xref: /freebsd/sys/fs/fuse/fuse_io.c (revision 724b4bfdf1306e4f2c451b6d146fe0fe0353b2c8)
1 /*
2  * Copyright (c) 2007-2009 Google Inc.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are
7  * met:
8  *
9  * * Redistributions of source code must retain the above copyright
10  *   notice, this list of conditions and the following disclaimer.
11  * * Redistributions in binary form must reproduce the above
12  *   copyright notice, this list of conditions and the following disclaimer
13  *   in the documentation and/or other materials provided with the
14  *   distribution.
15  * * Neither the name of Google Inc. nor the names of its
16  *   contributors may be used to endorse or promote products derived from
17  *   this software without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
22  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
23  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
25  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
29  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  *
31  * Copyright (C) 2005 Csaba Henk.
32  * All rights reserved.
33  *
34  * Redistribution and use in source and binary forms, with or without
35  * modification, are permitted provided that the following conditions
36  * are met:
37  * 1. Redistributions of source code must retain the above copyright
38  *    notice, this list of conditions and the following disclaimer.
39  * 2. Redistributions in binary form must reproduce the above copyright
40  *    notice, this list of conditions and the following disclaimer in the
41  *    documentation and/or other materials provided with the distribution.
42  *
43  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND
44  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
45  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
46  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
47  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
48  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
49  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
50  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
51  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
52  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
53  * SUCH DAMAGE.
54  */
55 
56 #include <sys/cdefs.h>
57 __FBSDID("$FreeBSD$");
58 
59 #include <sys/types.h>
60 #include <sys/module.h>
61 #include <sys/systm.h>
62 #include <sys/errno.h>
63 #include <sys/param.h>
64 #include <sys/kernel.h>
65 #include <sys/conf.h>
66 #include <sys/uio.h>
67 #include <sys/malloc.h>
68 #include <sys/queue.h>
69 #include <sys/lock.h>
70 #include <sys/sx.h>
71 #include <sys/mutex.h>
72 #include <sys/proc.h>
73 #include <sys/mount.h>
74 #include <sys/vnode.h>
75 #include <sys/stat.h>
76 #include <sys/unistd.h>
77 #include <sys/filedesc.h>
78 #include <sys/file.h>
79 #include <sys/fcntl.h>
80 #include <sys/bio.h>
81 #include <sys/buf.h>
82 #include <sys/sysctl.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_extern.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vnode_pager.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);
114 static int
115 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
116     struct ucred *cred, struct fuse_filehandle *fufh);
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);
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 		if (directio) {
159 			FS_DEBUG("direct write of vnode %ju via file handle %ju\n",
160 			    (uintmax_t)VTOILLU(vp), (uintmax_t)fufh->fh_id);
161 			err = fuse_write_directbackend(vp, uio, cred, fufh);
162 			fuse_invalidate_attr(vp);
163 		} else {
164 			FS_DEBUG("buffered write of vnode %ju\n",
165 			      (uintmax_t)VTOILLU(vp));
166 			err = fuse_write_biobackend(vp, uio, cred, fufh);
167 		}
168 		break;
169 	default:
170 		panic("uninterpreted mode passed to fuse_io_dispatch");
171 	}
172 
173 	return (err);
174 }
175 
176 static int
177 fuse_read_biobackend(struct vnode *vp, struct uio *uio,
178     struct ucred *cred, struct fuse_filehandle *fufh)
179 {
180 	struct buf *bp;
181 	daddr_t lbn;
182 	int bcount;
183 	int err = 0, n = 0, on = 0;
184 	off_t filesize;
185 
186 	const int biosize = fuse_iosize(vp);
187 
188 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
189 	    uio->uio_resid, uio->uio_offset, VTOFUD(vp)->filesize);
190 
191 	if (uio->uio_resid == 0)
192 		return (0);
193 	if (uio->uio_offset < 0)
194 		return (EINVAL);
195 
196 	bcount = MIN(MAXBSIZE, biosize);
197 	filesize = VTOFUD(vp)->filesize;
198 
199 	do {
200 		if (fuse_isdeadfs(vp)) {
201 			err = ENXIO;
202 			break;
203 		}
204 		lbn = uio->uio_offset / biosize;
205 		on = uio->uio_offset & (biosize - 1);
206 
207 		FS_DEBUG2G("biosize %d, lbn %d, on %d\n", biosize, (int)lbn, on);
208 
209 		/*
210 	         * Obtain the buffer cache block.  Figure out the buffer size
211 	         * when we are at EOF.  If we are modifying the size of the
212 	         * buffer based on an EOF condition we need to hold
213 	         * nfs_rslock() through obtaining the buffer to prevent
214 	         * a potential writer-appender from messing with n_size.
215 	         * Otherwise we may accidently truncate the buffer and
216 	         * lose dirty data.
217 	         *
218 	         * Note that bcount is *not* DEV_BSIZE aligned.
219 	         */
220 		if ((off_t)lbn * biosize >= filesize) {
221 			bcount = 0;
222 		} else if ((off_t)(lbn + 1) * biosize > filesize) {
223 			bcount = filesize - (off_t)lbn *biosize;
224 		}
225 		bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
226 
227 		if (!bp)
228 			return (EINTR);
229 
230 		/*
231 	         * If B_CACHE is not set, we must issue the read.  If this
232 	         * fails, we return an error.
233 	         */
234 
235 		if ((bp->b_flags & B_CACHE) == 0) {
236 			bp->b_iocmd = BIO_READ;
237 			vfs_busy_pages(bp, 0);
238 			err = fuse_io_strategy(vp, bp);
239 			if (err) {
240 				brelse(bp);
241 				return (err);
242 			}
243 		}
244 		/*
245 	         * on is the offset into the current bp.  Figure out how many
246 	         * bytes we can copy out of the bp.  Note that bcount is
247 	         * NOT DEV_BSIZE aligned.
248 	         *
249 	         * Then figure out how many bytes we can copy into the uio.
250 	         */
251 
252 		n = 0;
253 		if (on < bcount)
254 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
255 		if (n > 0) {
256 			FS_DEBUG2G("feeding buffeater with %d bytes of buffer %p,"
257 				" saying %d was asked for\n",
258 				n, bp->b_data + on, n + (int)bp->b_resid);
259 			err = uiomove(bp->b_data + on, n, uio);
260 		}
261 		brelse(bp);
262 		FS_DEBUG2G("end of turn, err %d, uio->uio_resid %zd, n %d\n",
263 		    err, uio->uio_resid, n);
264 	} while (err == 0 && uio->uio_resid > 0 && n > 0);
265 
266 	return (err);
267 }
268 
269 static int
270 fuse_read_directbackend(struct vnode *vp, struct uio *uio,
271     struct ucred *cred, struct fuse_filehandle *fufh)
272 {
273 	struct fuse_dispatcher fdi;
274 	struct fuse_read_in *fri;
275 	int err = 0;
276 
277 	if (uio->uio_resid == 0)
278 		return (0);
279 
280 	fdisp_init(&fdi, 0);
281 
282 	/*
283          * XXX In "normal" case we use an intermediate kernel buffer for
284          * transmitting data from daemon's context to ours. Eventually, we should
285          * get rid of this. Anyway, if the target uio lives in sysspace (we are
286          * called from pageops), and the input data doesn't need kernel-side
287          * processing (we are not called from readdir) we can already invoke
288          * an optimized, "peer-to-peer" I/O routine.
289          */
290 	while (uio->uio_resid > 0) {
291 		fdi.iosize = sizeof(*fri);
292 		fdisp_make_vp(&fdi, FUSE_READ, vp, uio->uio_td, cred);
293 		fri = fdi.indata;
294 		fri->fh = fufh->fh_id;
295 		fri->offset = uio->uio_offset;
296 		fri->size = MIN(uio->uio_resid,
297 		    fuse_get_mpdata(vp->v_mount)->max_read);
298 
299 		FS_DEBUG2G("fri->fh %ju, fri->offset %ju, fri->size %ju\n",
300 			(uintmax_t)fri->fh, (uintmax_t)fri->offset,
301 			(uintmax_t)fri->size);
302 
303 		if ((err = fdisp_wait_answ(&fdi)))
304 			goto out;
305 
306 		FS_DEBUG2G("complete: got iosize=%d, requested fri.size=%zd; "
307 			"resid=%zd offset=%ju\n",
308 			fri->size, fdi.iosize, uio->uio_resid,
309 			(uintmax_t)uio->uio_offset);
310 
311 		if ((err = uiomove(fdi.answ, MIN(fri->size, fdi.iosize), uio)))
312 			break;
313 		if (fdi.iosize < fri->size)
314 			break;
315 	}
316 
317 out:
318 	fdisp_destroy(&fdi);
319 	return (err);
320 }
321 
322 static int
323 fuse_write_directbackend(struct vnode *vp, struct uio *uio,
324     struct ucred *cred, struct fuse_filehandle *fufh)
325 {
326 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
327 	struct fuse_write_in *fwi;
328 	struct fuse_dispatcher fdi;
329 	size_t chunksize;
330 	int diff;
331 	int err = 0;
332 
333 	if (!uio->uio_resid)
334 		return (0);
335 
336 	fdisp_init(&fdi, 0);
337 
338 	while (uio->uio_resid > 0) {
339 		chunksize = MIN(uio->uio_resid,
340 		    fuse_get_mpdata(vp->v_mount)->max_write);
341 
342 		fdi.iosize = sizeof(*fwi) + chunksize;
343 		fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred);
344 
345 		fwi = fdi.indata;
346 		fwi->fh = fufh->fh_id;
347 		fwi->offset = uio->uio_offset;
348 		fwi->size = chunksize;
349 
350 		if ((err = uiomove((char *)fdi.indata + sizeof(*fwi),
351 		    chunksize, uio)))
352 			break;
353 
354 		if ((err = fdisp_wait_answ(&fdi)))
355 			break;
356 
357 		diff = chunksize - ((struct fuse_write_out *)fdi.answ)->size;
358 		if (diff < 0) {
359 			err = EINVAL;
360 			break;
361 		}
362 		uio->uio_resid += diff;
363 		uio->uio_offset -= diff;
364 		if (uio->uio_offset > fvdat->filesize)
365 			fuse_vnode_setsize(vp, cred, uio->uio_offset);
366 	}
367 
368 	fdisp_destroy(&fdi);
369 
370 	return (err);
371 }
372 
373 static int
374 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
375     struct ucred *cred, struct fuse_filehandle *fufh)
376 {
377 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
378 	struct buf *bp;
379 	daddr_t lbn;
380 	int bcount;
381 	int n, on, err = 0;
382 
383 	const int biosize = fuse_iosize(vp);
384 
385 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
386 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
387 	    uio->uio_resid, uio->uio_offset, fvdat->filesize);
388 	if (vp->v_type != VREG)
389 		return (EIO);
390 	if (uio->uio_offset < 0)
391 		return (EINVAL);
392 	if (uio->uio_resid == 0)
393 		return (0);
394 
395 	/*
396          * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
397          * would exceed the local maximum per-file write commit size when
398          * combined with those, we must decide whether to flush,
399          * go synchronous, or return err.  We don't bother checking
400          * IO_UNIT -- we just make all writes atomic anyway, as there's
401          * no point optimizing for something that really won't ever happen.
402          */
403 	do {
404 		if (fuse_isdeadfs(vp)) {
405 			err = ENXIO;
406 			break;
407 		}
408 		lbn = uio->uio_offset / biosize;
409 		on = uio->uio_offset & (biosize - 1);
410 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
411 
412 		FS_DEBUG2G("lbn %ju, on %d, n %d, uio offset %ju, uio resid %zd\n",
413 			(uintmax_t)lbn, on, n,
414 			(uintmax_t)uio->uio_offset, uio->uio_resid);
415 
416 again:
417 		/*
418 	         * Handle direct append and file extension cases, calculate
419 	         * unaligned buffer size.
420 	         */
421 		if (uio->uio_offset == fvdat->filesize && n) {
422 			/*
423 	                 * Get the buffer (in its pre-append state to maintain
424 	                 * B_CACHE if it was previously set).  Resize the
425 	                 * nfsnode after we have locked the buffer to prevent
426 	                 * readers from reading garbage.
427 	                 */
428 			bcount = on;
429 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
430 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
431 
432 			if (bp != NULL) {
433 				long save;
434 
435 				err = fuse_vnode_setsize(vp, cred,
436 							 uio->uio_offset + n);
437 				if (err) {
438 					brelse(bp);
439 					break;
440 				}
441 				save = bp->b_flags & B_CACHE;
442 				bcount += n;
443 				allocbuf(bp, bcount);
444 				bp->b_flags |= save;
445 			}
446 		} else {
447 			/*
448 	                 * Obtain the locked cache block first, and then
449 	                 * adjust the file's size as appropriate.
450 	                 */
451 			bcount = on + n;
452 			if ((off_t)lbn * biosize + bcount < fvdat->filesize) {
453 				if ((off_t)(lbn + 1) * biosize < fvdat->filesize)
454 					bcount = biosize;
455 				else
456 					bcount = fvdat->filesize -
457 					  (off_t)lbn *biosize;
458 			}
459 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
460 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
461 			if (bp && uio->uio_offset + n > fvdat->filesize) {
462 				err = fuse_vnode_setsize(vp, cred,
463 							 uio->uio_offset + n);
464 				if (err) {
465 					brelse(bp);
466 					break;
467 				}
468 			}
469 		}
470 
471 		if (!bp) {
472 			err = EINTR;
473 			break;
474 		}
475 		/*
476 	         * Issue a READ if B_CACHE is not set.  In special-append
477 	         * mode, B_CACHE is based on the buffer prior to the write
478 	         * op and is typically set, avoiding the read.  If a read
479 	         * is required in special append mode, the server will
480 	         * probably send us a short-read since we extended the file
481 	         * on our end, resulting in b_resid == 0 and, thusly,
482 	         * B_CACHE getting set.
483 	         *
484 	         * We can also avoid issuing the read if the write covers
485 	         * the entire buffer.  We have to make sure the buffer state
486 	         * is reasonable in this case since we will not be initiating
487 	         * I/O.  See the comments in kern/vfs_bio.c's getblk() for
488 	         * more information.
489 	         *
490 	         * B_CACHE may also be set due to the buffer being cached
491 	         * normally.
492 	         */
493 
494 		if (on == 0 && n == bcount) {
495 			bp->b_flags |= B_CACHE;
496 			bp->b_flags &= ~B_INVAL;
497 			bp->b_ioflags &= ~BIO_ERROR;
498 		}
499 		if ((bp->b_flags & B_CACHE) == 0) {
500 			bp->b_iocmd = BIO_READ;
501 			vfs_busy_pages(bp, 0);
502 			fuse_io_strategy(vp, bp);
503 			if ((err = bp->b_error)) {
504 				brelse(bp);
505 				break;
506 			}
507 		}
508 		if (bp->b_wcred == NOCRED)
509 			bp->b_wcred = crhold(cred);
510 
511 		/*
512 	         * If dirtyend exceeds file size, chop it down.  This should
513 	         * not normally occur but there is an append race where it
514 	         * might occur XXX, so we log it.
515 	         *
516 	         * If the chopping creates a reverse-indexed or degenerate
517 	         * situation with dirtyoff/end, we 0 both of them.
518 	         */
519 
520 		if (bp->b_dirtyend > bcount) {
521 			FS_DEBUG("FUSE append race @%lx:%d\n",
522 			    (long)bp->b_blkno * biosize,
523 			    bp->b_dirtyend - bcount);
524 			bp->b_dirtyend = bcount;
525 		}
526 		if (bp->b_dirtyoff >= bp->b_dirtyend)
527 			bp->b_dirtyoff = bp->b_dirtyend = 0;
528 
529 		/*
530 	         * If the new write will leave a contiguous dirty
531 	         * area, just update the b_dirtyoff and b_dirtyend,
532 	         * otherwise force a write rpc of the old dirty area.
533 	         *
534 	         * While it is possible to merge discontiguous writes due to
535 	         * our having a B_CACHE buffer ( and thus valid read data
536 	         * for the hole), we don't because it could lead to
537 	         * significant cache coherency problems with multiple clients,
538 	         * especially if locking is implemented later on.
539 	         *
540 	         * as an optimization we could theoretically maintain
541 	         * a linked list of discontinuous areas, but we would still
542 	         * have to commit them separately so there isn't much
543 	         * advantage to it except perhaps a bit of asynchronization.
544 	         */
545 
546 		if (bp->b_dirtyend > 0 &&
547 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
548 			/*
549 	                 * Yes, we mean it. Write out everything to "storage"
550 	                 * immediatly, without hesitation. (Apart from other
551 	                 * reasons: the only way to know if a write is valid
552 	                 * if its actually written out.)
553 	                 */
554 			bwrite(bp);
555 			if (bp->b_error == EINTR) {
556 				err = EINTR;
557 				break;
558 			}
559 			goto again;
560 		}
561 		err = uiomove((char *)bp->b_data + on, n, uio);
562 
563 		/*
564 	         * Since this block is being modified, it must be written
565 	         * again and not just committed.  Since write clustering does
566 	         * not work for the stage 1 data write, only the stage 2
567 	         * commit rpc, we have to clear B_CLUSTEROK as well.
568 	         */
569 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
570 
571 		if (err) {
572 			bp->b_ioflags |= BIO_ERROR;
573 			bp->b_error = err;
574 			brelse(bp);
575 			break;
576 		}
577 		/*
578 	         * Only update dirtyoff/dirtyend if not a degenerate
579 	         * condition.
580 	         */
581 		if (n) {
582 			if (bp->b_dirtyend > 0) {
583 				bp->b_dirtyoff = MIN(on, bp->b_dirtyoff);
584 				bp->b_dirtyend = MAX((on + n), bp->b_dirtyend);
585 			} else {
586 				bp->b_dirtyoff = on;
587 				bp->b_dirtyend = on + n;
588 			}
589 			vfs_bio_set_valid(bp, on, n);
590 		}
591 		err = bwrite(bp);
592 		if (err)
593 			break;
594 	} while (uio->uio_resid > 0 && n > 0);
595 
596 	if (fuse_sync_resize && (fvdat->flag & FN_SIZECHANGE) != 0)
597 		fuse_vnode_savesize(vp, cred);
598 
599 	return (err);
600 }
601 
602 int
603 fuse_io_strategy(struct vnode *vp, struct buf *bp)
604 {
605 	struct fuse_filehandle *fufh;
606 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
607 	struct ucred *cred;
608 	struct uio *uiop;
609 	struct uio uio;
610 	struct iovec io;
611 	int error = 0;
612 
613 	const int biosize = fuse_iosize(vp);
614 
615 	MPASS(vp->v_type == VREG);
616 	MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE);
617 	FS_DEBUG("inode=%ju offset=%jd resid=%ld\n",
618 	    (uintmax_t)VTOI(vp), (intmax_t)(((off_t)bp->b_blkno) * biosize),
619 	    bp->b_bcount);
620 
621 	error = fuse_filehandle_getrw(vp,
622 	    (bp->b_iocmd == BIO_READ) ? FUFH_RDONLY : FUFH_WRONLY, &fufh);
623 	if (error) {
624 		printf("FUSE: strategy: filehandles are closed\n");
625 		bp->b_ioflags |= BIO_ERROR;
626 		bp->b_error = error;
627 		return (error);
628 	}
629 	cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred;
630 
631 	uiop = &uio;
632 	uiop->uio_iov = &io;
633 	uiop->uio_iovcnt = 1;
634 	uiop->uio_segflg = UIO_SYSSPACE;
635 	uiop->uio_td = curthread;
636 
637 	/*
638          * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
639          * do this here so we do not have to do it in all the code that
640          * calls us.
641          */
642 	bp->b_flags &= ~B_INVAL;
643 	bp->b_ioflags &= ~BIO_ERROR;
644 
645 	KASSERT(!(bp->b_flags & B_DONE),
646 	    ("fuse_io_strategy: bp %p already marked done", bp));
647 	if (bp->b_iocmd == BIO_READ) {
648 		io.iov_len = uiop->uio_resid = bp->b_bcount;
649 		io.iov_base = bp->b_data;
650 		uiop->uio_rw = UIO_READ;
651 
652 		uiop->uio_offset = ((off_t)bp->b_blkno) * biosize;
653 		error = fuse_read_directbackend(vp, uiop, cred, fufh);
654 
655 		if ((!error && uiop->uio_resid) ||
656 		    (fsess_opt_brokenio(vnode_mount(vp)) && error == EIO &&
657 		    uiop->uio_offset < fvdat->filesize && fvdat->filesize > 0 &&
658 		    uiop->uio_offset >= fvdat->cached_attrs.va_size)) {
659 			/*
660 	                 * If we had a short read with no error, we must have
661 	                 * hit a file hole.  We should zero-fill the remainder.
662 	                 * This can also occur if the server hits the file EOF.
663 	                 *
664 	                 * Holes used to be able to occur due to pending
665 	                 * writes, but that is not possible any longer.
666 	                 */
667 			int nread = bp->b_bcount - uiop->uio_resid;
668 			int left = uiop->uio_resid;
669 
670 			if (error != 0) {
671 				printf("FUSE: Fix broken io: offset %ju, "
672 				       " resid %zd, file size %ju/%ju\n",
673 				       (uintmax_t)uiop->uio_offset,
674 				    uiop->uio_resid, fvdat->filesize,
675 				    fvdat->cached_attrs.va_size);
676 				error = 0;
677 			}
678 			if (left > 0)
679 				bzero((char *)bp->b_data + nread, left);
680 			uiop->uio_resid = 0;
681 		}
682 		if (error) {
683 			bp->b_ioflags |= BIO_ERROR;
684 			bp->b_error = error;
685 		}
686 	} else {
687 		/*
688 	         * If we only need to commit, try to commit
689 	         */
690 		if (bp->b_flags & B_NEEDCOMMIT) {
691 			FS_DEBUG("write: B_NEEDCOMMIT flags set\n");
692 		}
693 		/*
694 	         * Setup for actual write
695 	         */
696 		if ((off_t)bp->b_blkno * biosize + bp->b_dirtyend >
697 		    fvdat->filesize)
698 			bp->b_dirtyend = fvdat->filesize -
699 				(off_t)bp->b_blkno * biosize;
700 
701 		if (bp->b_dirtyend > bp->b_dirtyoff) {
702 			io.iov_len = uiop->uio_resid = bp->b_dirtyend
703 			    - bp->b_dirtyoff;
704 			uiop->uio_offset = (off_t)bp->b_blkno * biosize
705 			    + bp->b_dirtyoff;
706 			io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
707 			uiop->uio_rw = UIO_WRITE;
708 
709 			error = fuse_write_directbackend(vp, uiop, cred, fufh);
710 
711 			if (error == EINTR || error == ETIMEDOUT
712 			    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
713 
714 				bp->b_flags &= ~(B_INVAL | B_NOCACHE);
715 				if ((bp->b_flags & B_PAGING) == 0) {
716 					bdirty(bp);
717 					bp->b_flags &= ~B_DONE;
718 				}
719 				if ((error == EINTR || error == ETIMEDOUT) &&
720 				    (bp->b_flags & B_ASYNC) == 0)
721 					bp->b_flags |= B_EINTR;
722 			} else {
723 				if (error) {
724 					bp->b_ioflags |= BIO_ERROR;
725 					bp->b_flags |= B_INVAL;
726 					bp->b_error = error;
727 				}
728 				bp->b_dirtyoff = bp->b_dirtyend = 0;
729 			}
730 		} else {
731 			bp->b_resid = 0;
732 			bufdone(bp);
733 			return (0);
734 		}
735 	}
736 	bp->b_resid = uiop->uio_resid;
737 	bufdone(bp);
738 	return (error);
739 }
740 
741 int
742 fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td)
743 {
744 	struct vop_fsync_args a = {
745 		.a_vp = vp,
746 		.a_waitfor = waitfor,
747 		.a_td = td,
748 	};
749 
750 	return (vop_stdfsync(&a));
751 }
752 
753 /*
754  * Flush and invalidate all dirty buffers. If another process is already
755  * doing the flush, just wait for completion.
756  */
757 int
758 fuse_io_invalbuf(struct vnode *vp, struct thread *td)
759 {
760 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
761 	int error = 0;
762 
763 	if (vp->v_iflag & VI_DOOMED)
764 		return 0;
765 
766 	ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf");
767 
768 	while (fvdat->flag & FN_FLUSHINPROG) {
769 		struct proc *p = td->td_proc;
770 
771 		if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF)
772 			return EIO;
773 		fvdat->flag |= FN_FLUSHWANT;
774 		tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz);
775 		error = 0;
776 		if (p != NULL) {
777 			PROC_LOCK(p);
778 			if (SIGNOTEMPTY(p->p_siglist) ||
779 			    SIGNOTEMPTY(td->td_siglist))
780 				error = EINTR;
781 			PROC_UNLOCK(p);
782 		}
783 		if (error == EINTR)
784 			return EINTR;
785 	}
786 	fvdat->flag |= FN_FLUSHINPROG;
787 
788 	if (vp->v_bufobj.bo_object != NULL) {
789 		VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
790 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
791 		VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
792 	}
793 	error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
794 	while (error) {
795 		if (error == ERESTART || error == EINTR) {
796 			fvdat->flag &= ~FN_FLUSHINPROG;
797 			if (fvdat->flag & FN_FLUSHWANT) {
798 				fvdat->flag &= ~FN_FLUSHWANT;
799 				wakeup(&fvdat->flag);
800 			}
801 			return EINTR;
802 		}
803 		error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
804 	}
805 	fvdat->flag &= ~FN_FLUSHINPROG;
806 	if (fvdat->flag & FN_FLUSHWANT) {
807 		fvdat->flag &= ~FN_FLUSHWANT;
808 		wakeup(&fvdat->flag);
809 	}
810 	return (error);
811 }
812