xref: /freebsd/sys/fs/fuse/fuse_io.c (revision c6ec7d31830ab1c80edae95ad5e4b9dba10c47ac)
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 || 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 		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 		} else {
163 			FS_DEBUG("buffered write of vnode %ju\n",
164 			      (uintmax_t)VTOILLU(vp));
165 			err = fuse_write_biobackend(vp, uio, cred, fufh);
166 		}
167 		break;
168 	default:
169 		panic("uninterpreted mode passed to fuse_io_dispatch");
170 	}
171 
172 	return (err);
173 }
174 
175 static int
176 fuse_read_biobackend(struct vnode *vp, struct uio *uio,
177     struct ucred *cred, struct fuse_filehandle *fufh)
178 {
179 	struct buf *bp;
180 	daddr_t lbn;
181 	int bcount;
182 	int err = 0, n = 0, on = 0;
183 	off_t filesize;
184 
185 	const int biosize = fuse_iosize(vp);
186 
187 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
188 	    uio->uio_resid, uio->uio_offset, VTOFUD(vp)->filesize);
189 
190 	if (uio->uio_resid == 0)
191 		return (0);
192 	if (uio->uio_offset < 0)
193 		return (EINVAL);
194 
195 	bcount = MIN(MAXBSIZE, biosize);
196 	filesize = VTOFUD(vp)->filesize;
197 
198 	do {
199 		if (fuse_isdeadfs(vp)) {
200 			err = ENXIO;
201 			break;
202 		}
203 		lbn = uio->uio_offset / biosize;
204 		on = uio->uio_offset & (biosize - 1);
205 
206 		FS_DEBUG2G("biosize %d, lbn %d, on %d\n", biosize, (int)lbn, on);
207 
208 		/*
209 	         * Obtain the buffer cache block.  Figure out the buffer size
210 	         * when we are at EOF.  If we are modifying the size of the
211 	         * buffer based on an EOF condition we need to hold
212 	         * nfs_rslock() through obtaining the buffer to prevent
213 	         * a potential writer-appender from messing with n_size.
214 	         * Otherwise we may accidently truncate the buffer and
215 	         * lose dirty data.
216 	         *
217 	         * Note that bcount is *not* DEV_BSIZE aligned.
218 	         */
219 		if ((off_t)lbn * biosize >= filesize) {
220 			bcount = 0;
221 		} else if ((off_t)(lbn + 1) * biosize > filesize) {
222 			bcount = filesize - (off_t)lbn *biosize;
223 		}
224 		bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
225 
226 		if (!bp)
227 			return (EINTR);
228 
229 		/*
230 	         * If B_CACHE is not set, we must issue the read.  If this
231 	         * fails, we return an error.
232 	         */
233 
234 		if ((bp->b_flags & B_CACHE) == 0) {
235 			bp->b_iocmd = BIO_READ;
236 			vfs_busy_pages(bp, 0);
237 			err = fuse_io_strategy(vp, bp);
238 			if (err) {
239 				brelse(bp);
240 				return (err);
241 			}
242 		}
243 		/*
244 	         * on is the offset into the current bp.  Figure out how many
245 	         * bytes we can copy out of the bp.  Note that bcount is
246 	         * NOT DEV_BSIZE aligned.
247 	         *
248 	         * Then figure out how many bytes we can copy into the uio.
249 	         */
250 
251 		n = 0;
252 		if (on < bcount)
253 			n = MIN((unsigned)(bcount - on), uio->uio_resid);
254 		if (n > 0) {
255 			FS_DEBUG2G("feeding buffeater with %d bytes of buffer %p,"
256 				" saying %d was asked for\n",
257 				n, bp->b_data + on, n + (int)bp->b_resid);
258 			err = uiomove(bp->b_data + on, n, uio);
259 		}
260 		brelse(bp);
261 		FS_DEBUG2G("end of turn, err %d, uio->uio_resid %zd, n %d\n",
262 		    err, uio->uio_resid, n);
263 	} while (err == 0 && uio->uio_resid > 0 && n > 0);
264 
265 	return (err);
266 }
267 
268 static int
269 fuse_read_directbackend(struct vnode *vp, struct uio *uio,
270     struct ucred *cred, struct fuse_filehandle *fufh)
271 {
272 	struct fuse_dispatcher fdi;
273 	struct fuse_read_in *fri;
274 	int err = 0;
275 
276 	if (uio->uio_resid == 0)
277 		return (0);
278 
279 	fdisp_init(&fdi, 0);
280 
281 	/*
282          * XXX In "normal" case we use an intermediate kernel buffer for
283          * transmitting data from daemon's context to ours. Eventually, we should
284          * get rid of this. Anyway, if the target uio lives in sysspace (we are
285          * called from pageops), and the input data doesn't need kernel-side
286          * processing (we are not called from readdir) we can already invoke
287          * an optimized, "peer-to-peer" I/O routine.
288          */
289 	while (uio->uio_resid > 0) {
290 		fdi.iosize = sizeof(*fri);
291 		fdisp_make_vp(&fdi, FUSE_READ, vp, uio->uio_td, cred);
292 		fri = fdi.indata;
293 		fri->fh = fufh->fh_id;
294 		fri->offset = uio->uio_offset;
295 		fri->size = MIN(uio->uio_resid,
296 		    fuse_get_mpdata(vp->v_mount)->max_read);
297 
298 		FS_DEBUG2G("fri->fh %ju, fri->offset %ju, fri->size %ju\n",
299 			(uintmax_t)fri->fh, (uintmax_t)fri->offset,
300 			(uintmax_t)fri->size);
301 
302 		if ((err = fdisp_wait_answ(&fdi)))
303 			goto out;
304 
305 		FS_DEBUG2G("complete: got iosize=%d, requested fri.size=%zd; "
306 			"resid=%zd offset=%ju\n",
307 			fri->size, fdi.iosize, uio->uio_resid,
308 			(uintmax_t)uio->uio_offset);
309 
310 		if ((err = uiomove(fdi.answ, MIN(fri->size, fdi.iosize), uio)))
311 			break;
312 		if (fdi.iosize < fri->size)
313 			break;
314 	}
315 
316 out:
317 	fdisp_destroy(&fdi);
318 	return (err);
319 }
320 
321 static int
322 fuse_write_directbackend(struct vnode *vp, struct uio *uio,
323     struct ucred *cred, struct fuse_filehandle *fufh)
324 {
325 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
326 	struct fuse_write_in *fwi;
327 	struct fuse_dispatcher fdi;
328 	size_t chunksize;
329 	int diff;
330 	int err = 0;
331 
332 	if (!uio->uio_resid)
333 		return (0);
334 
335 	fdisp_init(&fdi, 0);
336 
337 	while (uio->uio_resid > 0) {
338 		chunksize = MIN(uio->uio_resid,
339 		    fuse_get_mpdata(vp->v_mount)->max_write);
340 
341 		fdi.iosize = sizeof(*fwi) + chunksize;
342 		fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred);
343 
344 		fwi = fdi.indata;
345 		fwi->fh = fufh->fh_id;
346 		fwi->offset = uio->uio_offset;
347 		fwi->size = chunksize;
348 
349 		if ((err = uiomove((char *)fdi.indata + sizeof(*fwi),
350 		    chunksize, uio)))
351 			break;
352 
353 		if ((err = fdisp_wait_answ(&fdi)))
354 			break;
355 
356 		diff = chunksize - ((struct fuse_write_out *)fdi.answ)->size;
357 		if (diff < 0) {
358 			err = EINVAL;
359 			break;
360 		}
361 		uio->uio_resid += diff;
362 		uio->uio_offset -= diff;
363 		if (uio->uio_offset > fvdat->filesize)
364 			fuse_vnode_setsize(vp, cred, uio->uio_offset);
365 	}
366 
367 	fdisp_destroy(&fdi);
368 
369 	return (err);
370 }
371 
372 static int
373 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
374     struct ucred *cred, struct fuse_filehandle *fufh)
375 {
376 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
377 	struct buf *bp;
378 	daddr_t lbn;
379 	int bcount;
380 	int n, on, err = 0;
381 
382 	const int biosize = fuse_iosize(vp);
383 
384 	KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
385 	FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
386 	    uio->uio_resid, uio->uio_offset, fvdat->filesize);
387 	if (vp->v_type != VREG)
388 		return (EIO);
389 	if (uio->uio_offset < 0)
390 		return (EINVAL);
391 	if (uio->uio_resid == 0)
392 		return (0);
393 
394 	/*
395          * Find all of this file's B_NEEDCOMMIT buffers.  If our writes
396          * would exceed the local maximum per-file write commit size when
397          * combined with those, we must decide whether to flush,
398          * go synchronous, or return err.  We don't bother checking
399          * IO_UNIT -- we just make all writes atomic anyway, as there's
400          * no point optimizing for something that really won't ever happen.
401          */
402 	do {
403 		if (fuse_isdeadfs(vp)) {
404 			err = ENXIO;
405 			break;
406 		}
407 		lbn = uio->uio_offset / biosize;
408 		on = uio->uio_offset & (biosize - 1);
409 		n = MIN((unsigned)(biosize - on), uio->uio_resid);
410 
411 		FS_DEBUG2G("lbn %ju, on %d, n %d, uio offset %ju, uio resid %zd\n",
412 			(uintmax_t)lbn, on, n,
413 			(uintmax_t)uio->uio_offset, uio->uio_resid);
414 
415 again:
416 		/*
417 	         * Handle direct append and file extension cases, calculate
418 	         * unaligned buffer size.
419 	         */
420 		if (uio->uio_offset == fvdat->filesize && n) {
421 			/*
422 	                 * Get the buffer (in its pre-append state to maintain
423 	                 * B_CACHE if it was previously set).  Resize the
424 	                 * nfsnode after we have locked the buffer to prevent
425 	                 * readers from reading garbage.
426 	                 */
427 			bcount = on;
428 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
429 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
430 
431 			if (bp != NULL) {
432 				long save;
433 
434 				err = fuse_vnode_setsize(vp, cred,
435 							 uio->uio_offset + n);
436 				if (err) {
437 					brelse(bp);
438 					break;
439 				}
440 				save = bp->b_flags & B_CACHE;
441 				bcount += n;
442 				allocbuf(bp, bcount);
443 				bp->b_flags |= save;
444 			}
445 		} else {
446 			/*
447 	                 * Obtain the locked cache block first, and then
448 	                 * adjust the file's size as appropriate.
449 	                 */
450 			bcount = on + n;
451 			if ((off_t)lbn * biosize + bcount < fvdat->filesize) {
452 				if ((off_t)(lbn + 1) * biosize < fvdat->filesize)
453 					bcount = biosize;
454 				else
455 					bcount = fvdat->filesize -
456 					  (off_t)lbn *biosize;
457 			}
458 			FS_DEBUG("getting block from OS, bcount %d\n", bcount);
459 			bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
460 			if (bp && uio->uio_offset + n > fvdat->filesize) {
461 				err = fuse_vnode_setsize(vp, cred,
462 							 uio->uio_offset + n);
463 				if (err) {
464 					brelse(bp);
465 					break;
466 				}
467 			}
468 		}
469 
470 		if (!bp) {
471 			err = EINTR;
472 			break;
473 		}
474 		/*
475 	         * Issue a READ if B_CACHE is not set.  In special-append
476 	         * mode, B_CACHE is based on the buffer prior to the write
477 	         * op and is typically set, avoiding the read.  If a read
478 	         * is required in special append mode, the server will
479 	         * probably send us a short-read since we extended the file
480 	         * on our end, resulting in b_resid == 0 and, thusly,
481 	         * B_CACHE getting set.
482 	         *
483 	         * We can also avoid issuing the read if the write covers
484 	         * the entire buffer.  We have to make sure the buffer state
485 	         * is reasonable in this case since we will not be initiating
486 	         * I/O.  See the comments in kern/vfs_bio.c's getblk() for
487 	         * more information.
488 	         *
489 	         * B_CACHE may also be set due to the buffer being cached
490 	         * normally.
491 	         */
492 
493 		if (on == 0 && n == bcount) {
494 			bp->b_flags |= B_CACHE;
495 			bp->b_flags &= ~B_INVAL;
496 			bp->b_ioflags &= ~BIO_ERROR;
497 		}
498 		if ((bp->b_flags & B_CACHE) == 0) {
499 			bp->b_iocmd = BIO_READ;
500 			vfs_busy_pages(bp, 0);
501 			fuse_io_strategy(vp, bp);
502 			if ((err = bp->b_error)) {
503 				brelse(bp);
504 				break;
505 			}
506 		}
507 		if (bp->b_wcred == NOCRED)
508 			bp->b_wcred = crhold(cred);
509 
510 		/*
511 	         * If dirtyend exceeds file size, chop it down.  This should
512 	         * not normally occur but there is an append race where it
513 	         * might occur XXX, so we log it.
514 	         *
515 	         * If the chopping creates a reverse-indexed or degenerate
516 	         * situation with dirtyoff/end, we 0 both of them.
517 	         */
518 
519 		if (bp->b_dirtyend > bcount) {
520 			FS_DEBUG("FUSE append race @%lx:%d\n",
521 			    (long)bp->b_blkno * biosize,
522 			    bp->b_dirtyend - bcount);
523 			bp->b_dirtyend = bcount;
524 		}
525 		if (bp->b_dirtyoff >= bp->b_dirtyend)
526 			bp->b_dirtyoff = bp->b_dirtyend = 0;
527 
528 		/*
529 	         * If the new write will leave a contiguous dirty
530 	         * area, just update the b_dirtyoff and b_dirtyend,
531 	         * otherwise force a write rpc of the old dirty area.
532 	         *
533 	         * While it is possible to merge discontiguous writes due to
534 	         * our having a B_CACHE buffer ( and thus valid read data
535 	         * for the hole), we don't because it could lead to
536 	         * significant cache coherency problems with multiple clients,
537 	         * especially if locking is implemented later on.
538 	         *
539 	         * as an optimization we could theoretically maintain
540 	         * a linked list of discontinuous areas, but we would still
541 	         * have to commit them separately so there isn't much
542 	         * advantage to it except perhaps a bit of asynchronization.
543 	         */
544 
545 		if (bp->b_dirtyend > 0 &&
546 		    (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
547 			/*
548 	                 * Yes, we mean it. Write out everything to "storage"
549 	                 * immediatly, without hesitation. (Apart from other
550 	                 * reasons: the only way to know if a write is valid
551 	                 * if its actually written out.)
552 	                 */
553 			bwrite(bp);
554 			if (bp->b_error == EINTR) {
555 				err = EINTR;
556 				break;
557 			}
558 			goto again;
559 		}
560 		err = uiomove((char *)bp->b_data + on, n, uio);
561 
562 		/*
563 	         * Since this block is being modified, it must be written
564 	         * again and not just committed.  Since write clustering does
565 	         * not work for the stage 1 data write, only the stage 2
566 	         * commit rpc, we have to clear B_CLUSTEROK as well.
567 	         */
568 		bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
569 
570 		if (err) {
571 			bp->b_ioflags |= BIO_ERROR;
572 			bp->b_error = err;
573 			brelse(bp);
574 			break;
575 		}
576 		/*
577 	         * Only update dirtyoff/dirtyend if not a degenerate
578 	         * condition.
579 	         */
580 		if (n) {
581 			if (bp->b_dirtyend > 0) {
582 				bp->b_dirtyoff = MIN(on, bp->b_dirtyoff);
583 				bp->b_dirtyend = MAX((on + n), bp->b_dirtyend);
584 			} else {
585 				bp->b_dirtyoff = on;
586 				bp->b_dirtyend = on + n;
587 			}
588 			vfs_bio_set_valid(bp, on, n);
589 		}
590 		err = bwrite(bp);
591 		if (err)
592 			break;
593 	} while (uio->uio_resid > 0 && n > 0);
594 
595 	if (fuse_sync_resize && (fvdat->flag & FN_SIZECHANGE) != 0)
596 		fuse_vnode_savesize(vp, cred);
597 
598 	return (err);
599 }
600 
601 int
602 fuse_io_strategy(struct vnode *vp, struct buf *bp)
603 {
604 	struct fuse_filehandle *fufh;
605 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
606 	struct ucred *cred;
607 	struct uio *uiop;
608 	struct uio uio;
609 	struct iovec io;
610 	int error = 0;
611 
612 	const int biosize = fuse_iosize(vp);
613 
614 	MPASS(vp->v_type == VREG || vp->v_type == VDIR);
615 	MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE);
616 	FS_DEBUG("inode=%ju offset=%jd resid=%ld\n",
617 	    (uintmax_t)VTOI(vp), (intmax_t)(((off_t)bp->b_blkno) * biosize),
618 	    bp->b_bcount);
619 
620 	error = fuse_filehandle_getrw(vp,
621 	    (bp->b_iocmd == BIO_READ) ? FUFH_RDONLY : FUFH_WRONLY, &fufh);
622 	if (error) {
623 		printf("FUSE: strategy: filehandles are closed\n");
624 		bp->b_ioflags |= BIO_ERROR;
625 		bp->b_error = error;
626 		return (error);
627 	}
628 	cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred;
629 
630 	uiop = &uio;
631 	uiop->uio_iov = &io;
632 	uiop->uio_iovcnt = 1;
633 	uiop->uio_segflg = UIO_SYSSPACE;
634 	uiop->uio_td = curthread;
635 
636 	/*
637          * clear BIO_ERROR and B_INVAL state prior to initiating the I/O.  We
638          * do this here so we do not have to do it in all the code that
639          * calls us.
640          */
641 	bp->b_flags &= ~B_INVAL;
642 	bp->b_ioflags &= ~BIO_ERROR;
643 
644 	KASSERT(!(bp->b_flags & B_DONE),
645 	    ("fuse_io_strategy: bp %p already marked done", bp));
646 	if (bp->b_iocmd == BIO_READ) {
647 		io.iov_len = uiop->uio_resid = bp->b_bcount;
648 		io.iov_base = bp->b_data;
649 		uiop->uio_rw = UIO_READ;
650 
651 		uiop->uio_offset = ((off_t)bp->b_blkno) * biosize;
652 		error = fuse_read_directbackend(vp, uiop, cred, fufh);
653 
654 		if ((!error && uiop->uio_resid) ||
655 		    (fsess_opt_brokenio(vnode_mount(vp)) && error == EIO &&
656 		    uiop->uio_offset < fvdat->filesize && fvdat->filesize > 0 &&
657 		    uiop->uio_offset >= fvdat->cached_attrs.va_size)) {
658 			/*
659 	                 * If we had a short read with no error, we must have
660 	                 * hit a file hole.  We should zero-fill the remainder.
661 	                 * This can also occur if the server hits the file EOF.
662 	                 *
663 	                 * Holes used to be able to occur due to pending
664 	                 * writes, but that is not possible any longer.
665 	                 */
666 			int nread = bp->b_bcount - uiop->uio_resid;
667 			int left = uiop->uio_resid;
668 
669 			if (error != 0) {
670 				printf("FUSE: Fix broken io: offset %ju, "
671 				       " resid %zd, file size %ju/%ju\n",
672 				       (uintmax_t)uiop->uio_offset,
673 				    uiop->uio_resid, fvdat->filesize,
674 				    fvdat->cached_attrs.va_size);
675 				error = 0;
676 			}
677 			if (left > 0)
678 				bzero((char *)bp->b_data + nread, left);
679 			uiop->uio_resid = 0;
680 		}
681 		if (error) {
682 			bp->b_ioflags |= BIO_ERROR;
683 			bp->b_error = error;
684 		}
685 	} else {
686 		/*
687 	         * If we only need to commit, try to commit
688 	         */
689 		if (bp->b_flags & B_NEEDCOMMIT) {
690 			FS_DEBUG("write: B_NEEDCOMMIT flags set\n");
691 		}
692 		/*
693 	         * Setup for actual write
694 	         */
695 		if ((off_t)bp->b_blkno * biosize + bp->b_dirtyend >
696 		    fvdat->filesize)
697 			bp->b_dirtyend = fvdat->filesize -
698 				(off_t)bp->b_blkno * biosize;
699 
700 		if (bp->b_dirtyend > bp->b_dirtyoff) {
701 			io.iov_len = uiop->uio_resid = bp->b_dirtyend
702 			    - bp->b_dirtyoff;
703 			uiop->uio_offset = (off_t)bp->b_blkno * biosize
704 			    + bp->b_dirtyoff;
705 			io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
706 			uiop->uio_rw = UIO_WRITE;
707 
708 			error = fuse_write_directbackend(vp, uiop, cred, fufh);
709 
710 			if (error == EINTR || error == ETIMEDOUT
711 			    || (!error && (bp->b_flags & B_NEEDCOMMIT))) {
712 
713 				bp->b_flags &= ~(B_INVAL | B_NOCACHE);
714 				if ((bp->b_flags & B_PAGING) == 0) {
715 					bdirty(bp);
716 					bp->b_flags &= ~B_DONE;
717 				}
718 				if ((error == EINTR || error == ETIMEDOUT) &&
719 				    (bp->b_flags & B_ASYNC) == 0)
720 					bp->b_flags |= B_EINTR;
721 			} else {
722 				if (error) {
723 					bp->b_ioflags |= BIO_ERROR;
724 					bp->b_flags |= B_INVAL;
725 					bp->b_error = error;
726 				}
727 				bp->b_dirtyoff = bp->b_dirtyend = 0;
728 			}
729 		} else {
730 			bp->b_resid = 0;
731 			bufdone(bp);
732 			return (0);
733 		}
734 	}
735 	bp->b_resid = uiop->uio_resid;
736 	bufdone(bp);
737 	return (error);
738 }
739 
740 int
741 fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td)
742 {
743 	struct vop_fsync_args a = {
744 		.a_vp = vp,
745 		.a_waitfor = waitfor,
746 		.a_td = td,
747 	};
748 
749 	return (vop_stdfsync(&a));
750 }
751 
752 /*
753  * Flush and invalidate all dirty buffers. If another process is already
754  * doing the flush, just wait for completion.
755  */
756 int
757 fuse_io_invalbuf(struct vnode *vp, struct thread *td)
758 {
759 	struct fuse_vnode_data *fvdat = VTOFUD(vp);
760 	int error = 0;
761 
762 	if (vp->v_iflag & VI_DOOMED)
763 		return 0;
764 
765 	ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf");
766 
767 	while (fvdat->flag & FN_FLUSHINPROG) {
768 		struct proc *p = td->td_proc;
769 
770 		if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF)
771 			return EIO;
772 		fvdat->flag |= FN_FLUSHWANT;
773 		tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz);
774 		error = 0;
775 		if (p != NULL) {
776 			PROC_LOCK(p);
777 			if (SIGNOTEMPTY(p->p_siglist) ||
778 			    SIGNOTEMPTY(td->td_siglist))
779 				error = EINTR;
780 			PROC_UNLOCK(p);
781 		}
782 		if (error == EINTR)
783 			return EINTR;
784 	}
785 	fvdat->flag |= FN_FLUSHINPROG;
786 
787 	if (vp->v_bufobj.bo_object != NULL) {
788 		VM_OBJECT_LOCK(vp->v_bufobj.bo_object);
789 		vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
790 		VM_OBJECT_UNLOCK(vp->v_bufobj.bo_object);
791 	}
792 	error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
793 	while (error) {
794 		if (error == ERESTART || error == EINTR) {
795 			fvdat->flag &= ~FN_FLUSHINPROG;
796 			if (fvdat->flag & FN_FLUSHWANT) {
797 				fvdat->flag &= ~FN_FLUSHWANT;
798 				wakeup(&fvdat->flag);
799 			}
800 			return EINTR;
801 		}
802 		error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
803 	}
804 	fvdat->flag &= ~FN_FLUSHINPROG;
805 	if (fvdat->flag & FN_FLUSHWANT) {
806 		fvdat->flag &= ~FN_FLUSHWANT;
807 		wakeup(&fvdat->flag);
808 	}
809 	return (error);
810 }
811