1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1989, 1993 5 * The Regents of the University of California. All rights reserved. 6 * (c) UNIX System Laboratories, Inc. 7 * All or some portions of this file are derived from material licensed 8 * to the University of California by American Telephone and Telegraph 9 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 10 * the permission of UNIX System Laboratories, Inc. 11 * 12 * Copyright (c) 2012 Konstantin Belousov <kib@FreeBSD.org> 13 * Copyright (c) 2013, 2014 The FreeBSD Foundation 14 * 15 * Portions of this software were developed by Konstantin Belousov 16 * under sponsorship from the FreeBSD Foundation. 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions 20 * are met: 21 * 1. Redistributions of source code must retain the above copyright 22 * notice, this list of conditions and the following disclaimer. 23 * 2. Redistributions in binary form must reproduce the above copyright 24 * notice, this list of conditions and the following disclaimer in the 25 * documentation and/or other materials provided with the distribution. 26 * 3. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 */ 42 43 #include "opt_hwpmc_hooks.h" 44 45 #include <sys/param.h> 46 #include <sys/systm.h> 47 #include <sys/buf.h> 48 #include <sys/disk.h> 49 #include <sys/dirent.h> 50 #include <sys/fail.h> 51 #include <sys/fcntl.h> 52 #include <sys/file.h> 53 #include <sys/filio.h> 54 #include <sys/ktr.h> 55 #include <sys/ktrace.h> 56 #include <sys/limits.h> 57 #include <sys/lock.h> 58 #include <sys/mman.h> 59 #include <sys/mount.h> 60 #include <sys/mutex.h> 61 #include <sys/namei.h> 62 #include <sys/priv.h> 63 #include <sys/prng.h> 64 #include <sys/proc.h> 65 #include <sys/rwlock.h> 66 #include <sys/sleepqueue.h> 67 #include <sys/stat.h> 68 #include <sys/sysctl.h> 69 #include <sys/unistd.h> 70 #include <sys/user.h> 71 #include <sys/vnode.h> 72 73 #include <security/audit/audit.h> 74 #include <security/mac/mac_framework.h> 75 76 #include <vm/vm.h> 77 #include <vm/vm_extern.h> 78 #include <vm/pmap.h> 79 #include <vm/vm_map.h> 80 #include <vm/vm_object.h> 81 #include <vm/vm_page.h> 82 #include <vm/vm_pager.h> 83 #include <vm/vnode_pager.h> 84 85 #ifdef HWPMC_HOOKS 86 #include <sys/pmckern.h> 87 #endif 88 89 static fo_rdwr_t vn_read; 90 static fo_rdwr_t vn_write; 91 static fo_rdwr_t vn_io_fault; 92 static fo_truncate_t vn_truncate; 93 static fo_ioctl_t vn_ioctl; 94 static fo_poll_t vn_poll; 95 static fo_kqfilter_t vn_kqfilter; 96 static fo_close_t vn_closefile; 97 static fo_mmap_t vn_mmap; 98 static fo_fallocate_t vn_fallocate; 99 static fo_fspacectl_t vn_fspacectl; 100 101 const struct fileops vnops = { 102 .fo_read = vn_io_fault, 103 .fo_write = vn_io_fault, 104 .fo_truncate = vn_truncate, 105 .fo_ioctl = vn_ioctl, 106 .fo_poll = vn_poll, 107 .fo_kqfilter = vn_kqfilter, 108 .fo_stat = vn_statfile, 109 .fo_close = vn_closefile, 110 .fo_chmod = vn_chmod, 111 .fo_chown = vn_chown, 112 .fo_sendfile = vn_sendfile, 113 .fo_seek = vn_seek, 114 .fo_fill_kinfo = vn_fill_kinfo, 115 .fo_mmap = vn_mmap, 116 .fo_fallocate = vn_fallocate, 117 .fo_fspacectl = vn_fspacectl, 118 .fo_cmp = vn_cmp, 119 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 120 }; 121 122 const u_int io_hold_cnt = 16; 123 static int vn_io_fault_enable = 1; 124 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN, 125 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); 126 static int vn_io_fault_prefault = 0; 127 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN, 128 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); 129 static int vn_io_pgcache_read_enable = 1; 130 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN, 131 &vn_io_pgcache_read_enable, 0, 132 "Enable copying from page cache for reads, avoiding fs"); 133 static u_long vn_io_faults_cnt; 134 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, 135 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); 136 137 static int vfs_allow_read_dir = 0; 138 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW, 139 &vfs_allow_read_dir, 0, 140 "Enable read(2) of directory by root for filesystems that support it"); 141 142 /* 143 * Returns true if vn_io_fault mode of handling the i/o request should 144 * be used. 145 */ 146 static bool 147 do_vn_io_fault(struct vnode *vp, struct uio *uio) 148 { 149 struct mount *mp; 150 151 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && 152 (mp = vp->v_mount) != NULL && 153 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); 154 } 155 156 /* 157 * Structure used to pass arguments to vn_io_fault1(), to do either 158 * file- or vnode-based I/O calls. 159 */ 160 struct vn_io_fault_args { 161 enum { 162 VN_IO_FAULT_FOP, 163 VN_IO_FAULT_VOP 164 } kind; 165 struct ucred *cred; 166 int flags; 167 union { 168 struct fop_args_tag { 169 struct file *fp; 170 fo_rdwr_t *doio; 171 } fop_args; 172 struct vop_args_tag { 173 struct vnode *vp; 174 } vop_args; 175 } args; 176 }; 177 178 static int vn_io_fault1(struct vnode *vp, struct uio *uio, 179 struct vn_io_fault_args *args, struct thread *td); 180 181 int 182 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) 183 { 184 struct thread *td = curthread; 185 186 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); 187 } 188 189 static uint64_t 190 open2nameif(int fmode, u_int vn_open_flags, uint64_t cn_flags) 191 { 192 uint64_t res; 193 194 res = ISOPEN | LOCKLEAF | cn_flags; 195 if ((fmode & O_RESOLVE_BENEATH) != 0) 196 res |= RBENEATH; 197 if ((fmode & O_EMPTY_PATH) != 0) 198 res |= EMPTYPATH; 199 if ((fmode & FREAD) != 0) 200 res |= OPENREAD; 201 if ((fmode & FWRITE) != 0) 202 res |= OPENWRITE; 203 if ((fmode & O_NAMEDATTR) != 0) 204 res |= OPENNAMED | CREATENAMED; 205 if ((fmode & O_NOFOLLOW) != 0) 206 res &= ~FOLLOW; 207 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0) 208 res |= AUDITVNODE1; 209 else 210 res &= ~AUDITVNODE1; 211 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0) 212 res |= NOCAPCHECK; 213 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0) 214 res |= WANTIOCTLCAPS; 215 216 return (res); 217 } 218 219 /* 220 * For the O_NAMEDATTR case, check for a valid use of it. 221 */ 222 static int 223 vfs_check_namedattr(struct vnode *vp) 224 { 225 int error; 226 short irflag; 227 228 error = 0; 229 irflag = vn_irflag_read(vp); 230 if ((vp->v_mount->mnt_flag & MNT_NAMEDATTR) == 0 || 231 ((irflag & VIRF_NAMEDATTR) != 0 && vp->v_type != VREG)) 232 error = EINVAL; 233 else if ((irflag & (VIRF_NAMEDDIR | VIRF_NAMEDATTR)) == 0) 234 error = ENOATTR; 235 return (error); 236 } 237 238 /* 239 * Common code for vnode open operations via a name lookup. 240 * Lookup the vnode and invoke VOP_CREATE if needed. 241 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. 242 * 243 * Note that this does NOT free nameidata for the successful case, 244 * due to the NDINIT being done elsewhere. 245 */ 246 int 247 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, 248 struct ucred *cred, struct file *fp) 249 { 250 struct vnode *vp; 251 struct mount *mp; 252 struct vattr vat; 253 struct vattr *vap = &vat; 254 int fmode, error; 255 bool first_open; 256 257 restart: 258 first_open = false; 259 fmode = *flagp; 260 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | 261 O_EXCL | O_DIRECTORY) || 262 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH)) 263 return (EINVAL); 264 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { 265 ndp->ni_cnd.cn_nameiop = CREATE; 266 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags, 267 ndp->ni_cnd.cn_flags); 268 269 /* 270 * Set NOCACHE to avoid flushing the cache when 271 * rolling in many files at once. 272 * 273 * Set NC_KEEPPOSENTRY to keep positive entries if they already 274 * exist despite NOCACHE. 275 */ 276 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY; 277 if ((fmode & O_EXCL) != 0) 278 ndp->ni_cnd.cn_flags &= ~FOLLOW; 279 if ((vn_open_flags & VN_OPEN_INVFS) == 0) 280 bwillwrite(); 281 if ((error = namei(ndp)) != 0) 282 return (error); 283 if (ndp->ni_vp == NULL) { 284 if ((fmode & O_NAMEDATTR) != 0 && 285 (ndp->ni_dvp->v_mount->mnt_flag & MNT_NAMEDATTR) == 286 0) { 287 error = EINVAL; 288 vp = ndp->ni_dvp; 289 ndp->ni_dvp = NULL; 290 goto bad; 291 } 292 VATTR_NULL(vap); 293 vap->va_type = VREG; 294 vap->va_mode = cmode; 295 if (fmode & O_EXCL) 296 vap->va_vaflags |= VA_EXCLUSIVE; 297 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { 298 NDFREE_PNBUF(ndp); 299 vput(ndp->ni_dvp); 300 if ((error = vn_start_write(NULL, &mp, 301 V_XSLEEP | V_PCATCH)) != 0) 302 return (error); 303 NDREINIT(ndp); 304 goto restart; 305 } 306 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) 307 ndp->ni_cnd.cn_flags |= MAKEENTRY; 308 #ifdef MAC 309 error = mac_vnode_check_create(cred, ndp->ni_dvp, 310 &ndp->ni_cnd, vap); 311 if (error == 0) 312 #endif 313 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, 314 &ndp->ni_cnd, vap); 315 vp = ndp->ni_vp; 316 if (error == 0 && (fmode & O_EXCL) != 0 && 317 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) { 318 VI_LOCK(vp); 319 vp->v_iflag |= VI_FOPENING; 320 VI_UNLOCK(vp); 321 first_open = true; 322 } 323 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL, 324 false); 325 vn_finished_write(mp); 326 if (error) { 327 NDFREE_PNBUF(ndp); 328 if (error == ERELOOKUP) { 329 NDREINIT(ndp); 330 goto restart; 331 } 332 return (error); 333 } 334 fmode &= ~O_TRUNC; 335 } else { 336 if (ndp->ni_dvp == ndp->ni_vp) 337 vrele(ndp->ni_dvp); 338 else 339 vput(ndp->ni_dvp); 340 ndp->ni_dvp = NULL; 341 vp = ndp->ni_vp; 342 if (fmode & O_EXCL) { 343 error = EEXIST; 344 goto bad; 345 } 346 if ((fmode & O_NAMEDATTR) != 0) { 347 error = vfs_check_namedattr(vp); 348 if (error != 0) 349 goto bad; 350 } else if (vp->v_type == VDIR) { 351 error = EISDIR; 352 goto bad; 353 } 354 fmode &= ~O_CREAT; 355 } 356 } else { 357 ndp->ni_cnd.cn_nameiop = LOOKUP; 358 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags, 359 ndp->ni_cnd.cn_flags); 360 if ((fmode & FWRITE) == 0) 361 ndp->ni_cnd.cn_flags |= LOCKSHARED; 362 if ((error = namei(ndp)) != 0) 363 return (error); 364 vp = ndp->ni_vp; 365 if ((fmode & O_NAMEDATTR) != 0) { 366 error = vfs_check_namedattr(vp); 367 if (error != 0) 368 goto bad; 369 } 370 } 371 error = vn_open_vnode(vp, fmode, cred, curthread, fp); 372 if (first_open) { 373 VI_LOCK(vp); 374 vp->v_iflag &= ~VI_FOPENING; 375 wakeup(vp); 376 VI_UNLOCK(vp); 377 } 378 if (error) 379 goto bad; 380 *flagp = fmode; 381 return (0); 382 bad: 383 NDFREE_PNBUF(ndp); 384 vput(vp); 385 *flagp = fmode; 386 ndp->ni_vp = NULL; 387 return (error); 388 } 389 390 static int 391 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) 392 { 393 struct flock lf; 394 int error, lock_flags, type; 395 396 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); 397 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) 398 return (0); 399 KASSERT(fp != NULL, ("open with flock requires fp")); 400 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) 401 return (EOPNOTSUPP); 402 403 lock_flags = VOP_ISLOCKED(vp); 404 VOP_UNLOCK(vp); 405 406 lf.l_whence = SEEK_SET; 407 lf.l_start = 0; 408 lf.l_len = 0; 409 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; 410 type = F_FLOCK; 411 if ((fmode & FNONBLOCK) == 0) 412 type |= F_WAIT; 413 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) 414 type |= F_FIRSTOPEN; 415 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); 416 if (error == 0) 417 fp->f_flag |= FHASLOCK; 418 419 vn_lock(vp, lock_flags | LK_RETRY); 420 return (error); 421 } 422 423 /* 424 * Common code for vnode open operations once a vnode is located. 425 * Check permissions, and call the VOP_OPEN routine. 426 */ 427 int 428 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, 429 struct thread *td, struct file *fp) 430 { 431 accmode_t accmode; 432 int error; 433 434 KASSERT((fmode & O_PATH) == 0 || (fmode & O_ACCMODE) == 0, 435 ("%s: O_PATH and O_ACCMODE are mutually exclusive", __func__)); 436 437 if (vp->v_type == VLNK) { 438 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0) 439 return (EMLINK); 440 } 441 if (vp->v_type != VDIR && fmode & O_DIRECTORY) 442 return (ENOTDIR); 443 444 accmode = 0; 445 if ((fmode & O_PATH) == 0) { 446 if (vp->v_type == VSOCK) 447 return (EOPNOTSUPP); 448 if ((fmode & (FWRITE | O_TRUNC)) != 0) { 449 if (vp->v_type == VDIR) 450 return (EISDIR); 451 accmode |= VWRITE; 452 } 453 if ((fmode & FREAD) != 0) 454 accmode |= VREAD; 455 if ((fmode & O_APPEND) && (fmode & FWRITE)) 456 accmode |= VAPPEND; 457 #ifdef MAC 458 if ((fmode & O_CREAT) != 0) 459 accmode |= VCREAT; 460 #endif 461 } 462 if ((fmode & FEXEC) != 0) 463 accmode |= VEXEC; 464 #ifdef MAC 465 if ((fmode & O_VERIFY) != 0) 466 accmode |= VVERIFY; 467 error = mac_vnode_check_open(cred, vp, accmode); 468 if (error != 0) 469 return (error); 470 471 accmode &= ~(VCREAT | VVERIFY); 472 #endif 473 if ((fmode & O_CREAT) == 0 && accmode != 0) { 474 error = VOP_ACCESS(vp, accmode, cred, td); 475 if (error != 0) 476 return (error); 477 } 478 if ((fmode & O_PATH) != 0) { 479 if (vp->v_type != VFIFO && vp->v_type != VSOCK && 480 VOP_ACCESS(vp, VREAD, cred, td) == 0) 481 fp->f_flag |= FKQALLOWED; 482 return (0); 483 } 484 485 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 486 vn_lock(vp, LK_UPGRADE | LK_RETRY); 487 error = VOP_OPEN(vp, fmode, cred, td, fp); 488 if (error != 0) 489 return (error); 490 491 error = vn_open_vnode_advlock(vp, fmode, fp); 492 if (error == 0 && (fmode & FWRITE) != 0) { 493 error = VOP_ADD_WRITECOUNT(vp, 1); 494 if (error == 0) { 495 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 496 __func__, vp, vp->v_writecount); 497 } 498 } 499 500 /* 501 * Error from advlock or VOP_ADD_WRITECOUNT() still requires 502 * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). 503 */ 504 if (error != 0) { 505 if (fp != NULL) { 506 /* 507 * Arrange the call by having fdrop() to use 508 * vn_closefile(). This is to satisfy 509 * filesystems like devfs or tmpfs, which 510 * override fo_close(). 511 */ 512 fp->f_flag |= FOPENFAILED; 513 fp->f_vnode = vp; 514 if (fp->f_ops == &badfileops) { 515 fp->f_type = DTYPE_VNODE; 516 fp->f_ops = &vnops; 517 } 518 vref(vp); 519 } else { 520 /* 521 * If there is no fp, due to kernel-mode open, 522 * we can call VOP_CLOSE() now. 523 */ 524 if ((vp->v_type == VFIFO || 525 !MNT_EXTENDED_SHARED(vp->v_mount)) && 526 VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 527 vn_lock(vp, LK_UPGRADE | LK_RETRY); 528 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC), 529 cred, td); 530 } 531 } 532 533 ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); 534 return (error); 535 536 } 537 538 /* 539 * Check for write permissions on the specified vnode. 540 * Prototype text segments cannot be written. 541 * It is racy. 542 */ 543 int 544 vn_writechk(struct vnode *vp) 545 { 546 547 ASSERT_VOP_LOCKED(vp, "vn_writechk"); 548 /* 549 * If there's shared text associated with 550 * the vnode, try to free it up once. If 551 * we fail, we can't allow writing. 552 */ 553 if (VOP_IS_TEXT(vp)) 554 return (ETXTBSY); 555 556 return (0); 557 } 558 559 /* 560 * Vnode close call 561 */ 562 static int 563 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, 564 struct thread *td, bool keep_ref) 565 { 566 struct mount *mp; 567 int error, lock_flags; 568 569 lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ? 570 LK_SHARED : LK_EXCLUSIVE; 571 572 vn_start_write(vp, &mp, V_WAIT); 573 vn_lock(vp, lock_flags | LK_RETRY); 574 AUDIT_ARG_VNODE1(vp); 575 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { 576 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 577 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 578 __func__, vp, vp->v_writecount); 579 } 580 error = VOP_CLOSE(vp, flags, file_cred, td); 581 if (keep_ref) 582 VOP_UNLOCK(vp); 583 else 584 vput(vp); 585 vn_finished_write(mp); 586 return (error); 587 } 588 589 int 590 vn_close(struct vnode *vp, int flags, struct ucred *file_cred, 591 struct thread *td) 592 { 593 594 return (vn_close1(vp, flags, file_cred, td, false)); 595 } 596 597 /* 598 * Heuristic to detect sequential operation. 599 */ 600 static int 601 sequential_heuristic(struct uio *uio, struct file *fp) 602 { 603 enum uio_rw rw; 604 605 ASSERT_VOP_LOCKED(fp->f_vnode, __func__); 606 607 rw = uio->uio_rw; 608 if (fp->f_flag & FRDAHEAD) 609 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 610 611 /* 612 * Offset 0 is handled specially. open() sets f_seqcount to 1 so 613 * that the first I/O is normally considered to be slightly 614 * sequential. Seeking to offset 0 doesn't change sequentiality 615 * unless previous seeks have reduced f_seqcount to 0, in which 616 * case offset 0 is not special. 617 */ 618 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) || 619 uio->uio_offset == fp->f_nextoff[rw]) { 620 /* 621 * f_seqcount is in units of fixed-size blocks so that it 622 * depends mainly on the amount of sequential I/O and not 623 * much on the number of sequential I/O's. The fixed size 624 * of 16384 is hard-coded here since it is (not quite) just 625 * a magic size that works well here. This size is more 626 * closely related to the best I/O size for real disks than 627 * to any block size used by software. 628 */ 629 if (uio->uio_resid >= IO_SEQMAX * 16384) 630 fp->f_seqcount[rw] = IO_SEQMAX; 631 else { 632 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384); 633 if (fp->f_seqcount[rw] > IO_SEQMAX) 634 fp->f_seqcount[rw] = IO_SEQMAX; 635 } 636 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 637 } 638 639 /* Not sequential. Quickly draw-down sequentiality. */ 640 if (fp->f_seqcount[rw] > 1) 641 fp->f_seqcount[rw] = 1; 642 else 643 fp->f_seqcount[rw] = 0; 644 return (0); 645 } 646 647 /* 648 * Package up an I/O request on a vnode into a uio and do it. 649 */ 650 int 651 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, 652 enum uio_seg segflg, int ioflg, struct ucred *active_cred, 653 struct ucred *file_cred, ssize_t *aresid, struct thread *td) 654 { 655 struct uio auio; 656 struct iovec aiov; 657 struct mount *mp; 658 struct ucred *cred; 659 void *rl_cookie; 660 struct vn_io_fault_args args; 661 int error, lock_flags; 662 663 if (offset < 0 && vp->v_type != VCHR) 664 return (EINVAL); 665 auio.uio_iov = &aiov; 666 auio.uio_iovcnt = 1; 667 aiov.iov_base = base; 668 aiov.iov_len = len; 669 auio.uio_resid = len; 670 auio.uio_offset = offset; 671 auio.uio_segflg = segflg; 672 auio.uio_rw = rw; 673 auio.uio_td = td; 674 error = 0; 675 676 if ((ioflg & IO_NODELOCKED) == 0) { 677 if ((ioflg & IO_RANGELOCKED) == 0) { 678 if (rw == UIO_READ) { 679 rl_cookie = vn_rangelock_rlock(vp, offset, 680 offset + len); 681 } else if ((ioflg & IO_APPEND) != 0) { 682 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 683 } else { 684 rl_cookie = vn_rangelock_wlock(vp, offset, 685 offset + len); 686 } 687 } else 688 rl_cookie = NULL; 689 mp = NULL; 690 if (rw == UIO_WRITE) { 691 if (vp->v_type != VCHR && 692 (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH)) 693 != 0) 694 goto out; 695 lock_flags = vn_lktype_write(mp, vp); 696 } else 697 lock_flags = LK_SHARED; 698 vn_lock(vp, lock_flags | LK_RETRY); 699 } else 700 rl_cookie = NULL; 701 702 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 703 #ifdef MAC 704 if ((ioflg & IO_NOMACCHECK) == 0) { 705 if (rw == UIO_READ) 706 error = mac_vnode_check_read(active_cred, file_cred, 707 vp); 708 else 709 error = mac_vnode_check_write(active_cred, file_cred, 710 vp); 711 } 712 #endif 713 if (error == 0) { 714 if (file_cred != NULL) 715 cred = file_cred; 716 else 717 cred = active_cred; 718 if (do_vn_io_fault(vp, &auio)) { 719 args.kind = VN_IO_FAULT_VOP; 720 args.cred = cred; 721 args.flags = ioflg; 722 args.args.vop_args.vp = vp; 723 error = vn_io_fault1(vp, &auio, &args, td); 724 } else if (rw == UIO_READ) { 725 error = VOP_READ(vp, &auio, ioflg, cred); 726 } else /* if (rw == UIO_WRITE) */ { 727 error = VOP_WRITE(vp, &auio, ioflg, cred); 728 } 729 } 730 if (aresid) 731 *aresid = auio.uio_resid; 732 else 733 if (auio.uio_resid && error == 0) 734 error = EIO; 735 if ((ioflg & IO_NODELOCKED) == 0) { 736 VOP_UNLOCK(vp); 737 if (mp != NULL) 738 vn_finished_write(mp); 739 } 740 out: 741 if (rl_cookie != NULL) 742 vn_rangelock_unlock(vp, rl_cookie); 743 return (error); 744 } 745 746 /* 747 * Package up an I/O request on a vnode into a uio and do it. The I/O 748 * request is split up into smaller chunks and we try to avoid saturating 749 * the buffer cache while potentially holding a vnode locked, so we 750 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() 751 * to give other processes a chance to lock the vnode (either other processes 752 * core'ing the same binary, or unrelated processes scanning the directory). 753 */ 754 int 755 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, 756 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, 757 struct ucred *file_cred, size_t *aresid, struct thread *td) 758 { 759 int error = 0; 760 ssize_t iaresid; 761 762 do { 763 int chunk; 764 765 /* 766 * Force `offset' to a multiple of MAXBSIZE except possibly 767 * for the first chunk, so that filesystems only need to 768 * write full blocks except possibly for the first and last 769 * chunks. 770 */ 771 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; 772 773 if (chunk > len) 774 chunk = len; 775 if (rw != UIO_READ && vp->v_type == VREG) 776 bwillwrite(); 777 iaresid = 0; 778 error = vn_rdwr(rw, vp, base, chunk, offset, segflg, 779 ioflg, active_cred, file_cred, &iaresid, td); 780 len -= chunk; /* aresid calc already includes length */ 781 if (error) 782 break; 783 offset += chunk; 784 base = (char *)base + chunk; 785 kern_yield(PRI_USER); 786 } while (len); 787 if (aresid) 788 *aresid = len + iaresid; 789 return (error); 790 } 791 792 #if OFF_MAX <= LONG_MAX 793 off_t 794 foffset_lock(struct file *fp, int flags) 795 { 796 volatile short *flagsp; 797 off_t res; 798 short state; 799 800 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 801 802 if ((flags & FOF_NOLOCK) != 0) 803 return (atomic_load_long(&fp->f_offset)); 804 805 /* 806 * According to McKusick the vn lock was protecting f_offset here. 807 * It is now protected by the FOFFSET_LOCKED flag. 808 */ 809 flagsp = &fp->f_vnread_flags; 810 if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED)) 811 return (atomic_load_long(&fp->f_offset)); 812 813 sleepq_lock(&fp->f_vnread_flags); 814 state = atomic_load_16(flagsp); 815 for (;;) { 816 if ((state & FOFFSET_LOCKED) == 0) { 817 if (!atomic_fcmpset_acq_16(flagsp, &state, 818 FOFFSET_LOCKED)) 819 continue; 820 break; 821 } 822 if ((state & FOFFSET_LOCK_WAITING) == 0) { 823 if (!atomic_fcmpset_acq_16(flagsp, &state, 824 state | FOFFSET_LOCK_WAITING)) 825 continue; 826 } 827 DROP_GIANT(); 828 sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0); 829 sleepq_wait(&fp->f_vnread_flags, PUSER -1); 830 PICKUP_GIANT(); 831 sleepq_lock(&fp->f_vnread_flags); 832 state = atomic_load_16(flagsp); 833 } 834 res = atomic_load_long(&fp->f_offset); 835 sleepq_release(&fp->f_vnread_flags); 836 return (res); 837 } 838 839 void 840 foffset_unlock(struct file *fp, off_t val, int flags) 841 { 842 volatile short *flagsp; 843 short state; 844 845 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 846 847 if ((flags & FOF_NOUPDATE) == 0) 848 atomic_store_long(&fp->f_offset, val); 849 if ((flags & FOF_NEXTOFF_R) != 0) 850 fp->f_nextoff[UIO_READ] = val; 851 if ((flags & FOF_NEXTOFF_W) != 0) 852 fp->f_nextoff[UIO_WRITE] = val; 853 854 if ((flags & FOF_NOLOCK) != 0) 855 return; 856 857 flagsp = &fp->f_vnread_flags; 858 state = atomic_load_16(flagsp); 859 if ((state & FOFFSET_LOCK_WAITING) == 0 && 860 atomic_cmpset_rel_16(flagsp, state, 0)) 861 return; 862 863 sleepq_lock(&fp->f_vnread_flags); 864 MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0); 865 MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0); 866 fp->f_vnread_flags = 0; 867 sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0); 868 sleepq_release(&fp->f_vnread_flags); 869 } 870 871 static off_t 872 foffset_read(struct file *fp) 873 { 874 875 return (atomic_load_long(&fp->f_offset)); 876 } 877 #else 878 off_t 879 foffset_lock(struct file *fp, int flags) 880 { 881 struct mtx *mtxp; 882 off_t res; 883 884 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 885 886 mtxp = mtx_pool_find(mtxpool_sleep, fp); 887 mtx_lock(mtxp); 888 if ((flags & FOF_NOLOCK) == 0) { 889 while (fp->f_vnread_flags & FOFFSET_LOCKED) { 890 fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; 891 msleep(&fp->f_vnread_flags, mtxp, PUSER -1, 892 "vofflock", 0); 893 } 894 fp->f_vnread_flags |= FOFFSET_LOCKED; 895 } 896 res = fp->f_offset; 897 mtx_unlock(mtxp); 898 return (res); 899 } 900 901 void 902 foffset_unlock(struct file *fp, off_t val, int flags) 903 { 904 struct mtx *mtxp; 905 906 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 907 908 mtxp = mtx_pool_find(mtxpool_sleep, fp); 909 mtx_lock(mtxp); 910 if ((flags & FOF_NOUPDATE) == 0) 911 fp->f_offset = val; 912 if ((flags & FOF_NEXTOFF_R) != 0) 913 fp->f_nextoff[UIO_READ] = val; 914 if ((flags & FOF_NEXTOFF_W) != 0) 915 fp->f_nextoff[UIO_WRITE] = val; 916 if ((flags & FOF_NOLOCK) == 0) { 917 KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, 918 ("Lost FOFFSET_LOCKED")); 919 if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) 920 wakeup(&fp->f_vnread_flags); 921 fp->f_vnread_flags = 0; 922 } 923 mtx_unlock(mtxp); 924 } 925 926 static off_t 927 foffset_read(struct file *fp) 928 { 929 930 return (foffset_lock(fp, FOF_NOLOCK)); 931 } 932 #endif 933 934 void 935 foffset_lock_pair(struct file *fp1, off_t *off1p, struct file *fp2, off_t *off2p, 936 int flags) 937 { 938 KASSERT(fp1 != fp2, ("foffset_lock_pair: fp1 == fp2")); 939 940 /* Lock in a consistent order to avoid deadlock. */ 941 if ((uintptr_t)fp1 > (uintptr_t)fp2) { 942 struct file *tmpfp; 943 off_t *tmpoffp; 944 945 tmpfp = fp1, fp1 = fp2, fp2 = tmpfp; 946 tmpoffp = off1p, off1p = off2p, off2p = tmpoffp; 947 } 948 if (fp1 != NULL) 949 *off1p = foffset_lock(fp1, flags); 950 if (fp2 != NULL) 951 *off2p = foffset_lock(fp2, flags); 952 } 953 954 void 955 foffset_lock_uio(struct file *fp, struct uio *uio, int flags) 956 { 957 958 if ((flags & FOF_OFFSET) == 0) 959 uio->uio_offset = foffset_lock(fp, flags); 960 } 961 962 void 963 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) 964 { 965 966 if ((flags & FOF_OFFSET) == 0) 967 foffset_unlock(fp, uio->uio_offset, flags); 968 } 969 970 static int 971 get_advice(struct file *fp, struct uio *uio) 972 { 973 struct mtx *mtxp; 974 int ret; 975 976 ret = POSIX_FADV_NORMAL; 977 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) 978 return (ret); 979 980 mtxp = mtx_pool_find(mtxpool_sleep, fp); 981 mtx_lock(mtxp); 982 if (fp->f_advice != NULL && 983 uio->uio_offset >= fp->f_advice->fa_start && 984 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) 985 ret = fp->f_advice->fa_advice; 986 mtx_unlock(mtxp); 987 return (ret); 988 } 989 990 static int 991 get_write_ioflag(struct file *fp) 992 { 993 int ioflag; 994 struct mount *mp; 995 struct vnode *vp; 996 997 ioflag = 0; 998 vp = fp->f_vnode; 999 mp = atomic_load_ptr(&vp->v_mount); 1000 1001 if ((fp->f_flag & O_DIRECT) != 0) 1002 ioflag |= IO_DIRECT; 1003 1004 if ((fp->f_flag & O_FSYNC) != 0 || 1005 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0)) 1006 ioflag |= IO_SYNC; 1007 1008 /* 1009 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE() 1010 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC 1011 * fall back to full O_SYNC behavior. 1012 */ 1013 if ((fp->f_flag & O_DSYNC) != 0) 1014 ioflag |= IO_SYNC | IO_DATASYNC; 1015 1016 return (ioflag); 1017 } 1018 1019 int 1020 vn_read_from_obj(struct vnode *vp, struct uio *uio) 1021 { 1022 vm_object_t obj; 1023 vm_page_t ma[io_hold_cnt + 2]; 1024 off_t off, vsz; 1025 ssize_t resid; 1026 int error, i, j; 1027 1028 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2)); 1029 obj = atomic_load_ptr(&vp->v_object); 1030 if (obj == NULL) 1031 return (EJUSTRETURN); 1032 1033 /* 1034 * Depends on type stability of vm_objects. 1035 */ 1036 vm_object_pip_add(obj, 1); 1037 if ((obj->flags & OBJ_DEAD) != 0) { 1038 /* 1039 * Note that object might be already reused from the 1040 * vnode, and the OBJ_DEAD flag cleared. This is fine, 1041 * we recheck for DOOMED vnode state after all pages 1042 * are busied, and retract then. 1043 * 1044 * But we check for OBJ_DEAD to ensure that we do not 1045 * busy pages while vm_object_terminate_pages() 1046 * processes the queue. 1047 */ 1048 error = EJUSTRETURN; 1049 goto out_pip; 1050 } 1051 1052 resid = uio->uio_resid; 1053 off = uio->uio_offset; 1054 for (i = 0; resid > 0; i++) { 1055 MPASS(i < io_hold_cnt + 2); 1056 ma[i] = vm_page_grab_unlocked(obj, atop(off), 1057 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | 1058 VM_ALLOC_NOWAIT); 1059 if (ma[i] == NULL) 1060 break; 1061 1062 /* 1063 * Skip invalid pages. Valid mask can be partial only 1064 * at EOF, and we clip later. 1065 */ 1066 if (vm_page_none_valid(ma[i])) { 1067 vm_page_sunbusy(ma[i]); 1068 break; 1069 } 1070 1071 resid -= PAGE_SIZE; 1072 off += PAGE_SIZE; 1073 } 1074 if (i == 0) { 1075 error = EJUSTRETURN; 1076 goto out_pip; 1077 } 1078 1079 /* 1080 * Check VIRF_DOOMED after we busied our pages. Since 1081 * vgonel() terminates the vnode' vm_object, it cannot 1082 * process past pages busied by us. 1083 */ 1084 if (VN_IS_DOOMED(vp)) { 1085 error = EJUSTRETURN; 1086 goto out; 1087 } 1088 1089 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1); 1090 if (resid > uio->uio_resid) 1091 resid = uio->uio_resid; 1092 1093 /* 1094 * Unlocked read of vnp_size is safe because truncation cannot 1095 * pass busied page. But we load vnp_size into a local 1096 * variable so that possible concurrent extension does not 1097 * break calculation. 1098 */ 1099 #if defined(__powerpc__) && !defined(__powerpc64__) 1100 vsz = obj->un_pager.vnp.vnp_size; 1101 #else 1102 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size); 1103 #endif 1104 if (uio->uio_offset >= vsz) { 1105 error = EJUSTRETURN; 1106 goto out; 1107 } 1108 if (uio->uio_offset + resid > vsz) 1109 resid = vsz - uio->uio_offset; 1110 1111 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio); 1112 1113 out: 1114 for (j = 0; j < i; j++) { 1115 if (error == 0) 1116 vm_page_reference(ma[j]); 1117 vm_page_sunbusy(ma[j]); 1118 } 1119 out_pip: 1120 vm_object_pip_wakeup(obj); 1121 if (error != 0) 1122 return (error); 1123 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN); 1124 } 1125 1126 /* 1127 * File table vnode read routine. 1128 */ 1129 static int 1130 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1131 struct thread *td) 1132 { 1133 struct vnode *vp; 1134 off_t orig_offset; 1135 int error, ioflag; 1136 int advice; 1137 1138 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1139 uio->uio_td, td)); 1140 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1141 vp = fp->f_vnode; 1142 ioflag = 0; 1143 if (fp->f_flag & FNONBLOCK) 1144 ioflag |= IO_NDELAY; 1145 if (fp->f_flag & O_DIRECT) 1146 ioflag |= IO_DIRECT; 1147 1148 /* 1149 * Try to read from page cache. VIRF_DOOMED check is racy but 1150 * allows us to avoid unneeded work outright. 1151 */ 1152 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() && 1153 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) { 1154 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred); 1155 if (error == 0) { 1156 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1157 return (0); 1158 } 1159 if (error != EJUSTRETURN) 1160 return (error); 1161 } 1162 1163 advice = get_advice(fp, uio); 1164 vn_lock(vp, LK_SHARED | LK_RETRY); 1165 1166 switch (advice) { 1167 case POSIX_FADV_NORMAL: 1168 case POSIX_FADV_SEQUENTIAL: 1169 case POSIX_FADV_NOREUSE: 1170 ioflag |= sequential_heuristic(uio, fp); 1171 break; 1172 case POSIX_FADV_RANDOM: 1173 /* Disable read-ahead for random I/O. */ 1174 break; 1175 } 1176 orig_offset = uio->uio_offset; 1177 1178 #ifdef MAC 1179 error = mac_vnode_check_read(active_cred, fp->f_cred, vp); 1180 if (error == 0) 1181 #endif 1182 error = VOP_READ(vp, uio, ioflag, fp->f_cred); 1183 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1184 VOP_UNLOCK(vp); 1185 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1186 orig_offset != uio->uio_offset) 1187 /* 1188 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1189 * for the backing file after a POSIX_FADV_NOREUSE 1190 * read(2). 1191 */ 1192 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1193 POSIX_FADV_DONTNEED); 1194 return (error); 1195 } 1196 1197 /* 1198 * File table vnode write routine. 1199 */ 1200 static int 1201 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1202 struct thread *td) 1203 { 1204 struct vnode *vp; 1205 struct mount *mp; 1206 off_t orig_offset; 1207 int error, ioflag; 1208 int advice; 1209 bool need_finished_write; 1210 1211 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1212 uio->uio_td, td)); 1213 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1214 vp = fp->f_vnode; 1215 if (vp->v_type == VREG) 1216 bwillwrite(); 1217 ioflag = IO_UNIT; 1218 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0) 1219 ioflag |= IO_APPEND; 1220 if ((fp->f_flag & FNONBLOCK) != 0) 1221 ioflag |= IO_NDELAY; 1222 ioflag |= get_write_ioflag(fp); 1223 1224 mp = NULL; 1225 need_finished_write = false; 1226 if (vp->v_type != VCHR) { 1227 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 1228 if (error != 0) 1229 goto unlock; 1230 need_finished_write = true; 1231 } 1232 1233 advice = get_advice(fp, uio); 1234 1235 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY); 1236 switch (advice) { 1237 case POSIX_FADV_NORMAL: 1238 case POSIX_FADV_SEQUENTIAL: 1239 case POSIX_FADV_NOREUSE: 1240 ioflag |= sequential_heuristic(uio, fp); 1241 break; 1242 case POSIX_FADV_RANDOM: 1243 /* XXX: Is this correct? */ 1244 break; 1245 } 1246 orig_offset = uio->uio_offset; 1247 1248 #ifdef MAC 1249 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1250 if (error == 0) 1251 #endif 1252 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); 1253 fp->f_nextoff[UIO_WRITE] = uio->uio_offset; 1254 VOP_UNLOCK(vp); 1255 if (need_finished_write) 1256 vn_finished_write(mp); 1257 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1258 orig_offset != uio->uio_offset) 1259 /* 1260 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1261 * for the backing file after a POSIX_FADV_NOREUSE 1262 * write(2). 1263 */ 1264 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1265 POSIX_FADV_DONTNEED); 1266 unlock: 1267 return (error); 1268 } 1269 1270 /* 1271 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to 1272 * prevent the following deadlock: 1273 * 1274 * Assume that the thread A reads from the vnode vp1 into userspace 1275 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is 1276 * currently not resident, then system ends up with the call chain 1277 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> 1278 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) 1279 * which establishes lock order vp1->vn_lock, then vp2->vn_lock. 1280 * If, at the same time, thread B reads from vnode vp2 into buffer buf2 1281 * backed by the pages of vnode vp1, and some page in buf2 is not 1282 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. 1283 * 1284 * To prevent the lock order reversal and deadlock, vn_io_fault() does 1285 * not allow page faults to happen during VOP_READ() or VOP_WRITE(). 1286 * Instead, it first tries to do the whole range i/o with pagefaults 1287 * disabled. If all pages in the i/o buffer are resident and mapped, 1288 * VOP will succeed (ignoring the genuine filesystem errors). 1289 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do 1290 * i/o in chunks, with all pages in the chunk prefaulted and held 1291 * using vm_fault_quick_hold_pages(). 1292 * 1293 * Filesystems using this deadlock avoidance scheme should use the 1294 * array of the held pages from uio, saved in the curthread->td_ma, 1295 * instead of doing uiomove(). A helper function 1296 * vn_io_fault_uiomove() converts uiomove request into 1297 * uiomove_fromphys() over td_ma array. 1298 * 1299 * Since vnode locks do not cover the whole i/o anymore, rangelocks 1300 * make the current i/o request atomic with respect to other i/os and 1301 * truncations. 1302 */ 1303 1304 /* 1305 * Decode vn_io_fault_args and perform the corresponding i/o. 1306 */ 1307 static int 1308 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, 1309 struct thread *td) 1310 { 1311 int error, save; 1312 1313 error = 0; 1314 save = vm_fault_disable_pagefaults(); 1315 switch (args->kind) { 1316 case VN_IO_FAULT_FOP: 1317 error = (args->args.fop_args.doio)(args->args.fop_args.fp, 1318 uio, args->cred, args->flags, td); 1319 break; 1320 case VN_IO_FAULT_VOP: 1321 switch (uio->uio_rw) { 1322 case UIO_READ: 1323 error = VOP_READ(args->args.vop_args.vp, uio, 1324 args->flags, args->cred); 1325 break; 1326 case UIO_WRITE: 1327 error = VOP_WRITE(args->args.vop_args.vp, uio, 1328 args->flags, args->cred); 1329 break; 1330 } 1331 break; 1332 default: 1333 panic("vn_io_fault_doio: unknown kind of io %d %d", 1334 args->kind, uio->uio_rw); 1335 } 1336 vm_fault_enable_pagefaults(save); 1337 return (error); 1338 } 1339 1340 static int 1341 vn_io_fault_touch(char *base, const struct uio *uio) 1342 { 1343 int r; 1344 1345 r = fubyte(base); 1346 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) 1347 return (EFAULT); 1348 return (0); 1349 } 1350 1351 static int 1352 vn_io_fault_prefault_user(const struct uio *uio) 1353 { 1354 char *base; 1355 const struct iovec *iov; 1356 size_t len; 1357 ssize_t resid; 1358 int error, i; 1359 1360 KASSERT(uio->uio_segflg == UIO_USERSPACE, 1361 ("vn_io_fault_prefault userspace")); 1362 1363 error = i = 0; 1364 iov = uio->uio_iov; 1365 resid = uio->uio_resid; 1366 base = iov->iov_base; 1367 len = iov->iov_len; 1368 while (resid > 0) { 1369 error = vn_io_fault_touch(base, uio); 1370 if (error != 0) 1371 break; 1372 if (len < PAGE_SIZE) { 1373 if (len != 0) { 1374 error = vn_io_fault_touch(base + len - 1, uio); 1375 if (error != 0) 1376 break; 1377 resid -= len; 1378 } 1379 if (++i >= uio->uio_iovcnt) 1380 break; 1381 iov = uio->uio_iov + i; 1382 base = iov->iov_base; 1383 len = iov->iov_len; 1384 } else { 1385 len -= PAGE_SIZE; 1386 base += PAGE_SIZE; 1387 resid -= PAGE_SIZE; 1388 } 1389 } 1390 return (error); 1391 } 1392 1393 /* 1394 * Common code for vn_io_fault(), agnostic to the kind of i/o request. 1395 * Uses vn_io_fault_doio() to make the call to an actual i/o function. 1396 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request 1397 * into args and call vn_io_fault1() to handle faults during the user 1398 * mode buffer accesses. 1399 */ 1400 static int 1401 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, 1402 struct thread *td) 1403 { 1404 vm_page_t ma[io_hold_cnt + 2]; 1405 struct uio *uio_clone, short_uio; 1406 struct iovec short_iovec[1]; 1407 vm_page_t *prev_td_ma; 1408 vm_prot_t prot; 1409 vm_offset_t addr, end; 1410 size_t len, resid; 1411 ssize_t adv; 1412 int error, cnt, saveheld, prev_td_ma_cnt; 1413 1414 if (vn_io_fault_prefault) { 1415 error = vn_io_fault_prefault_user(uio); 1416 if (error != 0) 1417 return (error); /* Or ignore ? */ 1418 } 1419 1420 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; 1421 1422 /* 1423 * The UFS follows IO_UNIT directive and replays back both 1424 * uio_offset and uio_resid if an error is encountered during the 1425 * operation. But, since the iovec may be already advanced, 1426 * uio is still in an inconsistent state. 1427 * 1428 * Cache a copy of the original uio, which is advanced to the redo 1429 * point using UIO_NOCOPY below. 1430 */ 1431 uio_clone = cloneuio(uio); 1432 resid = uio->uio_resid; 1433 1434 short_uio.uio_segflg = UIO_USERSPACE; 1435 short_uio.uio_rw = uio->uio_rw; 1436 short_uio.uio_td = uio->uio_td; 1437 1438 error = vn_io_fault_doio(args, uio, td); 1439 if (error != EFAULT) 1440 goto out; 1441 1442 atomic_add_long(&vn_io_faults_cnt, 1); 1443 uio_clone->uio_segflg = UIO_NOCOPY; 1444 uiomove(NULL, resid - uio->uio_resid, uio_clone); 1445 uio_clone->uio_segflg = uio->uio_segflg; 1446 1447 saveheld = curthread_pflags_set(TDP_UIOHELD); 1448 prev_td_ma = td->td_ma; 1449 prev_td_ma_cnt = td->td_ma_cnt; 1450 1451 while (uio_clone->uio_resid != 0) { 1452 len = uio_clone->uio_iov->iov_len; 1453 if (len == 0) { 1454 KASSERT(uio_clone->uio_iovcnt >= 1, 1455 ("iovcnt underflow")); 1456 uio_clone->uio_iov++; 1457 uio_clone->uio_iovcnt--; 1458 continue; 1459 } 1460 if (len > ptoa(io_hold_cnt)) 1461 len = ptoa(io_hold_cnt); 1462 addr = (uintptr_t)uio_clone->uio_iov->iov_base; 1463 end = round_page(addr + len); 1464 if (end < addr) { 1465 error = EFAULT; 1466 break; 1467 } 1468 /* 1469 * A perfectly misaligned address and length could cause 1470 * both the start and the end of the chunk to use partial 1471 * page. +2 accounts for such a situation. 1472 */ 1473 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, 1474 addr, len, prot, ma, io_hold_cnt + 2); 1475 if (cnt == -1) { 1476 error = EFAULT; 1477 break; 1478 } 1479 short_uio.uio_iov = &short_iovec[0]; 1480 short_iovec[0].iov_base = (void *)addr; 1481 short_uio.uio_iovcnt = 1; 1482 short_uio.uio_resid = short_iovec[0].iov_len = len; 1483 short_uio.uio_offset = uio_clone->uio_offset; 1484 td->td_ma = ma; 1485 td->td_ma_cnt = cnt; 1486 1487 error = vn_io_fault_doio(args, &short_uio, td); 1488 vm_page_unhold_pages(ma, cnt); 1489 adv = len - short_uio.uio_resid; 1490 1491 uio_clone->uio_iov->iov_base = 1492 (char *)uio_clone->uio_iov->iov_base + adv; 1493 uio_clone->uio_iov->iov_len -= adv; 1494 uio_clone->uio_resid -= adv; 1495 uio_clone->uio_offset += adv; 1496 1497 uio->uio_resid -= adv; 1498 uio->uio_offset += adv; 1499 1500 if (error != 0 || adv == 0) 1501 break; 1502 } 1503 td->td_ma = prev_td_ma; 1504 td->td_ma_cnt = prev_td_ma_cnt; 1505 curthread_pflags_restore(saveheld); 1506 out: 1507 freeuio(uio_clone); 1508 return (error); 1509 } 1510 1511 static int 1512 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, 1513 int flags, struct thread *td) 1514 { 1515 fo_rdwr_t *doio; 1516 struct vnode *vp; 1517 void *rl_cookie; 1518 struct vn_io_fault_args args; 1519 int error; 1520 bool do_io_fault, do_rangelock; 1521 1522 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; 1523 vp = fp->f_vnode; 1524 1525 /* 1526 * The ability to read(2) on a directory has historically been 1527 * allowed for all users, but this can and has been the source of 1528 * at least one security issue in the past. As such, it is now hidden 1529 * away behind a sysctl for those that actually need it to use it, and 1530 * restricted to root when it's turned on to make it relatively safe to 1531 * leave on for longer sessions of need. 1532 */ 1533 if (vp->v_type == VDIR) { 1534 KASSERT(uio->uio_rw == UIO_READ, 1535 ("illegal write attempted on a directory")); 1536 if (!vfs_allow_read_dir) 1537 return (EISDIR); 1538 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0) 1539 return (EISDIR); 1540 } 1541 1542 do_io_fault = do_vn_io_fault(vp, uio); 1543 do_rangelock = do_io_fault || (vn_irflag_read(vp) & VIRF_PGREAD) != 0; 1544 foffset_lock_uio(fp, uio, flags); 1545 if (do_rangelock) { 1546 if (uio->uio_rw == UIO_READ) { 1547 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, 1548 uio->uio_offset + uio->uio_resid); 1549 } else if ((fp->f_flag & O_APPEND) != 0 || 1550 (flags & FOF_OFFSET) == 0) { 1551 /* For appenders, punt and lock the whole range. */ 1552 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1553 } else { 1554 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, 1555 uio->uio_offset + uio->uio_resid); 1556 } 1557 } 1558 if (do_io_fault) { 1559 args.kind = VN_IO_FAULT_FOP; 1560 args.args.fop_args.fp = fp; 1561 args.args.fop_args.doio = doio; 1562 args.cred = active_cred; 1563 args.flags = flags | FOF_OFFSET; 1564 error = vn_io_fault1(vp, uio, &args, td); 1565 } else { 1566 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); 1567 } 1568 if (do_rangelock) 1569 vn_rangelock_unlock(vp, rl_cookie); 1570 foffset_unlock_uio(fp, uio, flags); 1571 return (error); 1572 } 1573 1574 /* 1575 * Helper function to perform the requested uiomove operation using 1576 * the held pages for io->uio_iov[0].iov_base buffer instead of 1577 * copyin/copyout. Access to the pages with uiomove_fromphys() 1578 * instead of iov_base prevents page faults that could occur due to 1579 * pmap_collect() invalidating the mapping created by 1580 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or 1581 * object cleanup revoking the write access from page mappings. 1582 * 1583 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() 1584 * instead of plain uiomove(). 1585 */ 1586 int 1587 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) 1588 { 1589 struct uio transp_uio; 1590 struct iovec transp_iov[1]; 1591 struct thread *td; 1592 size_t adv; 1593 int error, pgadv; 1594 1595 td = curthread; 1596 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1597 uio->uio_segflg != UIO_USERSPACE) 1598 return (uiomove(data, xfersize, uio)); 1599 1600 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1601 transp_iov[0].iov_base = data; 1602 transp_uio.uio_iov = &transp_iov[0]; 1603 transp_uio.uio_iovcnt = 1; 1604 if (xfersize > uio->uio_resid) 1605 xfersize = uio->uio_resid; 1606 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; 1607 transp_uio.uio_offset = 0; 1608 transp_uio.uio_segflg = UIO_SYSSPACE; 1609 /* 1610 * Since transp_iov points to data, and td_ma page array 1611 * corresponds to original uio->uio_iov, we need to invert the 1612 * direction of the i/o operation as passed to 1613 * uiomove_fromphys(). 1614 */ 1615 switch (uio->uio_rw) { 1616 case UIO_WRITE: 1617 transp_uio.uio_rw = UIO_READ; 1618 break; 1619 case UIO_READ: 1620 transp_uio.uio_rw = UIO_WRITE; 1621 break; 1622 } 1623 transp_uio.uio_td = uio->uio_td; 1624 error = uiomove_fromphys(td->td_ma, 1625 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, 1626 xfersize, &transp_uio); 1627 adv = xfersize - transp_uio.uio_resid; 1628 pgadv = 1629 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - 1630 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); 1631 td->td_ma += pgadv; 1632 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1633 pgadv)); 1634 td->td_ma_cnt -= pgadv; 1635 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; 1636 uio->uio_iov->iov_len -= adv; 1637 uio->uio_resid -= adv; 1638 uio->uio_offset += adv; 1639 return (error); 1640 } 1641 1642 int 1643 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, 1644 struct uio *uio) 1645 { 1646 struct thread *td; 1647 vm_offset_t iov_base; 1648 int cnt, pgadv; 1649 1650 td = curthread; 1651 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1652 uio->uio_segflg != UIO_USERSPACE) 1653 return (uiomove_fromphys(ma, offset, xfersize, uio)); 1654 1655 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1656 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; 1657 iov_base = (vm_offset_t)uio->uio_iov->iov_base; 1658 switch (uio->uio_rw) { 1659 case UIO_WRITE: 1660 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, 1661 offset, cnt); 1662 break; 1663 case UIO_READ: 1664 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, 1665 cnt); 1666 break; 1667 } 1668 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); 1669 td->td_ma += pgadv; 1670 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1671 pgadv)); 1672 td->td_ma_cnt -= pgadv; 1673 uio->uio_iov->iov_base = (char *)(iov_base + cnt); 1674 uio->uio_iov->iov_len -= cnt; 1675 uio->uio_resid -= cnt; 1676 uio->uio_offset += cnt; 1677 return (0); 1678 } 1679 1680 /* 1681 * File table truncate routine. 1682 */ 1683 static int 1684 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1685 struct thread *td) 1686 { 1687 struct mount *mp; 1688 struct vnode *vp; 1689 void *rl_cookie; 1690 int error; 1691 1692 vp = fp->f_vnode; 1693 1694 retry: 1695 /* 1696 * Lock the whole range for truncation. Otherwise split i/o 1697 * might happen partly before and partly after the truncation. 1698 */ 1699 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1700 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 1701 if (error) 1702 goto out1; 1703 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1704 AUDIT_ARG_VNODE1(vp); 1705 if (vp->v_type == VDIR) { 1706 error = EISDIR; 1707 goto out; 1708 } 1709 #ifdef MAC 1710 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1711 if (error) 1712 goto out; 1713 #endif 1714 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, 1715 fp->f_cred); 1716 out: 1717 VOP_UNLOCK(vp); 1718 vn_finished_write(mp); 1719 out1: 1720 vn_rangelock_unlock(vp, rl_cookie); 1721 if (error == ERELOOKUP) 1722 goto retry; 1723 return (error); 1724 } 1725 1726 /* 1727 * Truncate a file that is already locked. 1728 */ 1729 int 1730 vn_truncate_locked(struct vnode *vp, off_t length, bool sync, 1731 struct ucred *cred) 1732 { 1733 struct vattr vattr; 1734 int error; 1735 1736 error = VOP_ADD_WRITECOUNT(vp, 1); 1737 if (error == 0) { 1738 VATTR_NULL(&vattr); 1739 vattr.va_size = length; 1740 if (sync) 1741 vattr.va_vaflags |= VA_SYNC; 1742 error = VOP_SETATTR(vp, &vattr, cred); 1743 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1744 } 1745 return (error); 1746 } 1747 1748 /* 1749 * File table vnode stat routine. 1750 */ 1751 int 1752 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred) 1753 { 1754 struct vnode *vp = fp->f_vnode; 1755 int error; 1756 1757 vn_lock(vp, LK_SHARED | LK_RETRY); 1758 error = VOP_STAT(vp, sb, active_cred, fp->f_cred); 1759 VOP_UNLOCK(vp); 1760 1761 return (error); 1762 } 1763 1764 /* 1765 * File table vnode ioctl routine. 1766 */ 1767 static int 1768 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 1769 struct thread *td) 1770 { 1771 struct vnode *vp; 1772 struct fiobmap2_arg *bmarg; 1773 off_t size; 1774 int error; 1775 1776 vp = fp->f_vnode; 1777 switch (vp->v_type) { 1778 case VDIR: 1779 case VREG: 1780 switch (com) { 1781 case FIONREAD: 1782 error = vn_getsize(vp, &size, active_cred); 1783 if (error == 0) 1784 *(int *)data = size - fp->f_offset; 1785 return (error); 1786 case FIOBMAP2: 1787 bmarg = (struct fiobmap2_arg *)data; 1788 vn_lock(vp, LK_SHARED | LK_RETRY); 1789 #ifdef MAC 1790 error = mac_vnode_check_read(active_cred, fp->f_cred, 1791 vp); 1792 if (error == 0) 1793 #endif 1794 error = VOP_BMAP(vp, bmarg->bn, NULL, 1795 &bmarg->bn, &bmarg->runp, &bmarg->runb); 1796 VOP_UNLOCK(vp); 1797 return (error); 1798 case FIONBIO: 1799 case FIOASYNC: 1800 return (0); 1801 default: 1802 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1803 active_cred, td)); 1804 } 1805 break; 1806 case VCHR: 1807 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1808 active_cred, td)); 1809 default: 1810 return (ENOTTY); 1811 } 1812 } 1813 1814 /* 1815 * File table vnode poll routine. 1816 */ 1817 static int 1818 vn_poll(struct file *fp, int events, struct ucred *active_cred, 1819 struct thread *td) 1820 { 1821 struct vnode *vp; 1822 int error; 1823 1824 vp = fp->f_vnode; 1825 #if defined(MAC) || defined(AUDIT) 1826 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) { 1827 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1828 AUDIT_ARG_VNODE1(vp); 1829 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); 1830 VOP_UNLOCK(vp); 1831 if (error != 0) 1832 return (error); 1833 } 1834 #endif 1835 error = VOP_POLL(vp, events, fp->f_cred, td); 1836 return (error); 1837 } 1838 1839 /* 1840 * Acquire the requested lock and then check for validity. LK_RETRY 1841 * permits vn_lock to return doomed vnodes. 1842 */ 1843 static int __noinline 1844 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, 1845 int error) 1846 { 1847 1848 KASSERT((flags & LK_RETRY) == 0 || error == 0, 1849 ("vn_lock: error %d incompatible with flags %#x", error, flags)); 1850 1851 if (error == 0) 1852 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); 1853 1854 if ((flags & LK_RETRY) == 0) { 1855 if (error == 0) { 1856 VOP_UNLOCK(vp); 1857 error = ENOENT; 1858 } 1859 return (error); 1860 } 1861 1862 /* 1863 * LK_RETRY case. 1864 * 1865 * Nothing to do if we got the lock. 1866 */ 1867 if (error == 0) 1868 return (0); 1869 1870 /* 1871 * Interlock was dropped by the call in _vn_lock. 1872 */ 1873 flags &= ~LK_INTERLOCK; 1874 do { 1875 error = VOP_LOCK1(vp, flags, file, line); 1876 } while (error != 0); 1877 return (0); 1878 } 1879 1880 int 1881 _vn_lock(struct vnode *vp, int flags, const char *file, int line) 1882 { 1883 int error; 1884 1885 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 1886 ("vn_lock: no locktype (%d passed)", flags)); 1887 VNPASS(vp->v_holdcnt > 0, vp); 1888 error = VOP_LOCK1(vp, flags, file, line); 1889 if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) 1890 return (_vn_lock_fallback(vp, flags, file, line, error)); 1891 return (0); 1892 } 1893 1894 /* 1895 * File table vnode close routine. 1896 */ 1897 static int 1898 vn_closefile(struct file *fp, struct thread *td) 1899 { 1900 struct vnode *vp; 1901 struct flock lf; 1902 int error; 1903 bool ref; 1904 1905 vp = fp->f_vnode; 1906 fp->f_ops = &badfileops; 1907 ref = (fp->f_flag & FHASLOCK) != 0; 1908 1909 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); 1910 1911 if (__predict_false(ref)) { 1912 lf.l_whence = SEEK_SET; 1913 lf.l_start = 0; 1914 lf.l_len = 0; 1915 lf.l_type = F_UNLCK; 1916 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); 1917 vrele(vp); 1918 } 1919 return (error); 1920 } 1921 1922 /* 1923 * Preparing to start a filesystem write operation. If the operation is 1924 * permitted, then we bump the count of operations in progress and 1925 * proceed. If a suspend request is in progress, we wait until the 1926 * suspension is over, and then proceed. 1927 */ 1928 static int 1929 vn_start_write_refed(struct mount *mp, int flags, bool mplocked) 1930 { 1931 struct mount_pcpu *mpcpu; 1932 int error, mflags; 1933 1934 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && 1935 vfs_op_thread_enter(mp, mpcpu)) { 1936 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 1937 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1); 1938 vfs_op_thread_exit(mp, mpcpu); 1939 return (0); 1940 } 1941 1942 if (mplocked) 1943 mtx_assert(MNT_MTX(mp), MA_OWNED); 1944 else 1945 MNT_ILOCK(mp); 1946 1947 error = 0; 1948 1949 /* 1950 * Check on status of suspension. 1951 */ 1952 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || 1953 mp->mnt_susp_owner != curthread) { 1954 mflags = 0; 1955 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) { 1956 if (flags & V_PCATCH) 1957 mflags |= PCATCH; 1958 } 1959 mflags |= (PUSER - 1); 1960 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 1961 if ((flags & V_NOWAIT) != 0) { 1962 error = EWOULDBLOCK; 1963 goto unlock; 1964 } 1965 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, 1966 "suspfs", 0); 1967 if (error != 0) 1968 goto unlock; 1969 } 1970 } 1971 if ((flags & V_XSLEEP) != 0) 1972 goto unlock; 1973 mp->mnt_writeopcount++; 1974 unlock: 1975 if (error != 0 || (flags & V_XSLEEP) != 0) 1976 MNT_REL(mp); 1977 MNT_IUNLOCK(mp); 1978 return (error); 1979 } 1980 1981 int 1982 vn_start_write(struct vnode *vp, struct mount **mpp, int flags) 1983 { 1984 struct mount *mp; 1985 int error; 1986 1987 KASSERT((flags & ~V_VALID_FLAGS) == 0, 1988 ("%s: invalid flags passed %d\n", __func__, flags)); 1989 1990 error = 0; 1991 /* 1992 * If a vnode is provided, get and return the mount point that 1993 * to which it will write. 1994 */ 1995 if (vp != NULL) { 1996 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 1997 *mpp = NULL; 1998 if (error != EOPNOTSUPP) 1999 return (error); 2000 return (0); 2001 } 2002 } 2003 if ((mp = *mpp) == NULL) 2004 return (0); 2005 2006 /* 2007 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 2008 * a vfs_ref(). 2009 * As long as a vnode is not provided we need to acquire a 2010 * refcount for the provided mountpoint too, in order to 2011 * emulate a vfs_ref(). 2012 */ 2013 if (vp == NULL) 2014 vfs_ref(mp); 2015 2016 error = vn_start_write_refed(mp, flags, false); 2017 if (error != 0 && (flags & V_NOWAIT) == 0) 2018 *mpp = NULL; 2019 return (error); 2020 } 2021 2022 /* 2023 * Secondary suspension. Used by operations such as vop_inactive 2024 * routines that are needed by the higher level functions. These 2025 * are allowed to proceed until all the higher level functions have 2026 * completed (indicated by mnt_writeopcount dropping to zero). At that 2027 * time, these operations are halted until the suspension is over. 2028 */ 2029 int 2030 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) 2031 { 2032 struct mount *mp; 2033 int error, mflags; 2034 2035 KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0, 2036 ("%s: invalid flags passed %d\n", __func__, flags)); 2037 2038 retry: 2039 if (vp != NULL) { 2040 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 2041 *mpp = NULL; 2042 if (error != EOPNOTSUPP) 2043 return (error); 2044 return (0); 2045 } 2046 } 2047 /* 2048 * If we are not suspended or have not yet reached suspended 2049 * mode, then let the operation proceed. 2050 */ 2051 if ((mp = *mpp) == NULL) 2052 return (0); 2053 2054 /* 2055 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 2056 * a vfs_ref(). 2057 * As long as a vnode is not provided we need to acquire a 2058 * refcount for the provided mountpoint too, in order to 2059 * emulate a vfs_ref(). 2060 */ 2061 MNT_ILOCK(mp); 2062 if (vp == NULL) 2063 MNT_REF(mp); 2064 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { 2065 mp->mnt_secondary_writes++; 2066 mp->mnt_secondary_accwrites++; 2067 MNT_IUNLOCK(mp); 2068 return (0); 2069 } 2070 if ((flags & V_NOWAIT) != 0) { 2071 MNT_REL(mp); 2072 MNT_IUNLOCK(mp); 2073 *mpp = NULL; 2074 return (EWOULDBLOCK); 2075 } 2076 /* 2077 * Wait for the suspension to finish. 2078 */ 2079 mflags = 0; 2080 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) { 2081 if ((flags & V_PCATCH) != 0) 2082 mflags |= PCATCH; 2083 } 2084 mflags |= (PUSER - 1) | PDROP; 2085 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0); 2086 vfs_rel(mp); 2087 if (error == 0) 2088 goto retry; 2089 *mpp = NULL; 2090 return (error); 2091 } 2092 2093 /* 2094 * Filesystem write operation has completed. If we are suspending and this 2095 * operation is the last one, notify the suspender that the suspension is 2096 * now in effect. 2097 */ 2098 void 2099 vn_finished_write(struct mount *mp) 2100 { 2101 struct mount_pcpu *mpcpu; 2102 int c; 2103 2104 if (mp == NULL) 2105 return; 2106 2107 if (vfs_op_thread_enter(mp, mpcpu)) { 2108 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1); 2109 vfs_mp_count_sub_pcpu(mpcpu, ref, 1); 2110 vfs_op_thread_exit(mp, mpcpu); 2111 return; 2112 } 2113 2114 MNT_ILOCK(mp); 2115 vfs_assert_mount_counters(mp); 2116 MNT_REL(mp); 2117 c = --mp->mnt_writeopcount; 2118 if (mp->mnt_vfs_ops == 0) { 2119 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 2120 MNT_IUNLOCK(mp); 2121 return; 2122 } 2123 if (c < 0) 2124 vfs_dump_mount_counters(mp); 2125 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) 2126 wakeup(&mp->mnt_writeopcount); 2127 MNT_IUNLOCK(mp); 2128 } 2129 2130 /* 2131 * Filesystem secondary write operation has completed. If we are 2132 * suspending and this operation is the last one, notify the suspender 2133 * that the suspension is now in effect. 2134 */ 2135 void 2136 vn_finished_secondary_write(struct mount *mp) 2137 { 2138 if (mp == NULL) 2139 return; 2140 MNT_ILOCK(mp); 2141 MNT_REL(mp); 2142 mp->mnt_secondary_writes--; 2143 if (mp->mnt_secondary_writes < 0) 2144 panic("vn_finished_secondary_write: neg cnt"); 2145 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 2146 mp->mnt_secondary_writes <= 0) 2147 wakeup(&mp->mnt_secondary_writes); 2148 MNT_IUNLOCK(mp); 2149 } 2150 2151 /* 2152 * Request a filesystem to suspend write operations. 2153 */ 2154 int 2155 vfs_write_suspend(struct mount *mp, int flags) 2156 { 2157 int error; 2158 2159 vfs_op_enter(mp); 2160 2161 MNT_ILOCK(mp); 2162 vfs_assert_mount_counters(mp); 2163 if (mp->mnt_susp_owner == curthread) { 2164 vfs_op_exit_locked(mp); 2165 MNT_IUNLOCK(mp); 2166 return (EALREADY); 2167 } 2168 while (mp->mnt_kern_flag & MNTK_SUSPEND) 2169 msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); 2170 2171 /* 2172 * Unmount holds a write reference on the mount point. If we 2173 * own busy reference and drain for writers, we deadlock with 2174 * the reference draining in the unmount path. Callers of 2175 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if 2176 * vfs_busy() reference is owned and caller is not in the 2177 * unmount context. 2178 */ 2179 if ((flags & VS_SKIP_UNMOUNT) != 0 && 2180 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { 2181 vfs_op_exit_locked(mp); 2182 MNT_IUNLOCK(mp); 2183 return (EBUSY); 2184 } 2185 2186 mp->mnt_kern_flag |= MNTK_SUSPEND; 2187 mp->mnt_susp_owner = curthread; 2188 if (mp->mnt_writeopcount > 0) 2189 (void) msleep(&mp->mnt_writeopcount, 2190 MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); 2191 else 2192 MNT_IUNLOCK(mp); 2193 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { 2194 vfs_write_resume(mp, 0); 2195 /* vfs_write_resume does vfs_op_exit() for us */ 2196 } 2197 return (error); 2198 } 2199 2200 /* 2201 * Request a filesystem to resume write operations. 2202 */ 2203 void 2204 vfs_write_resume(struct mount *mp, int flags) 2205 { 2206 2207 MNT_ILOCK(mp); 2208 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 2209 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); 2210 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | 2211 MNTK_SUSPENDED); 2212 mp->mnt_susp_owner = NULL; 2213 wakeup(&mp->mnt_writeopcount); 2214 wakeup(&mp->mnt_flag); 2215 curthread->td_pflags &= ~TDP_IGNSUSP; 2216 if ((flags & VR_START_WRITE) != 0) { 2217 MNT_REF(mp); 2218 mp->mnt_writeopcount++; 2219 } 2220 MNT_IUNLOCK(mp); 2221 if ((flags & VR_NO_SUSPCLR) == 0) 2222 VFS_SUSP_CLEAN(mp); 2223 vfs_op_exit(mp); 2224 } else if ((flags & VR_START_WRITE) != 0) { 2225 MNT_REF(mp); 2226 vn_start_write_refed(mp, 0, true); 2227 } else { 2228 MNT_IUNLOCK(mp); 2229 } 2230 } 2231 2232 /* 2233 * Helper loop around vfs_write_suspend() for filesystem unmount VFS 2234 * methods. 2235 */ 2236 int 2237 vfs_write_suspend_umnt(struct mount *mp) 2238 { 2239 int error; 2240 2241 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, 2242 ("vfs_write_suspend_umnt: recursed")); 2243 2244 /* dounmount() already called vn_start_write(). */ 2245 for (;;) { 2246 vn_finished_write(mp); 2247 error = vfs_write_suspend(mp, 0); 2248 if (error != 0) { 2249 vn_start_write(NULL, &mp, V_WAIT); 2250 return (error); 2251 } 2252 MNT_ILOCK(mp); 2253 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) 2254 break; 2255 MNT_IUNLOCK(mp); 2256 vn_start_write(NULL, &mp, V_WAIT); 2257 } 2258 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); 2259 wakeup(&mp->mnt_flag); 2260 MNT_IUNLOCK(mp); 2261 curthread->td_pflags |= TDP_IGNSUSP; 2262 return (0); 2263 } 2264 2265 /* 2266 * Implement kqueues for files by translating it to vnode operation. 2267 */ 2268 static int 2269 vn_kqfilter(struct file *fp, struct knote *kn) 2270 { 2271 2272 return (VOP_KQFILTER(fp->f_vnode, kn)); 2273 } 2274 2275 int 2276 vn_kqfilter_opath(struct file *fp, struct knote *kn) 2277 { 2278 if ((fp->f_flag & FKQALLOWED) == 0) 2279 return (EBADF); 2280 return (vn_kqfilter(fp, kn)); 2281 } 2282 2283 /* 2284 * Simplified in-kernel wrapper calls for extended attribute access. 2285 * Both calls pass in a NULL credential, authorizing as "kernel" access. 2286 * Set IO_NODELOCKED in ioflg if the vnode is already locked. 2287 */ 2288 int 2289 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, 2290 const char *attrname, int *buflen, char *buf, struct thread *td) 2291 { 2292 struct uio auio; 2293 struct iovec iov; 2294 int error; 2295 2296 iov.iov_len = *buflen; 2297 iov.iov_base = buf; 2298 2299 auio.uio_iov = &iov; 2300 auio.uio_iovcnt = 1; 2301 auio.uio_rw = UIO_READ; 2302 auio.uio_segflg = UIO_SYSSPACE; 2303 auio.uio_td = td; 2304 auio.uio_offset = 0; 2305 auio.uio_resid = *buflen; 2306 2307 if ((ioflg & IO_NODELOCKED) == 0) 2308 vn_lock(vp, LK_SHARED | LK_RETRY); 2309 2310 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2311 2312 /* authorize attribute retrieval as kernel */ 2313 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, 2314 td); 2315 2316 if ((ioflg & IO_NODELOCKED) == 0) 2317 VOP_UNLOCK(vp); 2318 2319 if (error == 0) { 2320 *buflen = *buflen - auio.uio_resid; 2321 } 2322 2323 return (error); 2324 } 2325 2326 /* 2327 * XXX failure mode if partially written? 2328 */ 2329 int 2330 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, 2331 const char *attrname, int buflen, char *buf, struct thread *td) 2332 { 2333 struct uio auio; 2334 struct iovec iov; 2335 struct mount *mp; 2336 int error; 2337 2338 iov.iov_len = buflen; 2339 iov.iov_base = buf; 2340 2341 auio.uio_iov = &iov; 2342 auio.uio_iovcnt = 1; 2343 auio.uio_rw = UIO_WRITE; 2344 auio.uio_segflg = UIO_SYSSPACE; 2345 auio.uio_td = td; 2346 auio.uio_offset = 0; 2347 auio.uio_resid = buflen; 2348 2349 if ((ioflg & IO_NODELOCKED) == 0) { 2350 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2351 return (error); 2352 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2353 } 2354 2355 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2356 2357 /* authorize attribute setting as kernel */ 2358 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); 2359 2360 if ((ioflg & IO_NODELOCKED) == 0) { 2361 vn_finished_write(mp); 2362 VOP_UNLOCK(vp); 2363 } 2364 2365 return (error); 2366 } 2367 2368 int 2369 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, 2370 const char *attrname, struct thread *td) 2371 { 2372 struct mount *mp; 2373 int error; 2374 2375 if ((ioflg & IO_NODELOCKED) == 0) { 2376 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2377 return (error); 2378 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2379 } 2380 2381 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2382 2383 /* authorize attribute removal as kernel */ 2384 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); 2385 if (error == EOPNOTSUPP) 2386 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, 2387 NULL, td); 2388 2389 if ((ioflg & IO_NODELOCKED) == 0) { 2390 vn_finished_write(mp); 2391 VOP_UNLOCK(vp); 2392 } 2393 2394 return (error); 2395 } 2396 2397 static int 2398 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, 2399 struct vnode **rvp) 2400 { 2401 2402 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); 2403 } 2404 2405 int 2406 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) 2407 { 2408 2409 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, 2410 lkflags, rvp)); 2411 } 2412 2413 int 2414 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, 2415 int lkflags, struct vnode **rvp) 2416 { 2417 struct mount *mp; 2418 int ltype, error; 2419 2420 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); 2421 mp = vp->v_mount; 2422 ltype = VOP_ISLOCKED(vp); 2423 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, 2424 ("vn_vget_ino: vp not locked")); 2425 error = vfs_busy(mp, MBF_NOWAIT); 2426 if (error != 0) { 2427 vfs_ref(mp); 2428 VOP_UNLOCK(vp); 2429 error = vfs_busy(mp, 0); 2430 vn_lock(vp, ltype | LK_RETRY); 2431 vfs_rel(mp); 2432 if (error != 0) 2433 return (ENOENT); 2434 if (VN_IS_DOOMED(vp)) { 2435 vfs_unbusy(mp); 2436 return (ENOENT); 2437 } 2438 } 2439 VOP_UNLOCK(vp); 2440 error = alloc(mp, alloc_arg, lkflags, rvp); 2441 vfs_unbusy(mp); 2442 if (error != 0 || *rvp != vp) 2443 vn_lock(vp, ltype | LK_RETRY); 2444 if (VN_IS_DOOMED(vp)) { 2445 if (error == 0) { 2446 if (*rvp == vp) 2447 vunref(vp); 2448 else 2449 vput(*rvp); 2450 } 2451 error = ENOENT; 2452 } 2453 return (error); 2454 } 2455 2456 static void 2457 vn_send_sigxfsz(struct proc *p) 2458 { 2459 PROC_LOCK(p); 2460 kern_psignal(p, SIGXFSZ); 2461 PROC_UNLOCK(p); 2462 } 2463 2464 int 2465 vn_rlimit_trunc(u_quad_t size, struct thread *td) 2466 { 2467 if (size <= lim_cur(td, RLIMIT_FSIZE)) 2468 return (0); 2469 vn_send_sigxfsz(td->td_proc); 2470 return (EFBIG); 2471 } 2472 2473 static int 2474 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz, 2475 bool adj, struct thread *td) 2476 { 2477 off_t lim; 2478 bool ktr_write; 2479 2480 if (vp->v_type != VREG) 2481 return (0); 2482 2483 /* 2484 * Handle file system maximum file size. 2485 */ 2486 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) { 2487 if (!adj || uio->uio_offset >= maxfsz) 2488 return (EFBIG); 2489 uio->uio_resid = maxfsz - uio->uio_offset; 2490 } 2491 2492 /* 2493 * This is kernel write (e.g. vnode_pager) or accounting 2494 * write, ignore limit. 2495 */ 2496 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0) 2497 return (0); 2498 2499 /* 2500 * Calculate file size limit. 2501 */ 2502 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0; 2503 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim : 2504 lim_cur(td, RLIMIT_FSIZE); 2505 2506 /* 2507 * Is the limit reached? 2508 */ 2509 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim)) 2510 return (0); 2511 2512 /* 2513 * Prepared filesystems can handle writes truncated to the 2514 * file size limit. 2515 */ 2516 if (adj && (uoff_t)uio->uio_offset < lim) { 2517 uio->uio_resid = lim - (uoff_t)uio->uio_offset; 2518 return (0); 2519 } 2520 2521 if (!ktr_write || ktr_filesize_limit_signal) 2522 vn_send_sigxfsz(td->td_proc); 2523 return (EFBIG); 2524 } 2525 2526 /* 2527 * Helper for VOP_WRITE() implementations, the common code to 2528 * handle maximum supported file size on the filesystem, and 2529 * RLIMIT_FSIZE, except for special writes from accounting subsystem 2530 * and ktrace. 2531 * 2532 * For maximum file size (maxfsz argument): 2533 * - return EFBIG if uio_offset is beyond it 2534 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz. 2535 * 2536 * For RLIMIT_FSIZE: 2537 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit 2538 * - otherwise, clamp uio_resid if write would extend file beyond limit. 2539 * 2540 * If clamping occured, the adjustment for uio_resid is stored in 2541 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return 2542 * from the VOP. 2543 */ 2544 int 2545 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz, 2546 ssize_t *resid_adj, struct thread *td) 2547 { 2548 ssize_t resid_orig; 2549 int error; 2550 bool adj; 2551 2552 resid_orig = uio->uio_resid; 2553 adj = resid_adj != NULL; 2554 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td); 2555 if (adj) 2556 *resid_adj = resid_orig - uio->uio_resid; 2557 return (error); 2558 } 2559 2560 void 2561 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj) 2562 { 2563 uio->uio_resid += resid_adj; 2564 } 2565 2566 int 2567 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, 2568 struct thread *td) 2569 { 2570 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL, 2571 td)); 2572 } 2573 2574 int 2575 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 2576 struct thread *td) 2577 { 2578 struct vnode *vp; 2579 2580 vp = fp->f_vnode; 2581 #ifdef AUDIT 2582 vn_lock(vp, LK_SHARED | LK_RETRY); 2583 AUDIT_ARG_VNODE1(vp); 2584 VOP_UNLOCK(vp); 2585 #endif 2586 return (setfmode(td, active_cred, vp, mode)); 2587 } 2588 2589 int 2590 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 2591 struct thread *td) 2592 { 2593 struct vnode *vp; 2594 2595 vp = fp->f_vnode; 2596 #ifdef AUDIT 2597 vn_lock(vp, LK_SHARED | LK_RETRY); 2598 AUDIT_ARG_VNODE1(vp); 2599 VOP_UNLOCK(vp); 2600 #endif 2601 return (setfown(td, active_cred, vp, uid, gid)); 2602 } 2603 2604 /* 2605 * Remove pages in the range ["start", "end") from the vnode's VM object. If 2606 * "end" is 0, then the range extends to the end of the object. 2607 */ 2608 void 2609 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2610 { 2611 vm_object_t object; 2612 2613 if ((object = vp->v_object) == NULL) 2614 return; 2615 VM_OBJECT_WLOCK(object); 2616 vm_object_page_remove(object, start, end, 0); 2617 VM_OBJECT_WUNLOCK(object); 2618 } 2619 2620 /* 2621 * Like vn_pages_remove(), but skips invalid pages, which by definition are not 2622 * mapped into any process' address space. Filesystems may use this in 2623 * preference to vn_pages_remove() to avoid blocking on pages busied in 2624 * preparation for a VOP_GETPAGES. 2625 */ 2626 void 2627 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2628 { 2629 vm_object_t object; 2630 2631 if ((object = vp->v_object) == NULL) 2632 return; 2633 VM_OBJECT_WLOCK(object); 2634 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY); 2635 VM_OBJECT_WUNLOCK(object); 2636 } 2637 2638 int 2639 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off, 2640 struct ucred *cred) 2641 { 2642 off_t size; 2643 daddr_t bn, bnp; 2644 uint64_t bsize; 2645 off_t noff; 2646 int error; 2647 2648 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2649 ("%s: Wrong command %lu", __func__, cmd)); 2650 ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked"); 2651 2652 if (vp->v_type != VREG) { 2653 error = ENOTTY; 2654 goto out; 2655 } 2656 error = vn_getsize_locked(vp, &size, cred); 2657 if (error != 0) 2658 goto out; 2659 noff = *off; 2660 if (noff < 0 || noff >= size) { 2661 error = ENXIO; 2662 goto out; 2663 } 2664 2665 /* See the comment in ufs_bmap_seekdata(). */ 2666 vnode_pager_clean_sync(vp); 2667 2668 bsize = vp->v_mount->mnt_stat.f_iosize; 2669 for (bn = noff / bsize; noff < size; bn++, noff += bsize - 2670 noff % bsize) { 2671 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); 2672 if (error == EOPNOTSUPP) { 2673 error = ENOTTY; 2674 goto out; 2675 } 2676 if ((bnp == -1 && cmd == FIOSEEKHOLE) || 2677 (bnp != -1 && cmd == FIOSEEKDATA)) { 2678 noff = bn * bsize; 2679 if (noff < *off) 2680 noff = *off; 2681 goto out; 2682 } 2683 } 2684 if (noff > size) 2685 noff = size; 2686 /* noff == size. There is an implicit hole at the end of file. */ 2687 if (cmd == FIOSEEKDATA) 2688 error = ENXIO; 2689 out: 2690 if (error == 0) 2691 *off = noff; 2692 return (error); 2693 } 2694 2695 int 2696 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) 2697 { 2698 int error; 2699 2700 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2701 ("%s: Wrong command %lu", __func__, cmd)); 2702 2703 if (vn_lock(vp, LK_EXCLUSIVE) != 0) 2704 return (EBADF); 2705 error = vn_bmap_seekhole_locked(vp, cmd, off, cred); 2706 VOP_UNLOCK(vp); 2707 return (error); 2708 } 2709 2710 int 2711 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) 2712 { 2713 struct ucred *cred; 2714 struct vnode *vp; 2715 off_t foffset, fsize, size; 2716 int error, noneg; 2717 2718 cred = td->td_ucred; 2719 vp = fp->f_vnode; 2720 noneg = (vp->v_type != VCHR); 2721 /* 2722 * Try to dodge locking for common case of querying the offset. 2723 */ 2724 if (whence == L_INCR && offset == 0) { 2725 foffset = foffset_read(fp); 2726 if (__predict_false(foffset < 0 && noneg)) { 2727 return (EOVERFLOW); 2728 } 2729 td->td_uretoff.tdu_off = foffset; 2730 return (0); 2731 } 2732 foffset = foffset_lock(fp, 0); 2733 error = 0; 2734 switch (whence) { 2735 case L_INCR: 2736 if (noneg && 2737 (foffset < 0 || 2738 (offset > 0 && foffset > OFF_MAX - offset))) { 2739 error = EOVERFLOW; 2740 break; 2741 } 2742 offset += foffset; 2743 break; 2744 case L_XTND: 2745 error = vn_getsize(vp, &fsize, cred); 2746 if (error != 0) 2747 break; 2748 2749 /* 2750 * If the file references a disk device, then fetch 2751 * the media size and use that to determine the ending 2752 * offset. 2753 */ 2754 if (fsize == 0 && vp->v_type == VCHR && 2755 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) 2756 fsize = size; 2757 if (noneg && offset > 0 && fsize > OFF_MAX - offset) { 2758 error = EOVERFLOW; 2759 break; 2760 } 2761 offset += fsize; 2762 break; 2763 case L_SET: 2764 break; 2765 case SEEK_DATA: 2766 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); 2767 if (error == ENOTTY) 2768 error = EINVAL; 2769 break; 2770 case SEEK_HOLE: 2771 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); 2772 if (error == ENOTTY) 2773 error = EINVAL; 2774 break; 2775 default: 2776 error = EINVAL; 2777 } 2778 if (error == 0 && noneg && offset < 0) 2779 error = EINVAL; 2780 if (error != 0) 2781 goto drop; 2782 VFS_KNOTE_UNLOCKED(vp, 0); 2783 td->td_uretoff.tdu_off = offset; 2784 drop: 2785 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 2786 return (error); 2787 } 2788 2789 int 2790 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, 2791 struct thread *td) 2792 { 2793 int error; 2794 2795 /* 2796 * Grant permission if the caller is the owner of the file, or 2797 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on 2798 * on the file. If the time pointer is null, then write 2799 * permission on the file is also sufficient. 2800 * 2801 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: 2802 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES 2803 * will be allowed to set the times [..] to the current 2804 * server time. 2805 */ 2806 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); 2807 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) 2808 error = VOP_ACCESS(vp, VWRITE, cred, td); 2809 return (error); 2810 } 2811 2812 int 2813 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 2814 { 2815 struct vnode *vp; 2816 int error; 2817 2818 if (fp->f_type == DTYPE_FIFO) 2819 kif->kf_type = KF_TYPE_FIFO; 2820 else 2821 kif->kf_type = KF_TYPE_VNODE; 2822 vp = fp->f_vnode; 2823 vref(vp); 2824 FILEDESC_SUNLOCK(fdp); 2825 error = vn_fill_kinfo_vnode(vp, kif); 2826 vrele(vp); 2827 FILEDESC_SLOCK(fdp); 2828 return (error); 2829 } 2830 2831 static inline void 2832 vn_fill_junk(struct kinfo_file *kif) 2833 { 2834 size_t len, olen; 2835 2836 /* 2837 * Simulate vn_fullpath returning changing values for a given 2838 * vp during e.g. coredump. 2839 */ 2840 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; 2841 olen = strlen(kif->kf_path); 2842 if (len < olen) 2843 strcpy(&kif->kf_path[len - 1], "$"); 2844 else 2845 for (; olen < len; olen++) 2846 strcpy(&kif->kf_path[olen], "A"); 2847 } 2848 2849 int 2850 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) 2851 { 2852 struct vattr va; 2853 char *fullpath, *freepath; 2854 int error; 2855 2856 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); 2857 freepath = NULL; 2858 fullpath = "-"; 2859 error = vn_fullpath(vp, &fullpath, &freepath); 2860 if (error == 0) { 2861 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); 2862 } 2863 if (freepath != NULL) 2864 free(freepath, M_TEMP); 2865 2866 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, 2867 vn_fill_junk(kif); 2868 ); 2869 2870 /* 2871 * Retrieve vnode attributes. 2872 */ 2873 va.va_fsid = VNOVAL; 2874 va.va_rdev = NODEV; 2875 vn_lock(vp, LK_SHARED | LK_RETRY); 2876 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 2877 VOP_UNLOCK(vp); 2878 if (error != 0) 2879 return (error); 2880 if (va.va_fsid != VNOVAL) 2881 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; 2882 else 2883 kif->kf_un.kf_file.kf_file_fsid = 2884 vp->v_mount->mnt_stat.f_fsid.val[0]; 2885 kif->kf_un.kf_file.kf_file_fsid_freebsd11 = 2886 kif->kf_un.kf_file.kf_file_fsid; /* truncate */ 2887 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; 2888 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); 2889 kif->kf_un.kf_file.kf_file_size = va.va_size; 2890 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; 2891 kif->kf_un.kf_file.kf_file_rdev_freebsd11 = 2892 kif->kf_un.kf_file.kf_file_rdev; /* truncate */ 2893 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink; 2894 return (0); 2895 } 2896 2897 int 2898 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, 2899 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, 2900 struct thread *td) 2901 { 2902 #ifdef HWPMC_HOOKS 2903 struct pmckern_map_in pkm; 2904 #endif 2905 struct mount *mp; 2906 struct vnode *vp; 2907 vm_object_t object; 2908 vm_prot_t maxprot; 2909 boolean_t writecounted; 2910 int error; 2911 2912 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ 2913 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) 2914 /* 2915 * POSIX shared-memory objects are defined to have 2916 * kernel persistence, and are not defined to support 2917 * read(2)/write(2) -- or even open(2). Thus, we can 2918 * use MAP_ASYNC to trade on-disk coherence for speed. 2919 * The shm_open(3) library routine turns on the FPOSIXSHM 2920 * flag to request this behavior. 2921 */ 2922 if ((fp->f_flag & FPOSIXSHM) != 0) 2923 flags |= MAP_NOSYNC; 2924 #endif 2925 vp = fp->f_vnode; 2926 2927 /* 2928 * Ensure that file and memory protections are 2929 * compatible. Note that we only worry about 2930 * writability if mapping is shared; in this case, 2931 * current and max prot are dictated by the open file. 2932 * XXX use the vnode instead? Problem is: what 2933 * credentials do we use for determination? What if 2934 * proc does a setuid? 2935 */ 2936 mp = vp->v_mount; 2937 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { 2938 maxprot = VM_PROT_NONE; 2939 if ((prot & VM_PROT_EXECUTE) != 0) 2940 return (EACCES); 2941 } else 2942 maxprot = VM_PROT_EXECUTE; 2943 if ((fp->f_flag & FREAD) != 0) 2944 maxprot |= VM_PROT_READ; 2945 else if ((prot & VM_PROT_READ) != 0) 2946 return (EACCES); 2947 2948 /* 2949 * If we are sharing potential changes via MAP_SHARED and we 2950 * are trying to get write permission although we opened it 2951 * without asking for it, bail out. 2952 */ 2953 if ((flags & MAP_SHARED) != 0) { 2954 if ((fp->f_flag & FWRITE) != 0) 2955 maxprot |= VM_PROT_WRITE; 2956 else if ((prot & VM_PROT_WRITE) != 0) 2957 return (EACCES); 2958 } else { 2959 maxprot |= VM_PROT_WRITE; 2960 cap_maxprot |= VM_PROT_WRITE; 2961 } 2962 maxprot &= cap_maxprot; 2963 2964 /* 2965 * For regular files and shared memory, POSIX requires that 2966 * the value of foff be a legitimate offset within the data 2967 * object. In particular, negative offsets are invalid. 2968 * Blocking negative offsets and overflows here avoids 2969 * possible wraparound or user-level access into reserved 2970 * ranges of the data object later. In contrast, POSIX does 2971 * not dictate how offsets are used by device drivers, so in 2972 * the case of a device mapping a negative offset is passed 2973 * on. 2974 */ 2975 if ( 2976 #ifdef _LP64 2977 size > OFF_MAX || 2978 #endif 2979 foff > OFF_MAX - size) 2980 return (EINVAL); 2981 2982 writecounted = FALSE; 2983 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, 2984 &foff, &object, &writecounted); 2985 if (error != 0) 2986 return (error); 2987 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, 2988 foff, writecounted, td); 2989 if (error != 0) { 2990 /* 2991 * If this mapping was accounted for in the vnode's 2992 * writecount, then undo that now. 2993 */ 2994 if (writecounted) 2995 vm_pager_release_writecount(object, 0, size); 2996 vm_object_deallocate(object); 2997 } 2998 #ifdef HWPMC_HOOKS 2999 /* Inform hwpmc(4) if an executable is being mapped. */ 3000 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { 3001 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { 3002 pkm.pm_file = vp; 3003 pkm.pm_address = (uintptr_t) *addr; 3004 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); 3005 } 3006 } 3007 #endif 3008 return (error); 3009 } 3010 3011 void 3012 vn_fsid(struct vnode *vp, struct vattr *va) 3013 { 3014 fsid_t *f; 3015 3016 f = &vp->v_mount->mnt_stat.f_fsid; 3017 va->va_fsid = (uint32_t)f->val[1]; 3018 va->va_fsid <<= sizeof(f->val[1]) * NBBY; 3019 va->va_fsid += (uint32_t)f->val[0]; 3020 } 3021 3022 int 3023 vn_fsync_buf(struct vnode *vp, int waitfor) 3024 { 3025 struct buf *bp, *nbp; 3026 struct bufobj *bo; 3027 struct mount *mp; 3028 int error, maxretry; 3029 3030 error = 0; 3031 maxretry = 10000; /* large, arbitrarily chosen */ 3032 mp = NULL; 3033 if (vp->v_type == VCHR) { 3034 VI_LOCK(vp); 3035 mp = vp->v_rdev->si_mountpt; 3036 VI_UNLOCK(vp); 3037 } 3038 bo = &vp->v_bufobj; 3039 BO_LOCK(bo); 3040 loop1: 3041 /* 3042 * MARK/SCAN initialization to avoid infinite loops. 3043 */ 3044 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 3045 bp->b_vflags &= ~BV_SCANNED; 3046 bp->b_error = 0; 3047 } 3048 3049 /* 3050 * Flush all dirty buffers associated with a vnode. 3051 */ 3052 loop2: 3053 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 3054 if ((bp->b_vflags & BV_SCANNED) != 0) 3055 continue; 3056 bp->b_vflags |= BV_SCANNED; 3057 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { 3058 if (waitfor != MNT_WAIT) 3059 continue; 3060 if (BUF_LOCK(bp, 3061 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, 3062 BO_LOCKPTR(bo)) != 0) { 3063 BO_LOCK(bo); 3064 goto loop1; 3065 } 3066 BO_LOCK(bo); 3067 } 3068 BO_UNLOCK(bo); 3069 KASSERT(bp->b_bufobj == bo, 3070 ("bp %p wrong b_bufobj %p should be %p", 3071 bp, bp->b_bufobj, bo)); 3072 if ((bp->b_flags & B_DELWRI) == 0) 3073 panic("fsync: not dirty"); 3074 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { 3075 vfs_bio_awrite(bp); 3076 } else { 3077 bremfree(bp); 3078 bawrite(bp); 3079 } 3080 if (maxretry < 1000) 3081 pause("dirty", hz < 1000 ? 1 : hz / 1000); 3082 BO_LOCK(bo); 3083 goto loop2; 3084 } 3085 3086 /* 3087 * If synchronous the caller expects us to completely resolve all 3088 * dirty buffers in the system. Wait for in-progress I/O to 3089 * complete (which could include background bitmap writes), then 3090 * retry if dirty blocks still exist. 3091 */ 3092 if (waitfor == MNT_WAIT) { 3093 bufobj_wwait(bo, 0, 0); 3094 if (bo->bo_dirty.bv_cnt > 0) { 3095 /* 3096 * If we are unable to write any of these buffers 3097 * then we fail now rather than trying endlessly 3098 * to write them out. 3099 */ 3100 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 3101 if ((error = bp->b_error) != 0) 3102 break; 3103 if ((mp != NULL && mp->mnt_secondary_writes > 0) || 3104 (error == 0 && --maxretry >= 0)) 3105 goto loop1; 3106 if (error == 0) 3107 error = EAGAIN; 3108 } 3109 } 3110 BO_UNLOCK(bo); 3111 if (error != 0) 3112 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); 3113 3114 return (error); 3115 } 3116 3117 /* 3118 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() 3119 * or vn_generic_copy_file_range() after rangelocking the byte ranges, 3120 * to do the actual copy. 3121 * vn_generic_copy_file_range() is factored out, so it can be called 3122 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from 3123 * different file systems. 3124 */ 3125 int 3126 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, 3127 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, 3128 struct ucred *outcred, struct thread *fsize_td) 3129 { 3130 struct mount *inmp, *outmp; 3131 struct vnode *invpl, *outvpl; 3132 int error; 3133 size_t len; 3134 uint64_t uval; 3135 3136 invpl = outvpl = NULL; 3137 len = *lenp; 3138 *lenp = 0; /* For error returns. */ 3139 error = 0; 3140 3141 /* Do some sanity checks on the arguments. */ 3142 if (invp->v_type == VDIR || outvp->v_type == VDIR) 3143 error = EISDIR; 3144 else if (*inoffp < 0 || *outoffp < 0 || 3145 invp->v_type != VREG || outvp->v_type != VREG) 3146 error = EINVAL; 3147 if (error != 0) 3148 goto out; 3149 3150 /* Ensure offset + len does not wrap around. */ 3151 uval = *inoffp; 3152 uval += len; 3153 if (uval > INT64_MAX) 3154 len = INT64_MAX - *inoffp; 3155 uval = *outoffp; 3156 uval += len; 3157 if (uval > INT64_MAX) 3158 len = INT64_MAX - *outoffp; 3159 if (len == 0) 3160 goto out; 3161 3162 error = VOP_GETLOWVNODE(invp, &invpl, FREAD); 3163 if (error != 0) 3164 goto out; 3165 error = VOP_GETLOWVNODE(outvp, &outvpl, FWRITE); 3166 if (error != 0) 3167 goto out1; 3168 3169 inmp = invpl->v_mount; 3170 outmp = outvpl->v_mount; 3171 if (inmp == NULL || outmp == NULL) 3172 goto out2; 3173 3174 for (;;) { 3175 error = vfs_busy(inmp, 0); 3176 if (error != 0) 3177 goto out2; 3178 if (inmp == outmp) 3179 break; 3180 error = vfs_busy(outmp, MBF_NOWAIT); 3181 if (error != 0) { 3182 vfs_unbusy(inmp); 3183 error = vfs_busy(outmp, 0); 3184 if (error == 0) { 3185 vfs_unbusy(outmp); 3186 continue; 3187 } 3188 goto out2; 3189 } 3190 break; 3191 } 3192 3193 /* 3194 * If the two vnodes are for the same file system type, call 3195 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() 3196 * which can handle copies across multiple file system types. 3197 */ 3198 *lenp = len; 3199 if (inmp == outmp || inmp->mnt_vfc == outmp->mnt_vfc) 3200 error = VOP_COPY_FILE_RANGE(invpl, inoffp, outvpl, outoffp, 3201 lenp, flags, incred, outcred, fsize_td); 3202 else 3203 error = ENOSYS; 3204 if (error == ENOSYS) 3205 error = vn_generic_copy_file_range(invpl, inoffp, outvpl, 3206 outoffp, lenp, flags, incred, outcred, fsize_td); 3207 vfs_unbusy(outmp); 3208 if (inmp != outmp) 3209 vfs_unbusy(inmp); 3210 out2: 3211 if (outvpl != NULL) 3212 vrele(outvpl); 3213 out1: 3214 if (invpl != NULL) 3215 vrele(invpl); 3216 out: 3217 return (error); 3218 } 3219 3220 /* 3221 * Test len bytes of data starting at dat for all bytes == 0. 3222 * Return true if all bytes are zero, false otherwise. 3223 * Expects dat to be well aligned. 3224 */ 3225 static bool 3226 mem_iszero(void *dat, int len) 3227 { 3228 int i; 3229 const u_int *p; 3230 const char *cp; 3231 3232 for (p = dat; len > 0; len -= sizeof(*p), p++) { 3233 if (len >= sizeof(*p)) { 3234 if (*p != 0) 3235 return (false); 3236 } else { 3237 cp = (const char *)p; 3238 for (i = 0; i < len; i++, cp++) 3239 if (*cp != '\0') 3240 return (false); 3241 } 3242 } 3243 return (true); 3244 } 3245 3246 /* 3247 * Look for a hole in the output file and, if found, adjust *outoffp 3248 * and *xferp to skip past the hole. 3249 * *xferp is the entire hole length to be written and xfer2 is how many bytes 3250 * to be written as 0's upon return. 3251 */ 3252 static off_t 3253 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, 3254 off_t *dataoffp, off_t *holeoffp, struct ucred *cred) 3255 { 3256 int error; 3257 off_t delta; 3258 3259 if (*holeoffp == 0 || *holeoffp <= *outoffp) { 3260 *dataoffp = *outoffp; 3261 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, 3262 curthread); 3263 if (error == 0) { 3264 *holeoffp = *dataoffp; 3265 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, 3266 curthread); 3267 } 3268 if (error != 0 || *holeoffp == *dataoffp) { 3269 /* 3270 * Since outvp is unlocked, it may be possible for 3271 * another thread to do a truncate(), lseek(), write() 3272 * creating a hole at startoff between the above 3273 * VOP_IOCTL() calls, if the other thread does not do 3274 * rangelocking. 3275 * If that happens, *holeoffp == *dataoffp and finding 3276 * the hole has failed, so disable vn_skip_hole(). 3277 */ 3278 *holeoffp = -1; /* Disable use of vn_skip_hole(). */ 3279 return (xfer2); 3280 } 3281 KASSERT(*dataoffp >= *outoffp, 3282 ("vn_skip_hole: dataoff=%jd < outoff=%jd", 3283 (intmax_t)*dataoffp, (intmax_t)*outoffp)); 3284 KASSERT(*holeoffp > *dataoffp, 3285 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", 3286 (intmax_t)*holeoffp, (intmax_t)*dataoffp)); 3287 } 3288 3289 /* 3290 * If there is a hole before the data starts, advance *outoffp and 3291 * *xferp past the hole. 3292 */ 3293 if (*dataoffp > *outoffp) { 3294 delta = *dataoffp - *outoffp; 3295 if (delta >= *xferp) { 3296 /* Entire *xferp is a hole. */ 3297 *outoffp += *xferp; 3298 *xferp = 0; 3299 return (0); 3300 } 3301 *xferp -= delta; 3302 *outoffp += delta; 3303 xfer2 = MIN(xfer2, *xferp); 3304 } 3305 3306 /* 3307 * If a hole starts before the end of this xfer2, reduce this xfer2 so 3308 * that the write ends at the start of the hole. 3309 * *holeoffp should always be greater than *outoffp, but for the 3310 * non-INVARIANTS case, check this to make sure xfer2 remains a sane 3311 * value. 3312 */ 3313 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) 3314 xfer2 = *holeoffp - *outoffp; 3315 return (xfer2); 3316 } 3317 3318 /* 3319 * Write an xfer sized chunk to outvp in blksize blocks from dat. 3320 * dat is a maximum of blksize in length and can be written repeatedly in 3321 * the chunk. 3322 * If growfile == true, just grow the file via vn_truncate_locked() instead 3323 * of doing actual writes. 3324 * If checkhole == true, a hole is being punched, so skip over any hole 3325 * already in the output file. 3326 */ 3327 static int 3328 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, 3329 u_long blksize, bool growfile, bool checkhole, struct ucred *cred) 3330 { 3331 struct mount *mp; 3332 off_t dataoff, holeoff, xfer2; 3333 int error; 3334 3335 /* 3336 * Loop around doing writes of blksize until write has been completed. 3337 * Lock/unlock on each loop iteration so that a bwillwrite() can be 3338 * done for each iteration, since the xfer argument can be very 3339 * large if there is a large hole to punch in the output file. 3340 */ 3341 error = 0; 3342 holeoff = 0; 3343 do { 3344 xfer2 = MIN(xfer, blksize); 3345 if (checkhole) { 3346 /* 3347 * Punching a hole. Skip writing if there is 3348 * already a hole in the output file. 3349 */ 3350 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, 3351 &dataoff, &holeoff, cred); 3352 if (xfer == 0) 3353 break; 3354 if (holeoff < 0) 3355 checkhole = false; 3356 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", 3357 (intmax_t)xfer2)); 3358 } 3359 bwillwrite(); 3360 mp = NULL; 3361 error = vn_start_write(outvp, &mp, V_WAIT); 3362 if (error != 0) 3363 break; 3364 if (growfile) { 3365 error = vn_lock(outvp, LK_EXCLUSIVE); 3366 if (error == 0) { 3367 error = vn_truncate_locked(outvp, outoff + xfer, 3368 false, cred); 3369 VOP_UNLOCK(outvp); 3370 } 3371 } else { 3372 error = vn_lock(outvp, vn_lktype_write(mp, outvp)); 3373 if (error == 0) { 3374 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, 3375 outoff, UIO_SYSSPACE, IO_NODELOCKED, 3376 curthread->td_ucred, cred, NULL, curthread); 3377 outoff += xfer2; 3378 xfer -= xfer2; 3379 VOP_UNLOCK(outvp); 3380 } 3381 } 3382 if (mp != NULL) 3383 vn_finished_write(mp); 3384 } while (!growfile && xfer > 0 && error == 0); 3385 return (error); 3386 } 3387 3388 /* 3389 * Copy a byte range of one file to another. This function can handle the 3390 * case where invp and outvp are on different file systems. 3391 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there 3392 * is no better file system specific way to do it. 3393 */ 3394 int 3395 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, 3396 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, 3397 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) 3398 { 3399 struct vattr inva; 3400 struct mount *mp; 3401 off_t startoff, endoff, xfer, xfer2; 3402 u_long blksize; 3403 int error, interrupted; 3404 bool cantseek, readzeros, eof, first, lastblock, holetoeof, sparse; 3405 ssize_t aresid, r = 0; 3406 size_t copylen, len, savlen; 3407 off_t outsize; 3408 char *dat; 3409 long holein, holeout; 3410 struct timespec curts, endts; 3411 3412 holein = holeout = 0; 3413 savlen = len = *lenp; 3414 error = 0; 3415 interrupted = 0; 3416 dat = NULL; 3417 3418 error = vn_lock(invp, LK_SHARED); 3419 if (error != 0) 3420 goto out; 3421 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) 3422 holein = 0; 3423 error = VOP_GETATTR(invp, &inva, incred); 3424 if (error == 0 && inva.va_size > OFF_MAX) 3425 error = EFBIG; 3426 VOP_UNLOCK(invp); 3427 if (error != 0) 3428 goto out; 3429 3430 /* 3431 * Use va_bytes >= va_size as a hint that the file does not have 3432 * sufficient holes to justify the overhead of doing FIOSEEKHOLE. 3433 * This hint does not work well for file systems doing compression 3434 * and may fail when allocations for extended attributes increases 3435 * the value of va_bytes to >= va_size. 3436 */ 3437 sparse = true; 3438 if (holein != 0 && inva.va_bytes >= inva.va_size) { 3439 holein = 0; 3440 sparse = false; 3441 } 3442 3443 mp = NULL; 3444 error = vn_start_write(outvp, &mp, V_WAIT); 3445 if (error == 0) 3446 error = vn_lock(outvp, LK_EXCLUSIVE); 3447 if (error == 0) { 3448 /* 3449 * If fsize_td != NULL, do a vn_rlimit_fsizex() call, 3450 * now that outvp is locked. 3451 */ 3452 if (fsize_td != NULL) { 3453 struct uio io; 3454 3455 io.uio_offset = *outoffp; 3456 io.uio_resid = len; 3457 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td); 3458 len = savlen = io.uio_resid; 3459 /* 3460 * No need to call vn_rlimit_fsizex_res before return, 3461 * since the uio is local. 3462 */ 3463 } 3464 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) 3465 holeout = 0; 3466 /* 3467 * Holes that are past EOF do not need to be written as a block 3468 * of zero bytes. So, truncate the output file as far as 3469 * possible and then use size to decide if writing 0 3470 * bytes is necessary in the loop below. 3471 */ 3472 if (error == 0) 3473 error = vn_getsize_locked(outvp, &outsize, outcred); 3474 if (error == 0 && outsize > *outoffp && 3475 *outoffp <= OFF_MAX - len && outsize <= *outoffp + len && 3476 *inoffp < inva.va_size && 3477 *outoffp <= OFF_MAX - (inva.va_size - *inoffp) && 3478 outsize <= *outoffp + (inva.va_size - *inoffp)) { 3479 #ifdef MAC 3480 error = mac_vnode_check_write(curthread->td_ucred, 3481 outcred, outvp); 3482 if (error == 0) 3483 #endif 3484 error = vn_truncate_locked(outvp, *outoffp, 3485 false, outcred); 3486 if (error == 0) 3487 outsize = *outoffp; 3488 } 3489 VOP_UNLOCK(outvp); 3490 } 3491 if (mp != NULL) 3492 vn_finished_write(mp); 3493 if (error != 0) 3494 goto out; 3495 3496 if (sparse && holein == 0 && holeout > 0) { 3497 /* 3498 * For this special case, the input data will be scanned 3499 * for blocks of all 0 bytes. For these blocks, the 3500 * write can be skipped for the output file to create 3501 * an unallocated region. 3502 * Therefore, use the appropriate size for the output file. 3503 */ 3504 blksize = holeout; 3505 if (blksize <= 512) { 3506 /* 3507 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE 3508 * of 512, although it actually only creates 3509 * unallocated regions for blocks >= f_iosize. 3510 */ 3511 blksize = outvp->v_mount->mnt_stat.f_iosize; 3512 } 3513 } else { 3514 /* 3515 * Use the larger of the two f_iosize values. If they are 3516 * not the same size, one will normally be an exact multiple of 3517 * the other, since they are both likely to be a power of 2. 3518 */ 3519 blksize = MAX(invp->v_mount->mnt_stat.f_iosize, 3520 outvp->v_mount->mnt_stat.f_iosize); 3521 } 3522 3523 /* Clip to sane limits. */ 3524 if (blksize < 4096) 3525 blksize = 4096; 3526 else if (blksize > maxphys) 3527 blksize = maxphys; 3528 dat = malloc(blksize, M_TEMP, M_WAITOK); 3529 3530 /* 3531 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA 3532 * to find holes. Otherwise, just scan the read block for all 0s 3533 * in the inner loop where the data copying is done. 3534 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may 3535 * support holes on the server, but do not support FIOSEEKHOLE. 3536 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate 3537 * that this function should return after 1second with a partial 3538 * completion. 3539 */ 3540 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) { 3541 getnanouptime(&endts); 3542 endts.tv_sec++; 3543 } else 3544 timespecclear(&endts); 3545 first = true; 3546 holetoeof = eof = false; 3547 while (len > 0 && error == 0 && !eof && interrupted == 0) { 3548 endoff = 0; /* To shut up compilers. */ 3549 cantseek = true; 3550 startoff = *inoffp; 3551 copylen = len; 3552 3553 /* 3554 * Find the next data area. If there is just a hole to EOF, 3555 * FIOSEEKDATA should fail with ENXIO. 3556 * (I do not know if any file system will report a hole to 3557 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA 3558 * will fail for those file systems.) 3559 * 3560 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, 3561 * the code just falls through to the inner copy loop. 3562 */ 3563 error = EINVAL; 3564 if (holein > 0) { 3565 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, 3566 incred, curthread); 3567 if (error == ENXIO) { 3568 startoff = endoff = inva.va_size; 3569 eof = holetoeof = true; 3570 error = 0; 3571 } 3572 } 3573 if (error == 0 && !holetoeof) { 3574 endoff = startoff; 3575 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, 3576 incred, curthread); 3577 /* 3578 * Since invp is unlocked, it may be possible for 3579 * another thread to do a truncate(), lseek(), write() 3580 * creating a hole at startoff between the above 3581 * VOP_IOCTL() calls, if the other thread does not do 3582 * rangelocking. 3583 * If that happens, startoff == endoff and finding 3584 * the hole has failed, so set an error. 3585 */ 3586 if (error == 0 && startoff == endoff) 3587 error = EINVAL; /* Any error. Reset to 0. */ 3588 } 3589 if (error == 0) { 3590 if (startoff > *inoffp) { 3591 /* Found hole before data block. */ 3592 xfer = MIN(startoff - *inoffp, len); 3593 if (*outoffp < outsize) { 3594 /* Must write 0s to punch hole. */ 3595 xfer2 = MIN(outsize - *outoffp, 3596 xfer); 3597 memset(dat, 0, MIN(xfer2, blksize)); 3598 error = vn_write_outvp(outvp, dat, 3599 *outoffp, xfer2, blksize, false, 3600 holeout > 0, outcred); 3601 } 3602 3603 if (error == 0 && *outoffp + xfer > 3604 outsize && (xfer == len || holetoeof)) { 3605 /* Grow output file (hole at end). */ 3606 error = vn_write_outvp(outvp, dat, 3607 *outoffp, xfer, blksize, true, 3608 false, outcred); 3609 } 3610 if (error == 0) { 3611 *inoffp += xfer; 3612 *outoffp += xfer; 3613 len -= xfer; 3614 if (len < savlen) { 3615 interrupted = sig_intr(); 3616 if (timespecisset(&endts) && 3617 interrupted == 0) { 3618 getnanouptime(&curts); 3619 if (timespeccmp(&curts, 3620 &endts, >=)) 3621 interrupted = 3622 EINTR; 3623 } 3624 } 3625 } 3626 } 3627 copylen = MIN(len, endoff - startoff); 3628 cantseek = false; 3629 } else { 3630 cantseek = true; 3631 if (!sparse) 3632 cantseek = false; 3633 startoff = *inoffp; 3634 copylen = len; 3635 error = 0; 3636 } 3637 3638 xfer = blksize; 3639 if (cantseek) { 3640 /* 3641 * Set first xfer to end at a block boundary, so that 3642 * holes are more likely detected in the loop below via 3643 * the for all bytes 0 method. 3644 */ 3645 xfer -= (*inoffp % blksize); 3646 } 3647 3648 /* 3649 * Loop copying the data block. If this was our first attempt 3650 * to copy anything, allow a zero-length block so that the VOPs 3651 * get a chance to update metadata, specifically the atime. 3652 */ 3653 while (error == 0 && ((copylen > 0 && !eof) || first) && 3654 interrupted == 0) { 3655 if (copylen < xfer) 3656 xfer = copylen; 3657 first = false; 3658 error = vn_lock(invp, LK_SHARED); 3659 if (error != 0) 3660 goto out; 3661 error = vn_rdwr(UIO_READ, invp, dat, xfer, 3662 startoff, UIO_SYSSPACE, IO_NODELOCKED, 3663 curthread->td_ucred, incred, &aresid, 3664 curthread); 3665 VOP_UNLOCK(invp); 3666 lastblock = false; 3667 if (error == 0 && (xfer == 0 || aresid > 0)) { 3668 /* Stop the copy at EOF on the input file. */ 3669 xfer -= aresid; 3670 eof = true; 3671 lastblock = true; 3672 } 3673 if (error == 0) { 3674 /* 3675 * Skip the write for holes past the initial EOF 3676 * of the output file, unless this is the last 3677 * write of the output file at EOF. 3678 */ 3679 readzeros = cantseek ? mem_iszero(dat, xfer) : 3680 false; 3681 if (xfer == len) 3682 lastblock = true; 3683 if (!cantseek || *outoffp < outsize || 3684 lastblock || !readzeros) 3685 error = vn_write_outvp(outvp, dat, 3686 *outoffp, xfer, blksize, 3687 readzeros && lastblock && 3688 *outoffp >= outsize, false, 3689 outcred); 3690 if (error == 0) { 3691 *inoffp += xfer; 3692 startoff += xfer; 3693 *outoffp += xfer; 3694 copylen -= xfer; 3695 len -= xfer; 3696 if (len < savlen) { 3697 interrupted = sig_intr(); 3698 if (timespecisset(&endts) && 3699 interrupted == 0) { 3700 getnanouptime(&curts); 3701 if (timespeccmp(&curts, 3702 &endts, >=)) 3703 interrupted = 3704 EINTR; 3705 } 3706 } 3707 } 3708 } 3709 xfer = blksize; 3710 } 3711 } 3712 out: 3713 *lenp = savlen - len; 3714 free(dat, M_TEMP); 3715 return (error); 3716 } 3717 3718 static int 3719 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 3720 { 3721 struct mount *mp; 3722 struct vnode *vp; 3723 off_t olen, ooffset; 3724 int error; 3725 #ifdef AUDIT 3726 int audited_vnode1 = 0; 3727 #endif 3728 3729 vp = fp->f_vnode; 3730 if (vp->v_type != VREG) 3731 return (ENODEV); 3732 3733 /* Allocating blocks may take a long time, so iterate. */ 3734 for (;;) { 3735 olen = len; 3736 ooffset = offset; 3737 3738 bwillwrite(); 3739 mp = NULL; 3740 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 3741 if (error != 0) 3742 break; 3743 error = vn_lock(vp, LK_EXCLUSIVE); 3744 if (error != 0) { 3745 vn_finished_write(mp); 3746 break; 3747 } 3748 #ifdef AUDIT 3749 if (!audited_vnode1) { 3750 AUDIT_ARG_VNODE1(vp); 3751 audited_vnode1 = 1; 3752 } 3753 #endif 3754 #ifdef MAC 3755 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); 3756 if (error == 0) 3757 #endif 3758 error = VOP_ALLOCATE(vp, &offset, &len, 0, 3759 td->td_ucred); 3760 VOP_UNLOCK(vp); 3761 vn_finished_write(mp); 3762 3763 if (olen + ooffset != offset + len) { 3764 panic("offset + len changed from %jx/%jx to %jx/%jx", 3765 ooffset, olen, offset, len); 3766 } 3767 if (error != 0 || len == 0) 3768 break; 3769 KASSERT(olen > len, ("Iteration did not make progress?")); 3770 maybe_yield(); 3771 } 3772 3773 return (error); 3774 } 3775 3776 static int 3777 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags, 3778 int ioflag, struct ucred *cred, struct ucred *active_cred, 3779 struct ucred *file_cred) 3780 { 3781 struct mount *mp; 3782 void *rl_cookie; 3783 off_t off, len; 3784 int error; 3785 #ifdef AUDIT 3786 bool audited_vnode1 = false; 3787 #endif 3788 3789 rl_cookie = NULL; 3790 error = 0; 3791 mp = NULL; 3792 off = *offset; 3793 len = *length; 3794 3795 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0) 3796 rl_cookie = vn_rangelock_wlock(vp, off, off + len); 3797 while (len > 0 && error == 0) { 3798 /* 3799 * Try to deallocate the longest range in one pass. 3800 * In case a pass takes too long to be executed, it returns 3801 * partial result. The residue will be proceeded in the next 3802 * pass. 3803 */ 3804 3805 if ((ioflag & IO_NODELOCKED) == 0) { 3806 bwillwrite(); 3807 if ((error = vn_start_write(vp, &mp, 3808 V_WAIT | V_PCATCH)) != 0) 3809 goto out; 3810 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY); 3811 } 3812 #ifdef AUDIT 3813 if (!audited_vnode1) { 3814 AUDIT_ARG_VNODE1(vp); 3815 audited_vnode1 = true; 3816 } 3817 #endif 3818 3819 #ifdef MAC 3820 if ((ioflag & IO_NOMACCHECK) == 0) 3821 error = mac_vnode_check_write(active_cred, file_cred, 3822 vp); 3823 #endif 3824 if (error == 0) 3825 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag, 3826 cred); 3827 3828 if ((ioflag & IO_NODELOCKED) == 0) { 3829 VOP_UNLOCK(vp); 3830 if (mp != NULL) { 3831 vn_finished_write(mp); 3832 mp = NULL; 3833 } 3834 } 3835 if (error == 0 && len != 0) 3836 maybe_yield(); 3837 } 3838 out: 3839 if (rl_cookie != NULL) 3840 vn_rangelock_unlock(vp, rl_cookie); 3841 *offset = off; 3842 *length = len; 3843 return (error); 3844 } 3845 3846 /* 3847 * This function is supposed to be used in the situations where the deallocation 3848 * is not triggered by a user request. 3849 */ 3850 int 3851 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags, 3852 int ioflag, struct ucred *active_cred, struct ucred *file_cred) 3853 { 3854 struct ucred *cred; 3855 3856 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset || 3857 flags != 0) 3858 return (EINVAL); 3859 if (vp->v_type != VREG) 3860 return (ENODEV); 3861 3862 cred = file_cred != NOCRED ? file_cred : active_cred; 3863 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred, 3864 active_cred, file_cred)); 3865 } 3866 3867 static int 3868 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, 3869 struct ucred *active_cred, struct thread *td) 3870 { 3871 int error; 3872 struct vnode *vp; 3873 int ioflag; 3874 3875 KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd")); 3876 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, 3877 ("vn_fspacectl: non-zero flags")); 3878 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, 3879 ("vn_fspacectl: offset/length overflow or underflow")); 3880 vp = fp->f_vnode; 3881 3882 if (vp->v_type != VREG) 3883 return (ENODEV); 3884 3885 ioflag = get_write_ioflag(fp); 3886 3887 switch (cmd) { 3888 case SPACECTL_DEALLOC: 3889 error = vn_deallocate_impl(vp, offset, length, flags, ioflag, 3890 active_cred, active_cred, fp->f_cred); 3891 break; 3892 default: 3893 panic("vn_fspacectl: unknown cmd %d", cmd); 3894 } 3895 3896 return (error); 3897 } 3898 3899 /* 3900 * Keep this assert as long as sizeof(struct dirent) is used as the maximum 3901 * entry size. 3902 */ 3903 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent), 3904 "'struct dirent' size must be a multiple of its alignment " 3905 "(see _GENERIC_DIRLEN())"); 3906 3907 /* 3908 * Returns successive directory entries through some caller's provided buffer. 3909 * 3910 * This function automatically refills the provided buffer with calls to 3911 * VOP_READDIR() (after MAC permission checks). 3912 * 3913 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the 3914 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must 3915 * be properly aligned to access 'struct dirent' structures and 'dirbuflen' 3916 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning 3917 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always 3918 * be returned if this requirement is not verified). '*dpp' points to the 3919 * current directory entry in the buffer and '*len' contains the remaining 3920 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry). 3921 * 3922 * At first call (or when restarting the read), '*len' must have been set to 0, 3923 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no 3924 * more entries as soon as '*len' is 0 after a call that returned 0. Calling 3925 * again this function after such a condition is considered an error and EINVAL 3926 * will be returned. Other possible error codes are those of VOP_READDIR(), 3927 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL 3928 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off' 3929 * and '*eofflag') must be re-initialized before a subsequent call. On error 3930 * or at end of directory, '*dpp' is reset to NULL. 3931 * 3932 * '*len', '*off' and '*eofflag' are internal state the caller should not 3933 * tamper with except as explained above. '*off' is the next directory offset 3934 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last 3935 * internal call to VOP_READDIR() that returned without error, indicating 3936 * whether it reached the end of the directory, and to 2 by this function after 3937 * all entries have been read. 3938 */ 3939 int 3940 vn_dir_next_dirent(struct vnode *vp, struct thread *td, 3941 char *dirbuf, size_t dirbuflen, 3942 struct dirent **dpp, size_t *len, off_t *off, int *eofflag) 3943 { 3944 struct dirent *dp = NULL; 3945 int reclen; 3946 int error; 3947 struct uio uio; 3948 struct iovec iov; 3949 3950 ASSERT_VOP_LOCKED(vp, "vnode not locked"); 3951 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory")); 3952 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen, 3953 "Address space overflow"); 3954 3955 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) { 3956 /* Don't take any chances in this case */ 3957 error = EINVAL; 3958 goto out; 3959 } 3960 3961 if (*len != 0) { 3962 dp = *dpp; 3963 3964 /* 3965 * The caller continued to call us after an error (we set dp to 3966 * NULL in a previous iteration). Bail out right now. 3967 */ 3968 if (__predict_false(dp == NULL)) 3969 return (EINVAL); 3970 3971 MPASS(*len <= dirbuflen); 3972 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp && 3973 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen, 3974 "Filled range not inside buffer"); 3975 3976 reclen = dp->d_reclen; 3977 if (reclen >= *len) { 3978 /* End of buffer reached */ 3979 *len = 0; 3980 } else { 3981 dp = (struct dirent *)((char *)dp + reclen); 3982 *len -= reclen; 3983 } 3984 } 3985 3986 if (*len == 0) { 3987 dp = NULL; 3988 3989 /* Have to refill. */ 3990 switch (*eofflag) { 3991 case 0: 3992 break; 3993 3994 case 1: 3995 /* Nothing more to read. */ 3996 *eofflag = 2; /* Remember the caller reached EOF. */ 3997 goto success; 3998 3999 default: 4000 /* The caller didn't test for EOF. */ 4001 error = EINVAL; 4002 goto out; 4003 } 4004 4005 iov.iov_base = dirbuf; 4006 iov.iov_len = dirbuflen; 4007 4008 uio.uio_iov = &iov; 4009 uio.uio_iovcnt = 1; 4010 uio.uio_offset = *off; 4011 uio.uio_resid = dirbuflen; 4012 uio.uio_segflg = UIO_SYSSPACE; 4013 uio.uio_rw = UIO_READ; 4014 uio.uio_td = td; 4015 4016 #ifdef MAC 4017 error = mac_vnode_check_readdir(td->td_ucred, vp); 4018 if (error == 0) 4019 #endif 4020 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag, 4021 NULL, NULL); 4022 if (error != 0) 4023 goto out; 4024 4025 *len = dirbuflen - uio.uio_resid; 4026 *off = uio.uio_offset; 4027 4028 if (*len == 0) { 4029 /* Sanity check on INVARIANTS. */ 4030 MPASS(*eofflag != 0); 4031 *eofflag = 1; 4032 goto success; 4033 } 4034 4035 /* 4036 * Normalize the flag returned by VOP_READDIR(), since we use 2 4037 * as a sentinel value. 4038 */ 4039 if (*eofflag != 0) 4040 *eofflag = 1; 4041 4042 dp = (struct dirent *)dirbuf; 4043 } 4044 4045 if (__predict_false(*len < GENERIC_MINDIRSIZ || 4046 dp->d_reclen < GENERIC_MINDIRSIZ)) { 4047 error = EINTEGRITY; 4048 dp = NULL; 4049 goto out; 4050 } 4051 4052 success: 4053 error = 0; 4054 out: 4055 *dpp = dp; 4056 return (error); 4057 } 4058 4059 /* 4060 * Checks whether a directory is empty or not. 4061 * 4062 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other 4063 * values are genuine errors preventing the check. 4064 */ 4065 int 4066 vn_dir_check_empty(struct vnode *vp) 4067 { 4068 struct thread *const td = curthread; 4069 char *dirbuf; 4070 size_t dirbuflen, len; 4071 off_t off; 4072 int eofflag, error; 4073 struct dirent *dp; 4074 struct vattr va; 4075 4076 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 4077 VNPASS(vp->v_type == VDIR, vp); 4078 4079 error = VOP_GETATTR(vp, &va, td->td_ucred); 4080 if (error != 0) 4081 return (error); 4082 4083 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ); 4084 if (dirbuflen < va.va_blocksize) 4085 dirbuflen = va.va_blocksize; 4086 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK); 4087 4088 len = 0; 4089 off = 0; 4090 eofflag = 0; 4091 4092 for (;;) { 4093 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen, 4094 &dp, &len, &off, &eofflag); 4095 if (error != 0) 4096 goto end; 4097 4098 if (len == 0) { 4099 /* EOF */ 4100 error = 0; 4101 goto end; 4102 } 4103 4104 /* 4105 * Skip whiteouts. Unionfs operates on filesystems only and 4106 * not on hierarchies, so these whiteouts would be shadowed on 4107 * the system hierarchy but not for a union using the 4108 * filesystem of their directories as the upper layer. 4109 * Additionally, unionfs currently transparently exposes 4110 * union-specific metadata of its upper layer, meaning that 4111 * whiteouts can be seen through the union view in empty 4112 * directories. Taking into account these whiteouts would then 4113 * prevent mounting another filesystem on such effectively 4114 * empty directories. 4115 */ 4116 if (dp->d_type == DT_WHT) 4117 continue; 4118 4119 /* 4120 * Any file in the directory which is not '.' or '..' indicates 4121 * the directory is not empty. 4122 */ 4123 switch (dp->d_namlen) { 4124 case 2: 4125 if (dp->d_name[1] != '.') { 4126 /* Can't be '..' (nor '.') */ 4127 error = ENOTEMPTY; 4128 goto end; 4129 } 4130 /* FALLTHROUGH */ 4131 case 1: 4132 if (dp->d_name[0] != '.') { 4133 /* Can't be '..' nor '.' */ 4134 error = ENOTEMPTY; 4135 goto end; 4136 } 4137 break; 4138 4139 default: 4140 error = ENOTEMPTY; 4141 goto end; 4142 } 4143 } 4144 4145 end: 4146 free(dirbuf, M_TEMP); 4147 return (error); 4148 } 4149 4150 4151 static u_long vn_lock_pair_pause_cnt; 4152 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD, 4153 &vn_lock_pair_pause_cnt, 0, 4154 "Count of vn_lock_pair deadlocks"); 4155 4156 u_int vn_lock_pair_pause_max; 4157 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW, 4158 &vn_lock_pair_pause_max, 0, 4159 "Max ticks for vn_lock_pair deadlock avoidance sleep"); 4160 4161 static void 4162 vn_lock_pair_pause(const char *wmesg) 4163 { 4164 atomic_add_long(&vn_lock_pair_pause_cnt, 1); 4165 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max)); 4166 } 4167 4168 /* 4169 * Lock pair of (possibly same) vnodes vp1, vp2, avoiding lock order 4170 * reversal. vp1_locked indicates whether vp1 is locked; if not, vp1 4171 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes 4172 * can be NULL. 4173 * 4174 * The function returns with both vnodes exclusively or shared locked, 4175 * according to corresponding lkflags, and guarantees that it does not 4176 * create lock order reversal with other threads during its execution. 4177 * Both vnodes could be unlocked temporary (and reclaimed). 4178 * 4179 * If requesting shared locking, locked vnode lock must not be recursed. 4180 * 4181 * Only one of LK_SHARED and LK_EXCLUSIVE must be specified. 4182 * LK_NODDLKTREAT can be optionally passed. 4183 * 4184 * If vp1 == vp2, only one, most exclusive, lock is obtained on it. 4185 */ 4186 void 4187 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1, 4188 struct vnode *vp2, bool vp2_locked, int lkflags2) 4189 { 4190 int error, locked1; 4191 4192 MPASS((((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0)) || 4193 (vp1 == NULL && lkflags1 == 0)); 4194 MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0); 4195 MPASS((((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0)) || 4196 (vp2 == NULL && lkflags2 == 0)); 4197 MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0); 4198 4199 if (vp1 == NULL && vp2 == NULL) 4200 return; 4201 4202 if (vp1 == vp2) { 4203 MPASS(vp1_locked == vp2_locked); 4204 4205 /* Select the most exclusive mode for lock. */ 4206 if ((lkflags1 & LK_TYPE_MASK) != (lkflags2 & LK_TYPE_MASK)) 4207 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE; 4208 4209 if (vp1_locked) { 4210 ASSERT_VOP_LOCKED(vp1, "vp1"); 4211 4212 /* No need to relock if any lock is exclusive. */ 4213 if ((vp1->v_vnlock->lock_object.lo_flags & 4214 LK_NOSHARE) != 0) 4215 return; 4216 4217 locked1 = VOP_ISLOCKED(vp1); 4218 if (((lkflags1 & LK_SHARED) != 0 && 4219 locked1 != LK_EXCLUSIVE) || 4220 ((lkflags1 & LK_EXCLUSIVE) != 0 && 4221 locked1 == LK_EXCLUSIVE)) 4222 return; 4223 VOP_UNLOCK(vp1); 4224 } 4225 4226 ASSERT_VOP_UNLOCKED(vp1, "vp1"); 4227 vn_lock(vp1, lkflags1 | LK_RETRY); 4228 return; 4229 } 4230 4231 if (vp1 != NULL) { 4232 if ((lkflags1 & LK_SHARED) != 0 && 4233 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0) 4234 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE; 4235 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) { 4236 ASSERT_VOP_LOCKED(vp1, "vp1"); 4237 if ((lkflags1 & LK_EXCLUSIVE) != 0) { 4238 VOP_UNLOCK(vp1); 4239 ASSERT_VOP_UNLOCKED(vp1, 4240 "vp1 shared recursed"); 4241 vp1_locked = false; 4242 } 4243 } else if (!vp1_locked) 4244 ASSERT_VOP_UNLOCKED(vp1, "vp1"); 4245 } else { 4246 vp1_locked = true; 4247 } 4248 4249 if (vp2 != NULL) { 4250 if ((lkflags2 & LK_SHARED) != 0 && 4251 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0) 4252 lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE; 4253 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) { 4254 ASSERT_VOP_LOCKED(vp2, "vp2"); 4255 if ((lkflags2 & LK_EXCLUSIVE) != 0) { 4256 VOP_UNLOCK(vp2); 4257 ASSERT_VOP_UNLOCKED(vp2, 4258 "vp2 shared recursed"); 4259 vp2_locked = false; 4260 } 4261 } else if (!vp2_locked) 4262 ASSERT_VOP_UNLOCKED(vp2, "vp2"); 4263 } else { 4264 vp2_locked = true; 4265 } 4266 4267 if (!vp1_locked && !vp2_locked) { 4268 vn_lock(vp1, lkflags1 | LK_RETRY); 4269 vp1_locked = true; 4270 } 4271 4272 while (!vp1_locked || !vp2_locked) { 4273 if (vp1_locked && vp2 != NULL) { 4274 if (vp1 != NULL) { 4275 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT, 4276 __FILE__, __LINE__); 4277 if (error == 0) 4278 break; 4279 VOP_UNLOCK(vp1); 4280 vp1_locked = false; 4281 vn_lock_pair_pause("vlp1"); 4282 } 4283 vn_lock(vp2, lkflags2 | LK_RETRY); 4284 vp2_locked = true; 4285 } 4286 if (vp2_locked && vp1 != NULL) { 4287 if (vp2 != NULL) { 4288 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT, 4289 __FILE__, __LINE__); 4290 if (error == 0) 4291 break; 4292 VOP_UNLOCK(vp2); 4293 vp2_locked = false; 4294 vn_lock_pair_pause("vlp2"); 4295 } 4296 vn_lock(vp1, lkflags1 | LK_RETRY); 4297 vp1_locked = true; 4298 } 4299 } 4300 if (vp1 != NULL) { 4301 if (lkflags1 == LK_EXCLUSIVE) 4302 ASSERT_VOP_ELOCKED(vp1, "vp1 ret"); 4303 else 4304 ASSERT_VOP_LOCKED(vp1, "vp1 ret"); 4305 } 4306 if (vp2 != NULL) { 4307 if (lkflags2 == LK_EXCLUSIVE) 4308 ASSERT_VOP_ELOCKED(vp2, "vp2 ret"); 4309 else 4310 ASSERT_VOP_LOCKED(vp2, "vp2 ret"); 4311 } 4312 } 4313 4314 int 4315 vn_lktype_write(struct mount *mp, struct vnode *vp) 4316 { 4317 if (MNT_SHARED_WRITES(mp) || 4318 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) 4319 return (LK_SHARED); 4320 return (LK_EXCLUSIVE); 4321 } 4322 4323 int 4324 vn_cmp(struct file *fp1, struct file *fp2, struct thread *td) 4325 { 4326 if (fp2->f_type != DTYPE_VNODE) 4327 return (3); 4328 return (kcmp_cmp((uintptr_t)fp1->f_vnode, (uintptr_t)fp2->f_vnode)); 4329 } 4330