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 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 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 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 ssize_t r; 307 void *fwi_data; 308 off_t as_written_offset; 309 int diff; 310 int err = 0; 311 bool direct_io = fufh->fuse_open_flags & FOPEN_DIRECT_IO; 312 bool wrote_anything = false; 313 uint32_t write_flags; 314 315 data = fuse_get_mpdata(vp->v_mount); 316 317 /* 318 * Don't set FUSE_WRITE_LOCKOWNER in write_flags. It can't be set 319 * accurately when using POSIX AIO, libfuse doesn't use it, and I'm not 320 * aware of any file systems that do. It was an attempt to add 321 * Linux-style mandatory locking to the FUSE protocol, but mandatory 322 * locking is deprecated even on Linux. See Linux commit 323 * f33321141b273d60cbb3a8f56a5489baad82ba5e . 324 */ 325 /* 326 * Set FUSE_WRITE_CACHE whenever we don't know the uid, gid, and/or pid 327 * that originated a write. For example when writing from the 328 * writeback cache. I don't know of a single file system that cares, 329 * but the protocol says we're supposed to do this. 330 */ 331 write_flags = !pages && ( 332 (ioflag & IO_DIRECT) || 333 !fsess_opt_datacache(vnode_mount(vp)) || 334 !fsess_opt_writeback(vnode_mount(vp))) ? 0 : FUSE_WRITE_CACHE; 335 336 if (uio->uio_resid == 0) 337 return (0); 338 339 if (ioflag & IO_APPEND) 340 uio_setoffset(uio, filesize); 341 342 err = vn_rlimit_fsizex(vp, uio, 0, &r, uio->uio_td); 343 if (err != 0) { 344 vn_rlimit_fsizex_res(uio, r); 345 return (err); 346 } 347 348 fdisp_init(&fdi, 0); 349 350 while (uio->uio_resid > 0) { 351 size_t sizeof_fwi; 352 353 if (fuse_libabi_geq(data, 7, 9)) { 354 sizeof_fwi = sizeof(*fwi); 355 } else { 356 sizeof_fwi = FUSE_COMPAT_WRITE_IN_SIZE; 357 } 358 359 chunksize = MIN(uio->uio_resid, data->max_write); 360 361 fdi.iosize = sizeof_fwi + chunksize; 362 fdisp_make_vp(&fdi, FUSE_WRITE, vp, uio->uio_td, cred); 363 364 fwi = fdi.indata; 365 fwi->fh = fufh->fh_id; 366 fwi->offset = uio->uio_offset; 367 fwi->size = chunksize; 368 fwi->write_flags = write_flags; 369 if (fuse_libabi_geq(data, 7, 9)) { 370 fwi->flags = fufh_type_2_fflags(fufh->fufh_type); 371 } 372 fwi_data = (char *)fdi.indata + sizeof_fwi; 373 374 if ((err = uiomove(fwi_data, chunksize, uio))) 375 break; 376 377 retry: 378 err = fdisp_wait_answ(&fdi); 379 if (err == ERESTART || err == EINTR || err == EWOULDBLOCK) { 380 /* 381 * Rewind the uio so dofilewrite will know it's 382 * incomplete 383 */ 384 uio->uio_resid += fwi->size; 385 uio->uio_offset -= fwi->size; 386 /* 387 * Change ERESTART into EINTR because we can't rewind 388 * uio->uio_iov. Basically, once uiomove(9) has been 389 * called, it's impossible to restart a syscall. 390 */ 391 if (err == ERESTART) 392 err = EINTR; 393 break; 394 } else if (err) { 395 break; 396 } else { 397 wrote_anything = true; 398 } 399 400 fwo = ((struct fuse_write_out *)fdi.answ); 401 402 if (fwo->size > fwi->size) { 403 fuse_warn(data, FSESS_WARN_WROTE_LONG, 404 "wrote more data than we provided it."); 405 /* This is bonkers. Clear attr cache. */ 406 fvdat->flag &= ~FN_SIZECHANGE; 407 fuse_vnode_clear_attr_cache(vp); 408 err = EINVAL; 409 break; 410 } 411 412 /* Adjust the uio in the case of short writes */ 413 diff = fwi->size - fwo->size; 414 415 as_written_offset = uio->uio_offset - diff; 416 417 if (as_written_offset - diff > filesize) { 418 fuse_vnode_setsize(vp, as_written_offset, false); 419 getnanouptime(&fvdat->last_local_modify); 420 } 421 if (as_written_offset - diff >= filesize) 422 fvdat->flag &= ~FN_SIZECHANGE; 423 424 if (diff > 0) { 425 /* Short write */ 426 if (!direct_io) { 427 fuse_warn(data, FSESS_WARN_SHORT_WRITE, 428 "short writes are only allowed with " 429 "direct_io."); 430 } 431 if (ioflag & IO_DIRECT) { 432 /* Return early */ 433 uio->uio_resid += diff; 434 uio->uio_offset -= diff; 435 break; 436 } else { 437 /* Resend the unwritten portion of data */ 438 fdi.iosize = sizeof_fwi + diff; 439 /* Refresh fdi without clearing data buffer */ 440 fdisp_refresh_vp(&fdi, FUSE_WRITE, vp, 441 uio->uio_td, cred); 442 fwi = fdi.indata; 443 MPASS2(fwi == fdi.indata, "FUSE dispatcher " 444 "reallocated despite no increase in " 445 "size?"); 446 void *src = (char*)fwi_data + fwo->size; 447 memmove(fwi_data, src, diff); 448 fwi->fh = fufh->fh_id; 449 fwi->offset = as_written_offset; 450 fwi->size = diff; 451 fwi->write_flags = write_flags; 452 goto retry; 453 } 454 } 455 } 456 457 fdisp_destroy(&fdi); 458 459 if (wrote_anything) 460 fuse_vnode_undirty_cached_timestamps(vp, false); 461 462 vn_rlimit_fsizex_res(uio, r); 463 return (err); 464 } 465 466 SDT_PROBE_DEFINE6(fusefs, , io, write_biobackend_start, "int64_t", "int", "int", 467 "struct uio*", "int", "bool"); 468 SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_append_race, "long", "int"); 469 SDT_PROBE_DEFINE2(fusefs, , io, write_biobackend_issue, "int", "struct buf*"); 470 471 int 472 fuse_write_biobackend(struct vnode *vp, struct uio *uio, 473 struct ucred *cred, struct fuse_filehandle *fufh, int ioflag, pid_t pid) 474 { 475 struct fuse_vnode_data *fvdat = VTOFUD(vp); 476 struct buf *bp; 477 daddr_t lbn; 478 off_t filesize; 479 ssize_t r; 480 int bcount; 481 int n, on, seqcount, err = 0; 482 483 const int biosize = fuse_iosize(vp); 484 485 seqcount = ioflag >> IO_SEQSHIFT; 486 487 KASSERT(uio->uio_rw == UIO_WRITE, ("fuse_write_biobackend mode")); 488 if (vp->v_type != VREG) 489 return (EIO); 490 if (uio->uio_offset < 0) 491 return (EINVAL); 492 if (uio->uio_resid == 0) 493 return (0); 494 495 err = fuse_vnode_size(vp, &filesize, cred, curthread); 496 if (err) 497 return err; 498 499 if (ioflag & IO_APPEND) 500 uio_setoffset(uio, filesize); 501 502 err = vn_rlimit_fsizex(vp, uio, 0, &r, uio->uio_td); 503 if (err != 0) { 504 vn_rlimit_fsizex_res(uio, r); 505 return (err); 506 } 507 508 do { 509 bool direct_append, extending; 510 511 if (fuse_isdeadfs(vp)) { 512 err = ENXIO; 513 break; 514 } 515 lbn = uio->uio_offset / biosize; 516 on = uio->uio_offset & (biosize - 1); 517 n = MIN((unsigned)(biosize - on), uio->uio_resid); 518 519 again: 520 /* Get or create a buffer for the write */ 521 direct_append = uio->uio_offset == filesize && n; 522 if (uio->uio_offset + n < filesize) { 523 extending = false; 524 if ((off_t)(lbn + 1) * biosize < filesize) { 525 /* Not the file's last block */ 526 bcount = biosize; 527 } else { 528 /* The file's last block */ 529 bcount = filesize - (off_t)lbn * biosize; 530 } 531 } else { 532 extending = true; 533 bcount = on + n; 534 } 535 if (direct_append) { 536 /* 537 * Take care to preserve the buffer's B_CACHE state so 538 * as not to cause an unnecessary read. 539 */ 540 bp = getblk(vp, lbn, on, PCATCH, 0, 0); 541 if (bp != NULL) { 542 uint32_t save = bp->b_flags & B_CACHE; 543 allocbuf(bp, bcount); 544 bp->b_flags |= save; 545 } 546 } else { 547 bp = getblk(vp, lbn, bcount, PCATCH, 0, 0); 548 } 549 if (!bp) { 550 err = EINTR; 551 break; 552 } 553 if (extending) { 554 /* 555 * Extend file _after_ locking buffer so we won't race 556 * with other readers 557 */ 558 err = fuse_vnode_setsize(vp, uio->uio_offset + n, false); 559 filesize = uio->uio_offset + n; 560 getnanouptime(&fvdat->last_local_modify); 561 fvdat->flag |= FN_SIZECHANGE; 562 if (err) { 563 brelse(bp); 564 break; 565 } 566 } 567 568 SDT_PROBE6(fusefs, , io, write_biobackend_start, 569 lbn, on, n, uio, bcount, direct_append); 570 /* 571 * Issue a READ if B_CACHE is not set. In special-append 572 * mode, B_CACHE is based on the buffer prior to the write 573 * op and is typically set, avoiding the read. If a read 574 * is required in special append mode, the server will 575 * probably send us a short-read since we extended the file 576 * on our end, resulting in b_resid == 0 and, thusly, 577 * B_CACHE getting set. 578 * 579 * We can also avoid issuing the read if the write covers 580 * the entire buffer. We have to make sure the buffer state 581 * is reasonable in this case since we will not be initiating 582 * I/O. See the comments in kern/vfs_bio.c's getblk() for 583 * more information. 584 * 585 * B_CACHE may also be set due to the buffer being cached 586 * normally. 587 */ 588 589 if (on == 0 && n == bcount) { 590 bp->b_flags |= B_CACHE; 591 bp->b_flags &= ~B_INVAL; 592 bp->b_ioflags &= ~BIO_ERROR; 593 } 594 if ((bp->b_flags & B_CACHE) == 0) { 595 bp->b_iocmd = BIO_READ; 596 vfs_busy_pages(bp, 0); 597 fuse_io_strategy(vp, bp); 598 if ((err = bp->b_error)) { 599 brelse(bp); 600 break; 601 } 602 if (bp->b_resid > 0) { 603 /* 604 * Short read indicates EOF. Update file size 605 * from the server and try again. 606 */ 607 SDT_PROBE2(fusefs, , io, trace, 1, 608 "Short read during a RMW"); 609 brelse(bp); 610 err = fuse_vnode_size(vp, &filesize, cred, 611 curthread); 612 if (err) 613 break; 614 else 615 goto again; 616 } 617 } 618 if (bp->b_wcred == NOCRED) 619 bp->b_wcred = crhold(cred); 620 621 /* 622 * If dirtyend exceeds file size, chop it down. This should 623 * not normally occur but there is an append race where it 624 * might occur XXX, so we log it. 625 * 626 * If the chopping creates a reverse-indexed or degenerate 627 * situation with dirtyoff/end, we 0 both of them. 628 */ 629 if (bp->b_dirtyend > bcount) { 630 SDT_PROBE2(fusefs, , io, write_biobackend_append_race, 631 (long)bp->b_blkno * biosize, 632 bp->b_dirtyend - bcount); 633 bp->b_dirtyend = bcount; 634 } 635 if (bp->b_dirtyoff >= bp->b_dirtyend) 636 bp->b_dirtyoff = bp->b_dirtyend = 0; 637 638 /* 639 * If the new write will leave a contiguous dirty 640 * area, just update the b_dirtyoff and b_dirtyend, 641 * otherwise force a write rpc of the old dirty area. 642 * 643 * While it is possible to merge discontiguous writes due to 644 * our having a B_CACHE buffer ( and thus valid read data 645 * for the hole), we don't because it could lead to 646 * significant cache coherency problems with multiple clients, 647 * especially if locking is implemented later on. 648 * 649 * as an optimization we could theoretically maintain 650 * a linked list of discontinuous areas, but we would still 651 * have to commit them separately so there isn't much 652 * advantage to it except perhaps a bit of asynchronization. 653 */ 654 655 if (bp->b_dirtyend > 0 && 656 (on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) { 657 /* 658 * Yes, we mean it. Write out everything to "storage" 659 * immediately, without hesitation. (Apart from other 660 * reasons: the only way to know if a write is valid 661 * if its actually written out.) 662 */ 663 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 0, bp); 664 bwrite(bp); 665 if (bp->b_error == EINTR) { 666 err = EINTR; 667 break; 668 } 669 goto again; 670 } 671 err = uiomove((char *)bp->b_data + on, n, uio); 672 673 if (err) { 674 bp->b_ioflags |= BIO_ERROR; 675 bp->b_error = err; 676 brelse(bp); 677 break; 678 /* TODO: vfs_bio_clrbuf like ffs_write does? */ 679 } 680 /* 681 * Only update dirtyoff/dirtyend if not a degenerate 682 * condition. 683 */ 684 if (n) { 685 if (bp->b_dirtyend > 0) { 686 bp->b_dirtyoff = MIN(on, bp->b_dirtyoff); 687 bp->b_dirtyend = MAX((on + n), bp->b_dirtyend); 688 } else { 689 bp->b_dirtyoff = on; 690 bp->b_dirtyend = on + n; 691 } 692 vfs_bio_set_valid(bp, on, n); 693 } 694 695 vfs_bio_set_flags(bp, ioflag); 696 697 bp->b_flags |= B_FUSEFS_WRITE_CACHE; 698 if (ioflag & IO_SYNC) { 699 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 2, bp); 700 if (!(ioflag & IO_VMIO)) 701 bp->b_flags &= ~B_FUSEFS_WRITE_CACHE; 702 err = bwrite(bp); 703 } else if (vm_page_count_severe() || 704 buf_dirty_count_severe() || 705 (ioflag & IO_ASYNC)) { 706 bp->b_flags |= B_CLUSTEROK; 707 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 3, bp); 708 bawrite(bp); 709 } else if (on == 0 && n == bcount) { 710 if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERW) == 0) { 711 bp->b_flags |= B_CLUSTEROK; 712 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 713 4, bp); 714 cluster_write(vp, &fvdat->clusterw, bp, 715 filesize, seqcount, 0); 716 } else { 717 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 718 5, bp); 719 bawrite(bp); 720 } 721 } else if (ioflag & IO_DIRECT) { 722 bp->b_flags |= B_CLUSTEROK; 723 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 6, bp); 724 bawrite(bp); 725 } else { 726 bp->b_flags &= ~B_CLUSTEROK; 727 SDT_PROBE2(fusefs, , io, write_biobackend_issue, 7, bp); 728 bdwrite(bp); 729 } 730 if (err) 731 break; 732 } while (uio->uio_resid > 0 && n > 0); 733 734 vn_rlimit_fsizex_res(uio, r); 735 return (err); 736 } 737 738 int 739 fuse_io_strategy(struct vnode *vp, struct buf *bp) 740 { 741 struct fuse_vnode_data *fvdat = VTOFUD(vp); 742 struct fuse_filehandle *fufh; 743 struct ucred *cred; 744 struct uio *uiop; 745 struct uio uio; 746 struct iovec io; 747 off_t filesize; 748 int error = 0; 749 int fflag; 750 /* We don't know the true pid when we're dealing with the cache */ 751 pid_t pid = 0; 752 753 const int biosize = fuse_iosize(vp); 754 755 MPASS(vp->v_type == VREG || vp->v_type == VDIR); 756 MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE); 757 758 fflag = bp->b_iocmd == BIO_READ ? FREAD : FWRITE; 759 cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred; 760 error = fuse_filehandle_getrw(vp, fflag, &fufh, cred, pid); 761 if (bp->b_iocmd == BIO_READ && error == EBADF) { 762 /* 763 * This may be a read-modify-write operation on a cached file 764 * opened O_WRONLY. The FUSE protocol allows this. 765 */ 766 error = fuse_filehandle_get(vp, FWRITE, &fufh, cred, pid); 767 } 768 if (error) { 769 printf("FUSE: strategy: filehandles are closed\n"); 770 bp->b_ioflags |= BIO_ERROR; 771 bp->b_error = error; 772 bufdone(bp); 773 return (error); 774 } 775 776 uiop = &uio; 777 uiop->uio_iov = &io; 778 uiop->uio_iovcnt = 1; 779 uiop->uio_segflg = UIO_SYSSPACE; 780 uiop->uio_td = curthread; 781 782 /* 783 * clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We 784 * do this here so we do not have to do it in all the code that 785 * calls us. 786 */ 787 bp->b_flags &= ~B_INVAL; 788 bp->b_ioflags &= ~BIO_ERROR; 789 790 KASSERT(!(bp->b_flags & B_DONE), 791 ("fuse_io_strategy: bp %p already marked done", bp)); 792 if (bp->b_iocmd == BIO_READ) { 793 ssize_t left; 794 795 io.iov_len = uiop->uio_resid = bp->b_bcount; 796 io.iov_base = bp->b_data; 797 uiop->uio_rw = UIO_READ; 798 799 uiop->uio_offset = ((off_t)bp->b_lblkno) * biosize; 800 error = fuse_read_directbackend(vp, uiop, cred, fufh); 801 /* 802 * Store the amount we failed to read in the buffer's private 803 * field, so callers can truncate the file if necessary' 804 */ 805 806 if (!error && uiop->uio_resid) { 807 int nread = bp->b_bcount - uiop->uio_resid; 808 left = uiop->uio_resid; 809 bzero((char *)bp->b_data + nread, left); 810 811 if ((fvdat->flag & FN_SIZECHANGE) == 0) { 812 /* 813 * A short read with no error, when not using 814 * direct io, and when no writes are cached, 815 * indicates EOF caused by a server-side 816 * truncation. Clear the attr cache so we'll 817 * pick up the new file size and timestamps. 818 * 819 * We must still bzero the remaining buffer so 820 * uninitialized data doesn't get exposed by a 821 * future truncate that extends the file. 822 * 823 * To prevent lock order problems, we must 824 * truncate the file upstack, not here. 825 */ 826 SDT_PROBE2(fusefs, , io, trace, 1, 827 "Short read of a clean file"); 828 fuse_vnode_clear_attr_cache(vp); 829 } else { 830 /* 831 * If dirty writes _are_ cached beyond EOF, 832 * that indicates a newly created hole that the 833 * server doesn't know about. Those don't pose 834 * any problem. 835 * XXX: we don't currently track whether dirty 836 * writes are cached beyond EOF, before EOF, or 837 * both. 838 */ 839 SDT_PROBE2(fusefs, , io, trace, 1, 840 "Short read of a dirty file"); 841 uiop->uio_resid = 0; 842 } 843 } 844 if (error) { 845 bp->b_ioflags |= BIO_ERROR; 846 bp->b_error = error; 847 } 848 } else { 849 /* 850 * Setup for actual write 851 */ 852 /* 853 * If the file's size is cached, use that value, even if the 854 * cache is expired. At this point we're already committed to 855 * writing something. If the FUSE server has changed the 856 * file's size behind our back, it's too late for us to do 857 * anything about it. In particular, we can't invalidate any 858 * part of the file's buffers because VOP_STRATEGY is called 859 * with them already locked. 860 */ 861 filesize = fvdat->cached_attrs.va_size; 862 /* filesize must've been cached by fuse_vnop_open. */ 863 KASSERT(filesize != VNOVAL, ("filesize should've been cached")); 864 865 if ((off_t)bp->b_lblkno * biosize + bp->b_dirtyend > filesize) 866 bp->b_dirtyend = filesize - 867 (off_t)bp->b_lblkno * biosize; 868 869 if (bp->b_dirtyend > bp->b_dirtyoff) { 870 io.iov_len = uiop->uio_resid = bp->b_dirtyend 871 - bp->b_dirtyoff; 872 uiop->uio_offset = (off_t)bp->b_lblkno * biosize 873 + bp->b_dirtyoff; 874 io.iov_base = (char *)bp->b_data + bp->b_dirtyoff; 875 uiop->uio_rw = UIO_WRITE; 876 877 bool pages = bp->b_flags & B_FUSEFS_WRITE_CACHE; 878 error = fuse_write_directbackend(vp, uiop, cred, fufh, 879 filesize, 0, pages); 880 881 if (error == EINTR || error == ETIMEDOUT) { 882 bp->b_flags &= ~(B_INVAL | B_NOCACHE); 883 if ((bp->b_flags & B_PAGING) == 0) { 884 bdirty(bp); 885 bp->b_flags &= ~B_DONE; 886 } 887 if ((error == EINTR || error == ETIMEDOUT) && 888 (bp->b_flags & B_ASYNC) == 0) 889 bp->b_flags |= B_EINTR; 890 } else { 891 if (error) { 892 bp->b_ioflags |= BIO_ERROR; 893 bp->b_flags |= B_INVAL; 894 bp->b_error = error; 895 } 896 bp->b_dirtyoff = bp->b_dirtyend = 0; 897 } 898 } else { 899 bp->b_resid = 0; 900 bufdone(bp); 901 return (0); 902 } 903 } 904 bp->b_resid = uiop->uio_resid; 905 bufdone(bp); 906 return (error); 907 } 908 909 int 910 fuse_io_flushbuf(struct vnode *vp, int waitfor, struct thread *td) 911 { 912 913 return (vn_fsync_buf(vp, waitfor)); 914 } 915 916 /* 917 * Flush and invalidate all dirty buffers. If another process is already 918 * doing the flush, just wait for completion. 919 */ 920 int 921 fuse_io_invalbuf(struct vnode *vp, struct thread *td) 922 { 923 struct fuse_vnode_data *fvdat = VTOFUD(vp); 924 int error = 0; 925 926 if (VN_IS_DOOMED(vp)) 927 return 0; 928 929 ASSERT_VOP_ELOCKED(vp, "fuse_io_invalbuf"); 930 931 while (fvdat->flag & FN_FLUSHINPROG) { 932 struct proc *p = td->td_proc; 933 934 if (vp->v_mount->mnt_kern_flag & MNTK_UNMOUNTF) 935 return EIO; 936 fvdat->flag |= FN_FLUSHWANT; 937 tsleep(&fvdat->flag, PRIBIO + 2, "fusevinv", 2 * hz); 938 error = 0; 939 if (p != NULL) { 940 PROC_LOCK(p); 941 if (SIGNOTEMPTY(p->p_siglist) || 942 SIGNOTEMPTY(td->td_siglist)) 943 error = EINTR; 944 PROC_UNLOCK(p); 945 } 946 if (error == EINTR) 947 return EINTR; 948 } 949 fvdat->flag |= FN_FLUSHINPROG; 950 951 if (vp->v_bufobj.bo_object != NULL) { 952 VM_OBJECT_WLOCK(vp->v_bufobj.bo_object); 953 vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC); 954 VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object); 955 } 956 error = vinvalbuf(vp, V_SAVE, PCATCH, 0); 957 while (error) { 958 if (error == ERESTART || error == EINTR) { 959 fvdat->flag &= ~FN_FLUSHINPROG; 960 if (fvdat->flag & FN_FLUSHWANT) { 961 fvdat->flag &= ~FN_FLUSHWANT; 962 wakeup(&fvdat->flag); 963 } 964 return EINTR; 965 } 966 error = vinvalbuf(vp, V_SAVE, PCATCH, 0); 967 } 968 fvdat->flag &= ~FN_FLUSHINPROG; 969 if (fvdat->flag & FN_FLUSHWANT) { 970 fvdat->flag &= ~FN_FLUSHWANT; 971 wakeup(&fvdat->flag); 972 } 973 return (error); 974 } 975