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/module.h> 68 #include <sys/systm.h> 69 #include <sys/errno.h> 70 #include <sys/param.h> 71 #include <sys/kernel.h> 72 #include <sys/conf.h> 73 #include <sys/uio.h> 74 #include <sys/malloc.h> 75 #include <sys/queue.h> 76 #include <sys/lock.h> 77 #include <sys/sx.h> 78 #include <sys/mutex.h> 79 #include <sys/rwlock.h> 80 #include <sys/priv.h> 81 #include <sys/proc.h> 82 #include <sys/mount.h> 83 #include <sys/vnode.h> 84 #include <sys/stat.h> 85 #include <sys/unistd.h> 86 #include <sys/filedesc.h> 87 #include <sys/file.h> 88 #include <sys/fcntl.h> 89 #include <sys/bio.h> 90 #include <sys/buf.h> 91 #include <sys/sysctl.h> 92 #include <sys/vmmeter.h> 93 94 #include <vm/vm.h> 95 #include <vm/vm_extern.h> 96 #include <vm/pmap.h> 97 #include <vm/vm_map.h> 98 #include <vm/vm_page.h> 99 #include <vm/vm_object.h> 100 101 #include "fuse.h" 102 #include "fuse_file.h" 103 #include "fuse_node.h" 104 #include "fuse_internal.h" 105 #include "fuse_ipc.h" 106 #include "fuse_io.h" 107 108 /* 109 * Set in a struct buf to indicate that the write came from the buffer cache 110 * and the originating cred and pid are no longer known. 111 */ 112 #define B_FUSEFS_WRITE_CACHE B_FS_FLAG1 113 114 SDT_PROVIDER_DECLARE(fusefs); 115 /* 116 * Fuse trace probe: 117 * arg0: verbosity. Higher numbers give more verbose messages 118 * arg1: Textual message 119 */ 120 SDT_PROBE_DEFINE2(fusefs, , io, trace, "int", "char*"); 121 122 static int 123 fuse_inval_buf_range(struct vnode *vp, off_t filesize, off_t start, off_t end); 124 static void 125 fuse_io_clear_suid_on_write(struct vnode *vp, struct ucred *cred, 126 struct thread *td); 127 static int 128 fuse_read_directbackend(struct vnode *vp, struct uio *uio, 129 struct ucred *cred, struct fuse_filehandle *fufh); 130 static int 131 fuse_read_biobackend(struct vnode *vp, struct uio *uio, int ioflag, 132 struct ucred *cred, struct fuse_filehandle *fufh, pid_t pid); 133 static int 134 fuse_write_directbackend(struct vnode *vp, struct uio *uio, 135 struct ucred *cred, struct fuse_filehandle *fufh, off_t filesize, 136 int ioflag, bool pages); 137 static int 138 fuse_write_biobackend(struct vnode *vp, struct uio *uio, 139 struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid); 140 141 /* Invalidate a range of cached data, whether dirty of not */ 142 static int 143 fuse_inval_buf_range(struct vnode *vp, off_t filesize, off_t start, off_t end) 144 { 145 struct buf *bp; 146 daddr_t left_lbn, end_lbn, right_lbn; 147 off_t new_filesize; 148 int iosize, left_on, right_on, right_blksize; 149 150 iosize = fuse_iosize(vp); 151 left_lbn = start / iosize; 152 end_lbn = howmany(end, iosize); 153 left_on = start & (iosize - 1); 154 if (left_on != 0) { 155 bp = getblk(vp, left_lbn, iosize, PCATCH, 0, 0); 156 if ((bp->b_flags & B_CACHE) != 0 && bp->b_dirtyend >= left_on) { 157 /* 158 * Flush the dirty buffer, because we don't have a 159 * byte-granular way to record which parts of the 160 * buffer are valid. 161 */ 162 bwrite(bp); 163 if (bp->b_error) 164 return (bp->b_error); 165 } else { 166 brelse(bp); 167 } 168 } 169 right_on = end & (iosize - 1); 170 if (right_on != 0) { 171 right_lbn = end / iosize; 172 new_filesize = MAX(filesize, end); 173 right_blksize = MIN(iosize, new_filesize - iosize * right_lbn); 174 bp = getblk(vp, right_lbn, right_blksize, PCATCH, 0, 0); 175 if ((bp->b_flags & B_CACHE) != 0 && bp->b_dirtyoff < right_on) { 176 /* 177 * Flush the dirty buffer, because we don't have a 178 * byte-granular way to record which parts of the 179 * buffer are valid. 180 */ 181 bwrite(bp); 182 if (bp->b_error) 183 return (bp->b_error); 184 } else { 185 brelse(bp); 186 } 187 } 188 189 v_inval_buf_range(vp, left_lbn, end_lbn, iosize); 190 return (0); 191 } 192 193 /* 194 * FreeBSD clears the SUID and SGID bits on any write by a non-root user. 195 */ 196 static void 197 fuse_io_clear_suid_on_write(struct vnode *vp, struct ucred *cred, 198 struct thread *td) 199 { 200 struct fuse_data *data; 201 struct mount *mp; 202 struct vattr va; 203 int dataflags; 204 205 mp = vnode_mount(vp); 206 data = fuse_get_mpdata(mp); 207 dataflags = data->dataflags; 208 209 if (dataflags & FSESS_DEFAULT_PERMISSIONS) { 210 if (priv_check_cred(cred, PRIV_VFS_RETAINSUGID)) { 211 fuse_internal_getattr(vp, &va, cred, td); 212 if (va.va_mode & (S_ISUID | S_ISGID)) { 213 mode_t mode = va.va_mode & ~(S_ISUID | S_ISGID); 214 /* Clear all vattr fields except mode */ 215 vattr_null(&va); 216 va.va_mode = mode; 217 218 /* 219 * Ignore fuse_internal_setattr's return value, 220 * because at this point the write operation has 221 * already succeeded and we don't want to return 222 * failing status for that. 223 */ 224 (void)fuse_internal_setattr(vp, &va, td, NULL); 225 } 226 } 227 } 228 } 229 230 SDT_PROBE_DEFINE5(fusefs, , io, io_dispatch, "struct vnode*", "struct uio*", 231 "int", "struct ucred*", "struct fuse_filehandle*"); 232 SDT_PROBE_DEFINE4(fusefs, , io, io_dispatch_filehandles_closed, "struct vnode*", 233 "struct uio*", "int", "struct ucred*"); 234 int 235 fuse_io_dispatch(struct vnode *vp, struct uio *uio, int ioflag, 236 struct ucred *cred, pid_t pid) 237 { 238 struct fuse_filehandle *fufh; 239 int err, directio; 240 int fflag; 241 bool closefufh = false; 242 243 MPASS(vp->v_type == VREG || vp->v_type == VDIR); 244 245 fflag = (uio->uio_rw == UIO_READ) ? FREAD : FWRITE; 246 err = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); 247 if (err == EBADF && vnode_mount(vp)->mnt_flag & MNT_EXPORTED) { 248 /* 249 * nfsd will do I/O without first doing VOP_OPEN. We 250 * must implicitly open the file here 251 */ 252 err = fuse_filehandle_open(vp, fflag, &fufh, curthread, cred); 253 closefufh = true; 254 } 255 else if (err) { 256 SDT_PROBE4(fusefs, , io, io_dispatch_filehandles_closed, 257 vp, uio, ioflag, cred); 258 printf("FUSE: io dispatch: filehandles are closed\n"); 259 return err; 260 } 261 if (err) 262 goto out; 263 SDT_PROBE5(fusefs, , io, io_dispatch, vp, uio, ioflag, cred, fufh); 264 265 /* 266 * Ideally, when the daemon asks for direct io at open time, the 267 * standard file flag should be set according to this, so that would 268 * just change the default mode, which later on could be changed via 269 * fcntl(2). 270 * But this doesn't work, the O_DIRECT flag gets cleared at some point 271 * (don't know where). So to make any use of the Fuse direct_io option, 272 * we hardwire it into the file's private data (similarly to Linux, 273 * btw.). 274 */ 275 directio = (ioflag & IO_DIRECT) || !fsess_opt_datacache(vnode_mount(vp)); 276 277 switch (uio->uio_rw) { 278 case UIO_READ: 279 if (directio) { 280 SDT_PROBE2(fusefs, , io, trace, 1, 281 "direct read of vnode"); 282 err = fuse_read_directbackend(vp, uio, cred, fufh); 283 } else { 284 SDT_PROBE2(fusefs, , io, trace, 1, 285 "buffered read of vnode"); 286 err = fuse_read_biobackend(vp, uio, ioflag, cred, fufh, 287 pid); 288 } 289 break; 290 case UIO_WRITE: 291 fuse_vnode_update(vp, FN_MTIMECHANGE | FN_CTIMECHANGE); 292 if (directio) { 293 off_t start, end, filesize; 294 295 SDT_PROBE2(fusefs, , io, trace, 1, 296 "direct write of vnode"); 297 298 err = fuse_vnode_size(vp, &filesize, cred, curthread); 299 if (err) 300 goto out; 301 302 start = uio->uio_offset; 303 end = start + uio->uio_resid; 304 KASSERT((ioflag & (IO_VMIO | IO_DIRECT)) != 305 (IO_VMIO | IO_DIRECT), 306 ("IO_DIRECT used for a cache flush?")); 307 /* Invalidate the write cache when writing directly */ 308 err = fuse_inval_buf_range(vp, filesize, start, end); 309 if (err) 310 return (err); 311 err = fuse_write_directbackend(vp, uio, cred, fufh, 312 filesize, ioflag, false); 313 } else { 314 SDT_PROBE2(fusefs, , io, trace, 1, 315 "buffered write of vnode"); 316 if (!fsess_opt_writeback(vnode_mount(vp))) 317 ioflag |= IO_SYNC; 318 err = fuse_write_biobackend(vp, uio, cred, fufh, ioflag, 319 pid); 320 } 321 fuse_io_clear_suid_on_write(vp, cred, uio->uio_td); 322 break; 323 default: 324 panic("uninterpreted mode passed to fuse_io_dispatch"); 325 } 326 327 out: 328 if (closefufh) 329 fuse_filehandle_close(vp, fufh, curthread, cred); 330 331 return (err); 332 } 333 334 SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_start, "int", "int", "int", "int"); 335 SDT_PROBE_DEFINE2(fusefs, , io, read_bio_backend_feed, "int", "struct buf*"); 336 SDT_PROBE_DEFINE4(fusefs, , io, read_bio_backend_end, "int", "ssize_t", "int", 337 "struct buf*"); 338 static int 339 fuse_read_biobackend(struct vnode *vp, struct uio *uio, int ioflag, 340 struct ucred *cred, struct fuse_filehandle *fufh, pid_t pid) 341 { 342 struct buf *bp; 343 struct mount *mp; 344 struct fuse_data *data; 345 daddr_t lbn, nextlbn; 346 int bcount, nextsize; 347 int err, n = 0, on = 0, seqcount; 348 off_t filesize; 349 350 const int biosize = fuse_iosize(vp); 351 mp = vnode_mount(vp); 352 data = fuse_get_mpdata(mp); 353 354 if (uio->uio_offset < 0) 355 return (EINVAL); 356 357 seqcount = ioflag >> IO_SEQSHIFT; 358 359 err = fuse_vnode_size(vp, &filesize, cred, curthread); 360 if (err) 361 return err; 362 363 for (err = 0, bp = NULL; uio->uio_resid > 0; bp = NULL) { 364 if (fuse_isdeadfs(vp)) { 365 err = ENXIO; 366 break; 367 } 368 if (filesize - uio->uio_offset <= 0) 369 break; 370 lbn = uio->uio_offset / biosize; 371 on = uio->uio_offset & (biosize - 1); 372 373 if ((off_t)lbn * biosize >= filesize) { 374 bcount = 0; 375 } else if ((off_t)(lbn + 1) * biosize > filesize) { 376 bcount = filesize - (off_t)lbn *biosize; 377 } else { 378 bcount = biosize; 379 } 380 nextlbn = lbn + 1; 381 nextsize = MIN(biosize, filesize - nextlbn * biosize); 382 383 SDT_PROBE4(fusefs, , io, read_bio_backend_start, 384 biosize, (int)lbn, on, bcount); 385 386 if (bcount < biosize) { 387 /* If near EOF, don't do readahead */ 388 err = bread(vp, lbn, bcount, NOCRED, &bp); 389 } else if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERR) == 0) { 390 /* Try clustered read */ 391 long totread = uio->uio_resid + on; 392 seqcount = MIN(seqcount, 393 data->max_readahead_blocks + 1); 394 err = cluster_read(vp, filesize, lbn, bcount, NOCRED, 395 totread, seqcount, 0, &bp); 396 } else if (seqcount > 1 && data->max_readahead_blocks >= 1) { 397 /* Try non-clustered readahead */ 398 err = breadn(vp, lbn, bcount, &nextlbn, &nextsize, 1, 399 NOCRED, &bp); 400 } else { 401 /* Just read what was requested */ 402 err = bread(vp, lbn, bcount, NOCRED, &bp); 403 } 404 405 if (err) { 406 brelse(bp); 407 bp = NULL; 408 break; 409 } 410 411 /* 412 * on is the offset into the current bp. Figure out how many 413 * bytes we can copy out of the bp. Note that bcount is 414 * NOT DEV_BSIZE aligned. 415 * 416 * Then figure out how many bytes we can copy into the uio. 417 */ 418 419 n = 0; 420 if (on < bcount - bp->b_resid) 421 n = MIN((unsigned)(bcount - bp->b_resid - on), 422 uio->uio_resid); 423 if (n > 0) { 424 SDT_PROBE2(fusefs, , io, read_bio_backend_feed, n, bp); 425 err = uiomove(bp->b_data + on, n, uio); 426 } 427 vfs_bio_brelse(bp, ioflag); 428 SDT_PROBE4(fusefs, , io, read_bio_backend_end, err, 429 uio->uio_resid, n, bp); 430 if (bp->b_resid > 0) { 431 /* Short read indicates EOF */ 432 break; 433 } 434 } 435 436 return (err); 437 } 438 439 SDT_PROBE_DEFINE1(fusefs, , io, read_directbackend_start, 440 "struct fuse_read_in*"); 441 SDT_PROBE_DEFINE3(fusefs, , io, read_directbackend_complete, 442 "struct fuse_dispatcher*", "struct fuse_read_in*", "struct uio*"); 443 444 static int 445 fuse_read_directbackend(struct vnode *vp, struct uio *uio, 446 struct ucred *cred, struct fuse_filehandle *fufh) 447 { 448 struct fuse_data *data; 449 struct fuse_dispatcher fdi; 450 struct fuse_read_in *fri; 451 int err = 0; 452 453 data = fuse_get_mpdata(vp->v_mount); 454 455 if (uio->uio_resid == 0) 456 return (0); 457 458 fdisp_init(&fdi, 0); 459 460 /* 461 * XXX In "normal" case we use an intermediate kernel buffer for 462 * transmitting data from daemon's context to ours. Eventually, we should 463 * get rid of this. Anyway, if the target uio lives in sysspace (we are 464 * called from pageops), and the input data doesn't need kernel-side 465 * processing (we are not called from readdir) we can already invoke 466 * an optimized, "peer-to-peer" I/O routine. 467 */ 468 while (uio->uio_resid > 0) { 469 fdi.iosize = sizeof(*fri); 470 fdisp_make_vp(&fdi, FUSE_READ, vp, uio->uio_td, cred); 471 fri = fdi.indata; 472 fri->fh = fufh->fh_id; 473 fri->offset = uio->uio_offset; 474 fri->size = MIN(uio->uio_resid, 475 fuse_get_mpdata(vp->v_mount)->max_read); 476 if (fuse_libabi_geq(data, 7, 9)) { 477 /* See comment regarding FUSE_WRITE_LOCKOWNER */ 478 fri->read_flags = 0; 479 fri->flags = fufh_type_2_fflags(fufh->fufh_type); 480 } 481 482 SDT_PROBE1(fusefs, , io, read_directbackend_start, fri); 483 484 if ((err = fdisp_wait_answ(&fdi))) 485 goto out; 486 487 SDT_PROBE3(fusefs, , io, read_directbackend_complete, 488 &fdi, fri, uio); 489 490 if ((err = uiomove(fdi.answ, MIN(fri->size, fdi.iosize), uio))) 491 break; 492 if (fdi.iosize < fri->size) { 493 /* 494 * Short read. Should only happen at EOF or with 495 * direct io. 496 */ 497 break; 498 } 499 } 500 501 out: 502 fdisp_destroy(&fdi); 503 return (err); 504 } 505 506 static int 507 fuse_write_directbackend(struct vnode *vp, struct uio *uio, 508 struct ucred *cred, struct fuse_filehandle *fufh, off_t filesize, 509 int ioflag, bool pages) 510 { 511 struct fuse_vnode_data *fvdat = VTOFUD(vp); 512 struct fuse_data *data; 513 struct fuse_write_in *fwi; 514 struct fuse_write_out *fwo; 515 struct fuse_dispatcher fdi; 516 size_t chunksize; 517 void *fwi_data; 518 off_t as_written_offset; 519 int diff; 520 int err = 0; 521 bool direct_io = fufh->fuse_open_flags & FOPEN_DIRECT_IO; 522 bool wrote_anything = false; 523 uint32_t write_flags; 524 525 data = fuse_get_mpdata(vp->v_mount); 526 527 /* 528 * Don't set FUSE_WRITE_LOCKOWNER in write_flags. It can't be set 529 * accurately when using POSIX AIO, libfuse doesn't use it, and I'm not 530 * aware of any file systems that do. It was an attempt to add 531 * Linux-style mandatory locking to the FUSE protocol, but mandatory 532 * locking is deprecated even on Linux. See Linux commit 533 * f33321141b273d60cbb3a8f56a5489baad82ba5e . 534 */ 535 /* 536 * Set FUSE_WRITE_CACHE whenever we don't know the uid, gid, and/or pid 537 * that originated a write. For example when writing from the 538 * writeback cache. I don't know of a single file system that cares, 539 * but the protocol says we're supposed to do this. 540 */ 541 write_flags = !pages && ( 542 (ioflag & IO_DIRECT) || 543 !fsess_opt_datacache(vnode_mount(vp)) || 544 !fsess_opt_writeback(vnode_mount(vp))) ? 0 : FUSE_WRITE_CACHE; 545 546 if (uio->uio_resid == 0) 547 return (0); 548 549 if (ioflag & IO_APPEND) 550 uio_setoffset(uio, filesize); 551 552 if (vn_rlimit_fsize(vp, uio, uio->uio_td)) 553 return (EFBIG); 554 555 fdisp_init(&fdi, 0); 556 557 while (uio->uio_resid > 0) { 558 chunksize = MIN(uio->uio_resid, data->max_write); 559 560 fdi.iosize = sizeof(*fwi) + chunksize; 561 fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred); 562 563 fwi = fdi.indata; 564 fwi->fh = fufh->fh_id; 565 fwi->offset = uio->uio_offset; 566 fwi->size = chunksize; 567 fwi->write_flags = write_flags; 568 if (fuse_libabi_geq(data, 7, 9)) { 569 fwi->flags = fufh_type_2_fflags(fufh->fufh_type); 570 fwi_data = (char *)fdi.indata + sizeof(*fwi); 571 } else { 572 fwi_data = (char *)fdi.indata + 573 FUSE_COMPAT_WRITE_IN_SIZE; 574 } 575 576 if ((err = uiomove(fwi_data, chunksize, uio))) 577 break; 578 579 retry: 580 err = fdisp_wait_answ(&fdi); 581 if (err == ERESTART || err == EINTR || err == EWOULDBLOCK) { 582 /* 583 * Rewind the uio so dofilewrite will know it's 584 * incomplete 585 */ 586 uio->uio_resid += fwi->size; 587 uio->uio_offset -= fwi->size; 588 /* 589 * Change ERESTART into EINTR because we can't rewind 590 * uio->uio_iov. Basically, once uiomove(9) has been 591 * called, it's impossible to restart a syscall. 592 */ 593 if (err == ERESTART) 594 err = EINTR; 595 break; 596 } else if (err) { 597 break; 598 } else { 599 wrote_anything = true; 600 } 601 602 fwo = ((struct fuse_write_out *)fdi.answ); 603 604 /* Adjust the uio in the case of short writes */ 605 diff = fwi->size - fwo->size; 606 as_written_offset = uio->uio_offset - diff; 607 608 if (as_written_offset - diff > filesize) 609 fuse_vnode_setsize(vp, as_written_offset); 610 if (as_written_offset - diff >= filesize) 611 fvdat->flag &= ~FN_SIZECHANGE; 612 613 if (diff < 0) { 614 printf("WARNING: misbehaving FUSE filesystem " 615 "wrote more data than we provided it\n"); 616 err = EINVAL; 617 break; 618 } else if (diff > 0) { 619 /* Short write */ 620 if (!direct_io) { 621 printf("WARNING: misbehaving FUSE filesystem: " 622 "short writes are only allowed with " 623 "direct_io\n"); 624 } 625 if (ioflag & IO_DIRECT) { 626 /* Return early */ 627 uio->uio_resid += diff; 628 uio->uio_offset -= diff; 629 break; 630 } else { 631 /* Resend the unwritten portion of data */ 632 fdi.iosize = sizeof(*fwi) + diff; 633 /* Refresh fdi without clearing data buffer */ 634 fdisp_refresh_vp(&fdi, FUSE_WRITE, vp, 635 uio->uio_td, cred); 636 fwi = fdi.indata; 637 MPASS2(fwi == fdi.indata, "FUSE dispatcher " 638 "reallocated despite no increase in " 639 "size?"); 640 void *src = (char*)fwi_data + fwo->size; 641 memmove(fwi_data, src, diff); 642 fwi->fh = fufh->fh_id; 643 fwi->offset = as_written_offset; 644 fwi->size = diff; 645 fwi->write_flags = write_flags; 646 goto retry; 647 } 648 } 649 } 650 651 fdisp_destroy(&fdi); 652 653 if (wrote_anything) 654 fuse_vnode_undirty_cached_timestamps(vp); 655 656 return (err); 657 } 658 659 SDT_PROBE_DEFINE6(fusefs, , io, write_biobackend_start, "int64_t", "int", "int", 660 "struct uio*", "int", "bool"); 661 SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_append_race, "long", "int"); 662 SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_issue, "int", "struct buf*"); 663 664 static int 665 fuse_write_biobackend(struct vnode *vp, struct uio *uio, 666 struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid) 667 { 668 struct fuse_vnode_data *fvdat = VTOFUD(vp); 669 struct buf *bp; 670 daddr_t lbn; 671 off_t filesize; 672 int bcount; 673 int n, on, seqcount, err = 0; 674 bool last_page; 675 676 const int biosize = fuse_iosize(vp); 677 678 seqcount = ioflag >> IO_SEQSHIFT; 679 680 KASSERT(uio->uio_rw == UIO_WRITE, ("fuse_write_biobackend mode")); 681 if (vp->v_type != VREG) 682 return (EIO); 683 if (uio->uio_offset < 0) 684 return (EINVAL); 685 if (uio->uio_resid == 0) 686 return (0); 687 688 err = fuse_vnode_size(vp, &filesize, cred, curthread); 689 if (err) 690 return err; 691 692 if (ioflag & IO_APPEND) 693 uio_setoffset(uio, filesize); 694 695 if (vn_rlimit_fsize(vp, uio, uio->uio_td)) 696 return (EFBIG); 697 698 do { 699 bool direct_append, extending; 700 701 if (fuse_isdeadfs(vp)) { 702 err = ENXIO; 703 break; 704 } 705 lbn = uio->uio_offset / biosize; 706 on = uio->uio_offset & (biosize - 1); 707 n = MIN((unsigned)(biosize - on), uio->uio_resid); 708 709 again: 710 /* Get or create a buffer for the write */ 711 direct_append = uio->uio_offset == filesize && n; 712 if (uio->uio_offset + n < filesize) { 713 extending = false; 714 if ((off_t)(lbn + 1) * biosize < filesize) { 715 /* Not the file's last block */ 716 bcount = biosize; 717 } else { 718 /* The file's last block */ 719 bcount = filesize - (off_t)lbn * biosize; 720 } 721 } else { 722 extending = true; 723 bcount = on + n; 724 } 725 if (howmany(((off_t)lbn * biosize + on + n - 1), PAGE_SIZE) >= 726 howmany(filesize, PAGE_SIZE)) 727 last_page = true; 728 else 729 last_page = false; 730 if (direct_append) { 731 /* 732 * Take care to preserve the buffer's B_CACHE state so 733 * as not to cause an unnecessary read. 734 */ 735 bp = getblk(vp, lbn, on, PCATCH, 0, 0); 736 if (bp != NULL) { 737 uint32_t save = bp->b_flags & B_CACHE; 738 allocbuf(bp, bcount); 739 bp->b_flags |= save; 740 } 741 } else { 742 bp = getblk(vp, lbn, bcount, PCATCH, 0, 0); 743 } 744 if (!bp) { 745 err = EINTR; 746 break; 747 } 748 if (extending) { 749 /* 750 * Extend file _after_ locking buffer so we won't race 751 * with other readers 752 */ 753 err = fuse_vnode_setsize(vp, uio->uio_offset + n); 754 filesize = uio->uio_offset + n; 755 fvdat->flag |= FN_SIZECHANGE; 756 if (err) { 757 brelse(bp); 758 break; 759 } 760 } 761 762 SDT_PROBE6(fusefs, , io, write_biobackend_start, 763 lbn, on, n, uio, bcount, direct_append); 764 /* 765 * Issue a READ if B_CACHE is not set. In special-append 766 * mode, B_CACHE is based on the buffer prior to the write 767 * op and is typically set, avoiding the read. If a read 768 * is required in special append mode, the server will 769 * probably send us a short-read since we extended the file 770 * on our end, resulting in b_resid == 0 and, thusly, 771 * B_CACHE getting set. 772 * 773 * We can also avoid issuing the read if the write covers 774 * the entire buffer. We have to make sure the buffer state 775 * is reasonable in this case since we will not be initiating 776 * I/O. See the comments in kern/vfs_bio.c's getblk() for 777 * more information. 778 * 779 * B_CACHE may also be set due to the buffer being cached 780 * normally. 781 */ 782 783 if (on == 0 && n == bcount) { 784 bp->b_flags |= B_CACHE; 785 bp->b_flags &= ~B_INVAL; 786 bp->b_ioflags &= ~BIO_ERROR; 787 } 788 if ((bp->b_flags & B_CACHE) == 0) { 789 bp->b_iocmd = BIO_READ; 790 vfs_busy_pages(bp, 0); 791 fuse_io_strategy(vp, bp); 792 if ((err = bp->b_error)) { 793 brelse(bp); 794 break; 795 } 796 if (bp->b_resid > 0) { 797 /* 798 * Short read indicates EOF. Update file size 799 * from the server and try again. 800 */ 801 SDT_PROBE2(fusefs, , io, trace, 1, 802 "Short read during a RMW"); 803 brelse(bp); 804 err = fuse_vnode_size(vp, &filesize, cred, 805 curthread); 806 if (err) 807 break; 808 else 809 goto again; 810 } 811 } 812 if (bp->b_wcred == NOCRED) 813 bp->b_wcred = crhold(cred); 814 815 /* 816 * If dirtyend exceeds file size, chop it down. This should 817 * not normally occur but there is an append race where it 818 * might occur XXX, so we log it. 819 * 820 * If the chopping creates a reverse-indexed or degenerate 821 * situation with dirtyoff/end, we 0 both of them. 822 */ 823 if (bp->b_dirtyend > bcount) { 824 SDT_PROBE2(fusefs, , io, write_biobackend_append_race, 825 (long)bp->b_blkno * biosize, 826 bp->b_dirtyend - bcount); 827 bp->b_dirtyend = bcount; 828 } 829 if (bp->b_dirtyoff >= bp->b_dirtyend) 830 bp->b_dirtyoff = bp->b_dirtyend = 0; 831 832 /* 833 * If the new write will leave a contiguous dirty 834 * area, just update the b_dirtyoff and b_dirtyend, 835 * otherwise force a write rpc of the old dirty area. 836 * 837 * While it is possible to merge discontiguous writes due to 838 * our having a B_CACHE buffer ( and thus valid read data 839 * for the hole), we don't because it could lead to 840 * significant cache coherency problems with multiple clients, 841 * especially if locking is implemented later on. 842 * 843 * as an optimization we could theoretically maintain 844 * a linked list of discontinuous areas, but we would still 845 * have to commit them separately so there isn't much 846 * advantage to it except perhaps a bit of asynchronization. 847 */ 848 849 if (bp->b_dirtyend > 0 && 850 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 851 /* 852 * Yes, we mean it. Write out everything to "storage" 853 * immediately, without hesitation. (Apart from other 854 * reasons: the only way to know if a write is valid 855 * if its actually written out.) 856 */ 857 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 0, bp); 858 bwrite(bp); 859 if (bp->b_error == EINTR) { 860 err = EINTR; 861 break; 862 } 863 goto again; 864 } 865 err = uiomove((char *)bp->b_data + on, n, uio); 866 867 if (err) { 868 bp->b_ioflags |= BIO_ERROR; 869 bp->b_error = err; 870 brelse(bp); 871 break; 872 /* TODO: vfs_bio_clrbuf like ffs_write does? */ 873 } 874 /* 875 * Only update dirtyoff/dirtyend if not a degenerate 876 * condition. 877 */ 878 if (n) { 879 if (bp->b_dirtyend > 0) { 880 bp->b_dirtyoff = MIN(on, bp->b_dirtyoff); 881 bp->b_dirtyend = MAX((on + n), bp->b_dirtyend); 882 } else { 883 bp->b_dirtyoff = on; 884 bp->b_dirtyend = on + n; 885 } 886 vfs_bio_set_valid(bp, on, n); 887 } 888 889 vfs_bio_set_flags(bp, ioflag); 890 891 bp->b_flags |= B_FUSEFS_WRITE_CACHE; 892 if (ioflag & IO_SYNC) { 893 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 2, bp); 894 if (!(ioflag & IO_VMIO)) 895 bp->b_flags &= ~B_FUSEFS_WRITE_CACHE; 896 err = bwrite(bp); 897 } else if (vm_page_count_severe() || 898 buf_dirty_count_severe() || 899 (ioflag & IO_ASYNC)) { 900 bp->b_flags |= B_CLUSTEROK; 901 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 3, bp); 902 bawrite(bp); 903 } else if (on == 0 && n == bcount) { 904 if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERW) == 0) { 905 bp->b_flags |= B_CLUSTEROK; 906 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 907 4, bp); 908 cluster_write(vp, bp, filesize, seqcount, 0); 909 } else { 910 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 911 5, bp); 912 bawrite(bp); 913 } 914 } else if (ioflag & IO_DIRECT) { 915 bp->b_flags |= B_CLUSTEROK; 916 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 6, bp); 917 bawrite(bp); 918 } else { 919 bp->b_flags &= ~B_CLUSTEROK; 920 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 7, bp); 921 bdwrite(bp); 922 } 923 if (err) 924 break; 925 } while (uio->uio_resid > 0 && n > 0); 926 927 return (err); 928 } 929 930 int 931 fuse_io_strategy(struct vnode *vp, struct buf *bp) 932 { 933 struct fuse_vnode_data *fvdat = VTOFUD(vp); 934 struct fuse_filehandle *fufh; 935 struct ucred *cred; 936 struct uio *uiop; 937 struct uio uio; 938 struct iovec io; 939 off_t filesize; 940 int error = 0; 941 int fflag; 942 /* We don't know the true pid when we're dealing with the cache */ 943 pid_t pid = 0; 944 945 const int biosize = fuse_iosize(vp); 946 947 MPASS(vp->v_type == VREG || vp->v_type == VDIR); 948 MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE); 949 950 fflag = bp->b_iocmd == BIO_READ ? FREAD : FWRITE; 951 cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred; 952 error = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); 953 if (bp->b_iocmd == BIO_READ && error == EBADF) { 954 /* 955 * This may be a read-modify-write operation on a cached file 956 * opened O_WRONLY. The FUSE protocol allows this. 957 */ 958 error = fuse_filehandle_get(vp, FWRITE, &fufh, cred, pid); 959 } 960 if (error) { 961 printf("FUSE: strategy: filehandles are closed\n"); 962 bp->b_ioflags |= BIO_ERROR; 963 bp->b_error = error; 964 bufdone(bp); 965 return (error); 966 } 967 968 uiop = &uio; 969 uiop->uio_iov = &io; 970 uiop->uio_iovcnt = 1; 971 uiop->uio_segflg = UIO_SYSSPACE; 972 uiop->uio_td = curthread; 973 974 /* 975 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 976 * do this here so we do not have to do it in all the code that 977 * calls us. 978 */ 979 bp->b_flags &= ~B_INVAL; 980 bp->b_ioflags &= ~BIO_ERROR; 981 982 KASSERT(!(bp->b_flags & B_DONE), 983 ("fuse_io_strategy: bp %p already marked done", bp)); 984 if (bp->b_iocmd == BIO_READ) { 985 ssize_t left; 986 987 io.iov_len = uiop->uio_resid = bp->b_bcount; 988 io.iov_base = bp->b_data; 989 uiop->uio_rw = UIO_READ; 990 991 uiop->uio_offset = ((off_t)bp->b_lblkno) * biosize; 992 error = fuse_read_directbackend(vp, uiop, cred, fufh); 993 /* 994 * Store the amount we failed to read in the buffer's private 995 * field, so callers can truncate the file if necessary' 996 */ 997 998 if (!error && uiop->uio_resid) { 999 int nread = bp->b_bcount - uiop->uio_resid; 1000 left = uiop->uio_resid; 1001 bzero((char *)bp->b_data + nread, left); 1002 1003 if ((fvdat->flag & FN_SIZECHANGE) == 0) { 1004 /* 1005 * A short read with no error, when not using 1006 * direct io, and when no writes are cached, 1007 * indicates EOF caused by a server-side 1008 * truncation. Clear the attr cache so we'll 1009 * pick up the new file size and timestamps. 1010 * 1011 * We must still bzero the remaining buffer so 1012 * uninitialized data doesn't get exposed by a 1013 * future truncate that extends the file. 1014 * 1015 * To prevent lock order problems, we must 1016 * truncate the file upstack, not here. 1017 */ 1018 SDT_PROBE2(fusefs, , io, trace, 1, 1019 "Short read of a clean file"); 1020 fuse_vnode_clear_attr_cache(vp); 1021 } else { 1022 /* 1023 * If dirty writes _are_ cached beyond EOF, 1024 * that indicates a newly created hole that the 1025 * server doesn't know about. Those don't pose 1026 * any problem. 1027 * XXX: we don't currently track whether dirty 1028 * writes are cached beyond EOF, before EOF, or 1029 * both. 1030 */ 1031 SDT_PROBE2(fusefs, , io, trace, 1, 1032 "Short read of a dirty file"); 1033 uiop->uio_resid = 0; 1034 } 1035 1036 } 1037 if (error) { 1038 bp->b_ioflags |= BIO_ERROR; 1039 bp->b_error = error; 1040 } 1041 } else { 1042 /* 1043 * Setup for actual write 1044 */ 1045 error = fuse_vnode_size(vp, &filesize, cred, curthread); 1046 if (error) { 1047 bp->b_ioflags |= BIO_ERROR; 1048 bp->b_error = error; 1049 bufdone(bp); 1050 return (error); 1051 } 1052 1053 if ((off_t)bp->b_lblkno * biosize + bp->b_dirtyend > filesize) 1054 bp->b_dirtyend = filesize - 1055 (off_t)bp->b_lblkno * biosize; 1056 1057 if (bp->b_dirtyend > bp->b_dirtyoff) { 1058 io.iov_len = uiop->uio_resid = bp->b_dirtyend 1059 - bp->b_dirtyoff; 1060 uiop->uio_offset = (off_t)bp->b_lblkno * biosize 1061 + bp->b_dirtyoff; 1062 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 1063 uiop->uio_rw = UIO_WRITE; 1064 1065 bool pages = bp->b_flags & B_FUSEFS_WRITE_CACHE; 1066 error = fuse_write_directbackend(vp, uiop, cred, fufh, 1067 filesize, 0, pages); 1068 1069 if (error == EINTR || error == ETIMEDOUT) { 1070 bp->b_flags &= ~(B_INVAL | B_NOCACHE); 1071 if ((bp->b_flags & B_PAGING) == 0) { 1072 bdirty(bp); 1073 bp->b_flags &= ~B_DONE; 1074 } 1075 if ((error == EINTR || error == ETIMEDOUT) && 1076 (bp->b_flags & B_ASYNC) == 0) 1077 bp->b_flags |= B_EINTR; 1078 } else { 1079 if (error) { 1080 bp->b_ioflags |= BIO_ERROR; 1081 bp->b_flags |= B_INVAL; 1082 bp->b_error = error; 1083 } 1084 bp->b_dirtyoff = bp->b_dirtyend = 0; 1085 } 1086 } else { 1087 bp->b_resid = 0; 1088 bufdone(bp); 1089 return (0); 1090 } 1091 } 1092 bp->b_resid = uiop->uio_resid; 1093 bufdone(bp); 1094 return (error); 1095 } 1096 1097 int 1098 fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td) 1099 { 1100 1101 return (vn_fsync_buf(vp, waitfor)); 1102 } 1103 1104 /* 1105 * Flush and invalidate all dirty buffers. If another process is already 1106 * doing the flush, just wait for completion. 1107 */ 1108 int 1109 fuse_io_invalbuf(struct vnode *vp, struct thread *td) 1110 { 1111 struct fuse_vnode_data *fvdat = VTOFUD(vp); 1112 int error = 0; 1113 1114 if (vp->v_iflag & VI_DOOMED) 1115 return 0; 1116 1117 ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf"); 1118 1119 while (fvdat->flag & FN_FLUSHINPROG) { 1120 struct proc *p = td->td_proc; 1121 1122 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) 1123 return EIO; 1124 fvdat->flag |= FN_FLUSHWANT; 1125 tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz); 1126 error = 0; 1127 if (p != NULL) { 1128 PROC_LOCK(p); 1129 if (SIGNOTEMPTY(p->p_siglist) || 1130 SIGNOTEMPTY(td->td_siglist)) 1131 error = EINTR; 1132 PROC_UNLOCK(p); 1133 } 1134 if (error == EINTR) 1135 return EINTR; 1136 } 1137 fvdat->flag |= FN_FLUSHINPROG; 1138 1139 if (vp->v_bufobj.bo_object != NULL) { 1140 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); 1141 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 1142 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); 1143 } 1144 error = vinvalbuf(vp, V_SAVE, PCATCH, 0); 1145 while (error) { 1146 if (error == ERESTART || error == EINTR) { 1147 fvdat->flag &= ~FN_FLUSHINPROG; 1148 if (fvdat->flag & FN_FLUSHWANT) { 1149 fvdat->flag &= ~FN_FLUSHWANT; 1150 wakeup(&fvdat->flag); 1151 } 1152 return EINTR; 1153 } 1154 error = vinvalbuf(vp, V_SAVE, PCATCH, 0); 1155 } 1156 fvdat->flag &= ~FN_FLUSHINPROG; 1157 if (fvdat->flag & FN_FLUSHWANT) { 1158 fvdat->flag &= ~FN_FLUSHWANT; 1159 wakeup(&fvdat->flag); 1160 } 1161 return (error); 1162 } 1163