UIO_READ) ? FUFH_RDONLY : FUFH_WRONLY, &fufh);
129 	if (err) {
130 		printf("FUSE: io dispatch: filehandles are closed\n");
131 		return err;
132 	}
133 	/*
134          * Ideally, when the daemon asks for direct io at open time, the
135          * standard file flag should be set according to this, so that would
136          * just change the default mode, which later on could be changed via
137          * fcntl(2).
138          * But this doesn't work, the O_DIRECT flag gets cleared at some point
139          * (don't know where). So to make any use of the Fuse direct_io option,
140          * we hardwire it into the file's private data (similarly to Linux,
141          * btw.).
142          */
143 	directio = (ioflag & IO_DIRECT) || !fsess_opt_datacache(vnode_mount(vp));
144 
145 	switch (uio->uio_rw) {
146 	case UIO_READ:
147 		if (directio) {
148 			FS_DEBUG("direct read of vnode %ju via file handle %ju\n",
149 			    (uintmax_t)VTOILLU(vp), (uintmax_t)fufh->fh_id);
150 			err = fuse_read_directbackend(vp, uio, cred, fufh);
151 		} else {
152 			FS_DEBUG("buffered read of vnode %ju\n",
153 			      (uintmax_t)VTOILLU(vp));
154 			err = fuse_read_biobackend(vp, uio, cred, fufh);
155 		}
156 		break;
157 	case UIO_WRITE:
158 		/*
159 		 * Kludge: simulate write-through caching via write-around
160 		 * caching.  Same effect, as far as never caching dirty data,
161 		 * but slightly pessimal in that newly written data is not
162 		 * cached.
163 		 */
164 		if (directio || fuse_data_cache_mode == FUSE_CACHE_WT) {
165 			FS_DEBUG("direct write of vnode %ju via file handle %ju\n",
166 			    (uintmax_t)VTOILLU(vp), (uintmax_t)fufh->fh_id);
167 			err = fuse_write_directbackend(vp, uio, cred, fufh, ioflag);
168 		} else {
169 			FS_DEBUG("buffered write of vnode %ju\n",
170 			      (uintmax_t)VTOILLU(vp));
171 			err = fuse_write_biobackend(vp, uio, cred, fufh, ioflag);
172 		}
173 		break;
174 	default:
175 		panic("uninterpreted mode passed to fuse_io_dispatch");
176 	}
177 
178 	return (err);
179 }
180 
181 static int
182 fuse_read_biobackend(struct vnode *vp, struct uio *uio,
183     struct ucred *cred, struct fuse_filehandle *fufh)
184 {
185 	struct buf *bp;
186 	daddr_t lbn;
187 	int bcount;
188 	int err = 0, n = 0, on = 0;
189 	off_t filesize;
190 
191 	const int biosize = fuse_iosize(vp);
192 
193 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
194 	    uio->uio_resid, uio->uio_offset, VTOFUD(vp)->filesize);
195 
196 	if (uio->uio_resid == 0)
197 		return (0);
198 	if (uio->uio_offset < 0)
199 		return (EINVAL);
200 
201 	bcount = MIN(MAXBSIZE, biosize);
202 	filesize = VTOFUD(vp)->filesize;
203 
204 	do {
205 		if (fuse_isdeadfs(vp)) {
206 			err = ENXIO;
207 			break;
208 		}
209 		lbn = uio->uio_offset / biosize;
210 		on = uio->uio_offset & (biosize - 1);
211 
212 		FS_DEBUG2G("biosize %d, lbn %d, on %d\n", biosize, (int)lbn, on);
213 
214 		/*
215 	         * Obtain the buffer cache block.  Figure out the buffer size
216 	         * when we are at EOF.  If we are modifying the size of the
217 	         * buffer based on an EOF condition we need to hold
218 	         * nfs_rslock() through obtaining the buffer to prevent
219 	         * a potential writer-appender from messing with n_size.
220 	         * Otherwise we may accidentally truncate the buffer and
221 	         * lose dirty data.
222 	         *
223 	         * Note that bcount is *not* DEV_BSIZE aligned.
224 	         */
225 		if ((off_t)lbn * biosize >= filesize) {
226 			bcount = 0;
227 		} else if ((off_t)(lbn + 1) * biosize > filesize) {
228 			bcount = filesize - (off_t)lbn *biosize;
229 		}
230 		bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
231 
232 		if (!bp)
233 			return (EINTR);
234 
235 		/*
236 	         * If B_CACHE is not set, we must issue the read.  If this
237 	         * fails, we return an error.
238 	         */
239 
240 		if ((bp->b_flags & B_CACHE) == 0) {
241 			bp->b_iocmd = BIO_READ;
242 			vfs_busy_pages(bp, 0);
243 			err = fuse_io_strategy(vp, bp);
244 			if (err) {
245 				brelse(bp);
246 				return (err);
247 			}
248 		}
249 		/*
250 	         * on is the offset into the current bp.  Figure out how many
251 	         * bytes we can copy out of the bp.  Note that bcount is
252 	         * NOT DEV_BSIZE aligned.
253 	         *
254 	         * Then figure out how many bytes we can copy into the uio.
255 	         */
256 
257 		n = 0;
258 		if (on < bcount)
259 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
260 		if (n > 0) {
261 			FS_DEBUG2G("feeding buffeater with %d bytes of buffer %p,"
262 				" saying %d was asked for\n",
263 				n, bp->b_data + on, n + (int)bp->b_resid);
264 			err = uiomove(bp->b_data + on, n, uio);
265 		}
266 		brelse(bp);
267 		FS_DEBUG2G("end of turn, err %d, uio->uio_resid %zd, n %d\n",
268 		    err, uio->uio_resid, n);
269 	} while (err == 0 && uio->uio_resid > 0 && n > 0);
270 
271 	return (err);
272 }
273 
274 static int
275 fuse_read_directbackend(struct vnode *vp, struct uio *uio,
276     struct ucred *cred, struct fuse_filehandle *fufh)
277 {
278 	struct fuse_dispatcher fdi;
279 	struct fuse_read_in *fri;
280 	int err = 0;
281 
282 	if (uio->uio_resid == 0)
283 		return (0);
284 
285 	fdisp_init(&fdi, 0);
286 
287 	/*
288          * XXX In "normal" case we use an intermediate kernel buffer for
289          * transmitting data from daemon's context to ours. Eventually, we should
290          * get rid of this. Anyway, if the target uio lives in sysspace (we are
291          * called from pageops), and the input data doesn't need kernel-side
292          * processing (we are not called from readdir) we can already invoke
293          * an optimized, "peer-to-peer" I/O routine.
294          */
295 	while (uio->uio_resid > 0) {
296 		fdi.iosize = sizeof(*fri);
297 		fdisp_make_vp(&fdi, FUSE_READ, vp, uio->uio_td, cred);
298 		fri = fdi.indata;
299 		fri->fh = fufh->fh_id;
300 		fri->offset = uio->uio_offset;
301 		fri->size = MIN(uio->uio_resid,
302 		    fuse_get_mpdata(vp->v_mount)->max_read);
303 
304 		FS_DEBUG2G("fri->fh %ju, fri->offset %ju, fri->size %ju\n",
305 			(uintmax_t)fri->fh, (uintmax_t)fri->offset,
306 			(uintmax_t)fri->size);
307 
308 		if ((err = fdisp_wait_answ(&fdi)))
309 			goto out;
310 
311 		FS_DEBUG2G("complete: got iosize=%d, requested fri.size=%zd; "
312 			"resid=%zd offset=%ju\n",
313 			fri->size, fdi.iosize, uio->uio_resid,
314 			(uintmax_t)uio->uio_offset);
315 
316 		if ((err = uiomove(fdi.answ, MIN(fri->size, fdi.iosize), uio)))
317 			break;
318 		if (fdi.iosize < fri->size)
319 			break;
320 	}
321 
322 out:
323 	fdisp_destroy(&fdi);
324 	return (err);
325 }
326 
327 static int
328 fuse_write_directbackend(struct vnode *vp, struct uio *uio,
329     struct ucred *cred, struct fuse_filehandle *fufh, int ioflag)
330 {
331 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
332 	struct fuse_write_in *fwi;
333 	struct fuse_dispatcher fdi;
334 	size_t chunksize;
335 	int diff;
336 	int err = 0;
337 
338 	if (uio->uio_resid == 0)
339 		return (0);
340 	if (ioflag & IO_APPEND)
341 		uio_setoffset(uio, fvdat->filesize);
342 
343 	fdisp_init(&fdi, 0);
344 
345 	while (uio->uio_resid > 0) {
346 		chunksize = MIN(uio->uio_resid,
347 		    fuse_get_mpdata(vp->v_mount)->max_write);
348 
349 		fdi.iosize = sizeof(*fwi) + chunksize;
350 		fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred);
351 
352 		fwi = fdi.indata;
353 		fwi->fh = fufh->fh_id;
354 		fwi->offset = uio->uio_offset;
355 		fwi->size = chunksize;
356 
357 		if ((err = uiomove((char *)fdi.indata + sizeof(*fwi),
358 		    chunksize, uio)))
359 			break;
360 
361 		if ((err = fdisp_wait_answ(&fdi)))
362 			break;
363 
364 		diff = chunksize - ((struct fuse_write_out *)fdi.answ)->size;
365 		if (diff < 0) {
366 			err = EINVAL;
367 			break;
368 		}
369 		uio->uio_resid += diff;
370 		uio->uio_offset -= diff;
371 		if (uio->uio_offset > fvdat->filesize &&
372 		    fuse_data_cache_mode != FUSE_CACHE_UC) {
373 			fuse_vnode_setsize(vp, cred, uio->uio_offset);
374 			fvdat->flag &= ~FN_SIZECHANGE;
375 		}
376 	}
377 
378 	fdisp_destroy(&fdi);
379 
380 	return (err);
381 }
382 
383 static int
384 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
385     struct ucred *cred, struct fuse_filehandle *fufh, int ioflag)
386 {
387 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
388 	struct buf *bp;
389 	daddr_t lbn;
390 	int bcount;
391 	int n, on, err = 0;
392 
393 	const int biosize = fuse_iosize(vp);
394 
395 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
396 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
397 	    uio->uio_resid, uio->uio_offset, fvdat->filesize);
398 	if (vp->v_type != VREG)
399 		return (EIO);
400 	if (uio->uio_offset < 0)
401 		return (EINVAL);
402 	if (uio->uio_resid == 0)
403 		return (0);
404 	if (ioflag & IO_APPEND)
405 		uio_setoffset(uio, fvdat->filesize);
406 
407 	/*
408          * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
409          * would exceed the local maximum per-file write commit size when
410          * combined with those, we must decide whether to flush,
411          * go synchronous, or return err.  We don't bother checking
412          * IO_UNIT -- we just make all writes atomic anyway, as there's
413          * no point optimizing for something that really won't ever happen.
414          */
415 	do {
416 		if (fuse_isdeadfs(vp)) {
417 			err = ENXIO;
418 			break;
419 		}
420 		lbn = uio->uio_offset / biosize;
421 		on = uio->uio_offset & (biosize - 1);
422 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
423 
424 		FS_DEBUG2G("lbn %ju, on %d, n %d, uio offset %ju, uio resid %zd\n",
425 			(uintmax_t)lbn, on, n,
426 			(uintmax_t)uio->uio_offset, uio->uio_resid);
427 
428 again:
429 		/*
430 	         * Handle direct append and file extension cases, calculate
431 	         * unaligned buffer size.
432 	         */
433 		if (uio->uio_offset == fvdat->filesize && n) {
434 			/*
435 	                 * Get the buffer (in its pre-append state to maintain
436 	                 * B_CACHE if it was previously set).  Resize the
437 	                 * nfsnode after we have locked the buffer to prevent
438 	                 * readers from reading garbage.
439 	                 */
440 			bcount = on;
441 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
442 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
443 
444 			if (bp != NULL) {
445 				long save;
446 
447 				err = fuse_vnode_setsize(vp, cred,
448 							 uio->uio_offset + n);
449 				if (err) {
450 					brelse(bp);
451 					break;
452 				}
453 				save = bp->b_flags & B_CACHE;
454 				bcount += n;
455 				allocbuf(bp, bcount);
456 				bp->b_flags |= save;
457 			}
458 		} else {
459 			/*
460 	                 * Obtain the locked cache block first, and then
461 	                 * adjust the file's size as appropriate.
462 	                 */
463 			bcount = on + n;
464 			if ((off_t)lbn * biosize + bcount < fvdat->filesize) {
465 				if ((off_t)(lbn + 1) * biosize < fvdat->filesize)
466 					bcount = biosize;
467 				else
468 					bcount = fvdat->filesize -
469 					  (off_t)lbn *biosize;
470 			}
471 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
472 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
473 			if (bp && uio->uio_offset + n > fvdat->filesize) {
474 				err = fuse_vnode_setsize(vp, cred,
475 							 uio->uio_offset + n);
476 				if (err) {
477 					brelse(bp);
478 					break;
479 				}
480 			}
481 		}
482 
483 		if (!bp) {
484 			err = EINTR;
485 			break;
486 		}
487 		/*
488 	         * Issue a READ if B_CACHE is not set.  In special-append
489 	         * mode, B_CACHE is based on the buffer prior to the write
490 	         * op and is typically set, avoiding the read.  If a read
491 	         * is required in special append mode, the server will
492 	         * probably send us a short-read since we extended the file
493 	         * on our end, resulting in b_resid == 0 and, thusly,
494 	         * B_CACHE getting set.
495 	         *
496 	         * We can also avoid issuing the read if the write covers
497 	         * the entire buffer.  We have to make sure the buffer state
498 	         * is reasonable in this case since we will not be initiating
499 	         * I/O.  See the comments in kern/vfs_bio.c's getblk() for
500 	         * more information.
501 	         *
502 	         * B_CACHE may also be set due to the buffer being cached
503 	         * normally.
504 	         */
505 
506 		if (on == 0 && n == bcount) {
507 			bp->b_flags |= B_CACHE;
508 			bp->b_flags &= ~B_INVAL;
509 			bp->b_ioflags &= ~BIO_ERROR;
510 		}
511 		if ((bp->b_flags & B_CACHE) == 0) {
512 			bp->b_iocmd = BIO_READ;
513 			vfs_busy_pages(bp, 0);
514 			fuse_io_strategy(vp, bp);
515 			if ((err = bp->b_error)) {
516 				brelse(bp);
517 				break;
518 			}
519 		}
520 		if (bp->b_wcred == NOCRED)
521 			bp->b_wcred = crhold(cred);
522 
523 		/*
524 	         * If dirtyend exceeds file size, chop it down.  This should
525 	         * not normally occur but there is an append race where it
526 	         * might occur XXX, so we log it.
527 	         *
528 	         * If the chopping creates a reverse-indexed or degenerate
529 	         * situation with dirtyoff/end, we 0 both of them.
530 	         */
531 
532 		if (bp->b_dirtyend > bcount) {
533 			FS_DEBUG("FUSE append race @%lx:%d\n",
534 			    (long)bp->b_blkno * biosize,
535 			    bp->b_dirtyend - bcount);
536 			bp->b_dirtyend = bcount;
537 		}
538 		if (bp->b_dirtyoff >= bp->b_dirtyend)
539 			bp->b_dirtyoff = bp->b_dirtyend = 0;
540 
541 		/*
542 	         * If the new write will leave a contiguous dirty
543 	         * area, just update the b_dirtyoff and b_dirtyend,
544 	         * otherwise force a write rpc of the old dirty area.
545 	         *
546 	         * While it is possible to merge discontiguous writes due to
547 	         * our having a B_CACHE buffer ( and thus valid read data
548 	         * for the hole), we don't because it could lead to
549 	         * significant cache coherency problems with multiple clients,
550 	         * especially if locking is implemented later on.
551 	         *
552 	         * as an optimization we could theoretically maintain
553 	         * a linked list of discontinuous areas, but we would still
554 	         * have to commit them separately so there isn't much
555 	         * advantage to it except perhaps a bit of asynchronization.
556 	         */
557 
558 		if (bp->b_dirtyend > 0 &&
559 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
560 			/*
561 	                 * Yes, we mean it. Write out everything to "storage"
562 	                 * immediately, without hesitation. (Apart from other
563 	                 * reasons: the only way to know if a write is valid
564 	                 * if its actually written out.)
565 	                 */
566 			bwrite(bp);
567 			if (bp->b_error == EINTR) {
568 				err = EINTR;
569 				break;
570 			}
571 			goto again;
572 		}
573 		err = uiomove((char *)bp->b_data + on, n, uio);
574 
575 		/*
576 	         * Since this block is being modified, it must be written
577 	         * again and not just committed.  Since write clustering does
578 	         * not work for the stage 1 data write, only the stage 2
579 	         * commit rpc, we have to clear B_CLUSTEROK as well.
580 	         */
581 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
582 
583 		if (err) {
584 			bp->b_ioflags |= BIO_ERROR;
585 			bp->b_error = err;
586 			brelse(bp);
587 			break;
588 		}
589 		/*
590 	         * Only update dirtyoff/dirtyend if not a degenerate
591 	         * condition.
592 	         */
593 		if (n) {
594 			if (bp->b_dirtyend > 0) {
595 				bp->b_dirtyoff = MIN(on, bp->b_dirtyoff);
596 				bp->b_dirtyend = MAX((on + n), bp->b_dirtyend);
597 			} else {
598 				bp->b_dirtyoff = on;
599 				bp->b_dirtyend = on + n;
600 			}
601 			vfs_bio_set_valid(bp, on, n);
602 		}
603 		err = bwrite(bp);
604 		if (err)
605 			break;
606 	} while (uio->uio_resid > 0 && n > 0);
607 
608 	if (fuse_sync_resize && (fvdat->flag & FN_SIZECHANGE) != 0)
609 		fuse_vnode_savesize(vp, cred);
610 
611 	return (err);
612 }
613 
614 int
615 fuse_io_strategy(struct vnode *vp, struct buf *bp)
616 {
617 	struct fuse_filehandle *fufh;
618 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
619 	struct ucred *cred;
620 	struct uio *uiop;
621 	struct uio uio;
622 	struct iovec io;
623 	int error = 0;
624 
625 	const int biosize = fuse_iosize(vp);
626 
627 	MPASS(vp->v_type == VREG || vp->v_type == VDIR);
628 	MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE);
629 	FS_DEBUG("inode=%ju offset=%jd resid=%ld\n",
630 	    (uintmax_t)VTOI(vp), (intmax_t)(((off_t)bp->b_blkno) * biosize),
631 	    bp->b_bcount);
632 
633 	error = fuse_filehandle_getrw(vp,
634 	    (bp->b_iocmd == BIO_READ) ? FUFH_RDONLY : FUFH_WRONLY, &fufh);
635 	if (error) {
636 		printf("FUSE: strategy: filehandles are closed\n");
637 		bp->b_ioflags |= BIO_ERROR;
638 		bp->b_error = error;
639 		return (error);
640 	}
641 	cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred;
642 
643 	uiop = &uio;
644 	uiop->uio_iov = &io;
645 	uiop->uio_iovcnt = 1;
646 	uiop->uio_segflg = UIO_SYSSPACE;
647 	uiop->uio_td = curthread;
648 
649 	/*
650          * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
651          * do this here so we do not have to do it in all the code that
652          * calls us.
653          */
654 	bp->b_flags &= ~B_INVAL;
655 	bp->b_ioflags &= ~BIO_ERROR;
656 
657 	KASSERT(!(bp->b_flags & B_DONE),
658 	    ("fuse_io_strategy: bp %p already marked done", bp));
659 	if (bp->b_iocmd == BIO_READ) {
660 		io.iov_len = uiop->uio_resid = bp->b_bcount;
661 		io.iov_base = bp->b_data;
662 		uiop->uio_rw = UIO_READ;
663 
664 		uiop->uio_offset = ((off_t)bp->b_blkno) * biosize;
665 		error = fuse_read_directbackend(vp, uiop, cred, fufh);
666 
667 		/* XXXCEM: Potentially invalid access to cached_attrs here */
668 		if ((!error && uiop->uio_resid) ||
669 		    (fsess_opt_brokenio(vnode_mount(vp)) && error == EIO &&
670 		    uiop->uio_offset < fvdat->filesize && fvdat->filesize > 0 &&
671 		    uiop->uio_offset >= fvdat->cached_attrs.va_size)) {
672 			/*
673 	                 * If we had a short read with no error, we must have
674 	                 * hit a file hole.  We should zero-fill the remainder.
675 	                 * This can also occur if the server hits the file EOF.
676 	                 *
677 	                 * Holes used to be able to occur due to pending
678 	                 * writes, but that is not possible any longer.
679 	                 */
680 			int nread = bp->b_bcount - uiop->uio_resid;
681 			int left = uiop->uio_resid;
682 
683 			if (error != 0) {
684 				printf("FUSE: Fix broken io: offset %ju, "
685 				       " resid %zd, file size %ju/%ju\n",
686 				       (uintmax_t)uiop->uio_offset,
687 				    uiop->uio_resid, fvdat->filesize,
688 				    fvdat->cached_attrs.va_size);
689 				error = 0;
690 			}
691 			if (left > 0)
692 				bzero((char *)bp->b_data + nread, left);
693 			uiop->uio_resid = 0;
694 		}
695 		if (error) {
696 			bp->b_ioflags |= BIO_ERROR;
697 			bp->b_error = error;
698 		}
699 	} else {
700 		/*
701 	         * If we only need to commit, try to commit
702 	         */
703 		if (bp->b_flags & B_NEEDCOMMIT) {
704 			FS_DEBUG("write: B_NEEDCOMMIT flags set\n");
705 		}
706 		/*
707 	         * Setup for actual write
708 	         */
709 		if ((off_t)bp->b_blkno * biosize + bp->b_dirtyend >
710 		    fvdat->filesize)
711 			bp->b_dirtyend = fvdat->filesize -
712 				(off_t)bp->b_blkno * biosize;
713 
714 		if (bp->b_dirtyend > bp->b_dirtyoff) {
715 			io.iov_len = uiop->uio_resid = bp->b_dirtyend
716 			    - bp->b_dirtyoff;
717 			uiop->uio_offset = (off_t)bp->b_blkno * biosize
718 			    + bp->b_dirtyoff;
719 			io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
720 			uiop->uio_rw = UIO_WRITE;
721 
722 			error = fuse_write_directbackend(vp, uiop, cred, fufh, 0);
723 
724 			if (error == EINTR || error == ETIMEDOUT
725 			    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
726 
727 				bp->b_flags &= ~(B_INVAL | B_NOCACHE);
728 				if ((bp->b_flags & B_PAGING) == 0) {
729 					bdirty(bp);
730 					bp->b_flags &= ~B_DONE;
731 				}
732 				if ((error == EINTR || error == ETIMEDOUT) &&
733 				    (bp->b_flags & B_ASYNC) == 0)
734 					bp->b_flags |= B_EINTR;
735 			} else {
736 				if (error) {
737 					bp->b_ioflags |= BIO_ERROR;
738 					bp->b_flags |= B_INVAL;
739 					bp->b_error = error;
740 				}
741 				bp->b_dirtyoff = bp->b_dirtyend = 0;
742 			}
743 		} else {
744 			bp->b_resid = 0;
745 			bufdone(bp);
746 			return (0);
747 		}
748 	}
749 	bp->b_resid = uiop->uio_resid;
750 	bufdone(bp);
751 	return (error);
752 }
753 
754 int
755 fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td)
756 {
757 	struct vop_fsync_args a = {
758 		.a_vp = vp,
759 		.a_waitfor = waitfor,
760 		.a_td = td,
761 	};
762 
763 	return (vop_stdfsync(&a));
764 }
765 
766 /*
767  * Flush and invalidate all dirty buffers. If another process is already
768  * doing the flush, just wait for completion.
769  */
770 int
771 fuse_io_invalbuf(struct vnode *vp, struct thread *td)
772 {
773 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
774 	int error = 0;
775 
776 	if (vp->v_iflag & VI_DOOMED)
777 		return 0;
778 
779 	ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf");
780 
781 	while (fvdat->flag & FN_FLUSHINPROG) {
782 		struct proc *p = td->td_proc;
783 
784 		if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF)
785 			return EIO;
786 		fvdat->flag |= FN_FLUSHWANT;
787 		tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz);
788 		error = 0;
789 		if (p != NULL) {
790 			PROC_LOCK(p);
791 			if (SIGNOTEMPTY(p->p_siglist) ||
792 			    SIGNOTEMPTY(td->td_siglist))
793 				error = EINTR;
794 			PROC_UNLOCK(p);
795 		}
796 		if (error == EINTR)
797 			return EINTR;
798 	}
799 	fvdat->flag |= FN_FLUSHINPROG;
800 
801 	if (vp->v_bufobj.bo_object != NULL) {
802 		VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
803 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
804 		VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
805 	}
806 	error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
807 	while (error) {
808 		if (error == ERESTART || error == EINTR) {
809 			fvdat->flag &= ~FN_FLUSHINPROG;
810 			if (fvdat->flag & FN_FLUSHWANT) {
811 				fvdat->flag &= ~FN_FLUSHWANT;
812 				wakeup(&fvdat->flag);
813 			}
814 			return EINTR;
815 		}
816 		error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
817 	}
818 	fvdat->flag &= ~FN_FLUSHINPROG;
819 	if (fvdat->flag & FN_FLUSHWANT) {
820 		fvdat->flag &= ~FN_FLUSHWANT;
821 		wakeup(&fvdat->flag);
822 	}
823 	return (error);
824 }
825