xref: /freebsd/sys/fs/fuse/fuse_io.c (revision 80ff58b89dcacfe07fe20b045890df9db5ca0af0)
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 
91 #include "fuse.h"
92 #include "fuse_file.h"
93 #include "fuse_node.h"
94 #include "fuse_internal.h"
95 #include "fuse_ipc.h"
96 #include "fuse_io.h"
97 
98 #define FUSE_DEBUG_MODULE IO
99 #include "fuse_debug.h"
100 
101 
102 static int
103 fuse_read_directbackend(struct vnode *vp, struct uio *uio,
104     struct ucred *cred, struct fuse_filehandle *fufh);
105 static int
106 fuse_read_biobackend(struct vnode *vp, struct uio *uio,
107     struct ucred *cred, struct fuse_filehandle *fufh);
108 static int
109 fuse_write_directbackend(struct vnode *vp, struct uio *uio,
110     struct ucred *cred, struct fuse_filehandle *fufh);
111 static int
112 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
113     struct ucred *cred, struct fuse_filehandle *fufh, int ioflag);
114 
115 int
116 fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag,
117     struct ucred *cred)
118 {
119 	struct fuse_filehandle *fufh;
120 	int err, directio;
121 
122 	MPASS(vp->v_type == VREG || vp->v_type == VDIR);
123 
124 	err = fuse_filehandle_getrw(vp,
125 	    (uio->uio_rw == UIO_READ) ? FUFH_RDONLY : FUFH_WRONLY, &fufh);
126 	if (err) {
127 		printf("FUSE: io dispatch: filehandles are closed\n");
128 		return err;
129 	}
130 	/*
131          * Ideally, when the daemon asks for direct io at open time, the
132          * standard file flag should be set according to this, so that would
133          * just change the default mode, which later on could be changed via
134          * fcntl(2).
135          * But this doesn't work, the O_DIRECT flag gets cleared at some point
136          * (don't know where). So to make any use of the Fuse direct_io option,
137          * we hardwire it into the file's private data (similarly to Linux,
138          * btw.).
139          */
140 	directio = (ioflag & IO_DIRECT) || !fsess_opt_datacache(vnode_mount(vp));
141 
142 	switch (uio->uio_rw) {
143 	case UIO_READ:
144 		if (directio) {
145 			FS_DEBUG("direct read of vnode %ju via file handle %ju\n",
146 			    (uintmax_t)VTOILLU(vp), (uintmax_t)fufh->fh_id);
147 			err = fuse_read_directbackend(vp, uio, cred, fufh);
148 		} else {
149 			FS_DEBUG("buffered read of vnode %ju\n",
150 			      (uintmax_t)VTOILLU(vp));
151 			err = fuse_read_biobackend(vp, uio, cred, fufh);
152 		}
153 		break;
154 	case UIO_WRITE:
155 		if (directio) {
156 			FS_DEBUG("direct write of vnode %ju via file handle %ju\n",
157 			    (uintmax_t)VTOILLU(vp), (uintmax_t)fufh->fh_id);
158 			err = fuse_write_directbackend(vp, uio, cred, fufh);
159 		} else {
160 			FS_DEBUG("buffered write of vnode %ju\n",
161 			      (uintmax_t)VTOILLU(vp));
162 			err = fuse_write_biobackend(vp, uio, cred, fufh, ioflag);
163 		}
164 		break;
165 	default:
166 		panic("uninterpreted mode passed to fuse_io_dispatch");
167 	}
168 
169 	return (err);
170 }
171 
172 static int
173 fuse_read_biobackend(struct vnode *vp, struct uio *uio,
174     struct ucred *cred, struct fuse_filehandle *fufh)
175 {
176 	struct buf *bp;
177 	daddr_t lbn;
178 	int bcount;
179 	int err = 0, n = 0, on = 0;
180 	off_t filesize;
181 
182 	const int biosize = fuse_iosize(vp);
183 
184 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
185 	    uio->uio_resid, uio->uio_offset, VTOFUD(vp)->filesize);
186 
187 	if (uio->uio_resid == 0)
188 		return (0);
189 	if (uio->uio_offset < 0)
190 		return (EINVAL);
191 
192 	bcount = MIN(MAXBSIZE, biosize);
193 	filesize = VTOFUD(vp)->filesize;
194 
195 	do {
196 		if (fuse_isdeadfs(vp)) {
197 			err = ENXIO;
198 			break;
199 		}
200 		lbn = uio->uio_offset / biosize;
201 		on = uio->uio_offset & (biosize - 1);
202 
203 		FS_DEBUG2G("biosize %d, lbn %d, on %d\n", biosize, (int)lbn, on);
204 
205 		/*
206 	         * Obtain the buffer cache block.  Figure out the buffer size
207 	         * when we are at EOF.  If we are modifying the size of the
208 	         * buffer based on an EOF condition we need to hold
209 	         * nfs_rslock() through obtaining the buffer to prevent
210 	         * a potential writer-appender from messing with n_size.
211 	         * Otherwise we may accidently truncate the buffer and
212 	         * lose dirty data.
213 	         *
214 	         * Note that bcount is *not* DEV_BSIZE aligned.
215 	         */
216 		if ((off_t)lbn * biosize >= filesize) {
217 			bcount = 0;
218 		} else if ((off_t)(lbn + 1) * biosize > filesize) {
219 			bcount = filesize - (off_t)lbn *biosize;
220 		}
221 		bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
222 
223 		if (!bp)
224 			return (EINTR);
225 
226 		/*
227 	         * If B_CACHE is not set, we must issue the read.  If this
228 	         * fails, we return an error.
229 	         */
230 
231 		if ((bp->b_flags & B_CACHE) == 0) {
232 			bp->b_iocmd = BIO_READ;
233 			vfs_busy_pages(bp, 0);
234 			err = fuse_io_strategy(vp, bp);
235 			if (err) {
236 				brelse(bp);
237 				return (err);
238 			}
239 		}
240 		/*
241 	         * on is the offset into the current bp.  Figure out how many
242 	         * bytes we can copy out of the bp.  Note that bcount is
243 	         * NOT DEV_BSIZE aligned.
244 	         *
245 	         * Then figure out how many bytes we can copy into the uio.
246 	         */
247 
248 		n = 0;
249 		if (on < bcount)
250 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
251 		if (n > 0) {
252 			FS_DEBUG2G("feeding buffeater with %d bytes of buffer %p,"
253 				" saying %d was asked for\n",
254 				n, bp->b_data + on, n + (int)bp->b_resid);
255 			err = uiomove(bp->b_data + on, n, uio);
256 		}
257 		brelse(bp);
258 		FS_DEBUG2G("end of turn, err %d, uio->uio_resid %zd, n %d\n",
259 		    err, uio->uio_resid, n);
260 	} while (err == 0 && uio->uio_resid > 0 && n > 0);
261 
262 	return (err);
263 }
264 
265 static int
266 fuse_read_directbackend(struct vnode *vp, struct uio *uio,
267     struct ucred *cred, struct fuse_filehandle *fufh)
268 {
269 	struct fuse_dispatcher fdi;
270 	struct fuse_read_in *fri;
271 	int err = 0;
272 
273 	if (uio->uio_resid == 0)
274 		return (0);
275 
276 	fdisp_init(&fdi, 0);
277 
278 	/*
279          * XXX In "normal" case we use an intermediate kernel buffer for
280          * transmitting data from daemon's context to ours. Eventually, we should
281          * get rid of this. Anyway, if the target uio lives in sysspace (we are
282          * called from pageops), and the input data doesn't need kernel-side
283          * processing (we are not called from readdir) we can already invoke
284          * an optimized, "peer-to-peer" I/O routine.
285          */
286 	while (uio->uio_resid > 0) {
287 		fdi.iosize = sizeof(*fri);
288 		fdisp_make_vp(&fdi, FUSE_READ, vp, uio->uio_td, cred);
289 		fri = fdi.indata;
290 		fri->fh = fufh->fh_id;
291 		fri->offset = uio->uio_offset;
292 		fri->size = MIN(uio->uio_resid,
293 		    fuse_get_mpdata(vp->v_mount)->max_read);
294 
295 		FS_DEBUG2G("fri->fh %ju, fri->offset %ju, fri->size %ju\n",
296 			(uintmax_t)fri->fh, (uintmax_t)fri->offset,
297 			(uintmax_t)fri->size);
298 
299 		if ((err = fdisp_wait_answ(&fdi)))
300 			goto out;
301 
302 		FS_DEBUG2G("complete: got iosize=%d, requested fri.size=%zd; "
303 			"resid=%zd offset=%ju\n",
304 			fri->size, fdi.iosize, uio->uio_resid,
305 			(uintmax_t)uio->uio_offset);
306 
307 		if ((err = uiomove(fdi.answ, MIN(fri->size, fdi.iosize), uio)))
308 			break;
309 		if (fdi.iosize < fri->size)
310 			break;
311 	}
312 
313 out:
314 	fdisp_destroy(&fdi);
315 	return (err);
316 }
317 
318 static int
319 fuse_write_directbackend(struct vnode *vp, struct uio *uio,
320     struct ucred *cred, struct fuse_filehandle *fufh)
321 {
322 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
323 	struct fuse_write_in *fwi;
324 	struct fuse_dispatcher fdi;
325 	size_t chunksize;
326 	int diff;
327 	int err = 0;
328 
329 	if (!uio->uio_resid)
330 		return (0);
331 
332 	fdisp_init(&fdi, 0);
333 
334 	while (uio->uio_resid > 0) {
335 		chunksize = MIN(uio->uio_resid,
336 		    fuse_get_mpdata(vp->v_mount)->max_write);
337 
338 		fdi.iosize = sizeof(*fwi) + chunksize;
339 		fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred);
340 
341 		fwi = fdi.indata;
342 		fwi->fh = fufh->fh_id;
343 		fwi->offset = uio->uio_offset;
344 		fwi->size = chunksize;
345 
346 		if ((err = uiomove((char *)fdi.indata + sizeof(*fwi),
347 		    chunksize, uio)))
348 			break;
349 
350 		if ((err = fdisp_wait_answ(&fdi)))
351 			break;
352 
353 		diff = chunksize - ((struct fuse_write_out *)fdi.answ)->size;
354 		if (diff < 0) {
355 			err = EINVAL;
356 			break;
357 		}
358 		uio->uio_resid += diff;
359 		uio->uio_offset -= diff;
360 		if (uio->uio_offset > fvdat->filesize)
361 			fuse_vnode_setsize(vp, cred, uio->uio_offset);
362 	}
363 
364 	fdisp_destroy(&fdi);
365 
366 	return (err);
367 }
368 
369 static int
370 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
371     struct ucred *cred, struct fuse_filehandle *fufh, int ioflag)
372 {
373 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
374 	struct buf *bp;
375 	daddr_t lbn;
376 	int bcount;
377 	int n, on, err = 0;
378 
379 	const int biosize = fuse_iosize(vp);
380 
381 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
382 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
383 	    uio->uio_resid, uio->uio_offset, fvdat->filesize);
384 	if (vp->v_type != VREG)
385 		return (EIO);
386 	if (uio->uio_offset < 0)
387 		return (EINVAL);
388 	if (uio->uio_resid == 0)
389 		return (0);
390 	if (ioflag & IO_APPEND)
391 		uio_setoffset(uio, fvdat->filesize);
392 
393 	/*
394          * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
395          * would exceed the local maximum per-file write commit size when
396          * combined with those, we must decide whether to flush,
397          * go synchronous, or return err.  We don't bother checking
398          * IO_UNIT -- we just make all writes atomic anyway, as there's
399          * no point optimizing for something that really won't ever happen.
400          */
401 	do {
402 		if (fuse_isdeadfs(vp)) {
403 			err = ENXIO;
404 			break;
405 		}
406 		lbn = uio->uio_offset / biosize;
407 		on = uio->uio_offset & (biosize - 1);
408 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
409 
410 		FS_DEBUG2G("lbn %ju, on %d, n %d, uio offset %ju, uio resid %zd\n",
411 			(uintmax_t)lbn, on, n,
412 			(uintmax_t)uio->uio_offset, uio->uio_resid);
413 
414 again:
415 		/*
416 	         * Handle direct append and file extension cases, calculate
417 	         * unaligned buffer size.
418 	         */
419 		if (uio->uio_offset == fvdat->filesize && n) {
420 			/*
421 	                 * Get the buffer (in its pre-append state to maintain
422 	                 * B_CACHE if it was previously set).  Resize the
423 	                 * nfsnode after we have locked the buffer to prevent
424 	                 * readers from reading garbage.
425 	                 */
426 			bcount = on;
427 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
428 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
429 
430 			if (bp != NULL) {
431 				long save;
432 
433 				err = fuse_vnode_setsize(vp, cred,
434 							 uio->uio_offset + n);
435 				if (err) {
436 					brelse(bp);
437 					break;
438 				}
439 				save = bp->b_flags & B_CACHE;
440 				bcount += n;
441 				allocbuf(bp, bcount);
442 				bp->b_flags |= save;
443 			}
444 		} else {
445 			/*
446 	                 * Obtain the locked cache block first, and then
447 	                 * adjust the file's size as appropriate.
448 	                 */
449 			bcount = on + n;
450 			if ((off_t)lbn * biosize + bcount < fvdat->filesize) {
451 				if ((off_t)(lbn + 1) * biosize < fvdat->filesize)
452 					bcount = biosize;
453 				else
454 					bcount = fvdat->filesize -
455 					  (off_t)lbn *biosize;
456 			}
457 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
458 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
459 			if (bp && uio->uio_offset + n > fvdat->filesize) {
460 				err = fuse_vnode_setsize(vp, cred,
461 							 uio->uio_offset + n);
462 				if (err) {
463 					brelse(bp);
464 					break;
465 				}
466 			}
467 		}
468 
469 		if (!bp) {
470 			err = EINTR;
471 			break;
472 		}
473 		/*
474 	         * Issue a READ if B_CACHE is not set.  In special-append
475 	         * mode, B_CACHE is based on the buffer prior to the write
476 	         * op and is typically set, avoiding the read.  If a read
477 	         * is required in special append mode, the server will
478 	         * probably send us a short-read since we extended the file
479 	         * on our end, resulting in b_resid == 0 and, thusly,
480 	         * B_CACHE getting set.
481 	         *
482 	         * We can also avoid issuing the read if the write covers
483 	         * the entire buffer.  We have to make sure the buffer state
484 	         * is reasonable in this case since we will not be initiating
485 	         * I/O.  See the comments in kern/vfs_bio.c's getblk() for
486 	         * more information.
487 	         *
488 	         * B_CACHE may also be set due to the buffer being cached
489 	         * normally.
490 	         */
491 
492 		if (on == 0 && n == bcount) {
493 			bp->b_flags |= B_CACHE;
494 			bp->b_flags &= ~B_INVAL;
495 			bp->b_ioflags &= ~BIO_ERROR;
496 		}
497 		if ((bp->b_flags & B_CACHE) == 0) {
498 			bp->b_iocmd = BIO_READ;
499 			vfs_busy_pages(bp, 0);
500 			fuse_io_strategy(vp, bp);
501 			if ((err = bp->b_error)) {
502 				brelse(bp);
503 				break;
504 			}
505 		}
506 		if (bp->b_wcred == NOCRED)
507 			bp->b_wcred = crhold(cred);
508 
509 		/*
510 	         * If dirtyend exceeds file size, chop it down.  This should
511 	         * not normally occur but there is an append race where it
512 	         * might occur XXX, so we log it.
513 	         *
514 	         * If the chopping creates a reverse-indexed or degenerate
515 	         * situation with dirtyoff/end, we 0 both of them.
516 	         */
517 
518 		if (bp->b_dirtyend > bcount) {
519 			FS_DEBUG("FUSE append race @%lx:%d\n",
520 			    (long)bp->b_blkno * biosize,
521 			    bp->b_dirtyend - bcount);
522 			bp->b_dirtyend = bcount;
523 		}
524 		if (bp->b_dirtyoff >= bp->b_dirtyend)
525 			bp->b_dirtyoff = bp->b_dirtyend = 0;
526 
527 		/*
528 	         * If the new write will leave a contiguous dirty
529 	         * area, just update the b_dirtyoff and b_dirtyend,
530 	         * otherwise force a write rpc of the old dirty area.
531 	         *
532 	         * While it is possible to merge discontiguous writes due to
533 	         * our having a B_CACHE buffer ( and thus valid read data
534 	         * for the hole), we don't because it could lead to
535 	         * significant cache coherency problems with multiple clients,
536 	         * especially if locking is implemented later on.
537 	         *
538 	         * as an optimization we could theoretically maintain
539 	         * a linked list of discontinuous areas, but we would still
540 	         * have to commit them separately so there isn't much
541 	         * advantage to it except perhaps a bit of asynchronization.
542 	         */
543 
544 		if (bp->b_dirtyend > 0 &&
545 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
546 			/*
547 	                 * Yes, we mean it. Write out everything to "storage"
548 	                 * immediatly, without hesitation. (Apart from other
549 	                 * reasons: the only way to know if a write is valid
550 	                 * if its actually written out.)
551 	                 */
552 			bwrite(bp);
553 			if (bp->b_error == EINTR) {
554 				err = EINTR;
555 				break;
556 			}
557 			goto again;
558 		}
559 		err = uiomove((char *)bp->b_data + on, n, uio);
560 
561 		/*
562 	         * Since this block is being modified, it must be written
563 	         * again and not just committed.  Since write clustering does
564 	         * not work for the stage 1 data write, only the stage 2
565 	         * commit rpc, we have to clear B_CLUSTEROK as well.
566 	         */
567 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
568 
569 		if (err) {
570 			bp->b_ioflags |= BIO_ERROR;
571 			bp->b_error = err;
572 			brelse(bp);
573 			break;
574 		}
575 		/*
576 	         * Only update dirtyoff/dirtyend if not a degenerate
577 	         * condition.
578 	         */
579 		if (n) {
580 			if (bp->b_dirtyend > 0) {
581 				bp->b_dirtyoff = MIN(on, bp->b_dirtyoff);
582 				bp->b_dirtyend = MAX((on + n), bp->b_dirtyend);
583 			} else {
584 				bp->b_dirtyoff = on;
585 				bp->b_dirtyend = on + n;
586 			}
587 			vfs_bio_set_valid(bp, on, n);
588 		}
589 		err = bwrite(bp);
590 		if (err)
591 			break;
592 	} while (uio->uio_resid > 0 && n > 0);
593 
594 	if (fuse_sync_resize && (fvdat->flag & FN_SIZECHANGE) != 0)
595 		fuse_vnode_savesize(vp, cred);
596 
597 	return (err);
598 }
599 
600 int
601 fuse_io_strategy(struct vnode *vp, struct buf *bp)
602 {
603 	struct fuse_filehandle *fufh;
604 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
605 	struct ucred *cred;
606 	struct uio *uiop;
607 	struct uio uio;
608 	struct iovec io;
609 	int error = 0;
610 
611 	const int biosize = fuse_iosize(vp);
612 
613 	MPASS(vp->v_type == VREG || vp->v_type == VDIR);
614 	MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE);
615 	FS_DEBUG("inode=%ju offset=%jd resid=%ld\n",
616 	    (uintmax_t)VTOI(vp), (intmax_t)(((off_t)bp->b_blkno) * biosize),
617 	    bp->b_bcount);
618 
619 	error = fuse_filehandle_getrw(vp,
620 	    (bp->b_iocmd == BIO_READ) ? FUFH_RDONLY : FUFH_WRONLY, &fufh);
621 	if (error) {
622 		printf("FUSE: strategy: filehandles are closed\n");
623 		bp->b_ioflags |= BIO_ERROR;
624 		bp->b_error = error;
625 		return (error);
626 	}
627 	cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred;
628 
629 	uiop = &uio;
630 	uiop->uio_iov = &io;
631 	uiop->uio_iovcnt = 1;
632 	uiop->uio_segflg = UIO_SYSSPACE;
633 	uiop->uio_td = curthread;
634 
635 	/*
636          * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
637          * do this here so we do not have to do it in all the code that
638          * calls us.
639          */
640 	bp->b_flags &= ~B_INVAL;
641 	bp->b_ioflags &= ~BIO_ERROR;
642 
643 	KASSERT(!(bp->b_flags & B_DONE),
644 	    ("fuse_io_strategy: bp %p already marked done", bp));
645 	if (bp->b_iocmd == BIO_READ) {
646 		io.iov_len = uiop->uio_resid = bp->b_bcount;
647 		io.iov_base = bp->b_data;
648 		uiop->uio_rw = UIO_READ;
649 
650 		uiop->uio_offset = ((off_t)bp->b_blkno) * biosize;
651 		error = fuse_read_directbackend(vp, uiop, cred, fufh);
652 
653 		if ((!error && uiop->uio_resid) ||
654 		    (fsess_opt_brokenio(vnode_mount(vp)) && error == EIO &&
655 		    uiop->uio_offset < fvdat->filesize && fvdat->filesize > 0 &&
656 		    uiop->uio_offset >= fvdat->cached_attrs.va_size)) {
657 			/*
658 	                 * If we had a short read with no error, we must have
659 	                 * hit a file hole.  We should zero-fill the remainder.
660 	                 * This can also occur if the server hits the file EOF.
661 	                 *
662 	                 * Holes used to be able to occur due to pending
663 	                 * writes, but that is not possible any longer.
664 	                 */
665 			int nread = bp->b_bcount - uiop->uio_resid;
666 			int left = uiop->uio_resid;
667 
668 			if (error != 0) {
669 				printf("FUSE: Fix broken io: offset %ju, "
670 				       " resid %zd, file size %ju/%ju\n",
671 				       (uintmax_t)uiop->uio_offset,
672 				    uiop->uio_resid, fvdat->filesize,
673 				    fvdat->cached_attrs.va_size);
674 				error = 0;
675 			}
676 			if (left > 0)
677 				bzero((char *)bp->b_data + nread, left);
678 			uiop->uio_resid = 0;
679 		}
680 		if (error) {
681 			bp->b_ioflags |= BIO_ERROR;
682 			bp->b_error = error;
683 		}
684 	} else {
685 		/*
686 	         * If we only need to commit, try to commit
687 	         */
688 		if (bp->b_flags & B_NEEDCOMMIT) {
689 			FS_DEBUG("write: B_NEEDCOMMIT flags set\n");
690 		}
691 		/*
692 	         * Setup for actual write
693 	         */
694 		if ((off_t)bp->b_blkno * biosize + bp->b_dirtyend >
695 		    fvdat->filesize)
696 			bp->b_dirtyend = fvdat->filesize -
697 				(off_t)bp->b_blkno * biosize;
698 
699 		if (bp->b_dirtyend > bp->b_dirtyoff) {
700 			io.iov_len = uiop->uio_resid = bp->b_dirtyend
701 			    - bp->b_dirtyoff;
702 			uiop->uio_offset = (off_t)bp->b_blkno * biosize
703 			    + bp->b_dirtyoff;
704 			io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
705 			uiop->uio_rw = UIO_WRITE;
706 
707 			error = fuse_write_directbackend(vp, uiop, cred, fufh);
708 
709 			if (error == EINTR || error == ETIMEDOUT
710 			    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
711 
712 				bp->b_flags &= ~(B_INVAL | B_NOCACHE);
713 				if ((bp->b_flags & B_PAGING) == 0) {
714 					bdirty(bp);
715 					bp->b_flags &= ~B_DONE;
716 				}
717 				if ((error == EINTR || error == ETIMEDOUT) &&
718 				    (bp->b_flags & B_ASYNC) == 0)
719 					bp->b_flags |= B_EINTR;
720 			} else {
721 				if (error) {
722 					bp->b_ioflags |= BIO_ERROR;
723 					bp->b_flags |= B_INVAL;
724 					bp->b_error = error;
725 				}
726 				bp->b_dirtyoff = bp->b_dirtyend = 0;
727 			}
728 		} else {
729 			bp->b_resid = 0;
730 			bufdone(bp);
731 			return (0);
732 		}
733 	}
734 	bp->b_resid = uiop->uio_resid;
735 	bufdone(bp);
736 	return (error);
737 }
738 
739 int
740 fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td)
741 {
742 	struct vop_fsync_args a = {
743 		.a_vp = vp,
744 		.a_waitfor = waitfor,
745 		.a_td = td,
746 	};
747 
748 	return (vop_stdfsync(&a));
749 }
750 
751 /*
752  * Flush and invalidate all dirty buffers. If another process is already
753  * doing the flush, just wait for completion.
754  */
755 int
756 fuse_io_invalbuf(struct vnode *vp, struct thread *td)
757 {
758 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
759 	int error = 0;
760 
761 	if (vp->v_iflag & VI_DOOMED)
762 		return 0;
763 
764 	ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf");
765 
766 	while (fvdat->flag & FN_FLUSHINPROG) {
767 		struct proc *p = td->td_proc;
768 
769 		if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF)
770 			return EIO;
771 		fvdat->flag |= FN_FLUSHWANT;
772 		tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz);
773 		error = 0;
774 		if (p != NULL) {
775 			PROC_LOCK(p);
776 			if (SIGNOTEMPTY(p->p_siglist) ||
777 			    SIGNOTEMPTY(td->td_siglist))
778 				error = EINTR;
779 			PROC_UNLOCK(p);
780 		}
781 		if (error == EINTR)
782 			return EINTR;
783 	}
784 	fvdat->flag |= FN_FLUSHINPROG;
785 
786 	if (vp->v_bufobj.bo_object != NULL) {
787 		VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
788 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
789 		VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
790 	}
791 	error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
792 	while (error) {
793 		if (error == ERESTART || error == EINTR) {
794 			fvdat->flag &= ~FN_FLUSHINPROG;
795 			if (fvdat->flag & FN_FLUSHWANT) {
796 				fvdat->flag &= ~FN_FLUSHWANT;
797 				wakeup(&fvdat->flag);
798 			}
799 			return EINTR;
800 		}
801 		error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
802 	}
803 	fvdat->flag &= ~FN_FLUSHINPROG;
804 	if (fvdat->flag & FN_FLUSHWANT) {
805 		fvdat->flag &= ~FN_FLUSHWANT;
806 		wakeup(&fvdat->flag);
807 	}
808 	return (error);
809 }
810