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 * Copyright (c) 2019 The FreeBSD Foundation
37 *
38 * Portions of this software were developed by BFF Storage Systems, LLC under
39 * sponsorship from the FreeBSD Foundation.
40 *
41 * Redistribution and use in source and binary forms, with or without
42 * modification, are permitted provided that the following conditions
43 * are met:
44 * 1. Redistributions of source code must retain the above copyright
45 * notice, this list of conditions and the following disclaimer.
46 * 2. Redistributions in binary form must reproduce the above copyright
47 * notice, this list of conditions and the following disclaimer in the
48 * documentation and/or other materials provided with the distribution.
49 *
50 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND
51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
53 * ARE DISCLAIMED. IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
60 * SUCH DAMAGE.
61 */
62
63 #include <sys/cdefs.h>
64 __FBSDID("$FreeBSD$");
65
66 #include <sys/types.h>
67 #include <sys/param.h>
68 #include <sys/module.h>
69 #include <sys/systm.h>
70 #include <sys/errno.h>
71 #include <sys/param.h>
72 #include <sys/kernel.h>
73 #include <sys/conf.h>
74 #include <sys/uio.h>
75 #include <sys/malloc.h>
76 #include <sys/queue.h>
77 #include <sys/lock.h>
78 #include <sys/sx.h>
79 #include <sys/mutex.h>
80 #include <sys/rwlock.h>
81 #include <sys/priv.h>
82 #include <sys/proc.h>
83 #include <sys/mount.h>
84 #include <sys/vnode.h>
85 #include <sys/stat.h>
86 #include <sys/unistd.h>
87 #include <sys/filedesc.h>
88 #include <sys/file.h>
89 #include <sys/fcntl.h>
90 #include <sys/bio.h>
91 #include <sys/buf.h>
92 #include <sys/sysctl.h>
93 #include <sys/vmmeter.h>
94
95 #include <vm/vm.h>
96 #include <vm/vm_extern.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_object.h>
101
102 #include "fuse.h"
103 #include "fuse_file.h"
104 #include "fuse_node.h"
105 #include "fuse_internal.h"
106 #include "fuse_ipc.h"
107 #include "fuse_io.h"
108
109 /*
110 * Set in a struct buf to indicate that the write came from the buffer cache
111 * and the originating cred and pid are no longer known.
112 */
113 #define B_FUSEFS_WRITE_CACHE B_FS_FLAG1
114
115 SDT_PROVIDER_DECLARE(fusefs);
116 /*
117 * Fuse trace probe:
118 * arg0: verbosity. Higher numbers give more verbose messages
119 * arg1: Textual message
120 */
121 SDT_PROBE_DEFINE2(fusefs, , io, trace, "int", "char*");
122
123 SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_start, "int", "int", "int", "int");
124 SDT_PROBE_DEFINE2(fusefs, , io, read_bio_backend_feed, "int", "struct buf*");
125 SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_end, "int", "ssize_t", "int",
126 "struct buf*");
127 int
fuse_read_biobackend(struct vnode * vp,struct uio * uio,int ioflag,struct ucred * cred,struct fuse_filehandle * fufh,pid_t pid)128 fuse_read_biobackend(struct vnode *vp, struct uio *uio, int ioflag,
129 struct ucred *cred, struct fuse_filehandle *fufh, pid_t pid)
130 {
131 struct buf *bp;
132 struct mount *mp;
133 struct fuse_data *data;
134 daddr_t lbn, nextlbn;
135 int bcount, nextsize;
136 int err, n = 0, on = 0, seqcount;
137 off_t filesize;
138
139 const int biosize = fuse_iosize(vp);
140 mp = vnode_mount(vp);
141 data = fuse_get_mpdata(mp);
142
143 if (uio->uio_offset < 0)
144 return (EINVAL);
145
146 seqcount = ioflag >> IO_SEQSHIFT;
147
148 err = fuse_vnode_size(vp, &filesize, cred, curthread);
149 if (err)
150 return err;
151
152 for (err = 0, bp = NULL; uio->uio_resid > 0; bp = NULL) {
153 if (fuse_isdeadfs(vp)) {
154 err = ENXIO;
155 break;
156 }
157 if (filesize - uio->uio_offset <= 0)
158 break;
159 lbn = uio->uio_offset / biosize;
160 on = uio->uio_offset & (biosize - 1);
161
162 if ((off_t)lbn * biosize >= filesize) {
163 bcount = 0;
164 } else if ((off_t)(lbn + 1) * biosize > filesize) {
165 bcount = filesize - (off_t)lbn *biosize;
166 } else {
167 bcount = biosize;
168 }
169 nextlbn = lbn + 1;
170 nextsize = MIN(biosize, filesize - nextlbn * biosize);
171
172 SDT_PROBE4(fusefs, , io, read_bio_backend_start,
173 biosize, (int)lbn, on, bcount);
174
175 if (bcount < biosize) {
176 /* If near EOF, don't do readahead */
177 err = bread(vp, lbn, bcount, NOCRED, &bp);
178 } else if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERR) == 0) {
179 /* Try clustered read */
180 long totread = uio->uio_resid + on;
181 seqcount = MIN(seqcount,
182 data->max_readahead_blocks + 1);
183 err = cluster_read(vp, filesize, lbn, bcount, NOCRED,
184 totread, seqcount, 0, &bp);
185 } else if (seqcount > 1 && data->max_readahead_blocks >= 1) {
186 /* Try non-clustered readahead */
187 err = breadn(vp, lbn, bcount, &nextlbn, &nextsize, 1,
188 NOCRED, &bp);
189 } else {
190 /* Just read what was requested */
191 err = bread(vp, lbn, bcount, NOCRED, &bp);
192 }
193
194 if (err) {
195 brelse(bp);
196 bp = NULL;
197 break;
198 }
199
200 /*
201 * on is the offset into the current bp. Figure out how many
202 * bytes we can copy out of the bp. Note that bcount is
203 * NOT DEV_BSIZE aligned.
204 *
205 * Then figure out how many bytes we can copy into the uio.
206 */
207
208 n = 0;
209 if (on < bcount - bp->b_resid)
210 n = MIN((unsigned)(bcount - bp->b_resid - on),
211 uio->uio_resid);
212 if (n > 0) {
213 SDT_PROBE2(fusefs, , io, read_bio_backend_feed, n, bp);
214 err = uiomove(bp->b_data + on, n, uio);
215 }
216 vfs_bio_brelse(bp, ioflag);
217 SDT_PROBE4(fusefs, , io, read_bio_backend_end, err,
218 uio->uio_resid, n, bp);
219 if (bp->b_resid > 0) {
220 /* Short read indicates EOF */
221 break;
222 }
223 }
224
225 return (err);
226 }
227
228 SDT_PROBE_DEFINE1(fusefs, , io, read_directbackend_start,
229 "struct fuse_read_in*");
230 SDT_PROBE_DEFINE3(fusefs, , io, read_directbackend_complete,
231 "struct fuse_dispatcher*", "struct fuse_read_in*", "struct uio*");
232
233 int
fuse_read_directbackend(struct vnode * vp,struct uio * uio,struct ucred * cred,struct fuse_filehandle * fufh)234 fuse_read_directbackend(struct vnode *vp, struct uio *uio,
235 struct ucred *cred, struct fuse_filehandle *fufh)
236 {
237 struct fuse_data *data;
238 struct fuse_dispatcher fdi;
239 struct fuse_read_in *fri;
240 int err = 0;
241
242 data = fuse_get_mpdata(vp->v_mount);
243
244 if (uio->uio_resid == 0)
245 return (0);
246
247 fdisp_init(&fdi, 0);
248
249 /*
250 * XXX In "normal" case we use an intermediate kernel buffer for
251 * transmitting data from daemon's context to ours. Eventually, we should
252 * get rid of this. Anyway, if the target uio lives in sysspace (we are
253 * called from pageops), and the input data doesn't need kernel-side
254 * processing (we are not called from readdir) we can already invoke
255 * an optimized, "peer-to-peer" I/O routine.
256 */
257 while (uio->uio_resid > 0) {
258 fdi.iosize = sizeof(*fri);
259 fdisp_make_vp(&fdi, FUSE_READ, vp, uio->uio_td, cred);
260 fri = fdi.indata;
261 fri->fh = fufh->fh_id;
262 fri->offset = uio->uio_offset;
263 fri->size = MIN(uio->uio_resid,
264 fuse_get_mpdata(vp->v_mount)->max_read);
265 if (fuse_libabi_geq(data, 7, 9)) {
266 /* See comment regarding FUSE_WRITE_LOCKOWNER */
267 fri->read_flags = 0;
268 fri->flags = fufh_type_2_fflags(fufh->fufh_type);
269 }
270
271 SDT_PROBE1(fusefs, , io, read_directbackend_start, fri);
272
273 if ((err = fdisp_wait_answ(&fdi)))
274 goto out;
275
276 SDT_PROBE3(fusefs, , io, read_directbackend_complete,
277 &fdi, fri, uio);
278
279 if ((err = uiomove(fdi.answ, MIN(fri->size, fdi.iosize), uio)))
280 break;
281 if (fdi.iosize < fri->size) {
282 /*
283 * Short read. Should only happen at EOF or with
284 * direct io.
285 */
286 break;
287 }
288 }
289
290 out:
291 fdisp_destroy(&fdi);
292 return (err);
293 }
294
295 int
fuse_write_directbackend(struct vnode * vp,struct uio * uio,struct ucred * cred,struct fuse_filehandle * fufh,off_t filesize,int ioflag,bool pages)296 fuse_write_directbackend(struct vnode *vp, struct uio *uio,
297 struct ucred *cred, struct fuse_filehandle *fufh, off_t filesize,
298 int ioflag, bool pages)
299 {
300 struct fuse_vnode_data *fvdat = VTOFUD(vp);
301 struct fuse_data *data;
302 struct fuse_write_in *fwi;
303 struct fuse_write_out *fwo;
304 struct fuse_dispatcher fdi;
305 size_t chunksize;
306 void *fwi_data;
307 off_t as_written_offset;
308 int diff;
309 int err = 0;
310 bool direct_io = fufh->fuse_open_flags & FOPEN_DIRECT_IO;
311 bool wrote_anything = false;
312 uint32_t write_flags;
313
314 data = fuse_get_mpdata(vp->v_mount);
315
316 /*
317 * Don't set FUSE_WRITE_LOCKOWNER in write_flags. It can't be set
318 * accurately when using POSIX AIO, libfuse doesn't use it, and I'm not
319 * aware of any file systems that do. It was an attempt to add
320 * Linux-style mandatory locking to the FUSE protocol, but mandatory
321 * locking is deprecated even on Linux. See Linux commit
322 * f33321141b273d60cbb3a8f56a5489baad82ba5e .
323 */
324 /*
325 * Set FUSE_WRITE_CACHE whenever we don't know the uid, gid, and/or pid
326 * that originated a write. For example when writing from the
327 * writeback cache. I don't know of a single file system that cares,
328 * but the protocol says we're supposed to do this.
329 */
330 write_flags = !pages && (
331 (ioflag & IO_DIRECT) ||
332 !fsess_opt_datacache(vnode_mount(vp)) ||
333 !fsess_opt_writeback(vnode_mount(vp))) ? 0 : FUSE_WRITE_CACHE;
334
335 if (uio->uio_resid == 0)
336 return (0);
337
338 if (ioflag & IO_APPEND)
339 uio_setoffset(uio, filesize);
340
341 if (vn_rlimit_fsize(vp, uio, uio->uio_td))
342 return (EFBIG);
343
344 fdisp_init(&fdi, 0);
345
346 while (uio->uio_resid > 0) {
347 size_t sizeof_fwi;
348
349 if (fuse_libabi_geq(data, 7, 9)) {
350 sizeof_fwi = sizeof(*fwi);
351 } else {
352 sizeof_fwi = FUSE_COMPAT_WRITE_IN_SIZE;
353 }
354
355 chunksize = MIN(uio->uio_resid, data->max_write);
356
357 fdi.iosize = sizeof_fwi + chunksize;
358 fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred);
359
360 fwi = fdi.indata;
361 fwi->fh = fufh->fh_id;
362 fwi->offset = uio->uio_offset;
363 fwi->size = chunksize;
364 fwi->write_flags = write_flags;
365 if (fuse_libabi_geq(data, 7, 9)) {
366 fwi->flags = fufh_type_2_fflags(fufh->fufh_type);
367 }
368 fwi_data = (char *)fdi.indata + sizeof_fwi;
369
370 if ((err = uiomove(fwi_data, chunksize, uio)))
371 break;
372
373 retry:
374 err = fdisp_wait_answ(&fdi);
375 if (err == ERESTART || err == EINTR || err == EWOULDBLOCK) {
376 /*
377 * Rewind the uio so dofilewrite will know it's
378 * incomplete
379 */
380 uio->uio_resid += fwi->size;
381 uio->uio_offset -= fwi->size;
382 /*
383 * Change ERESTART into EINTR because we can't rewind
384 * uio->uio_iov. Basically, once uiomove(9) has been
385 * called, it's impossible to restart a syscall.
386 */
387 if (err == ERESTART)
388 err = EINTR;
389 break;
390 } else if (err) {
391 break;
392 } else {
393 wrote_anything = true;
394 }
395
396 fwo = ((struct fuse_write_out *)fdi.answ);
397
398 /* Adjust the uio in the case of short writes */
399 diff = fwi->size - fwo->size;
400 as_written_offset = uio->uio_offset - diff;
401
402 if (as_written_offset - diff > filesize) {
403 fuse_vnode_setsize(vp, as_written_offset, false);
404 getnanouptime(&fvdat->last_local_modify);
405 }
406 if (as_written_offset - diff >= filesize)
407 fvdat->flag &= ~FN_SIZECHANGE;
408
409 if (diff < 0) {
410 fuse_warn(data, FSESS_WARN_WROTE_LONG,
411 "wrote more data than we provided it.");
412 err = EINVAL;
413 break;
414 } else if (diff > 0) {
415 /* Short write */
416 if (!direct_io) {
417 fuse_warn(data, FSESS_WARN_SHORT_WRITE,
418 "short writes are only allowed with "
419 "direct_io.");
420 }
421 if (ioflag & IO_DIRECT) {
422 /* Return early */
423 uio->uio_resid += diff;
424 uio->uio_offset -= diff;
425 break;
426 } else {
427 /* Resend the unwritten portion of data */
428 fdi.iosize = sizeof_fwi + diff;
429 /* Refresh fdi without clearing data buffer */
430 fdisp_refresh_vp(&fdi, FUSE_WRITE, vp,
431 uio->uio_td, cred);
432 fwi = fdi.indata;
433 MPASS2(fwi == fdi.indata, "FUSE dispatcher "
434 "reallocated despite no increase in "
435 "size?");
436 void *src = (char*)fwi_data + fwo->size;
437 memmove(fwi_data, src, diff);
438 fwi->fh = fufh->fh_id;
439 fwi->offset = as_written_offset;
440 fwi->size = diff;
441 fwi->write_flags = write_flags;
442 goto retry;
443 }
444 }
445 }
446
447 fdisp_destroy(&fdi);
448
449 if (wrote_anything)
450 fuse_vnode_undirty_cached_timestamps(vp, false);
451
452 return (err);
453 }
454
455 SDT_PROBE_DEFINE6(fusefs, , io, write_biobackend_start, "int64_t", "int", "int",
456 "struct uio*", "int", "bool");
457 SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_append_race, "long", "int");
458 SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_issue, "int", "struct buf*");
459
460 int
fuse_write_biobackend(struct vnode * vp,struct uio * uio,struct ucred * cred,struct fuse_filehandle * fufh,int ioflag,pid_t pid)461 fuse_write_biobackend(struct vnode *vp, struct uio *uio,
462 struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid)
463 {
464 struct fuse_vnode_data *fvdat = VTOFUD(vp);
465 struct buf *bp;
466 daddr_t lbn;
467 off_t filesize;
468 int bcount;
469 int n, on, seqcount, err = 0;
470
471 const int biosize = fuse_iosize(vp);
472
473 seqcount = ioflag >> IO_SEQSHIFT;
474
475 KASSERT(uio->uio_rw == UIO_WRITE, ("fuse_write_biobackend mode"));
476 if (vp->v_type != VREG)
477 return (EIO);
478 if (uio->uio_offset < 0)
479 return (EINVAL);
480 if (uio->uio_resid == 0)
481 return (0);
482
483 err = fuse_vnode_size(vp, &filesize, cred, curthread);
484 if (err)
485 return err;
486
487 if (ioflag & IO_APPEND)
488 uio_setoffset(uio, filesize);
489
490 if (vn_rlimit_fsize(vp, uio, uio->uio_td))
491 return (EFBIG);
492
493 do {
494 bool direct_append, extending;
495
496 if (fuse_isdeadfs(vp)) {
497 err = ENXIO;
498 break;
499 }
500 lbn = uio->uio_offset / biosize;
501 on = uio->uio_offset & (biosize - 1);
502 n = MIN((unsigned)(biosize - on), uio->uio_resid);
503
504 again:
505 /* Get or create a buffer for the write */
506 direct_append = uio->uio_offset == filesize && n;
507 if (uio->uio_offset + n < filesize) {
508 extending = false;
509 if ((off_t)(lbn + 1) * biosize < filesize) {
510 /* Not the file's last block */
511 bcount = biosize;
512 } else {
513 /* The file's last block */
514 bcount = filesize - (off_t)lbn * biosize;
515 }
516 } else {
517 extending = true;
518 bcount = on + n;
519 }
520 if (direct_append) {
521 /*
522 * Take care to preserve the buffer's B_CACHE state so
523 * as not to cause an unnecessary read.
524 */
525 bp = getblk(vp, lbn, on, PCATCH, 0, 0);
526 if (bp != NULL) {
527 uint32_t save = bp->b_flags & B_CACHE;
528 allocbuf(bp, bcount);
529 bp->b_flags |= save;
530 }
531 } else {
532 bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
533 }
534 if (!bp) {
535 err = EINTR;
536 break;
537 }
538 if (extending) {
539 /*
540 * Extend file _after_ locking buffer so we won't race
541 * with other readers
542 */
543 err = fuse_vnode_setsize(vp, uio->uio_offset + n, false);
544 filesize = uio->uio_offset + n;
545 getnanouptime(&fvdat->last_local_modify);
546 fvdat->flag |= FN_SIZECHANGE;
547 if (err) {
548 brelse(bp);
549 break;
550 }
551 }
552
553 SDT_PROBE6(fusefs, , io, write_biobackend_start,
554 lbn, on, n, uio, bcount, direct_append);
555 /*
556 * Issue a READ if B_CACHE is not set. In special-append
557 * mode, B_CACHE is based on the buffer prior to the write
558 * op and is typically set, avoiding the read. If a read
559 * is required in special append mode, the server will
560 * probably send us a short-read since we extended the file
561 * on our end, resulting in b_resid == 0 and, thusly,
562 * B_CACHE getting set.
563 *
564 * We can also avoid issuing the read if the write covers
565 * the entire buffer. We have to make sure the buffer state
566 * is reasonable in this case since we will not be initiating
567 * I/O. See the comments in kern/vfs_bio.c's getblk() for
568 * more information.
569 *
570 * B_CACHE may also be set due to the buffer being cached
571 * normally.
572 */
573
574 if (on == 0 && n == bcount) {
575 bp->b_flags |= B_CACHE;
576 bp->b_flags &= ~B_INVAL;
577 bp->b_ioflags &= ~BIO_ERROR;
578 }
579 if ((bp->b_flags & B_CACHE) == 0) {
580 bp->b_iocmd = BIO_READ;
581 vfs_busy_pages(bp, 0);
582 fuse_io_strategy(vp, bp);
583 if ((err = bp->b_error)) {
584 brelse(bp);
585 break;
586 }
587 if (bp->b_resid > 0) {
588 /*
589 * Short read indicates EOF. Update file size
590 * from the server and try again.
591 */
592 SDT_PROBE2(fusefs, , io, trace, 1,
593 "Short read during a RMW");
594 brelse(bp);
595 err = fuse_vnode_size(vp, &filesize, cred,
596 curthread);
597 if (err)
598 break;
599 else
600 goto again;
601 }
602 }
603 if (bp->b_wcred == NOCRED)
604 bp->b_wcred = crhold(cred);
605
606 /*
607 * If dirtyend exceeds file size, chop it down. This should
608 * not normally occur but there is an append race where it
609 * might occur XXX, so we log it.
610 *
611 * If the chopping creates a reverse-indexed or degenerate
612 * situation with dirtyoff/end, we 0 both of them.
613 */
614 if (bp->b_dirtyend > bcount) {
615 SDT_PROBE2(fusefs, , io, write_biobackend_append_race,
616 (long)bp->b_blkno * biosize,
617 bp->b_dirtyend - bcount);
618 bp->b_dirtyend = bcount;
619 }
620 if (bp->b_dirtyoff >= bp->b_dirtyend)
621 bp->b_dirtyoff = bp->b_dirtyend = 0;
622
623 /*
624 * If the new write will leave a contiguous dirty
625 * area, just update the b_dirtyoff and b_dirtyend,
626 * otherwise force a write rpc of the old dirty area.
627 *
628 * While it is possible to merge discontiguous writes due to
629 * our having a B_CACHE buffer ( and thus valid read data
630 * for the hole), we don't because it could lead to
631 * significant cache coherency problems with multiple clients,
632 * especially if locking is implemented later on.
633 *
634 * as an optimization we could theoretically maintain
635 * a linked list of discontinuous areas, but we would still
636 * have to commit them separately so there isn't much
637 * advantage to it except perhaps a bit of asynchronization.
638 */
639
640 if (bp->b_dirtyend > 0 &&
641 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
642 /*
643 * Yes, we mean it. Write out everything to "storage"
644 * immediately, without hesitation. (Apart from other
645 * reasons: the only way to know if a write is valid
646 * if its actually written out.)
647 */
648 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 0, bp);
649 bwrite(bp);
650 if (bp->b_error == EINTR) {
651 err = EINTR;
652 break;
653 }
654 goto again;
655 }
656 err = uiomove((char *)bp->b_data + on, n, uio);
657
658 if (err) {
659 bp->b_ioflags |= BIO_ERROR;
660 bp->b_error = err;
661 brelse(bp);
662 break;
663 /* TODO: vfs_bio_clrbuf like ffs_write does? */
664 }
665 /*
666 * Only update dirtyoff/dirtyend if not a degenerate
667 * condition.
668 */
669 if (n) {
670 if (bp->b_dirtyend > 0) {
671 bp->b_dirtyoff = MIN(on, bp->b_dirtyoff);
672 bp->b_dirtyend = MAX((on + n), bp->b_dirtyend);
673 } else {
674 bp->b_dirtyoff = on;
675 bp->b_dirtyend = on + n;
676 }
677 vfs_bio_set_valid(bp, on, n);
678 }
679
680 vfs_bio_set_flags(bp, ioflag);
681
682 bp->b_flags |= B_FUSEFS_WRITE_CACHE;
683 if (ioflag & IO_SYNC) {
684 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 2, bp);
685 if (!(ioflag & IO_VMIO))
686 bp->b_flags &= ~B_FUSEFS_WRITE_CACHE;
687 err = bwrite(bp);
688 } else if (vm_page_count_severe() ||
689 buf_dirty_count_severe() ||
690 (ioflag & IO_ASYNC)) {
691 bp->b_flags |= B_CLUSTEROK;
692 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 3, bp);
693 bawrite(bp);
694 } else if (on == 0 && n == bcount) {
695 if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERW) == 0) {
696 bp->b_flags |= B_CLUSTEROK;
697 SDT_PROBE2(fusefs, , io, write_biobackend_issue,
698 4, bp);
699 cluster_write(vp, bp, filesize, seqcount, 0);
700 } else {
701 SDT_PROBE2(fusefs, , io, write_biobackend_issue,
702 5, bp);
703 bawrite(bp);
704 }
705 } else if (ioflag & IO_DIRECT) {
706 bp->b_flags |= B_CLUSTEROK;
707 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 6, bp);
708 bawrite(bp);
709 } else {
710 bp->b_flags &= ~B_CLUSTEROK;
711 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 7, bp);
712 bdwrite(bp);
713 }
714 if (err)
715 break;
716 } while (uio->uio_resid > 0 && n > 0);
717
718 return (err);
719 }
720
721 int
fuse_io_strategy(struct vnode * vp,struct buf * bp)722 fuse_io_strategy(struct vnode *vp, struct buf *bp)
723 {
724 struct fuse_vnode_data *fvdat = VTOFUD(vp);
725 struct fuse_filehandle *fufh;
726 struct ucred *cred;
727 struct uio *uiop;
728 struct uio uio;
729 struct iovec io;
730 off_t filesize;
731 int error = 0;
732 int fflag;
733 /* We don't know the true pid when we're dealing with the cache */
734 pid_t pid = 0;
735
736 const int biosize = fuse_iosize(vp);
737
738 MPASS(vp->v_type == VREG || vp->v_type == VDIR);
739 MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE);
740
741 fflag = bp->b_iocmd == BIO_READ ? FREAD : FWRITE;
742 cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred;
743 error = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid);
744 if (bp->b_iocmd == BIO_READ && error == EBADF) {
745 /*
746 * This may be a read-modify-write operation on a cached file
747 * opened O_WRONLY. The FUSE protocol allows this.
748 */
749 error = fuse_filehandle_get(vp, FWRITE, &fufh, cred, pid);
750 }
751 if (error) {
752 printf("FUSE: strategy: filehandles are closed\n");
753 bp->b_ioflags |= BIO_ERROR;
754 bp->b_error = error;
755 bufdone(bp);
756 return (error);
757 }
758
759 uiop = &uio;
760 uiop->uio_iov = &io;
761 uiop->uio_iovcnt = 1;
762 uiop->uio_segflg = UIO_SYSSPACE;
763 uiop->uio_td = curthread;
764
765 /*
766 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
767 * do this here so we do not have to do it in all the code that
768 * calls us.
769 */
770 bp->b_flags &= ~B_INVAL;
771 bp->b_ioflags &= ~BIO_ERROR;
772
773 KASSERT(!(bp->b_flags & B_DONE),
774 ("fuse_io_strategy: bp %p already marked done", bp));
775 if (bp->b_iocmd == BIO_READ) {
776 ssize_t left;
777
778 io.iov_len = uiop->uio_resid = bp->b_bcount;
779 io.iov_base = bp->b_data;
780 uiop->uio_rw = UIO_READ;
781
782 uiop->uio_offset = ((off_t)bp->b_lblkno) * biosize;
783 error = fuse_read_directbackend(vp, uiop, cred, fufh);
784 /*
785 * Store the amount we failed to read in the buffer's private
786 * field, so callers can truncate the file if necessary'
787 */
788
789 if (!error && uiop->uio_resid) {
790 int nread = bp->b_bcount - uiop->uio_resid;
791 left = uiop->uio_resid;
792 bzero((char *)bp->b_data + nread, left);
793
794 if ((fvdat->flag & FN_SIZECHANGE) == 0) {
795 /*
796 * A short read with no error, when not using
797 * direct io, and when no writes are cached,
798 * indicates EOF caused by a server-side
799 * truncation. Clear the attr cache so we'll
800 * pick up the new file size and timestamps.
801 *
802 * We must still bzero the remaining buffer so
803 * uninitialized data doesn't get exposed by a
804 * future truncate that extends the file.
805 *
806 * To prevent lock order problems, we must
807 * truncate the file upstack, not here.
808 */
809 SDT_PROBE2(fusefs, , io, trace, 1,
810 "Short read of a clean file");
811 fuse_vnode_clear_attr_cache(vp);
812 } else {
813 /*
814 * If dirty writes _are_ cached beyond EOF,
815 * that indicates a newly created hole that the
816 * server doesn't know about. Those don't pose
817 * any problem.
818 * XXX: we don't currently track whether dirty
819 * writes are cached beyond EOF, before EOF, or
820 * both.
821 */
822 SDT_PROBE2(fusefs, , io, trace, 1,
823 "Short read of a dirty file");
824 uiop->uio_resid = 0;
825 }
826 }
827 if (error) {
828 bp->b_ioflags |= BIO_ERROR;
829 bp->b_error = error;
830 }
831 } else {
832 /*
833 * Setup for actual write
834 */
835 /*
836 * If the file's size is cached, use that value, even if the
837 * cache is expired. At this point we're already committed to
838 * writing something. If the FUSE server has changed the
839 * file's size behind our back, it's too late for us to do
840 * anything about it. In particular, we can't invalidate any
841 * part of the file's buffers because VOP_STRATEGY is called
842 * with them already locked.
843 */
844 filesize = fvdat->cached_attrs.va_size;
845 /* filesize must've been cached by fuse_vnop_open. */
846 KASSERT(filesize != VNOVAL, ("filesize should've been cached"));
847
848 if ((off_t)bp->b_lblkno * biosize + bp->b_dirtyend > filesize)
849 bp->b_dirtyend = filesize -
850 (off_t)bp->b_lblkno * biosize;
851
852 if (bp->b_dirtyend > bp->b_dirtyoff) {
853 io.iov_len = uiop->uio_resid = bp->b_dirtyend
854 - bp->b_dirtyoff;
855 uiop->uio_offset = (off_t)bp->b_lblkno * biosize
856 + bp->b_dirtyoff;
857 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
858 uiop->uio_rw = UIO_WRITE;
859
860 bool pages = bp->b_flags & B_FUSEFS_WRITE_CACHE;
861 error = fuse_write_directbackend(vp, uiop, cred, fufh,
862 filesize, 0, pages);
863
864 if (error == EINTR || error == ETIMEDOUT) {
865 bp->b_flags &= ~(B_INVAL | B_NOCACHE);
866 if ((bp->b_flags & B_PAGING) == 0) {
867 bdirty(bp);
868 bp->b_flags &= ~B_DONE;
869 }
870 if ((error == EINTR || error == ETIMEDOUT) &&
871 (bp->b_flags & B_ASYNC) == 0)
872 bp->b_flags |= B_EINTR;
873 } else {
874 if (error) {
875 bp->b_ioflags |= BIO_ERROR;
876 bp->b_flags |= B_INVAL;
877 bp->b_error = error;
878 }
879 bp->b_dirtyoff = bp->b_dirtyend = 0;
880 }
881 } else {
882 bp->b_resid = 0;
883 bufdone(bp);
884 return (0);
885 }
886 }
887 bp->b_resid = uiop->uio_resid;
888 bufdone(bp);
889 return (error);
890 }
891
892 int
fuse_io_flushbuf(struct vnode * vp,int waitfor,struct thread * td)893 fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td)
894 {
895
896 return (vn_fsync_buf(vp, waitfor));
897 }
898
899 /*
900 * Flush and invalidate all dirty buffers. If another process is already
901 * doing the flush, just wait for completion.
902 */
903 int
fuse_io_invalbuf(struct vnode * vp,struct thread * td)904 fuse_io_invalbuf(struct vnode *vp, struct thread *td)
905 {
906 struct fuse_vnode_data *fvdat = VTOFUD(vp);
907 int error = 0;
908
909 if (VN_IS_DOOMED(vp))
910 return 0;
911
912 ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf");
913
914 while (fvdat->flag & FN_FLUSHINPROG) {
915 struct proc *p = td->td_proc;
916
917 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF)
918 return EIO;
919 fvdat->flag |= FN_FLUSHWANT;
920 tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz);
921 error = 0;
922 if (p != NULL) {
923 PROC_LOCK(p);
924 if (SIGNOTEMPTY(p->p_siglist) ||
925 SIGNOTEMPTY(td->td_siglist))
926 error = EINTR;
927 PROC_UNLOCK(p);
928 }
929 if (error == EINTR)
930 return EINTR;
931 }
932 fvdat->flag |= FN_FLUSHINPROG;
933
934 if (vp->v_bufobj.bo_object != NULL) {
935 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
936 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
937 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
938 }
939 error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
940 while (error) {
941 if (error == ERESTART || error == EINTR) {
942 fvdat->flag &= ~FN_FLUSHINPROG;
943 if (fvdat->flag & FN_FLUSHWANT) {
944 fvdat->flag &= ~FN_FLUSHWANT;
945 wakeup(&fvdat->flag);
946 }
947 return EINTR;
948 }
949 error = vinvalbuf(vp, V_SAVE, PCATCH, 0);
950 }
951 fvdat->flag &= ~FN_FLUSHINPROG;
952 if (fvdat->flag & FN_FLUSHWANT) {
953 fvdat->flag &= ~FN_FLUSHWANT;
954 wakeup(&fvdat->flag);
955 }
956 return (error);
957 }
958