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 #include "opt_hwt_hooks.h" 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/buf.h> 49 #include <sys/disk.h> 50 #include <sys/dirent.h> 51 #include <sys/fail.h> 52 #include <sys/fcntl.h> 53 #include <sys/file.h> 54 #include <sys/filio.h> 55 #include <sys/inotify.h> 56 #include <sys/ktr.h> 57 #include <sys/ktrace.h> 58 #include <sys/limits.h> 59 #include <sys/lock.h> 60 #include <sys/mman.h> 61 #include <sys/mount.h> 62 #include <sys/mutex.h> 63 #include <sys/namei.h> 64 #include <sys/priv.h> 65 #include <sys/prng.h> 66 #include <sys/proc.h> 67 #include <sys/rwlock.h> 68 #include <sys/sleepqueue.h> 69 #include <sys/stat.h> 70 #include <sys/sysctl.h> 71 #include <sys/unistd.h> 72 #include <sys/user.h> 73 #include <sys/vnode.h> 74 75 #include <security/audit/audit.h> 76 #include <security/mac/mac_framework.h> 77 78 #include <vm/vm.h> 79 #include <vm/vm_extern.h> 80 #include <vm/pmap.h> 81 #include <vm/vm_map.h> 82 #include <vm/vm_object.h> 83 #include <vm/vm_page.h> 84 #include <vm/vm_pager.h> 85 #include <vm/vnode_pager.h> 86 87 #ifdef HWPMC_HOOKS 88 #include <sys/pmckern.h> 89 #endif 90 91 #ifdef HWT_HOOKS 92 #include <dev/hwt/hwt_hook.h> 93 #endif 94 95 static fo_rdwr_t vn_read; 96 static fo_rdwr_t vn_write; 97 static fo_rdwr_t vn_io_fault; 98 static fo_truncate_t vn_truncate; 99 static fo_ioctl_t vn_ioctl; 100 static fo_poll_t vn_poll; 101 static fo_kqfilter_t vn_kqfilter; 102 static fo_close_t vn_closefile; 103 static fo_mmap_t vn_mmap; 104 static fo_fallocate_t vn_fallocate; 105 static fo_fspacectl_t vn_fspacectl; 106 107 const struct fileops vnops = { 108 .fo_read = vn_io_fault, 109 .fo_write = vn_io_fault, 110 .fo_truncate = vn_truncate, 111 .fo_ioctl = vn_ioctl, 112 .fo_poll = vn_poll, 113 .fo_kqfilter = vn_kqfilter, 114 .fo_stat = vn_statfile, 115 .fo_close = vn_closefile, 116 .fo_chmod = vn_chmod, 117 .fo_chown = vn_chown, 118 .fo_sendfile = vn_sendfile, 119 .fo_seek = vn_seek, 120 .fo_fill_kinfo = vn_fill_kinfo, 121 .fo_mmap = vn_mmap, 122 .fo_fallocate = vn_fallocate, 123 .fo_fspacectl = vn_fspacectl, 124 .fo_cmp = vn_cmp, 125 .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE 126 }; 127 128 const u_int io_hold_cnt = 16; 129 static int vn_io_fault_enable = 1; 130 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RWTUN, 131 &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); 132 static int vn_io_fault_prefault = 0; 133 SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RWTUN, 134 &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); 135 static int vn_io_pgcache_read_enable = 1; 136 SYSCTL_INT(_debug, OID_AUTO, vn_io_pgcache_read_enable, CTLFLAG_RWTUN, 137 &vn_io_pgcache_read_enable, 0, 138 "Enable copying from page cache for reads, avoiding fs"); 139 static u_long vn_io_faults_cnt; 140 SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, 141 &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); 142 143 static int vfs_allow_read_dir = 0; 144 SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW, 145 &vfs_allow_read_dir, 0, 146 "Enable read(2) of directory by root for filesystems that support it"); 147 148 /* 149 * Returns true if vn_io_fault mode of handling the i/o request should 150 * be used. 151 */ 152 static bool 153 do_vn_io_fault(struct vnode *vp, struct uio *uio) 154 { 155 struct mount *mp; 156 157 return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && 158 (mp = vp->v_mount) != NULL && 159 (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); 160 } 161 162 /* 163 * Structure used to pass arguments to vn_io_fault1(), to do either 164 * file- or vnode-based I/O calls. 165 */ 166 struct vn_io_fault_args { 167 enum { 168 VN_IO_FAULT_FOP, 169 VN_IO_FAULT_VOP 170 } kind; 171 struct ucred *cred; 172 int flags; 173 union { 174 struct fop_args_tag { 175 struct file *fp; 176 fo_rdwr_t *doio; 177 } fop_args; 178 struct vop_args_tag { 179 struct vnode *vp; 180 } vop_args; 181 } args; 182 }; 183 184 static int vn_io_fault1(struct vnode *vp, struct uio *uio, 185 struct vn_io_fault_args *args, struct thread *td); 186 187 int 188 vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) 189 { 190 struct thread *td = curthread; 191 192 return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); 193 } 194 195 static uint64_t 196 open2nameif(int fmode, u_int vn_open_flags, uint64_t cn_flags) 197 { 198 uint64_t res; 199 200 res = ISOPEN | LOCKLEAF | cn_flags; 201 if ((fmode & O_RESOLVE_BENEATH) != 0) 202 res |= RBENEATH; 203 if ((fmode & O_EMPTY_PATH) != 0) 204 res |= EMPTYPATH; 205 if ((fmode & FREAD) != 0) 206 res |= OPENREAD; 207 if ((fmode & FWRITE) != 0) 208 res |= OPENWRITE; 209 if ((fmode & O_NAMEDATTR) != 0) 210 res |= OPENNAMED | CREATENAMED; 211 if ((fmode & O_NOFOLLOW) != 0) 212 res &= ~FOLLOW; 213 if ((vn_open_flags & VN_OPEN_NOAUDIT) == 0) 214 res |= AUDITVNODE1; 215 else 216 res &= ~AUDITVNODE1; 217 if ((vn_open_flags & VN_OPEN_NOCAPCHECK) != 0) 218 res |= NOCAPCHECK; 219 if ((vn_open_flags & VN_OPEN_WANTIOCTLCAPS) != 0) 220 res |= WANTIOCTLCAPS; 221 222 return (res); 223 } 224 225 /* 226 * For the O_NAMEDATTR case, check for a valid use of it. 227 */ 228 static int 229 vfs_check_namedattr(struct vnode *vp) 230 { 231 int error; 232 short irflag; 233 234 error = 0; 235 irflag = vn_irflag_read(vp); 236 if ((vp->v_mount->mnt_flag & MNT_NAMEDATTR) == 0 || 237 ((irflag & VIRF_NAMEDATTR) != 0 && vp->v_type != VREG)) 238 error = EINVAL; 239 else if ((irflag & (VIRF_NAMEDDIR | VIRF_NAMEDATTR)) == 0) 240 error = ENOATTR; 241 return (error); 242 } 243 244 /* 245 * Common code for vnode open operations via a name lookup. 246 * Lookup the vnode and invoke VOP_CREATE if needed. 247 * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. 248 * 249 * Note that this does NOT free nameidata for the successful case, 250 * due to the NDINIT being done elsewhere. 251 */ 252 int 253 vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, 254 struct ucred *cred, struct file *fp) 255 { 256 struct vnode *vp; 257 struct mount *mp; 258 struct vattr vat; 259 struct vattr *vap = &vat; 260 int fmode, error; 261 bool first_open; 262 263 restart: 264 first_open = false; 265 fmode = *flagp; 266 if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | 267 O_EXCL | O_DIRECTORY) || 268 (fmode & (O_CREAT | O_EMPTY_PATH)) == (O_CREAT | O_EMPTY_PATH)) 269 return (EINVAL); 270 else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { 271 ndp->ni_cnd.cn_nameiop = CREATE; 272 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags, 273 ndp->ni_cnd.cn_flags); 274 275 /* 276 * Set NOCACHE to avoid flushing the cache when 277 * rolling in many files at once. 278 * 279 * Set NC_KEEPPOSENTRY to keep positive entries if they already 280 * exist despite NOCACHE. 281 */ 282 ndp->ni_cnd.cn_flags |= LOCKPARENT | NOCACHE | NC_KEEPPOSENTRY; 283 if ((fmode & O_EXCL) != 0) 284 ndp->ni_cnd.cn_flags &= ~FOLLOW; 285 if ((vn_open_flags & VN_OPEN_INVFS) == 0) 286 bwillwrite(); 287 if ((error = namei(ndp)) != 0) 288 return (error); 289 if (ndp->ni_vp == NULL) { 290 if ((fmode & O_NAMEDATTR) != 0 && 291 (ndp->ni_dvp->v_mount->mnt_flag & MNT_NAMEDATTR) == 292 0) { 293 error = EINVAL; 294 vp = ndp->ni_dvp; 295 ndp->ni_dvp = NULL; 296 goto bad; 297 } 298 VATTR_NULL(vap); 299 vap->va_type = VREG; 300 vap->va_mode = cmode; 301 if (fmode & O_EXCL) 302 vap->va_vaflags |= VA_EXCLUSIVE; 303 if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { 304 NDFREE_PNBUF(ndp); 305 vput(ndp->ni_dvp); 306 if ((error = vn_start_write(NULL, &mp, 307 V_XSLEEP | V_PCATCH)) != 0) 308 return (error); 309 NDREINIT(ndp); 310 goto restart; 311 } 312 if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0 || 313 (vn_irflag_read(ndp->ni_dvp) & VIRF_INOTIFY) != 0) 314 ndp->ni_cnd.cn_flags |= MAKEENTRY; 315 #ifdef MAC 316 error = mac_vnode_check_create(cred, ndp->ni_dvp, 317 &ndp->ni_cnd, vap); 318 if (error == 0) 319 #endif 320 error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, 321 &ndp->ni_cnd, vap); 322 vp = ndp->ni_vp; 323 if (error == 0 && (fmode & O_EXCL) != 0 && 324 (fmode & (O_EXLOCK | O_SHLOCK)) != 0) { 325 VI_LOCK(vp); 326 vp->v_iflag |= VI_FOPENING; 327 VI_UNLOCK(vp); 328 first_open = true; 329 } 330 VOP_VPUT_PAIR(ndp->ni_dvp, error == 0 ? &vp : NULL, 331 false); 332 vn_finished_write(mp); 333 if (error) { 334 NDFREE_PNBUF(ndp); 335 if (error == ERELOOKUP) { 336 NDREINIT(ndp); 337 goto restart; 338 } 339 return (error); 340 } 341 fmode &= ~O_TRUNC; 342 } else { 343 if (ndp->ni_dvp == ndp->ni_vp) 344 vrele(ndp->ni_dvp); 345 else 346 vput(ndp->ni_dvp); 347 ndp->ni_dvp = NULL; 348 vp = ndp->ni_vp; 349 if (fmode & O_EXCL) { 350 error = EEXIST; 351 goto bad; 352 } 353 if ((fmode & O_NAMEDATTR) != 0) { 354 error = vfs_check_namedattr(vp); 355 if (error != 0) 356 goto bad; 357 } else if (vp->v_type == VDIR) { 358 error = EISDIR; 359 goto bad; 360 } 361 fmode &= ~O_CREAT; 362 } 363 } else { 364 ndp->ni_cnd.cn_nameiop = LOOKUP; 365 ndp->ni_cnd.cn_flags = open2nameif(fmode, vn_open_flags, 366 ndp->ni_cnd.cn_flags); 367 if ((fmode & FWRITE) == 0) 368 ndp->ni_cnd.cn_flags |= LOCKSHARED; 369 if ((error = namei(ndp)) != 0) 370 return (error); 371 vp = ndp->ni_vp; 372 if ((fmode & O_NAMEDATTR) != 0) { 373 error = vfs_check_namedattr(vp); 374 if (error != 0) 375 goto bad; 376 } 377 } 378 error = vn_open_vnode(vp, fmode, cred, curthread, fp); 379 if (first_open) { 380 VI_LOCK(vp); 381 vp->v_iflag &= ~VI_FOPENING; 382 wakeup(vp); 383 VI_UNLOCK(vp); 384 } 385 if (error) 386 goto bad; 387 *flagp = fmode; 388 return (0); 389 bad: 390 NDFREE_PNBUF(ndp); 391 vput(vp); 392 *flagp = fmode; 393 ndp->ni_vp = NULL; 394 return (error); 395 } 396 397 static int 398 vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) 399 { 400 struct flock lf; 401 int error, lock_flags, type; 402 403 ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); 404 if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) 405 return (0); 406 KASSERT(fp != NULL, ("open with flock requires fp")); 407 if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) 408 return (EOPNOTSUPP); 409 410 lock_flags = VOP_ISLOCKED(vp); 411 VOP_UNLOCK(vp); 412 413 lf.l_whence = SEEK_SET; 414 lf.l_start = 0; 415 lf.l_len = 0; 416 lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; 417 type = F_FLOCK; 418 if ((fmode & FNONBLOCK) == 0) 419 type |= F_WAIT; 420 if ((fmode & (O_CREAT | O_EXCL)) == (O_CREAT | O_EXCL)) 421 type |= F_FIRSTOPEN; 422 error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); 423 if (error == 0) 424 fp->f_flag |= FHASLOCK; 425 426 vn_lock(vp, lock_flags | LK_RETRY); 427 return (error); 428 } 429 430 /* 431 * Common code for vnode open operations once a vnode is located. 432 * Check permissions, and call the VOP_OPEN routine. 433 */ 434 int 435 vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, 436 struct thread *td, struct file *fp) 437 { 438 accmode_t accmode; 439 int error; 440 441 KASSERT((fmode & O_PATH) == 0 || (fmode & O_ACCMODE) == 0, 442 ("%s: O_PATH and O_ACCMODE are mutually exclusive", __func__)); 443 444 if (vp->v_type == VLNK) { 445 if ((fmode & O_PATH) == 0 || (fmode & FEXEC) != 0) 446 return (EMLINK); 447 } 448 if (vp->v_type != VDIR && fmode & O_DIRECTORY) 449 return (ENOTDIR); 450 451 accmode = 0; 452 if ((fmode & O_PATH) == 0) { 453 if (vp->v_type == VSOCK) 454 return (EOPNOTSUPP); 455 if ((fmode & (FWRITE | O_TRUNC)) != 0) { 456 if (vp->v_type == VDIR) 457 return (EISDIR); 458 accmode |= VWRITE; 459 } 460 if ((fmode & FREAD) != 0) 461 accmode |= VREAD; 462 if ((fmode & O_APPEND) && (fmode & FWRITE)) 463 accmode |= VAPPEND; 464 #ifdef MAC 465 if ((fmode & O_CREAT) != 0) 466 accmode |= VCREAT; 467 #endif 468 } 469 if ((fmode & FEXEC) != 0) 470 accmode |= VEXEC; 471 #ifdef MAC 472 if ((fmode & O_VERIFY) != 0) 473 accmode |= VVERIFY; 474 error = mac_vnode_check_open(cred, vp, accmode); 475 if (error != 0) 476 return (error); 477 478 accmode &= ~(VCREAT | VVERIFY); 479 #endif 480 if ((fmode & O_CREAT) == 0 && accmode != 0) { 481 error = VOP_ACCESS(vp, accmode, cred, td); 482 if (error != 0) 483 return (error); 484 } 485 if ((fmode & O_PATH) != 0) { 486 if (vp->v_type != VFIFO && vp->v_type != VSOCK && 487 VOP_ACCESS(vp, VREAD, cred, td) == 0) 488 fp->f_flag |= FKQALLOWED; 489 INOTIFY(vp, IN_OPEN); 490 return (0); 491 } 492 493 if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 494 vn_lock(vp, LK_UPGRADE | LK_RETRY); 495 error = VOP_OPEN(vp, fmode, cred, td, fp); 496 if (error != 0) 497 return (error); 498 499 error = vn_open_vnode_advlock(vp, fmode, fp); 500 if (error == 0 && (fmode & FWRITE) != 0) { 501 error = VOP_ADD_WRITECOUNT(vp, 1); 502 if (error == 0) { 503 CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", 504 __func__, vp, vp->v_writecount); 505 } 506 } 507 508 /* 509 * Error from advlock or VOP_ADD_WRITECOUNT() still requires 510 * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). 511 */ 512 if (error != 0) { 513 if (fp != NULL) { 514 /* 515 * Arrange the call by having fdrop() to use 516 * vn_closefile(). This is to satisfy 517 * filesystems like devfs or tmpfs, which 518 * override fo_close(). 519 */ 520 fp->f_flag |= FOPENFAILED; 521 fp->f_vnode = vp; 522 if (fp->f_ops == &badfileops) { 523 fp->f_type = DTYPE_VNODE; 524 fp->f_ops = &vnops; 525 } 526 vref(vp); 527 } else { 528 /* 529 * If there is no fp, due to kernel-mode open, 530 * we can call VOP_CLOSE() now. 531 */ 532 if ((vp->v_type == VFIFO || 533 !MNT_EXTENDED_SHARED(vp->v_mount)) && 534 VOP_ISLOCKED(vp) != LK_EXCLUSIVE) 535 vn_lock(vp, LK_UPGRADE | LK_RETRY); 536 (void)VOP_CLOSE(vp, fmode & (FREAD | FWRITE | FEXEC), 537 cred, td); 538 } 539 } 540 541 ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); 542 return (error); 543 544 } 545 546 /* 547 * Check for write permissions on the specified vnode. 548 * Prototype text segments cannot be written. 549 * It is racy. 550 */ 551 int 552 vn_writechk(struct vnode *vp) 553 { 554 555 ASSERT_VOP_LOCKED(vp, "vn_writechk"); 556 /* 557 * If there's shared text associated with 558 * the vnode, try to free it up once. If 559 * we fail, we can't allow writing. 560 */ 561 if (VOP_IS_TEXT(vp)) 562 return (ETXTBSY); 563 564 return (0); 565 } 566 567 /* 568 * Vnode close call 569 */ 570 static int 571 vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, 572 struct thread *td, bool keep_ref) 573 { 574 struct mount *mp; 575 int error, lock_flags; 576 577 lock_flags = vp->v_type != VFIFO && MNT_EXTENDED_SHARED(vp->v_mount) ? 578 LK_SHARED : LK_EXCLUSIVE; 579 580 vn_start_write(vp, &mp, V_WAIT); 581 vn_lock(vp, lock_flags | LK_RETRY); 582 AUDIT_ARG_VNODE1(vp); 583 if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { 584 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 585 CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", 586 __func__, vp, vp->v_writecount); 587 } 588 error = VOP_CLOSE(vp, flags, file_cred, td); 589 if (keep_ref) 590 VOP_UNLOCK(vp); 591 else 592 vput(vp); 593 vn_finished_write(mp); 594 return (error); 595 } 596 597 int 598 vn_close(struct vnode *vp, int flags, struct ucred *file_cred, 599 struct thread *td) 600 { 601 602 return (vn_close1(vp, flags, file_cred, td, false)); 603 } 604 605 /* 606 * Heuristic to detect sequential operation. 607 */ 608 static int 609 sequential_heuristic(struct uio *uio, struct file *fp) 610 { 611 enum uio_rw rw; 612 613 ASSERT_VOP_LOCKED(fp->f_vnode, __func__); 614 615 rw = uio->uio_rw; 616 if (fp->f_flag & FRDAHEAD) 617 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 618 619 /* 620 * Offset 0 is handled specially. open() sets f_seqcount to 1 so 621 * that the first I/O is normally considered to be slightly 622 * sequential. Seeking to offset 0 doesn't change sequentiality 623 * unless previous seeks have reduced f_seqcount to 0, in which 624 * case offset 0 is not special. 625 */ 626 if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) || 627 uio->uio_offset == fp->f_nextoff[rw]) { 628 /* 629 * f_seqcount is in units of fixed-size blocks so that it 630 * depends mainly on the amount of sequential I/O and not 631 * much on the number of sequential I/O's. The fixed size 632 * of 16384 is hard-coded here since it is (not quite) just 633 * a magic size that works well here. This size is more 634 * closely related to the best I/O size for real disks than 635 * to any block size used by software. 636 */ 637 if (uio->uio_resid >= IO_SEQMAX * 16384) 638 fp->f_seqcount[rw] = IO_SEQMAX; 639 else { 640 fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384); 641 if (fp->f_seqcount[rw] > IO_SEQMAX) 642 fp->f_seqcount[rw] = IO_SEQMAX; 643 } 644 return (fp->f_seqcount[rw] << IO_SEQSHIFT); 645 } 646 647 /* Not sequential. Quickly draw-down sequentiality. */ 648 if (fp->f_seqcount[rw] > 1) 649 fp->f_seqcount[rw] = 1; 650 else 651 fp->f_seqcount[rw] = 0; 652 return (0); 653 } 654 655 /* 656 * Package up an I/O request on a vnode into a uio and do it. 657 */ 658 int 659 vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, 660 enum uio_seg segflg, int ioflg, struct ucred *active_cred, 661 struct ucred *file_cred, ssize_t *aresid, struct thread *td) 662 { 663 struct uio auio; 664 struct iovec aiov; 665 struct mount *mp; 666 struct ucred *cred; 667 void *rl_cookie; 668 struct vn_io_fault_args args; 669 int error, lock_flags; 670 671 if (offset < 0 && vp->v_type != VCHR) 672 return (EINVAL); 673 auio.uio_iov = &aiov; 674 auio.uio_iovcnt = 1; 675 aiov.iov_base = base; 676 aiov.iov_len = len; 677 auio.uio_resid = len; 678 auio.uio_offset = offset; 679 auio.uio_segflg = segflg; 680 auio.uio_rw = rw; 681 auio.uio_td = td; 682 error = 0; 683 684 if ((ioflg & IO_NODELOCKED) == 0) { 685 if ((ioflg & IO_RANGELOCKED) == 0) { 686 if (rw == UIO_READ) { 687 rl_cookie = vn_rangelock_rlock(vp, offset, 688 offset + len); 689 } else if ((ioflg & IO_APPEND) != 0) { 690 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 691 } else { 692 rl_cookie = vn_rangelock_wlock(vp, offset, 693 offset + len); 694 } 695 } else 696 rl_cookie = NULL; 697 mp = NULL; 698 if (rw == UIO_WRITE) { 699 if (vp->v_type != VCHR && 700 (error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH)) 701 != 0) 702 goto out; 703 lock_flags = vn_lktype_write(mp, vp); 704 } else 705 lock_flags = LK_SHARED; 706 vn_lock(vp, lock_flags | LK_RETRY); 707 } else 708 rl_cookie = NULL; 709 710 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 711 #ifdef MAC 712 if ((ioflg & IO_NOMACCHECK) == 0) { 713 if (rw == UIO_READ) 714 error = mac_vnode_check_read(active_cred, file_cred, 715 vp); 716 else 717 error = mac_vnode_check_write(active_cred, file_cred, 718 vp); 719 } 720 #endif 721 if (error == 0) { 722 if (file_cred != NULL) 723 cred = file_cred; 724 else 725 cred = active_cred; 726 if (do_vn_io_fault(vp, &auio)) { 727 args.kind = VN_IO_FAULT_VOP; 728 args.cred = cred; 729 args.flags = ioflg; 730 args.args.vop_args.vp = vp; 731 error = vn_io_fault1(vp, &auio, &args, td); 732 } else if (rw == UIO_READ) { 733 error = VOP_READ(vp, &auio, ioflg, cred); 734 } else /* if (rw == UIO_WRITE) */ { 735 error = VOP_WRITE(vp, &auio, ioflg, cred); 736 } 737 } 738 if (aresid) 739 *aresid = auio.uio_resid; 740 else 741 if (auio.uio_resid && error == 0) 742 error = EIO; 743 if ((ioflg & IO_NODELOCKED) == 0) { 744 VOP_UNLOCK(vp); 745 if (mp != NULL) 746 vn_finished_write(mp); 747 } 748 out: 749 if (rl_cookie != NULL) 750 vn_rangelock_unlock(vp, rl_cookie); 751 return (error); 752 } 753 754 /* 755 * Package up an I/O request on a vnode into a uio and do it. The I/O 756 * request is split up into smaller chunks and we try to avoid saturating 757 * the buffer cache while potentially holding a vnode locked, so we 758 * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() 759 * to give other processes a chance to lock the vnode (either other processes 760 * core'ing the same binary, or unrelated processes scanning the directory). 761 */ 762 int 763 vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, 764 off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, 765 struct ucred *file_cred, size_t *aresid, struct thread *td) 766 { 767 int error = 0; 768 ssize_t iaresid; 769 770 do { 771 int chunk; 772 773 /* 774 * Force `offset' to a multiple of MAXBSIZE except possibly 775 * for the first chunk, so that filesystems only need to 776 * write full blocks except possibly for the first and last 777 * chunks. 778 */ 779 chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; 780 781 if (chunk > len) 782 chunk = len; 783 if (rw != UIO_READ && vp->v_type == VREG) 784 bwillwrite(); 785 iaresid = 0; 786 error = vn_rdwr(rw, vp, base, chunk, offset, segflg, 787 ioflg, active_cred, file_cred, &iaresid, td); 788 len -= chunk; /* aresid calc already includes length */ 789 if (error) 790 break; 791 offset += chunk; 792 base = (char *)base + chunk; 793 kern_yield(PRI_USER); 794 } while (len); 795 if (aresid) 796 *aresid = len + iaresid; 797 return (error); 798 } 799 800 #if OFF_MAX <= LONG_MAX 801 static void 802 file_v_lock(struct file *fp, short lock_bit, short lock_wait_bit) 803 { 804 short *flagsp; 805 short state; 806 807 flagsp = &fp->f_vflags; 808 state = atomic_load_16(flagsp); 809 if ((state & lock_bit) == 0 && 810 atomic_cmpset_acq_16(flagsp, state, state | lock_bit)) 811 return; 812 813 sleepq_lock(flagsp); 814 state = atomic_load_16(flagsp); 815 for (;;) { 816 if ((state & lock_bit) == 0) { 817 if (!atomic_fcmpset_acq_16(flagsp, &state, 818 state | lock_bit)) 819 continue; 820 break; 821 } 822 if ((state & lock_wait_bit) == 0) { 823 if (!atomic_fcmpset_acq_16(flagsp, &state, 824 state | lock_wait_bit)) 825 continue; 826 } 827 DROP_GIANT(); 828 sleepq_add(flagsp, NULL, "vofflock", 0, 0); 829 sleepq_wait(flagsp, PRI_MAX_KERN); 830 PICKUP_GIANT(); 831 sleepq_lock(flagsp); 832 state = atomic_load_16(flagsp); 833 } 834 sleepq_release(flagsp); 835 } 836 837 static void 838 file_v_unlock(struct file *fp, short lock_bit, short lock_wait_bit) 839 { 840 short *flagsp; 841 short state; 842 843 flagsp = &fp->f_vflags; 844 state = atomic_load_16(flagsp); 845 if ((state & lock_wait_bit) == 0 && 846 atomic_cmpset_rel_16(flagsp, state, state & ~lock_bit)) 847 return; 848 849 sleepq_lock(flagsp); 850 MPASS((*flagsp & lock_bit) != 0); 851 MPASS((*flagsp & lock_wait_bit) != 0); 852 atomic_clear_16(flagsp, lock_bit | lock_wait_bit); 853 sleepq_broadcast(flagsp, SLEEPQ_SLEEP, 0, 0); 854 sleepq_release(flagsp); 855 } 856 857 off_t 858 foffset_lock(struct file *fp, int flags) 859 { 860 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 861 862 if ((flags & FOF_NOLOCK) == 0) { 863 file_v_lock(fp, FILE_V_FOFFSET_LOCKED, 864 FILE_V_FOFFSET_LOCK_WAITING); 865 } 866 867 /* 868 * According to McKusick the vn lock was protecting f_offset here. 869 * It is now protected by the FOFFSET_LOCKED flag. 870 */ 871 return (atomic_load_long(&fp->f_offset)); 872 } 873 874 void 875 foffset_unlock(struct file *fp, off_t val, int flags) 876 { 877 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 878 879 if ((flags & FOF_NOUPDATE) == 0) 880 atomic_store_long(&fp->f_offset, val); 881 if ((flags & FOF_NEXTOFF_R) != 0) 882 fp->f_nextoff[UIO_READ] = val; 883 if ((flags & FOF_NEXTOFF_W) != 0) 884 fp->f_nextoff[UIO_WRITE] = val; 885 886 if ((flags & FOF_NOLOCK) == 0) { 887 file_v_unlock(fp, FILE_V_FOFFSET_LOCKED, 888 FILE_V_FOFFSET_LOCK_WAITING); 889 } 890 } 891 892 static off_t 893 foffset_read(struct file *fp) 894 { 895 896 return (atomic_load_long(&fp->f_offset)); 897 } 898 899 void 900 fsetfl_lock(struct file *fp) 901 { 902 file_v_lock(fp, FILE_V_SETFL_LOCKED, FILE_V_SETFL_LOCK_WAITING); 903 } 904 905 void 906 fsetfl_unlock(struct file *fp) 907 { 908 file_v_unlock(fp, FILE_V_SETFL_LOCKED, FILE_V_SETFL_LOCK_WAITING); 909 } 910 911 #else /* OFF_MAX <= LONG_MAX */ 912 913 static void 914 file_v_lock_mtxp(struct file *fp, struct mtx *mtxp, short lock_bit, 915 short lock_wait_bit) 916 { 917 mtx_assert(mtxp, MA_OWNED); 918 919 while ((fp->f_vflags & lock_bit) != 0) { 920 fp->f_vflags |= lock_wait_bit; 921 msleep(&fp->f_vflags, mtxp, PRI_MAX_KERN, 922 "vofflock", 0); 923 } 924 fp->f_vflags |= lock_bit; 925 } 926 927 static void 928 file_v_unlock_mtxp(struct file *fp, struct mtx *mtxp, short lock_bit, 929 short lock_wait_bit) 930 { 931 mtx_assert(mtxp, MA_OWNED); 932 933 KASSERT((fp->f_vflags & lock_bit) != 0, ("Lost lock_bit")); 934 if ((fp->f_vflags & lock_wait_bit) != 0) 935 wakeup(&fp->f_vflags); 936 fp->f_vflags &= ~(lock_bit | lock_wait_bit); 937 } 938 939 off_t 940 foffset_lock(struct file *fp, int flags) 941 { 942 struct mtx *mtxp; 943 off_t res; 944 945 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 946 947 mtxp = mtx_pool_find(mtxpool_sleep, fp); 948 mtx_lock(mtxp); 949 if ((flags & FOF_NOLOCK) == 0) { 950 file_v_lock_mtxp(fp, mtxp, FILE_V_FOFFSET_LOCKED, 951 FILE_V_FOFFSET_LOCK_WAITING); 952 } 953 res = fp->f_offset; 954 mtx_unlock(mtxp); 955 return (res); 956 } 957 958 void 959 foffset_unlock(struct file *fp, off_t val, int flags) 960 { 961 struct mtx *mtxp; 962 963 KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); 964 965 mtxp = mtx_pool_find(mtxpool_sleep, fp); 966 mtx_lock(mtxp); 967 if ((flags & FOF_NOUPDATE) == 0) 968 fp->f_offset = val; 969 if ((flags & FOF_NEXTOFF_R) != 0) 970 fp->f_nextoff[UIO_READ] = val; 971 if ((flags & FOF_NEXTOFF_W) != 0) 972 fp->f_nextoff[UIO_WRITE] = val; 973 if ((flags & FOF_NOLOCK) == 0) { 974 file_v_unlock_mtxp(fp, mtxp, FILE_V_FOFFSET_LOCKED, 975 FILE_V_FOFFSET_LOCK_WAITING); 976 } 977 mtx_unlock(mtxp); 978 } 979 980 static off_t 981 foffset_read(struct file *fp) 982 { 983 984 return (foffset_lock(fp, FOF_NOLOCK)); 985 } 986 987 void 988 fsetfl_lock(struct file *fp) 989 { 990 struct mtx *mtxp; 991 992 mtxp = mtx_pool_find(mtxpool_sleep, fp); 993 mtx_lock(mtxp); 994 file_v_lock_mtxp(fp, mtxp, FILE_V_SETFL_LOCKED, 995 FILE_V_SETFL_LOCK_WAITING); 996 mtx_unlock(mtxp); 997 } 998 999 void 1000 fsetfl_unlock(struct file *fp) 1001 { 1002 struct mtx *mtxp; 1003 1004 mtxp = mtx_pool_find(mtxpool_sleep, fp); 1005 mtx_lock(mtxp); 1006 file_v_unlock_mtxp(fp, mtxp, FILE_V_SETFL_LOCKED, 1007 FILE_V_SETFL_LOCK_WAITING); 1008 mtx_unlock(mtxp); 1009 } 1010 #endif 1011 1012 void 1013 foffset_lock_pair(struct file *fp1, off_t *off1p, struct file *fp2, off_t *off2p, 1014 int flags) 1015 { 1016 KASSERT(fp1 != fp2, ("foffset_lock_pair: fp1 == fp2")); 1017 1018 /* Lock in a consistent order to avoid deadlock. */ 1019 if ((uintptr_t)fp1 > (uintptr_t)fp2) { 1020 struct file *tmpfp; 1021 off_t *tmpoffp; 1022 1023 tmpfp = fp1, fp1 = fp2, fp2 = tmpfp; 1024 tmpoffp = off1p, off1p = off2p, off2p = tmpoffp; 1025 } 1026 if (fp1 != NULL) 1027 *off1p = foffset_lock(fp1, flags); 1028 if (fp2 != NULL) 1029 *off2p = foffset_lock(fp2, flags); 1030 } 1031 1032 void 1033 foffset_lock_uio(struct file *fp, struct uio *uio, int flags) 1034 { 1035 1036 if ((flags & FOF_OFFSET) == 0) 1037 uio->uio_offset = foffset_lock(fp, flags); 1038 } 1039 1040 void 1041 foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) 1042 { 1043 1044 if ((flags & FOF_OFFSET) == 0) 1045 foffset_unlock(fp, uio->uio_offset, flags); 1046 } 1047 1048 static int 1049 get_advice(struct file *fp, struct uio *uio) 1050 { 1051 struct mtx *mtxp; 1052 int ret; 1053 1054 ret = POSIX_FADV_NORMAL; 1055 if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) 1056 return (ret); 1057 1058 mtxp = mtx_pool_find(mtxpool_sleep, fp); 1059 mtx_lock(mtxp); 1060 if (fp->f_advice != NULL && 1061 uio->uio_offset >= fp->f_advice->fa_start && 1062 uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) 1063 ret = fp->f_advice->fa_advice; 1064 mtx_unlock(mtxp); 1065 return (ret); 1066 } 1067 1068 static int 1069 get_write_ioflag(struct file *fp) 1070 { 1071 int ioflag; 1072 struct mount *mp; 1073 struct vnode *vp; 1074 1075 ioflag = 0; 1076 vp = fp->f_vnode; 1077 mp = atomic_load_ptr(&vp->v_mount); 1078 1079 if ((fp->f_flag & O_DIRECT) != 0) 1080 ioflag |= IO_DIRECT; 1081 1082 if ((fp->f_flag & O_FSYNC) != 0 || 1083 (mp != NULL && (mp->mnt_flag & MNT_SYNCHRONOUS) != 0)) 1084 ioflag |= IO_SYNC; 1085 1086 /* 1087 * For O_DSYNC we set both IO_SYNC and IO_DATASYNC, so that VOP_WRITE() 1088 * or VOP_DEALLOCATE() implementations that don't understand IO_DATASYNC 1089 * fall back to full O_SYNC behavior. 1090 */ 1091 if ((fp->f_flag & O_DSYNC) != 0) 1092 ioflag |= IO_SYNC | IO_DATASYNC; 1093 1094 return (ioflag); 1095 } 1096 1097 int 1098 vn_read_from_obj(struct vnode *vp, struct uio *uio) 1099 { 1100 vm_object_t obj; 1101 vm_page_t ma[io_hold_cnt + 2]; 1102 off_t off, vsz; 1103 ssize_t resid; 1104 int error, i, j; 1105 1106 MPASS(uio->uio_resid <= ptoa(io_hold_cnt + 2)); 1107 obj = atomic_load_ptr(&vp->v_object); 1108 if (obj == NULL) 1109 return (EJUSTRETURN); 1110 1111 /* 1112 * Depends on type stability of vm_objects. 1113 */ 1114 vm_object_pip_add(obj, 1); 1115 if ((obj->flags & OBJ_DEAD) != 0) { 1116 /* 1117 * Note that object might be already reused from the 1118 * vnode, and the OBJ_DEAD flag cleared. This is fine, 1119 * we recheck for DOOMED vnode state after all pages 1120 * are busied, and retract then. 1121 * 1122 * But we check for OBJ_DEAD to ensure that we do not 1123 * busy pages while vm_object_terminate_pages() 1124 * processes the queue. 1125 */ 1126 error = EJUSTRETURN; 1127 goto out_pip; 1128 } 1129 1130 resid = uio->uio_resid; 1131 off = uio->uio_offset; 1132 for (i = 0; resid > 0; i++) { 1133 MPASS(i < io_hold_cnt + 2); 1134 ma[i] = vm_page_grab_unlocked(obj, atop(off), 1135 VM_ALLOC_NOCREAT | VM_ALLOC_SBUSY | VM_ALLOC_IGN_SBUSY | 1136 VM_ALLOC_NOWAIT); 1137 if (ma[i] == NULL) 1138 break; 1139 1140 /* 1141 * Skip invalid pages. Valid mask can be partial only 1142 * at EOF, and we clip later. 1143 */ 1144 if (vm_page_none_valid(ma[i])) { 1145 vm_page_sunbusy(ma[i]); 1146 break; 1147 } 1148 1149 resid -= PAGE_SIZE; 1150 off += PAGE_SIZE; 1151 } 1152 if (i == 0) { 1153 error = EJUSTRETURN; 1154 goto out_pip; 1155 } 1156 1157 /* 1158 * Check VIRF_DOOMED after we busied our pages. Since 1159 * vgonel() terminates the vnode' vm_object, it cannot 1160 * process past pages busied by us. 1161 */ 1162 if (VN_IS_DOOMED(vp)) { 1163 error = EJUSTRETURN; 1164 goto out; 1165 } 1166 1167 resid = PAGE_SIZE - (uio->uio_offset & PAGE_MASK) + ptoa(i - 1); 1168 if (resid > uio->uio_resid) 1169 resid = uio->uio_resid; 1170 1171 /* 1172 * Unlocked read of vnp_size is safe because truncation cannot 1173 * pass busied page. But we load vnp_size into a local 1174 * variable so that possible concurrent extension does not 1175 * break calculation. 1176 */ 1177 #if defined(__powerpc__) && !defined(__powerpc64__) 1178 vsz = obj->un_pager.vnp.vnp_size; 1179 #else 1180 vsz = atomic_load_64(&obj->un_pager.vnp.vnp_size); 1181 #endif 1182 if (uio->uio_offset >= vsz) { 1183 error = EJUSTRETURN; 1184 goto out; 1185 } 1186 if (uio->uio_offset + resid > vsz) 1187 resid = vsz - uio->uio_offset; 1188 1189 error = vn_io_fault_pgmove(ma, uio->uio_offset & PAGE_MASK, resid, uio); 1190 1191 out: 1192 for (j = 0; j < i; j++) { 1193 if (error == 0) 1194 vm_page_reference(ma[j]); 1195 vm_page_sunbusy(ma[j]); 1196 } 1197 out_pip: 1198 vm_object_pip_wakeup(obj); 1199 if (error != 0) 1200 return (error); 1201 return (uio->uio_resid == 0 ? 0 : EJUSTRETURN); 1202 } 1203 1204 /* 1205 * File table vnode read routine. 1206 */ 1207 static int 1208 vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1209 struct thread *td) 1210 { 1211 struct vnode *vp; 1212 off_t orig_offset; 1213 int error, ioflag; 1214 int advice; 1215 1216 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1217 uio->uio_td, td)); 1218 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1219 vp = fp->f_vnode; 1220 ioflag = 0; 1221 if (fp->f_flag & FNONBLOCK) 1222 ioflag |= IO_NDELAY; 1223 if (fp->f_flag & O_DIRECT) 1224 ioflag |= IO_DIRECT; 1225 1226 /* 1227 * Try to read from page cache. VIRF_DOOMED check is racy but 1228 * allows us to avoid unneeded work outright. 1229 */ 1230 if (vn_io_pgcache_read_enable && !mac_vnode_check_read_enabled() && 1231 (vn_irflag_read(vp) & (VIRF_DOOMED | VIRF_PGREAD)) == VIRF_PGREAD) { 1232 error = VOP_READ_PGCACHE(vp, uio, ioflag, fp->f_cred); 1233 if (error == 0) { 1234 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1235 return (0); 1236 } 1237 if (error != EJUSTRETURN) 1238 return (error); 1239 } 1240 1241 advice = get_advice(fp, uio); 1242 vn_lock(vp, LK_SHARED | LK_RETRY); 1243 1244 switch (advice) { 1245 case POSIX_FADV_NORMAL: 1246 case POSIX_FADV_SEQUENTIAL: 1247 case POSIX_FADV_NOREUSE: 1248 ioflag |= sequential_heuristic(uio, fp); 1249 break; 1250 case POSIX_FADV_RANDOM: 1251 /* Disable read-ahead for random I/O. */ 1252 break; 1253 } 1254 orig_offset = uio->uio_offset; 1255 1256 #ifdef MAC 1257 error = mac_vnode_check_read(active_cred, fp->f_cred, vp); 1258 if (error == 0) 1259 #endif 1260 error = VOP_READ(vp, uio, ioflag, fp->f_cred); 1261 fp->f_nextoff[UIO_READ] = uio->uio_offset; 1262 VOP_UNLOCK(vp); 1263 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1264 orig_offset != uio->uio_offset) 1265 /* 1266 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1267 * for the backing file after a POSIX_FADV_NOREUSE 1268 * read(2). 1269 */ 1270 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1271 POSIX_FADV_DONTNEED); 1272 return (error); 1273 } 1274 1275 /* 1276 * File table vnode write routine. 1277 */ 1278 static int 1279 vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, 1280 struct thread *td) 1281 { 1282 struct vnode *vp; 1283 struct mount *mp; 1284 off_t orig_offset; 1285 int error, ioflag; 1286 int advice; 1287 bool need_finished_write; 1288 1289 KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", 1290 uio->uio_td, td)); 1291 KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); 1292 vp = fp->f_vnode; 1293 if (vp->v_type == VREG) 1294 bwillwrite(); 1295 ioflag = IO_UNIT; 1296 if (vp->v_type == VREG && (fp->f_flag & O_APPEND) != 0) 1297 ioflag |= IO_APPEND; 1298 if ((fp->f_flag & FNONBLOCK) != 0) 1299 ioflag |= IO_NDELAY; 1300 ioflag |= get_write_ioflag(fp); 1301 1302 mp = NULL; 1303 need_finished_write = false; 1304 if (vp->v_type != VCHR) { 1305 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 1306 if (error != 0) 1307 goto unlock; 1308 need_finished_write = true; 1309 } 1310 1311 advice = get_advice(fp, uio); 1312 1313 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY); 1314 switch (advice) { 1315 case POSIX_FADV_NORMAL: 1316 case POSIX_FADV_SEQUENTIAL: 1317 case POSIX_FADV_NOREUSE: 1318 ioflag |= sequential_heuristic(uio, fp); 1319 break; 1320 case POSIX_FADV_RANDOM: 1321 /* XXX: Is this correct? */ 1322 break; 1323 } 1324 orig_offset = uio->uio_offset; 1325 1326 #ifdef MAC 1327 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1328 if (error == 0) 1329 #endif 1330 error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); 1331 fp->f_nextoff[UIO_WRITE] = uio->uio_offset; 1332 VOP_UNLOCK(vp); 1333 if (need_finished_write) 1334 vn_finished_write(mp); 1335 if (error == 0 && advice == POSIX_FADV_NOREUSE && 1336 orig_offset != uio->uio_offset) 1337 /* 1338 * Use POSIX_FADV_DONTNEED to flush pages and buffers 1339 * for the backing file after a POSIX_FADV_NOREUSE 1340 * write(2). 1341 */ 1342 error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, 1343 POSIX_FADV_DONTNEED); 1344 unlock: 1345 return (error); 1346 } 1347 1348 /* 1349 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to 1350 * prevent the following deadlock: 1351 * 1352 * Assume that the thread A reads from the vnode vp1 into userspace 1353 * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is 1354 * currently not resident, then system ends up with the call chain 1355 * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> 1356 * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) 1357 * which establishes lock order vp1->vn_lock, then vp2->vn_lock. 1358 * If, at the same time, thread B reads from vnode vp2 into buffer buf2 1359 * backed by the pages of vnode vp1, and some page in buf2 is not 1360 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. 1361 * 1362 * To prevent the lock order reversal and deadlock, vn_io_fault() does 1363 * not allow page faults to happen during VOP_READ() or VOP_WRITE(). 1364 * Instead, it first tries to do the whole range i/o with pagefaults 1365 * disabled. If all pages in the i/o buffer are resident and mapped, 1366 * VOP will succeed (ignoring the genuine filesystem errors). 1367 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do 1368 * i/o in chunks, with all pages in the chunk prefaulted and held 1369 * using vm_fault_quick_hold_pages(). 1370 * 1371 * Filesystems using this deadlock avoidance scheme should use the 1372 * array of the held pages from uio, saved in the curthread->td_ma, 1373 * instead of doing uiomove(). A helper function 1374 * vn_io_fault_uiomove() converts uiomove request into 1375 * uiomove_fromphys() over td_ma array. 1376 * 1377 * Since vnode locks do not cover the whole i/o anymore, rangelocks 1378 * make the current i/o request atomic with respect to other i/os and 1379 * truncations. 1380 */ 1381 1382 /* 1383 * Decode vn_io_fault_args and perform the corresponding i/o. 1384 */ 1385 static int 1386 vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, 1387 struct thread *td) 1388 { 1389 int error, save; 1390 1391 error = 0; 1392 save = vm_fault_disable_pagefaults(); 1393 switch (args->kind) { 1394 case VN_IO_FAULT_FOP: 1395 error = (args->args.fop_args.doio)(args->args.fop_args.fp, 1396 uio, args->cred, args->flags, td); 1397 break; 1398 case VN_IO_FAULT_VOP: 1399 switch (uio->uio_rw) { 1400 case UIO_READ: 1401 error = VOP_READ(args->args.vop_args.vp, uio, 1402 args->flags, args->cred); 1403 break; 1404 case UIO_WRITE: 1405 error = VOP_WRITE(args->args.vop_args.vp, uio, 1406 args->flags, args->cred); 1407 break; 1408 } 1409 break; 1410 default: 1411 panic("vn_io_fault_doio: unknown kind of io %d %d", 1412 args->kind, uio->uio_rw); 1413 } 1414 vm_fault_enable_pagefaults(save); 1415 return (error); 1416 } 1417 1418 static int 1419 vn_io_fault_touch(char *base, const struct uio *uio) 1420 { 1421 int r; 1422 1423 r = fubyte(base); 1424 if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) 1425 return (EFAULT); 1426 return (0); 1427 } 1428 1429 static int 1430 vn_io_fault_prefault_user(const struct uio *uio) 1431 { 1432 char *base; 1433 const struct iovec *iov; 1434 size_t len; 1435 ssize_t resid; 1436 int error, i; 1437 1438 KASSERT(uio->uio_segflg == UIO_USERSPACE, 1439 ("vn_io_fault_prefault userspace")); 1440 1441 error = i = 0; 1442 iov = uio->uio_iov; 1443 resid = uio->uio_resid; 1444 base = iov->iov_base; 1445 len = iov->iov_len; 1446 while (resid > 0) { 1447 error = vn_io_fault_touch(base, uio); 1448 if (error != 0) 1449 break; 1450 if (len < PAGE_SIZE) { 1451 if (len != 0) { 1452 error = vn_io_fault_touch(base + len - 1, uio); 1453 if (error != 0) 1454 break; 1455 resid -= len; 1456 } 1457 if (++i >= uio->uio_iovcnt) 1458 break; 1459 iov = uio->uio_iov + i; 1460 base = iov->iov_base; 1461 len = iov->iov_len; 1462 } else { 1463 len -= PAGE_SIZE; 1464 base += PAGE_SIZE; 1465 resid -= PAGE_SIZE; 1466 } 1467 } 1468 return (error); 1469 } 1470 1471 /* 1472 * Common code for vn_io_fault(), agnostic to the kind of i/o request. 1473 * Uses vn_io_fault_doio() to make the call to an actual i/o function. 1474 * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request 1475 * into args and call vn_io_fault1() to handle faults during the user 1476 * mode buffer accesses. 1477 */ 1478 static int 1479 vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, 1480 struct thread *td) 1481 { 1482 vm_page_t ma[io_hold_cnt + 2]; 1483 struct uio *uio_clone, short_uio; 1484 struct iovec short_iovec[1]; 1485 vm_page_t *prev_td_ma; 1486 vm_prot_t prot; 1487 vm_offset_t addr, end; 1488 size_t len, resid; 1489 ssize_t adv; 1490 int error, cnt, saveheld, prev_td_ma_cnt; 1491 1492 if (vn_io_fault_prefault) { 1493 error = vn_io_fault_prefault_user(uio); 1494 if (error != 0) 1495 return (error); /* Or ignore ? */ 1496 } 1497 1498 prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; 1499 1500 /* 1501 * The UFS follows IO_UNIT directive and replays back both 1502 * uio_offset and uio_resid if an error is encountered during the 1503 * operation. But, since the iovec may be already advanced, 1504 * uio is still in an inconsistent state. 1505 * 1506 * Cache a copy of the original uio, which is advanced to the redo 1507 * point using UIO_NOCOPY below. 1508 */ 1509 uio_clone = cloneuio(uio); 1510 resid = uio->uio_resid; 1511 1512 short_uio.uio_segflg = UIO_USERSPACE; 1513 short_uio.uio_rw = uio->uio_rw; 1514 short_uio.uio_td = uio->uio_td; 1515 1516 error = vn_io_fault_doio(args, uio, td); 1517 if (error != EFAULT) 1518 goto out; 1519 1520 atomic_add_long(&vn_io_faults_cnt, 1); 1521 uio_clone->uio_segflg = UIO_NOCOPY; 1522 uiomove(NULL, resid - uio->uio_resid, uio_clone); 1523 uio_clone->uio_segflg = uio->uio_segflg; 1524 1525 saveheld = curthread_pflags_set(TDP_UIOHELD); 1526 prev_td_ma = td->td_ma; 1527 prev_td_ma_cnt = td->td_ma_cnt; 1528 1529 while (uio_clone->uio_resid != 0) { 1530 len = uio_clone->uio_iov->iov_len; 1531 if (len == 0) { 1532 KASSERT(uio_clone->uio_iovcnt >= 1, 1533 ("iovcnt underflow")); 1534 uio_clone->uio_iov++; 1535 uio_clone->uio_iovcnt--; 1536 continue; 1537 } 1538 if (len > ptoa(io_hold_cnt)) 1539 len = ptoa(io_hold_cnt); 1540 addr = (uintptr_t)uio_clone->uio_iov->iov_base; 1541 end = round_page(addr + len); 1542 if (end < addr) { 1543 error = EFAULT; 1544 break; 1545 } 1546 /* 1547 * A perfectly misaligned address and length could cause 1548 * both the start and the end of the chunk to use partial 1549 * page. +2 accounts for such a situation. 1550 */ 1551 cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, 1552 addr, len, prot, ma, io_hold_cnt + 2); 1553 if (cnt == -1) { 1554 error = EFAULT; 1555 break; 1556 } 1557 short_uio.uio_iov = &short_iovec[0]; 1558 short_iovec[0].iov_base = (void *)addr; 1559 short_uio.uio_iovcnt = 1; 1560 short_uio.uio_resid = short_iovec[0].iov_len = len; 1561 short_uio.uio_offset = uio_clone->uio_offset; 1562 td->td_ma = ma; 1563 td->td_ma_cnt = cnt; 1564 1565 error = vn_io_fault_doio(args, &short_uio, td); 1566 vm_page_unhold_pages(ma, cnt); 1567 adv = len - short_uio.uio_resid; 1568 1569 uio_clone->uio_iov->iov_base = 1570 (char *)uio_clone->uio_iov->iov_base + adv; 1571 uio_clone->uio_iov->iov_len -= adv; 1572 uio_clone->uio_resid -= adv; 1573 uio_clone->uio_offset += adv; 1574 1575 uio->uio_resid -= adv; 1576 uio->uio_offset += adv; 1577 1578 if (error != 0 || adv == 0) 1579 break; 1580 } 1581 td->td_ma = prev_td_ma; 1582 td->td_ma_cnt = prev_td_ma_cnt; 1583 curthread_pflags_restore(saveheld); 1584 out: 1585 freeuio(uio_clone); 1586 return (error); 1587 } 1588 1589 static int 1590 vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, 1591 int flags, struct thread *td) 1592 { 1593 fo_rdwr_t *doio; 1594 struct vnode *vp; 1595 void *rl_cookie; 1596 struct vn_io_fault_args args; 1597 int error; 1598 bool do_io_fault, do_rangelock; 1599 1600 doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; 1601 vp = fp->f_vnode; 1602 1603 /* 1604 * The ability to read(2) on a directory has historically been 1605 * allowed for all users, but this can and has been the source of 1606 * at least one security issue in the past. As such, it is now hidden 1607 * away behind a sysctl for those that actually need it to use it, and 1608 * restricted to root when it's turned on to make it relatively safe to 1609 * leave on for longer sessions of need. 1610 */ 1611 if (vp->v_type == VDIR) { 1612 KASSERT(uio->uio_rw == UIO_READ, 1613 ("illegal write attempted on a directory")); 1614 if (!vfs_allow_read_dir) 1615 return (EISDIR); 1616 if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0) 1617 return (EISDIR); 1618 } 1619 1620 do_io_fault = do_vn_io_fault(vp, uio); 1621 do_rangelock = do_io_fault || (vn_irflag_read(vp) & VIRF_PGREAD) != 0; 1622 foffset_lock_uio(fp, uio, flags); 1623 if (do_rangelock) { 1624 if (uio->uio_rw == UIO_READ) { 1625 rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, 1626 uio->uio_offset + uio->uio_resid); 1627 } else if ((fp->f_flag & O_APPEND) != 0 || 1628 (flags & FOF_OFFSET) == 0) { 1629 /* For appenders, punt and lock the whole range. */ 1630 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1631 } else { 1632 rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, 1633 uio->uio_offset + uio->uio_resid); 1634 } 1635 } 1636 if (do_io_fault) { 1637 args.kind = VN_IO_FAULT_FOP; 1638 args.args.fop_args.fp = fp; 1639 args.args.fop_args.doio = doio; 1640 args.cred = active_cred; 1641 args.flags = flags | FOF_OFFSET; 1642 error = vn_io_fault1(vp, uio, &args, td); 1643 } else { 1644 error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); 1645 } 1646 if (do_rangelock) 1647 vn_rangelock_unlock(vp, rl_cookie); 1648 foffset_unlock_uio(fp, uio, flags); 1649 return (error); 1650 } 1651 1652 /* 1653 * Helper function to perform the requested uiomove operation using 1654 * the held pages for io->uio_iov[0].iov_base buffer instead of 1655 * copyin/copyout. Access to the pages with uiomove_fromphys() 1656 * instead of iov_base prevents page faults that could occur due to 1657 * pmap_collect() invalidating the mapping created by 1658 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or 1659 * object cleanup revoking the write access from page mappings. 1660 * 1661 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() 1662 * instead of plain uiomove(). 1663 */ 1664 int 1665 vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) 1666 { 1667 struct uio transp_uio; 1668 struct iovec transp_iov[1]; 1669 struct thread *td; 1670 size_t adv; 1671 int error, pgadv; 1672 1673 td = curthread; 1674 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1675 uio->uio_segflg != UIO_USERSPACE) 1676 return (uiomove(data, xfersize, uio)); 1677 1678 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1679 transp_iov[0].iov_base = data; 1680 transp_uio.uio_iov = &transp_iov[0]; 1681 transp_uio.uio_iovcnt = 1; 1682 if (xfersize > uio->uio_resid) 1683 xfersize = uio->uio_resid; 1684 transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; 1685 transp_uio.uio_offset = 0; 1686 transp_uio.uio_segflg = UIO_SYSSPACE; 1687 /* 1688 * Since transp_iov points to data, and td_ma page array 1689 * corresponds to original uio->uio_iov, we need to invert the 1690 * direction of the i/o operation as passed to 1691 * uiomove_fromphys(). 1692 */ 1693 switch (uio->uio_rw) { 1694 case UIO_WRITE: 1695 transp_uio.uio_rw = UIO_READ; 1696 break; 1697 case UIO_READ: 1698 transp_uio.uio_rw = UIO_WRITE; 1699 break; 1700 } 1701 transp_uio.uio_td = uio->uio_td; 1702 error = uiomove_fromphys(td->td_ma, 1703 ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, 1704 xfersize, &transp_uio); 1705 adv = xfersize - transp_uio.uio_resid; 1706 pgadv = 1707 (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - 1708 (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); 1709 td->td_ma += pgadv; 1710 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1711 pgadv)); 1712 td->td_ma_cnt -= pgadv; 1713 uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; 1714 uio->uio_iov->iov_len -= adv; 1715 uio->uio_resid -= adv; 1716 uio->uio_offset += adv; 1717 return (error); 1718 } 1719 1720 int 1721 vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, 1722 struct uio *uio) 1723 { 1724 struct thread *td; 1725 vm_offset_t iov_base; 1726 int cnt, pgadv; 1727 1728 td = curthread; 1729 if ((td->td_pflags & TDP_UIOHELD) == 0 || 1730 uio->uio_segflg != UIO_USERSPACE) 1731 return (uiomove_fromphys(ma, offset, xfersize, uio)); 1732 1733 KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); 1734 cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; 1735 iov_base = (vm_offset_t)uio->uio_iov->iov_base; 1736 switch (uio->uio_rw) { 1737 case UIO_WRITE: 1738 pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, 1739 offset, cnt); 1740 break; 1741 case UIO_READ: 1742 pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, 1743 cnt); 1744 break; 1745 } 1746 pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); 1747 td->td_ma += pgadv; 1748 KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, 1749 pgadv)); 1750 td->td_ma_cnt -= pgadv; 1751 uio->uio_iov->iov_base = (char *)(iov_base + cnt); 1752 uio->uio_iov->iov_len -= cnt; 1753 uio->uio_resid -= cnt; 1754 uio->uio_offset += cnt; 1755 return (0); 1756 } 1757 1758 /* 1759 * File table truncate routine. 1760 */ 1761 static int 1762 vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, 1763 struct thread *td) 1764 { 1765 struct mount *mp; 1766 struct vnode *vp; 1767 void *rl_cookie; 1768 int error; 1769 1770 vp = fp->f_vnode; 1771 1772 retry: 1773 /* 1774 * Lock the whole range for truncation. Otherwise split i/o 1775 * might happen partly before and partly after the truncation. 1776 */ 1777 rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); 1778 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 1779 if (error) 1780 goto out1; 1781 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1782 AUDIT_ARG_VNODE1(vp); 1783 if (vp->v_type == VDIR) { 1784 error = EISDIR; 1785 goto out; 1786 } 1787 #ifdef MAC 1788 error = mac_vnode_check_write(active_cred, fp->f_cred, vp); 1789 if (error) 1790 goto out; 1791 #endif 1792 error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, 1793 fp->f_cred); 1794 out: 1795 VOP_UNLOCK(vp); 1796 vn_finished_write(mp); 1797 out1: 1798 vn_rangelock_unlock(vp, rl_cookie); 1799 if (error == ERELOOKUP) 1800 goto retry; 1801 return (error); 1802 } 1803 1804 /* 1805 * Truncate a file that is already locked. 1806 */ 1807 int 1808 vn_truncate_locked(struct vnode *vp, off_t length, bool sync, 1809 struct ucred *cred) 1810 { 1811 struct vattr vattr; 1812 int error; 1813 1814 error = VOP_ADD_WRITECOUNT(vp, 1); 1815 if (error == 0) { 1816 VATTR_NULL(&vattr); 1817 vattr.va_size = length; 1818 if (sync) 1819 vattr.va_vaflags |= VA_SYNC; 1820 error = VOP_SETATTR(vp, &vattr, cred); 1821 VOP_ADD_WRITECOUNT_CHECKED(vp, -1); 1822 if (error == 0) 1823 INOTIFY(vp, IN_MODIFY); 1824 } 1825 return (error); 1826 } 1827 1828 /* 1829 * File table vnode stat routine. 1830 */ 1831 int 1832 vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred) 1833 { 1834 struct vnode *vp = fp->f_vnode; 1835 int error; 1836 1837 vn_lock(vp, LK_SHARED | LK_RETRY); 1838 error = VOP_STAT(vp, sb, active_cred, fp->f_cred); 1839 VOP_UNLOCK(vp); 1840 1841 return (error); 1842 } 1843 1844 /* 1845 * File table vnode ioctl routine. 1846 */ 1847 static int 1848 vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, 1849 struct thread *td) 1850 { 1851 struct vnode *vp; 1852 struct fiobmap2_arg *bmarg; 1853 off_t size; 1854 int error; 1855 1856 vp = fp->f_vnode; 1857 switch (vp->v_type) { 1858 case VDIR: 1859 case VREG: 1860 switch (com) { 1861 case FIONREAD: 1862 error = vn_getsize(vp, &size, active_cred); 1863 if (error == 0) 1864 *(int *)data = size - fp->f_offset; 1865 return (error); 1866 case FIOBMAP2: 1867 bmarg = (struct fiobmap2_arg *)data; 1868 vn_lock(vp, LK_SHARED | LK_RETRY); 1869 #ifdef MAC 1870 error = mac_vnode_check_read(active_cred, fp->f_cred, 1871 vp); 1872 if (error == 0) 1873 #endif 1874 error = VOP_BMAP(vp, bmarg->bn, NULL, 1875 &bmarg->bn, &bmarg->runp, &bmarg->runb); 1876 VOP_UNLOCK(vp); 1877 return (error); 1878 case FIONBIO: 1879 case FIOASYNC: 1880 return (0); 1881 default: 1882 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1883 active_cred, td)); 1884 } 1885 break; 1886 case VCHR: 1887 return (VOP_IOCTL(vp, com, data, fp->f_flag, 1888 active_cred, td)); 1889 default: 1890 return (ENOTTY); 1891 } 1892 } 1893 1894 /* 1895 * File table vnode poll routine. 1896 */ 1897 static int 1898 vn_poll(struct file *fp, int events, struct ucred *active_cred, 1899 struct thread *td) 1900 { 1901 struct vnode *vp; 1902 int error; 1903 1904 vp = fp->f_vnode; 1905 #if defined(MAC) || defined(AUDIT) 1906 if (AUDITING_TD(td) || mac_vnode_check_poll_enabled()) { 1907 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 1908 AUDIT_ARG_VNODE1(vp); 1909 error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); 1910 VOP_UNLOCK(vp); 1911 if (error != 0) 1912 return (error); 1913 } 1914 #endif 1915 error = VOP_POLL(vp, events, fp->f_cred, td); 1916 return (error); 1917 } 1918 1919 /* 1920 * Acquire the requested lock and then check for validity. LK_RETRY 1921 * permits vn_lock to return doomed vnodes. 1922 */ 1923 static int __noinline 1924 _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, 1925 int error) 1926 { 1927 1928 KASSERT((flags & LK_RETRY) == 0 || error == 0, 1929 ("vn_lock: error %d incompatible with flags %#x", error, flags)); 1930 1931 if (error == 0) 1932 VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); 1933 1934 if ((flags & LK_RETRY) == 0) { 1935 if (error == 0) { 1936 VOP_UNLOCK(vp); 1937 error = ENOENT; 1938 } 1939 return (error); 1940 } 1941 1942 /* 1943 * LK_RETRY case. 1944 * 1945 * Nothing to do if we got the lock. 1946 */ 1947 if (error == 0) 1948 return (0); 1949 1950 /* 1951 * Interlock was dropped by the call in _vn_lock. 1952 */ 1953 flags &= ~LK_INTERLOCK; 1954 do { 1955 error = VOP_LOCK1(vp, flags, file, line); 1956 } while (error != 0); 1957 return (0); 1958 } 1959 1960 int 1961 _vn_lock(struct vnode *vp, int flags, const char *file, int line) 1962 { 1963 int error; 1964 1965 VNASSERT((flags & LK_TYPE_MASK) != 0, vp, 1966 ("vn_lock: no locktype (%d passed)", flags)); 1967 VNPASS(vp->v_holdcnt > 0, vp); 1968 error = VOP_LOCK1(vp, flags, file, line); 1969 if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) 1970 return (_vn_lock_fallback(vp, flags, file, line, error)); 1971 return (0); 1972 } 1973 1974 /* 1975 * File table vnode close routine. 1976 */ 1977 static int 1978 vn_closefile(struct file *fp, struct thread *td) 1979 { 1980 struct vnode *vp; 1981 struct flock lf; 1982 int error; 1983 bool ref; 1984 1985 vp = fp->f_vnode; 1986 fp->f_ops = &badfileops; 1987 ref = (fp->f_flag & FHASLOCK) != 0; 1988 1989 error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); 1990 1991 if (__predict_false(ref)) { 1992 lf.l_whence = SEEK_SET; 1993 lf.l_start = 0; 1994 lf.l_len = 0; 1995 lf.l_type = F_UNLCK; 1996 (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); 1997 vrele(vp); 1998 } 1999 return (error); 2000 } 2001 2002 /* 2003 * Preparing to start a filesystem write operation. If the operation is 2004 * permitted, then we bump the count of operations in progress and 2005 * proceed. If a suspend request is in progress, we wait until the 2006 * suspension is over, and then proceed. 2007 */ 2008 static int 2009 vn_start_write_refed(struct mount *mp, int flags, bool mplocked) 2010 { 2011 struct mount_pcpu *mpcpu; 2012 int error, mflags; 2013 2014 if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && 2015 vfs_op_thread_enter(mp, mpcpu)) { 2016 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 2017 vfs_mp_count_add_pcpu(mpcpu, writeopcount, 1); 2018 vfs_op_thread_exit(mp, mpcpu); 2019 return (0); 2020 } 2021 2022 if (mplocked) 2023 mtx_assert(MNT_MTX(mp), MA_OWNED); 2024 else 2025 MNT_ILOCK(mp); 2026 2027 error = 0; 2028 2029 /* 2030 * Check on status of suspension. 2031 */ 2032 if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || 2033 mp->mnt_susp_owner != curthread) { 2034 mflags = 0; 2035 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) { 2036 if (flags & V_PCATCH) 2037 mflags |= PCATCH; 2038 } 2039 mflags |= PRI_MAX_KERN; 2040 while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 2041 if ((flags & V_NOWAIT) != 0) { 2042 error = EWOULDBLOCK; 2043 goto unlock; 2044 } 2045 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, 2046 "suspfs", 0); 2047 if (error != 0) 2048 goto unlock; 2049 } 2050 } 2051 if ((flags & V_XSLEEP) != 0) 2052 goto unlock; 2053 mp->mnt_writeopcount++; 2054 unlock: 2055 if (error != 0 || (flags & V_XSLEEP) != 0) 2056 MNT_REL(mp); 2057 MNT_IUNLOCK(mp); 2058 return (error); 2059 } 2060 2061 int 2062 vn_start_write(struct vnode *vp, struct mount **mpp, int flags) 2063 { 2064 struct mount *mp; 2065 int error; 2066 2067 KASSERT((flags & ~V_VALID_FLAGS) == 0, 2068 ("%s: invalid flags passed %d\n", __func__, flags)); 2069 2070 error = 0; 2071 /* 2072 * If a vnode is provided, get and return the mount point that 2073 * to which it will write. 2074 */ 2075 if (vp != NULL) { 2076 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 2077 *mpp = NULL; 2078 if (error != EOPNOTSUPP) 2079 return (error); 2080 return (0); 2081 } 2082 } 2083 if ((mp = *mpp) == NULL) 2084 return (0); 2085 2086 /* 2087 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 2088 * a vfs_ref(). 2089 * As long as a vnode is not provided we need to acquire a 2090 * refcount for the provided mountpoint too, in order to 2091 * emulate a vfs_ref(). 2092 */ 2093 if (vp == NULL) 2094 vfs_ref(mp); 2095 2096 error = vn_start_write_refed(mp, flags, false); 2097 if (error != 0 && (flags & V_NOWAIT) == 0) 2098 *mpp = NULL; 2099 return (error); 2100 } 2101 2102 /* 2103 * Secondary suspension. Used by operations such as vop_inactive 2104 * routines that are needed by the higher level functions. These 2105 * are allowed to proceed until all the higher level functions have 2106 * completed (indicated by mnt_writeopcount dropping to zero). At that 2107 * time, these operations are halted until the suspension is over. 2108 */ 2109 int 2110 vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) 2111 { 2112 struct mount *mp; 2113 int error, mflags; 2114 2115 KASSERT((flags & (~V_VALID_FLAGS | V_XSLEEP)) == 0, 2116 ("%s: invalid flags passed %d\n", __func__, flags)); 2117 2118 retry: 2119 if (vp != NULL) { 2120 if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { 2121 *mpp = NULL; 2122 if (error != EOPNOTSUPP) 2123 return (error); 2124 return (0); 2125 } 2126 } 2127 /* 2128 * If we are not suspended or have not yet reached suspended 2129 * mode, then let the operation proceed. 2130 */ 2131 if ((mp = *mpp) == NULL) 2132 return (0); 2133 2134 /* 2135 * VOP_GETWRITEMOUNT() returns with the mp refcount held through 2136 * a vfs_ref(). 2137 * As long as a vnode is not provided we need to acquire a 2138 * refcount for the provided mountpoint too, in order to 2139 * emulate a vfs_ref(). 2140 */ 2141 MNT_ILOCK(mp); 2142 if (vp == NULL) 2143 MNT_REF(mp); 2144 if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { 2145 mp->mnt_secondary_writes++; 2146 mp->mnt_secondary_accwrites++; 2147 MNT_IUNLOCK(mp); 2148 return (0); 2149 } 2150 if ((flags & V_NOWAIT) != 0) { 2151 MNT_REL(mp); 2152 MNT_IUNLOCK(mp); 2153 *mpp = NULL; 2154 return (EWOULDBLOCK); 2155 } 2156 /* 2157 * Wait for the suspension to finish. 2158 */ 2159 mflags = 0; 2160 if ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0) { 2161 if ((flags & V_PCATCH) != 0) 2162 mflags |= PCATCH; 2163 } 2164 mflags |= PRI_MAX_KERN | PDROP; 2165 error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0); 2166 vfs_rel(mp); 2167 if (error == 0) 2168 goto retry; 2169 *mpp = NULL; 2170 return (error); 2171 } 2172 2173 /* 2174 * Filesystem write operation has completed. If we are suspending and this 2175 * operation is the last one, notify the suspender that the suspension is 2176 * now in effect. 2177 */ 2178 void 2179 vn_finished_write(struct mount *mp) 2180 { 2181 struct mount_pcpu *mpcpu; 2182 int c; 2183 2184 if (mp == NULL) 2185 return; 2186 2187 if (vfs_op_thread_enter(mp, mpcpu)) { 2188 vfs_mp_count_sub_pcpu(mpcpu, writeopcount, 1); 2189 vfs_mp_count_sub_pcpu(mpcpu, ref, 1); 2190 vfs_op_thread_exit(mp, mpcpu); 2191 return; 2192 } 2193 2194 MNT_ILOCK(mp); 2195 vfs_assert_mount_counters(mp); 2196 MNT_REL(mp); 2197 c = --mp->mnt_writeopcount; 2198 if (mp->mnt_vfs_ops == 0) { 2199 MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); 2200 MNT_IUNLOCK(mp); 2201 return; 2202 } 2203 if (c < 0) 2204 vfs_dump_mount_counters(mp); 2205 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) 2206 wakeup(&mp->mnt_writeopcount); 2207 MNT_IUNLOCK(mp); 2208 } 2209 2210 /* 2211 * Filesystem secondary write operation has completed. If we are 2212 * suspending and this operation is the last one, notify the suspender 2213 * that the suspension is now in effect. 2214 */ 2215 void 2216 vn_finished_secondary_write(struct mount *mp) 2217 { 2218 if (mp == NULL) 2219 return; 2220 MNT_ILOCK(mp); 2221 MNT_REL(mp); 2222 mp->mnt_secondary_writes--; 2223 if (mp->mnt_secondary_writes < 0) 2224 panic("vn_finished_secondary_write: neg cnt"); 2225 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && 2226 mp->mnt_secondary_writes <= 0) 2227 wakeup(&mp->mnt_secondary_writes); 2228 MNT_IUNLOCK(mp); 2229 } 2230 2231 /* 2232 * Request a filesystem to suspend write operations. 2233 */ 2234 int 2235 vfs_write_suspend(struct mount *mp, int flags) 2236 { 2237 int error; 2238 2239 vfs_op_enter(mp); 2240 2241 MNT_ILOCK(mp); 2242 vfs_assert_mount_counters(mp); 2243 if (mp->mnt_susp_owner == curthread) { 2244 vfs_op_exit_locked(mp); 2245 MNT_IUNLOCK(mp); 2246 return (EALREADY); 2247 } 2248 while (mp->mnt_kern_flag & MNTK_SUSPEND) 2249 msleep(&mp->mnt_flag, MNT_MTX(mp), PRI_MAX_KERN, "wsuspfs", 0); 2250 2251 /* 2252 * Unmount holds a write reference on the mount point. If we 2253 * own busy reference and drain for writers, we deadlock with 2254 * the reference draining in the unmount path. Callers of 2255 * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if 2256 * vfs_busy() reference is owned and caller is not in the 2257 * unmount context. 2258 */ 2259 if ((flags & VS_SKIP_UNMOUNT) != 0 && 2260 (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { 2261 vfs_op_exit_locked(mp); 2262 MNT_IUNLOCK(mp); 2263 return (EBUSY); 2264 } 2265 2266 mp->mnt_kern_flag |= MNTK_SUSPEND; 2267 mp->mnt_susp_owner = curthread; 2268 if (mp->mnt_writeopcount > 0) 2269 (void) msleep(&mp->mnt_writeopcount, 2270 MNT_MTX(mp), PRI_MAX_KERN | PDROP, "suspwt", 0); 2271 else 2272 MNT_IUNLOCK(mp); 2273 if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { 2274 vfs_write_resume(mp, 0); 2275 /* vfs_write_resume does vfs_op_exit() for us */ 2276 } 2277 return (error); 2278 } 2279 2280 /* 2281 * Request a filesystem to resume write operations. 2282 */ 2283 void 2284 vfs_write_resume(struct mount *mp, int flags) 2285 { 2286 2287 MNT_ILOCK(mp); 2288 if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { 2289 KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); 2290 mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | 2291 MNTK_SUSPENDED); 2292 mp->mnt_susp_owner = NULL; 2293 wakeup(&mp->mnt_writeopcount); 2294 wakeup(&mp->mnt_flag); 2295 curthread->td_pflags &= ~TDP_IGNSUSP; 2296 if ((flags & VR_START_WRITE) != 0) { 2297 MNT_REF(mp); 2298 mp->mnt_writeopcount++; 2299 } 2300 MNT_IUNLOCK(mp); 2301 if ((flags & VR_NO_SUSPCLR) == 0) 2302 VFS_SUSP_CLEAN(mp); 2303 vfs_op_exit(mp); 2304 } else if ((flags & VR_START_WRITE) != 0) { 2305 MNT_REF(mp); 2306 vn_start_write_refed(mp, 0, true); 2307 } else { 2308 MNT_IUNLOCK(mp); 2309 } 2310 } 2311 2312 /* 2313 * Helper loop around vfs_write_suspend() for filesystem unmount VFS 2314 * methods. 2315 */ 2316 int 2317 vfs_write_suspend_umnt(struct mount *mp) 2318 { 2319 int error; 2320 2321 KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, 2322 ("vfs_write_suspend_umnt: recursed")); 2323 2324 /* dounmount() already called vn_start_write(). */ 2325 for (;;) { 2326 vn_finished_write(mp); 2327 error = vfs_write_suspend(mp, 0); 2328 if (error != 0) { 2329 vn_start_write(NULL, &mp, V_WAIT); 2330 return (error); 2331 } 2332 MNT_ILOCK(mp); 2333 if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) 2334 break; 2335 MNT_IUNLOCK(mp); 2336 vn_start_write(NULL, &mp, V_WAIT); 2337 } 2338 mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); 2339 wakeup(&mp->mnt_flag); 2340 MNT_IUNLOCK(mp); 2341 curthread->td_pflags |= TDP_IGNSUSP; 2342 return (0); 2343 } 2344 2345 /* 2346 * Implement kqueues for files by translating it to vnode operation. 2347 */ 2348 static int 2349 vn_kqfilter(struct file *fp, struct knote *kn) 2350 { 2351 2352 return (VOP_KQFILTER(fp->f_vnode, kn)); 2353 } 2354 2355 int 2356 vn_kqfilter_opath(struct file *fp, struct knote *kn) 2357 { 2358 if ((fp->f_flag & FKQALLOWED) == 0) 2359 return (EBADF); 2360 return (vn_kqfilter(fp, kn)); 2361 } 2362 2363 /* 2364 * Simplified in-kernel wrapper calls for extended attribute access. 2365 * Both calls pass in a NULL credential, authorizing as "kernel" access. 2366 * Set IO_NODELOCKED in ioflg if the vnode is already locked. 2367 */ 2368 int 2369 vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, 2370 const char *attrname, int *buflen, char *buf, struct thread *td) 2371 { 2372 struct uio auio; 2373 struct iovec iov; 2374 int error; 2375 2376 iov.iov_len = *buflen; 2377 iov.iov_base = buf; 2378 2379 auio.uio_iov = &iov; 2380 auio.uio_iovcnt = 1; 2381 auio.uio_rw = UIO_READ; 2382 auio.uio_segflg = UIO_SYSSPACE; 2383 auio.uio_td = td; 2384 auio.uio_offset = 0; 2385 auio.uio_resid = *buflen; 2386 2387 if ((ioflg & IO_NODELOCKED) == 0) 2388 vn_lock(vp, LK_SHARED | LK_RETRY); 2389 2390 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2391 2392 /* authorize attribute retrieval as kernel */ 2393 error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, 2394 td); 2395 2396 if ((ioflg & IO_NODELOCKED) == 0) 2397 VOP_UNLOCK(vp); 2398 2399 if (error == 0) { 2400 *buflen = *buflen - auio.uio_resid; 2401 } 2402 2403 return (error); 2404 } 2405 2406 /* 2407 * XXX failure mode if partially written? 2408 */ 2409 int 2410 vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, 2411 const char *attrname, int buflen, char *buf, struct thread *td) 2412 { 2413 struct uio auio; 2414 struct iovec iov; 2415 struct mount *mp; 2416 int error; 2417 2418 iov.iov_len = buflen; 2419 iov.iov_base = buf; 2420 2421 auio.uio_iov = &iov; 2422 auio.uio_iovcnt = 1; 2423 auio.uio_rw = UIO_WRITE; 2424 auio.uio_segflg = UIO_SYSSPACE; 2425 auio.uio_td = td; 2426 auio.uio_offset = 0; 2427 auio.uio_resid = buflen; 2428 2429 if ((ioflg & IO_NODELOCKED) == 0) { 2430 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2431 return (error); 2432 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2433 } 2434 2435 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2436 2437 /* authorize attribute setting as kernel */ 2438 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); 2439 2440 if ((ioflg & IO_NODELOCKED) == 0) { 2441 vn_finished_write(mp); 2442 VOP_UNLOCK(vp); 2443 } 2444 2445 return (error); 2446 } 2447 2448 int 2449 vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, 2450 const char *attrname, struct thread *td) 2451 { 2452 struct mount *mp; 2453 int error; 2454 2455 if ((ioflg & IO_NODELOCKED) == 0) { 2456 if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) 2457 return (error); 2458 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); 2459 } 2460 2461 ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); 2462 2463 /* authorize attribute removal as kernel */ 2464 error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); 2465 if (error == EOPNOTSUPP) 2466 error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, 2467 NULL, td); 2468 2469 if ((ioflg & IO_NODELOCKED) == 0) { 2470 vn_finished_write(mp); 2471 VOP_UNLOCK(vp); 2472 } 2473 2474 return (error); 2475 } 2476 2477 static int 2478 vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, 2479 struct vnode **rvp) 2480 { 2481 2482 return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); 2483 } 2484 2485 int 2486 vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) 2487 { 2488 2489 return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, 2490 lkflags, rvp)); 2491 } 2492 2493 int 2494 vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, 2495 int lkflags, struct vnode **rvp) 2496 { 2497 struct mount *mp; 2498 int ltype, error; 2499 2500 ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); 2501 mp = vp->v_mount; 2502 ltype = VOP_ISLOCKED(vp); 2503 KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, 2504 ("vn_vget_ino: vp not locked")); 2505 error = vfs_busy(mp, MBF_NOWAIT); 2506 if (error != 0) { 2507 vfs_ref(mp); 2508 VOP_UNLOCK(vp); 2509 error = vfs_busy(mp, 0); 2510 vn_lock(vp, ltype | LK_RETRY); 2511 vfs_rel(mp); 2512 if (error != 0) 2513 return (ENOENT); 2514 if (VN_IS_DOOMED(vp)) { 2515 vfs_unbusy(mp); 2516 return (ENOENT); 2517 } 2518 } 2519 VOP_UNLOCK(vp); 2520 error = alloc(mp, alloc_arg, lkflags, rvp); 2521 vfs_unbusy(mp); 2522 if (error != 0 || *rvp != vp) 2523 vn_lock(vp, ltype | LK_RETRY); 2524 if (VN_IS_DOOMED(vp)) { 2525 if (error == 0) { 2526 if (*rvp == vp) 2527 vunref(vp); 2528 else 2529 vput(*rvp); 2530 } 2531 error = ENOENT; 2532 } 2533 return (error); 2534 } 2535 2536 static void 2537 vn_send_sigxfsz(struct proc *p) 2538 { 2539 PROC_LOCK(p); 2540 kern_psignal(p, SIGXFSZ); 2541 PROC_UNLOCK(p); 2542 } 2543 2544 int 2545 vn_rlimit_trunc(u_quad_t size, struct thread *td) 2546 { 2547 if (size <= lim_cur(td, RLIMIT_FSIZE)) 2548 return (0); 2549 vn_send_sigxfsz(td->td_proc); 2550 return (EFBIG); 2551 } 2552 2553 static int 2554 vn_rlimit_fsizex1(const struct vnode *vp, struct uio *uio, off_t maxfsz, 2555 bool adj, struct thread *td) 2556 { 2557 off_t lim; 2558 bool ktr_write; 2559 2560 if (vp->v_type != VREG) 2561 return (0); 2562 2563 /* 2564 * Handle file system maximum file size. 2565 */ 2566 if (maxfsz != 0 && uio->uio_offset + uio->uio_resid > maxfsz) { 2567 if (!adj || uio->uio_offset >= maxfsz) 2568 return (EFBIG); 2569 uio->uio_resid = maxfsz - uio->uio_offset; 2570 } 2571 2572 /* 2573 * This is kernel write (e.g. vnode_pager) or accounting 2574 * write, ignore limit. 2575 */ 2576 if (td == NULL || (td->td_pflags2 & TDP2_ACCT) != 0) 2577 return (0); 2578 2579 /* 2580 * Calculate file size limit. 2581 */ 2582 ktr_write = (td->td_pflags & TDP_INKTRACE) != 0; 2583 lim = __predict_false(ktr_write) ? td->td_ktr_io_lim : 2584 lim_cur(td, RLIMIT_FSIZE); 2585 2586 /* 2587 * Is the limit reached? 2588 */ 2589 if (__predict_true((uoff_t)uio->uio_offset + uio->uio_resid <= lim)) 2590 return (0); 2591 2592 /* 2593 * Prepared filesystems can handle writes truncated to the 2594 * file size limit. 2595 */ 2596 if (adj && (uoff_t)uio->uio_offset < lim) { 2597 uio->uio_resid = lim - (uoff_t)uio->uio_offset; 2598 return (0); 2599 } 2600 2601 if (!ktr_write || ktr_filesize_limit_signal) 2602 vn_send_sigxfsz(td->td_proc); 2603 return (EFBIG); 2604 } 2605 2606 /* 2607 * Helper for VOP_WRITE() implementations, the common code to 2608 * handle maximum supported file size on the filesystem, and 2609 * RLIMIT_FSIZE, except for special writes from accounting subsystem 2610 * and ktrace. 2611 * 2612 * For maximum file size (maxfsz argument): 2613 * - return EFBIG if uio_offset is beyond it 2614 * - otherwise, clamp uio_resid if write would extend file beyond maxfsz. 2615 * 2616 * For RLIMIT_FSIZE: 2617 * - return EFBIG and send SIGXFSZ if uio_offset is beyond the limit 2618 * - otherwise, clamp uio_resid if write would extend file beyond limit. 2619 * 2620 * If clamping occured, the adjustment for uio_resid is stored in 2621 * *resid_adj, to be re-applied by vn_rlimit_fsizex_res() on return 2622 * from the VOP. 2623 */ 2624 int 2625 vn_rlimit_fsizex(const struct vnode *vp, struct uio *uio, off_t maxfsz, 2626 ssize_t *resid_adj, struct thread *td) 2627 { 2628 ssize_t resid_orig; 2629 int error; 2630 bool adj; 2631 2632 resid_orig = uio->uio_resid; 2633 adj = resid_adj != NULL; 2634 error = vn_rlimit_fsizex1(vp, uio, maxfsz, adj, td); 2635 if (adj) 2636 *resid_adj = resid_orig - uio->uio_resid; 2637 return (error); 2638 } 2639 2640 void 2641 vn_rlimit_fsizex_res(struct uio *uio, ssize_t resid_adj) 2642 { 2643 uio->uio_resid += resid_adj; 2644 } 2645 2646 int 2647 vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, 2648 struct thread *td) 2649 { 2650 return (vn_rlimit_fsizex(vp, __DECONST(struct uio *, uio), 0, NULL, 2651 td)); 2652 } 2653 2654 int 2655 vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, 2656 struct thread *td) 2657 { 2658 struct vnode *vp; 2659 2660 vp = fp->f_vnode; 2661 #ifdef AUDIT 2662 vn_lock(vp, LK_SHARED | LK_RETRY); 2663 AUDIT_ARG_VNODE1(vp); 2664 VOP_UNLOCK(vp); 2665 #endif 2666 return (setfmode(td, active_cred, vp, mode)); 2667 } 2668 2669 int 2670 vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, 2671 struct thread *td) 2672 { 2673 struct vnode *vp; 2674 2675 vp = fp->f_vnode; 2676 #ifdef AUDIT 2677 vn_lock(vp, LK_SHARED | LK_RETRY); 2678 AUDIT_ARG_VNODE1(vp); 2679 VOP_UNLOCK(vp); 2680 #endif 2681 return (setfown(td, active_cred, vp, uid, gid)); 2682 } 2683 2684 /* 2685 * Remove pages in the range ["start", "end") from the vnode's VM object. If 2686 * "end" is 0, then the range extends to the end of the object. 2687 */ 2688 void 2689 vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2690 { 2691 vm_object_t object; 2692 2693 if ((object = vp->v_object) == NULL) 2694 return; 2695 VM_OBJECT_WLOCK(object); 2696 vm_object_page_remove(object, start, end, 0); 2697 VM_OBJECT_WUNLOCK(object); 2698 } 2699 2700 /* 2701 * Like vn_pages_remove(), but skips invalid pages, which by definition are not 2702 * mapped into any process' address space. Filesystems may use this in 2703 * preference to vn_pages_remove() to avoid blocking on pages busied in 2704 * preparation for a VOP_GETPAGES. 2705 */ 2706 void 2707 vn_pages_remove_valid(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) 2708 { 2709 vm_object_t object; 2710 2711 if ((object = vp->v_object) == NULL) 2712 return; 2713 VM_OBJECT_WLOCK(object); 2714 vm_object_page_remove(object, start, end, OBJPR_VALIDONLY); 2715 VM_OBJECT_WUNLOCK(object); 2716 } 2717 2718 int 2719 vn_bmap_seekhole_locked(struct vnode *vp, u_long cmd, off_t *off, 2720 struct ucred *cred) 2721 { 2722 off_t size; 2723 daddr_t bn, bnp; 2724 uint64_t bsize; 2725 off_t noff; 2726 int error; 2727 2728 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2729 ("%s: Wrong command %lu", __func__, cmd)); 2730 ASSERT_VOP_ELOCKED(vp, "vn_bmap_seekhole_locked"); 2731 2732 if (vp->v_type != VREG) { 2733 error = ENOTTY; 2734 goto out; 2735 } 2736 error = vn_getsize_locked(vp, &size, cred); 2737 if (error != 0) 2738 goto out; 2739 noff = *off; 2740 if (noff < 0 || noff >= size) { 2741 error = ENXIO; 2742 goto out; 2743 } 2744 2745 /* See the comment in ufs_bmap_seekdata(). */ 2746 vnode_pager_clean_sync(vp); 2747 2748 bsize = vp->v_mount->mnt_stat.f_iosize; 2749 for (bn = noff / bsize; noff < size; bn++, noff += bsize - 2750 noff % bsize) { 2751 error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); 2752 if (error == EOPNOTSUPP) { 2753 error = ENOTTY; 2754 goto out; 2755 } 2756 if ((bnp == -1 && cmd == FIOSEEKHOLE) || 2757 (bnp != -1 && cmd == FIOSEEKDATA)) { 2758 noff = bn * bsize; 2759 if (noff < *off) 2760 noff = *off; 2761 goto out; 2762 } 2763 } 2764 if (noff > size) 2765 noff = size; 2766 /* noff == size. There is an implicit hole at the end of file. */ 2767 if (cmd == FIOSEEKDATA) 2768 error = ENXIO; 2769 out: 2770 if (error == 0) 2771 *off = noff; 2772 return (error); 2773 } 2774 2775 int 2776 vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) 2777 { 2778 int error; 2779 2780 KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, 2781 ("%s: Wrong command %lu", __func__, cmd)); 2782 2783 if (vn_lock(vp, LK_EXCLUSIVE) != 0) 2784 return (EBADF); 2785 error = vn_bmap_seekhole_locked(vp, cmd, off, cred); 2786 VOP_UNLOCK(vp); 2787 return (error); 2788 } 2789 2790 int 2791 vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) 2792 { 2793 struct ucred *cred; 2794 struct vnode *vp; 2795 off_t foffset, fsize, size; 2796 int error, noneg; 2797 2798 cred = td->td_ucred; 2799 vp = fp->f_vnode; 2800 noneg = (vp->v_type != VCHR); 2801 /* 2802 * Try to dodge locking for common case of querying the offset. 2803 */ 2804 if (whence == L_INCR && offset == 0) { 2805 foffset = foffset_read(fp); 2806 if (__predict_false(foffset < 0 && noneg)) { 2807 return (EOVERFLOW); 2808 } 2809 td->td_uretoff.tdu_off = foffset; 2810 return (0); 2811 } 2812 foffset = foffset_lock(fp, 0); 2813 error = 0; 2814 switch (whence) { 2815 case L_INCR: 2816 if (noneg && 2817 (foffset < 0 || 2818 (offset > 0 && foffset > OFF_MAX - offset))) { 2819 error = EOVERFLOW; 2820 break; 2821 } 2822 offset += foffset; 2823 break; 2824 case L_XTND: 2825 error = vn_getsize(vp, &fsize, cred); 2826 if (error != 0) 2827 break; 2828 2829 /* 2830 * If the file references a disk device, then fetch 2831 * the media size and use that to determine the ending 2832 * offset. 2833 */ 2834 if (fsize == 0 && vp->v_type == VCHR && 2835 fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) 2836 fsize = size; 2837 if (noneg && offset > 0 && fsize > OFF_MAX - offset) { 2838 error = EOVERFLOW; 2839 break; 2840 } 2841 offset += fsize; 2842 break; 2843 case L_SET: 2844 break; 2845 case SEEK_DATA: 2846 error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); 2847 if (error == ENOTTY) 2848 error = EINVAL; 2849 break; 2850 case SEEK_HOLE: 2851 error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); 2852 if (error == ENOTTY) 2853 error = EINVAL; 2854 break; 2855 default: 2856 error = EINVAL; 2857 } 2858 if (error == 0 && noneg && offset < 0) 2859 error = EINVAL; 2860 if (error != 0) 2861 goto drop; 2862 VFS_KNOTE_UNLOCKED(vp, 0); 2863 td->td_uretoff.tdu_off = offset; 2864 drop: 2865 foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); 2866 return (error); 2867 } 2868 2869 int 2870 vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, 2871 struct thread *td) 2872 { 2873 int error; 2874 2875 /* 2876 * Grant permission if the caller is the owner of the file, or 2877 * the super-user, or has ACL_WRITE_ATTRIBUTES permission on 2878 * on the file. If the time pointer is null, then write 2879 * permission on the file is also sufficient. 2880 * 2881 * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: 2882 * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES 2883 * will be allowed to set the times [..] to the current 2884 * server time. 2885 */ 2886 error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); 2887 if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) 2888 error = VOP_ACCESS(vp, VWRITE, cred, td); 2889 return (error); 2890 } 2891 2892 int 2893 vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) 2894 { 2895 struct vnode *vp; 2896 int error; 2897 2898 if (fp->f_type == DTYPE_FIFO) 2899 kif->kf_type = KF_TYPE_FIFO; 2900 else 2901 kif->kf_type = KF_TYPE_VNODE; 2902 vp = fp->f_vnode; 2903 vref(vp); 2904 FILEDESC_SUNLOCK(fdp); 2905 error = vn_fill_kinfo_vnode(vp, kif); 2906 vrele(vp); 2907 FILEDESC_SLOCK(fdp); 2908 return (error); 2909 } 2910 2911 static inline void 2912 vn_fill_junk(struct kinfo_file *kif) 2913 { 2914 size_t len, olen; 2915 2916 /* 2917 * Simulate vn_fullpath returning changing values for a given 2918 * vp during e.g. coredump. 2919 */ 2920 len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; 2921 olen = strlen(kif->kf_path); 2922 if (len < olen) 2923 strcpy(&kif->kf_path[len - 1], "$"); 2924 else 2925 for (; olen < len; olen++) 2926 strcpy(&kif->kf_path[olen], "A"); 2927 } 2928 2929 int 2930 vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) 2931 { 2932 struct vattr va; 2933 char *fullpath, *freepath; 2934 int error; 2935 2936 kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); 2937 freepath = NULL; 2938 fullpath = "-"; 2939 error = vn_fullpath(vp, &fullpath, &freepath); 2940 if (error == 0) { 2941 strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); 2942 } 2943 if (freepath != NULL) 2944 free(freepath, M_TEMP); 2945 2946 KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, 2947 vn_fill_junk(kif); 2948 ); 2949 2950 /* 2951 * Retrieve vnode attributes. 2952 */ 2953 va.va_fsid = VNOVAL; 2954 va.va_rdev = NODEV; 2955 vn_lock(vp, LK_SHARED | LK_RETRY); 2956 error = VOP_GETATTR(vp, &va, curthread->td_ucred); 2957 VOP_UNLOCK(vp); 2958 if (error != 0) 2959 return (error); 2960 if (va.va_fsid != VNOVAL) 2961 kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; 2962 else 2963 kif->kf_un.kf_file.kf_file_fsid = 2964 vp->v_mount->mnt_stat.f_fsid.val[0]; 2965 kif->kf_un.kf_file.kf_file_fsid_freebsd11 = 2966 kif->kf_un.kf_file.kf_file_fsid; /* truncate */ 2967 kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; 2968 kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); 2969 kif->kf_un.kf_file.kf_file_size = va.va_size; 2970 kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; 2971 kif->kf_un.kf_file.kf_file_rdev_freebsd11 = 2972 kif->kf_un.kf_file.kf_file_rdev; /* truncate */ 2973 kif->kf_un.kf_file.kf_file_nlink = va.va_nlink; 2974 return (0); 2975 } 2976 2977 int 2978 vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, 2979 vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, 2980 struct thread *td) 2981 { 2982 #ifdef HWPMC_HOOKS 2983 struct pmckern_map_in pkm; 2984 #endif 2985 struct mount *mp; 2986 struct vnode *vp; 2987 vm_object_t object; 2988 vm_prot_t maxprot; 2989 boolean_t writecounted; 2990 int error; 2991 2992 #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ 2993 defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) 2994 /* 2995 * POSIX shared-memory objects are defined to have 2996 * kernel persistence, and are not defined to support 2997 * read(2)/write(2) -- or even open(2). Thus, we can 2998 * use MAP_ASYNC to trade on-disk coherence for speed. 2999 * The shm_open(3) library routine turns on the FPOSIXSHM 3000 * flag to request this behavior. 3001 */ 3002 if ((fp->f_flag & FPOSIXSHM) != 0) 3003 flags |= MAP_NOSYNC; 3004 #endif 3005 vp = fp->f_vnode; 3006 3007 /* 3008 * Ensure that file and memory protections are 3009 * compatible. Note that we only worry about 3010 * writability if mapping is shared; in this case, 3011 * current and max prot are dictated by the open file. 3012 * XXX use the vnode instead? Problem is: what 3013 * credentials do we use for determination? What if 3014 * proc does a setuid? 3015 */ 3016 mp = vp->v_mount; 3017 if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { 3018 maxprot = VM_PROT_NONE; 3019 if ((prot & VM_PROT_EXECUTE) != 0) 3020 return (EACCES); 3021 } else 3022 maxprot = VM_PROT_EXECUTE; 3023 if ((fp->f_flag & FREAD) != 0) 3024 maxprot |= VM_PROT_READ; 3025 else if ((prot & VM_PROT_READ) != 0) 3026 return (EACCES); 3027 3028 /* 3029 * If we are sharing potential changes via MAP_SHARED and we 3030 * are trying to get write permission although we opened it 3031 * without asking for it, bail out. 3032 */ 3033 if ((flags & MAP_SHARED) != 0) { 3034 if ((fp->f_flag & FWRITE) != 0) 3035 maxprot |= VM_PROT_WRITE; 3036 else if ((prot & VM_PROT_WRITE) != 0) 3037 return (EACCES); 3038 } else { 3039 maxprot |= VM_PROT_WRITE; 3040 cap_maxprot |= VM_PROT_WRITE; 3041 } 3042 maxprot &= cap_maxprot; 3043 3044 /* 3045 * For regular files and shared memory, POSIX requires that 3046 * the value of foff be a legitimate offset within the data 3047 * object. In particular, negative offsets are invalid. 3048 * Blocking negative offsets and overflows here avoids 3049 * possible wraparound or user-level access into reserved 3050 * ranges of the data object later. In contrast, POSIX does 3051 * not dictate how offsets are used by device drivers, so in 3052 * the case of a device mapping a negative offset is passed 3053 * on. 3054 */ 3055 if ( 3056 #ifdef _LP64 3057 size > OFF_MAX || 3058 #endif 3059 foff > OFF_MAX - size) 3060 return (EINVAL); 3061 3062 writecounted = FALSE; 3063 error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, 3064 &foff, &object, &writecounted); 3065 if (error != 0) 3066 return (error); 3067 error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, 3068 foff, writecounted, td); 3069 if (error != 0) { 3070 /* 3071 * If this mapping was accounted for in the vnode's 3072 * writecount, then undo that now. 3073 */ 3074 if (writecounted) 3075 vm_pager_release_writecount(object, 0, size); 3076 vm_object_deallocate(object); 3077 } 3078 #ifdef HWPMC_HOOKS 3079 /* Inform hwpmc(4) if an executable is being mapped. */ 3080 if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { 3081 if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { 3082 pkm.pm_file = vp; 3083 pkm.pm_address = (uintptr_t) *addr; 3084 PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); 3085 } 3086 } 3087 #endif 3088 3089 #ifdef HWT_HOOKS 3090 if (HWT_HOOK_INSTALLED && (prot & VM_PROT_EXECUTE) != 0 && 3091 error == 0) { 3092 struct hwt_record_entry ent; 3093 char *fullpath; 3094 char *freepath; 3095 3096 if (vn_fullpath(vp, &fullpath, &freepath) == 0) { 3097 ent.fullpath = fullpath; 3098 ent.addr = (uintptr_t) *addr; 3099 ent.record_type = HWT_RECORD_MMAP; 3100 HWT_CALL_HOOK(td, HWT_MMAP, &ent); 3101 free(freepath, M_TEMP); 3102 } 3103 } 3104 #endif 3105 3106 return (error); 3107 } 3108 3109 void 3110 vn_fsid(struct vnode *vp, struct vattr *va) 3111 { 3112 fsid_t *f; 3113 3114 f = &vp->v_mount->mnt_stat.f_fsid; 3115 va->va_fsid = (uint32_t)f->val[1]; 3116 va->va_fsid <<= sizeof(f->val[1]) * NBBY; 3117 va->va_fsid += (uint32_t)f->val[0]; 3118 } 3119 3120 int 3121 vn_fsync_buf(struct vnode *vp, int waitfor) 3122 { 3123 struct buf *bp, *nbp; 3124 struct bufobj *bo; 3125 struct mount *mp; 3126 int error, maxretry; 3127 3128 error = 0; 3129 maxretry = 10000; /* large, arbitrarily chosen */ 3130 mp = NULL; 3131 if (vp->v_type == VCHR) { 3132 VI_LOCK(vp); 3133 mp = vp->v_rdev->si_mountpt; 3134 VI_UNLOCK(vp); 3135 } 3136 bo = &vp->v_bufobj; 3137 BO_LOCK(bo); 3138 loop1: 3139 /* 3140 * MARK/SCAN initialization to avoid infinite loops. 3141 */ 3142 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { 3143 bp->b_vflags &= ~BV_SCANNED; 3144 bp->b_error = 0; 3145 } 3146 3147 /* 3148 * Flush all dirty buffers associated with a vnode. 3149 */ 3150 loop2: 3151 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { 3152 if ((bp->b_vflags & BV_SCANNED) != 0) 3153 continue; 3154 bp->b_vflags |= BV_SCANNED; 3155 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { 3156 if (waitfor != MNT_WAIT) 3157 continue; 3158 if (BUF_LOCK(bp, 3159 LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, 3160 BO_LOCKPTR(bo)) != 0) { 3161 BO_LOCK(bo); 3162 goto loop1; 3163 } 3164 BO_LOCK(bo); 3165 } 3166 BO_UNLOCK(bo); 3167 KASSERT(bp->b_bufobj == bo, 3168 ("bp %p wrong b_bufobj %p should be %p", 3169 bp, bp->b_bufobj, bo)); 3170 if ((bp->b_flags & B_DELWRI) == 0) 3171 panic("fsync: not dirty"); 3172 if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { 3173 vfs_bio_awrite(bp); 3174 } else { 3175 bremfree(bp); 3176 bawrite(bp); 3177 } 3178 if (maxretry < 1000) 3179 pause("dirty", hz < 1000 ? 1 : hz / 1000); 3180 BO_LOCK(bo); 3181 goto loop2; 3182 } 3183 3184 /* 3185 * If synchronous the caller expects us to completely resolve all 3186 * dirty buffers in the system. Wait for in-progress I/O to 3187 * complete (which could include background bitmap writes), then 3188 * retry if dirty blocks still exist. 3189 */ 3190 if (waitfor == MNT_WAIT) { 3191 bufobj_wwait(bo, 0, 0); 3192 if (bo->bo_dirty.bv_cnt > 0) { 3193 /* 3194 * If we are unable to write any of these buffers 3195 * then we fail now rather than trying endlessly 3196 * to write them out. 3197 */ 3198 TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) 3199 if ((error = bp->b_error) != 0) 3200 break; 3201 if ((mp != NULL && mp->mnt_secondary_writes > 0) || 3202 (error == 0 && --maxretry >= 0)) 3203 goto loop1; 3204 if (error == 0) 3205 error = EAGAIN; 3206 } 3207 } 3208 BO_UNLOCK(bo); 3209 if (error != 0) 3210 vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); 3211 3212 return (error); 3213 } 3214 3215 /* 3216 * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() 3217 * or vn_generic_copy_file_range() after rangelocking the byte ranges, 3218 * to do the actual copy. 3219 * vn_generic_copy_file_range() is factored out, so it can be called 3220 * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from 3221 * different file systems. 3222 */ 3223 int 3224 vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, 3225 off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, 3226 struct ucred *outcred, struct thread *fsize_td) 3227 { 3228 struct mount *inmp, *outmp; 3229 struct vnode *invpl, *outvpl; 3230 int error; 3231 size_t len; 3232 uint64_t uval; 3233 3234 invpl = outvpl = NULL; 3235 len = *lenp; 3236 *lenp = 0; /* For error returns. */ 3237 error = 0; 3238 3239 /* Do some sanity checks on the arguments. */ 3240 if (invp->v_type == VDIR || outvp->v_type == VDIR) 3241 error = EISDIR; 3242 else if (*inoffp < 0 || *outoffp < 0 || 3243 invp->v_type != VREG || outvp->v_type != VREG) 3244 error = EINVAL; 3245 if (error != 0) 3246 goto out; 3247 3248 /* Ensure offset + len does not wrap around. */ 3249 uval = *inoffp; 3250 uval += len; 3251 if (uval > INT64_MAX) 3252 len = INT64_MAX - *inoffp; 3253 uval = *outoffp; 3254 uval += len; 3255 if (uval > INT64_MAX) 3256 len = INT64_MAX - *outoffp; 3257 if (len == 0) 3258 goto out; 3259 3260 error = VOP_GETLOWVNODE(invp, &invpl, FREAD); 3261 if (error != 0) 3262 goto out; 3263 error = VOP_GETLOWVNODE(outvp, &outvpl, FWRITE); 3264 if (error != 0) 3265 goto out1; 3266 3267 inmp = invpl->v_mount; 3268 outmp = outvpl->v_mount; 3269 if (inmp == NULL || outmp == NULL) 3270 goto out2; 3271 3272 for (;;) { 3273 error = vfs_busy(inmp, 0); 3274 if (error != 0) 3275 goto out2; 3276 if (inmp == outmp) 3277 break; 3278 error = vfs_busy(outmp, MBF_NOWAIT); 3279 if (error != 0) { 3280 vfs_unbusy(inmp); 3281 error = vfs_busy(outmp, 0); 3282 if (error == 0) { 3283 vfs_unbusy(outmp); 3284 continue; 3285 } 3286 goto out2; 3287 } 3288 break; 3289 } 3290 3291 /* 3292 * If the two vnodes are for the same file system type, call 3293 * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() 3294 * which can handle copies across multiple file system types. 3295 */ 3296 *lenp = len; 3297 if (inmp == outmp || inmp->mnt_vfc == outmp->mnt_vfc) 3298 error = VOP_COPY_FILE_RANGE(invpl, inoffp, outvpl, outoffp, 3299 lenp, flags, incred, outcred, fsize_td); 3300 else 3301 error = ENOSYS; 3302 if (error == ENOSYS) 3303 error = vn_generic_copy_file_range(invpl, inoffp, outvpl, 3304 outoffp, lenp, flags, incred, outcred, fsize_td); 3305 vfs_unbusy(outmp); 3306 if (inmp != outmp) 3307 vfs_unbusy(inmp); 3308 out2: 3309 if (outvpl != NULL) 3310 vrele(outvpl); 3311 out1: 3312 if (invpl != NULL) 3313 vrele(invpl); 3314 out: 3315 return (error); 3316 } 3317 3318 /* 3319 * Test len bytes of data starting at dat for all bytes == 0. 3320 * Return true if all bytes are zero, false otherwise. 3321 * Expects dat to be well aligned. 3322 */ 3323 static bool 3324 mem_iszero(void *dat, int len) 3325 { 3326 int i; 3327 const u_int *p; 3328 const char *cp; 3329 3330 for (p = dat; len > 0; len -= sizeof(*p), p++) { 3331 if (len >= sizeof(*p)) { 3332 if (*p != 0) 3333 return (false); 3334 } else { 3335 cp = (const char *)p; 3336 for (i = 0; i < len; i++, cp++) 3337 if (*cp != '\0') 3338 return (false); 3339 } 3340 } 3341 return (true); 3342 } 3343 3344 /* 3345 * Look for a hole in the output file and, if found, adjust *outoffp 3346 * and *xferp to skip past the hole. 3347 * *xferp is the entire hole length to be written and xfer2 is how many bytes 3348 * to be written as 0's upon return. 3349 */ 3350 static off_t 3351 vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, 3352 off_t *dataoffp, off_t *holeoffp, struct ucred *cred) 3353 { 3354 int error; 3355 off_t delta; 3356 3357 if (*holeoffp == 0 || *holeoffp <= *outoffp) { 3358 *dataoffp = *outoffp; 3359 error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, 3360 curthread); 3361 if (error == 0) { 3362 *holeoffp = *dataoffp; 3363 error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, 3364 curthread); 3365 } 3366 if (error != 0 || *holeoffp == *dataoffp) { 3367 /* 3368 * Since outvp is unlocked, it may be possible for 3369 * another thread to do a truncate(), lseek(), write() 3370 * creating a hole at startoff between the above 3371 * VOP_IOCTL() calls, if the other thread does not do 3372 * rangelocking. 3373 * If that happens, *holeoffp == *dataoffp and finding 3374 * the hole has failed, so disable vn_skip_hole(). 3375 */ 3376 *holeoffp = -1; /* Disable use of vn_skip_hole(). */ 3377 return (xfer2); 3378 } 3379 KASSERT(*dataoffp >= *outoffp, 3380 ("vn_skip_hole: dataoff=%jd < outoff=%jd", 3381 (intmax_t)*dataoffp, (intmax_t)*outoffp)); 3382 KASSERT(*holeoffp > *dataoffp, 3383 ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", 3384 (intmax_t)*holeoffp, (intmax_t)*dataoffp)); 3385 } 3386 3387 /* 3388 * If there is a hole before the data starts, advance *outoffp and 3389 * *xferp past the hole. 3390 */ 3391 if (*dataoffp > *outoffp) { 3392 delta = *dataoffp - *outoffp; 3393 if (delta >= *xferp) { 3394 /* Entire *xferp is a hole. */ 3395 *outoffp += *xferp; 3396 *xferp = 0; 3397 return (0); 3398 } 3399 *xferp -= delta; 3400 *outoffp += delta; 3401 xfer2 = MIN(xfer2, *xferp); 3402 } 3403 3404 /* 3405 * If a hole starts before the end of this xfer2, reduce this xfer2 so 3406 * that the write ends at the start of the hole. 3407 * *holeoffp should always be greater than *outoffp, but for the 3408 * non-INVARIANTS case, check this to make sure xfer2 remains a sane 3409 * value. 3410 */ 3411 if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) 3412 xfer2 = *holeoffp - *outoffp; 3413 return (xfer2); 3414 } 3415 3416 /* 3417 * Write an xfer sized chunk to outvp in blksize blocks from dat. 3418 * dat is a maximum of blksize in length and can be written repeatedly in 3419 * the chunk. 3420 * If growfile == true, just grow the file via vn_truncate_locked() instead 3421 * of doing actual writes. 3422 * If checkhole == true, a hole is being punched, so skip over any hole 3423 * already in the output file. 3424 */ 3425 static int 3426 vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, 3427 u_long blksize, bool growfile, bool checkhole, struct ucred *cred) 3428 { 3429 struct mount *mp; 3430 off_t dataoff, holeoff, xfer2; 3431 int error; 3432 3433 /* 3434 * Loop around doing writes of blksize until write has been completed. 3435 * Lock/unlock on each loop iteration so that a bwillwrite() can be 3436 * done for each iteration, since the xfer argument can be very 3437 * large if there is a large hole to punch in the output file. 3438 */ 3439 error = 0; 3440 holeoff = 0; 3441 do { 3442 xfer2 = MIN(xfer, blksize); 3443 if (checkhole) { 3444 /* 3445 * Punching a hole. Skip writing if there is 3446 * already a hole in the output file. 3447 */ 3448 xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, 3449 &dataoff, &holeoff, cred); 3450 if (xfer == 0) 3451 break; 3452 if (holeoff < 0) 3453 checkhole = false; 3454 KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", 3455 (intmax_t)xfer2)); 3456 } 3457 bwillwrite(); 3458 mp = NULL; 3459 error = vn_start_write(outvp, &mp, V_WAIT); 3460 if (error != 0) 3461 break; 3462 if (growfile) { 3463 error = vn_lock(outvp, LK_EXCLUSIVE); 3464 if (error == 0) { 3465 error = vn_truncate_locked(outvp, outoff + xfer, 3466 false, cred); 3467 VOP_UNLOCK(outvp); 3468 } 3469 } else { 3470 error = vn_lock(outvp, vn_lktype_write(mp, outvp)); 3471 if (error == 0) { 3472 error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, 3473 outoff, UIO_SYSSPACE, IO_NODELOCKED, 3474 curthread->td_ucred, cred, NULL, curthread); 3475 outoff += xfer2; 3476 xfer -= xfer2; 3477 VOP_UNLOCK(outvp); 3478 } 3479 } 3480 if (mp != NULL) 3481 vn_finished_write(mp); 3482 } while (!growfile && xfer > 0 && error == 0); 3483 return (error); 3484 } 3485 3486 /* 3487 * Copy a byte range of one file to another. This function can handle the 3488 * case where invp and outvp are on different file systems. 3489 * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there 3490 * is no better file system specific way to do it. 3491 */ 3492 int 3493 vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, 3494 struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, 3495 struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) 3496 { 3497 struct vattr inva; 3498 struct mount *mp; 3499 off_t startoff, endoff, xfer, xfer2; 3500 u_long blksize; 3501 int error, interrupted; 3502 bool cantseek, readzeros, eof, first, lastblock, holetoeof, sparse; 3503 ssize_t aresid, r = 0; 3504 size_t copylen, len, savlen; 3505 off_t outsize; 3506 char *dat; 3507 long holein, holeout; 3508 struct timespec curts, endts; 3509 3510 holein = holeout = 0; 3511 savlen = len = *lenp; 3512 error = 0; 3513 interrupted = 0; 3514 dat = NULL; 3515 3516 if ((flags & COPY_FILE_RANGE_CLONE) != 0) { 3517 error = EOPNOTSUPP; 3518 goto out; 3519 } 3520 3521 error = vn_lock(invp, LK_SHARED); 3522 if (error != 0) 3523 goto out; 3524 if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) 3525 holein = 0; 3526 error = VOP_GETATTR(invp, &inva, incred); 3527 if (error == 0 && inva.va_size > OFF_MAX) 3528 error = EFBIG; 3529 VOP_UNLOCK(invp); 3530 if (error != 0) 3531 goto out; 3532 3533 /* 3534 * Use va_bytes >= va_size as a hint that the file does not have 3535 * sufficient holes to justify the overhead of doing FIOSEEKHOLE. 3536 * This hint does not work well for file systems doing compression 3537 * and may fail when allocations for extended attributes increases 3538 * the value of va_bytes to >= va_size. 3539 */ 3540 sparse = true; 3541 if (holein != 0 && inva.va_bytes >= inva.va_size) { 3542 holein = 0; 3543 sparse = false; 3544 } 3545 3546 mp = NULL; 3547 error = vn_start_write(outvp, &mp, V_WAIT); 3548 if (error == 0) 3549 error = vn_lock(outvp, LK_EXCLUSIVE); 3550 if (error == 0) { 3551 /* 3552 * If fsize_td != NULL, do a vn_rlimit_fsizex() call, 3553 * now that outvp is locked. 3554 */ 3555 if (fsize_td != NULL) { 3556 struct uio io; 3557 3558 io.uio_offset = *outoffp; 3559 io.uio_resid = len; 3560 error = vn_rlimit_fsizex(outvp, &io, 0, &r, fsize_td); 3561 len = savlen = io.uio_resid; 3562 /* 3563 * No need to call vn_rlimit_fsizex_res before return, 3564 * since the uio is local. 3565 */ 3566 } 3567 if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) 3568 holeout = 0; 3569 /* 3570 * Holes that are past EOF do not need to be written as a block 3571 * of zero bytes. So, truncate the output file as far as 3572 * possible and then use size to decide if writing 0 3573 * bytes is necessary in the loop below. 3574 */ 3575 if (error == 0) 3576 error = vn_getsize_locked(outvp, &outsize, outcred); 3577 if (error == 0 && outsize > *outoffp && 3578 *outoffp <= OFF_MAX - len && outsize <= *outoffp + len && 3579 *inoffp < inva.va_size && 3580 *outoffp <= OFF_MAX - (inva.va_size - *inoffp) && 3581 outsize <= *outoffp + (inva.va_size - *inoffp)) { 3582 #ifdef MAC 3583 error = mac_vnode_check_write(curthread->td_ucred, 3584 outcred, outvp); 3585 if (error == 0) 3586 #endif 3587 error = vn_truncate_locked(outvp, *outoffp, 3588 false, outcred); 3589 if (error == 0) 3590 outsize = *outoffp; 3591 } 3592 VOP_UNLOCK(outvp); 3593 } 3594 if (mp != NULL) 3595 vn_finished_write(mp); 3596 if (error != 0) 3597 goto out; 3598 3599 if (sparse && holein == 0 && holeout > 0) { 3600 /* 3601 * For this special case, the input data will be scanned 3602 * for blocks of all 0 bytes. For these blocks, the 3603 * write can be skipped for the output file to create 3604 * an unallocated region. 3605 * Therefore, use the appropriate size for the output file. 3606 */ 3607 blksize = holeout; 3608 if (blksize <= 512) { 3609 /* 3610 * Use f_iosize, since ZFS reports a _PC_MIN_HOLE_SIZE 3611 * of 512, although it actually only creates 3612 * unallocated regions for blocks >= f_iosize. 3613 */ 3614 blksize = outvp->v_mount->mnt_stat.f_iosize; 3615 } 3616 } else { 3617 /* 3618 * Use the larger of the two f_iosize values. If they are 3619 * not the same size, one will normally be an exact multiple of 3620 * the other, since they are both likely to be a power of 2. 3621 */ 3622 blksize = MAX(invp->v_mount->mnt_stat.f_iosize, 3623 outvp->v_mount->mnt_stat.f_iosize); 3624 } 3625 3626 /* Clip to sane limits. */ 3627 if (blksize < 4096) 3628 blksize = 4096; 3629 else if (blksize > maxphys) 3630 blksize = maxphys; 3631 dat = malloc(blksize, M_TEMP, M_WAITOK); 3632 3633 /* 3634 * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA 3635 * to find holes. Otherwise, just scan the read block for all 0s 3636 * in the inner loop where the data copying is done. 3637 * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may 3638 * support holes on the server, but do not support FIOSEEKHOLE. 3639 * The kernel flag COPY_FILE_RANGE_TIMEO1SEC is used to indicate 3640 * that this function should return after 1second with a partial 3641 * completion. 3642 */ 3643 if ((flags & COPY_FILE_RANGE_TIMEO1SEC) != 0) { 3644 getnanouptime(&endts); 3645 endts.tv_sec++; 3646 } else 3647 timespecclear(&endts); 3648 first = true; 3649 holetoeof = eof = false; 3650 while (len > 0 && error == 0 && !eof && interrupted == 0) { 3651 endoff = 0; /* To shut up compilers. */ 3652 cantseek = true; 3653 startoff = *inoffp; 3654 copylen = len; 3655 3656 /* 3657 * Find the next data area. If there is just a hole to EOF, 3658 * FIOSEEKDATA should fail with ENXIO. 3659 * (I do not know if any file system will report a hole to 3660 * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA 3661 * will fail for those file systems.) 3662 * 3663 * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, 3664 * the code just falls through to the inner copy loop. 3665 */ 3666 error = EINVAL; 3667 if (holein > 0) { 3668 error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, 3669 incred, curthread); 3670 if (error == ENXIO) { 3671 startoff = endoff = inva.va_size; 3672 eof = holetoeof = true; 3673 error = 0; 3674 } 3675 } 3676 if (error == 0 && !holetoeof) { 3677 endoff = startoff; 3678 error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, 3679 incred, curthread); 3680 /* 3681 * Since invp is unlocked, it may be possible for 3682 * another thread to do a truncate(), lseek(), write() 3683 * creating a hole at startoff between the above 3684 * VOP_IOCTL() calls, if the other thread does not do 3685 * rangelocking. 3686 * If that happens, startoff == endoff and finding 3687 * the hole has failed, so set an error. 3688 */ 3689 if (error == 0 && startoff == endoff) 3690 error = EINVAL; /* Any error. Reset to 0. */ 3691 } 3692 if (error == 0) { 3693 if (startoff > *inoffp) { 3694 /* Found hole before data block. */ 3695 xfer = MIN(startoff - *inoffp, len); 3696 if (*outoffp < outsize) { 3697 /* Must write 0s to punch hole. */ 3698 xfer2 = MIN(outsize - *outoffp, 3699 xfer); 3700 memset(dat, 0, MIN(xfer2, blksize)); 3701 error = vn_write_outvp(outvp, dat, 3702 *outoffp, xfer2, blksize, false, 3703 holeout > 0, outcred); 3704 } 3705 3706 if (error == 0 && *outoffp + xfer > 3707 outsize && (xfer == len || holetoeof)) { 3708 /* Grow output file (hole at end). */ 3709 error = vn_write_outvp(outvp, dat, 3710 *outoffp, xfer, blksize, true, 3711 false, outcred); 3712 } 3713 if (error == 0) { 3714 *inoffp += xfer; 3715 *outoffp += xfer; 3716 len -= xfer; 3717 if (len < savlen) { 3718 interrupted = sig_intr(); 3719 if (timespecisset(&endts) && 3720 interrupted == 0) { 3721 getnanouptime(&curts); 3722 if (timespeccmp(&curts, 3723 &endts, >=)) 3724 interrupted = 3725 EINTR; 3726 } 3727 } 3728 } 3729 } 3730 copylen = MIN(len, endoff - startoff); 3731 cantseek = false; 3732 } else { 3733 cantseek = true; 3734 if (!sparse) 3735 cantseek = false; 3736 startoff = *inoffp; 3737 copylen = len; 3738 error = 0; 3739 } 3740 3741 xfer = blksize; 3742 if (cantseek) { 3743 /* 3744 * Set first xfer to end at a block boundary, so that 3745 * holes are more likely detected in the loop below via 3746 * the for all bytes 0 method. 3747 */ 3748 xfer -= (*inoffp % blksize); 3749 } 3750 3751 /* 3752 * Loop copying the data block. If this was our first attempt 3753 * to copy anything, allow a zero-length block so that the VOPs 3754 * get a chance to update metadata, specifically the atime. 3755 */ 3756 while (error == 0 && ((copylen > 0 && !eof) || first) && 3757 interrupted == 0) { 3758 if (copylen < xfer) 3759 xfer = copylen; 3760 first = false; 3761 error = vn_lock(invp, LK_SHARED); 3762 if (error != 0) 3763 goto out; 3764 error = vn_rdwr(UIO_READ, invp, dat, xfer, 3765 startoff, UIO_SYSSPACE, IO_NODELOCKED, 3766 curthread->td_ucred, incred, &aresid, 3767 curthread); 3768 VOP_UNLOCK(invp); 3769 lastblock = false; 3770 if (error == 0 && (xfer == 0 || aresid > 0)) { 3771 /* Stop the copy at EOF on the input file. */ 3772 xfer -= aresid; 3773 eof = true; 3774 lastblock = true; 3775 } 3776 if (error == 0) { 3777 /* 3778 * Skip the write for holes past the initial EOF 3779 * of the output file, unless this is the last 3780 * write of the output file at EOF. 3781 */ 3782 readzeros = cantseek ? mem_iszero(dat, xfer) : 3783 false; 3784 if (xfer == len) 3785 lastblock = true; 3786 if (!cantseek || *outoffp < outsize || 3787 lastblock || !readzeros) 3788 error = vn_write_outvp(outvp, dat, 3789 *outoffp, xfer, blksize, 3790 readzeros && lastblock && 3791 *outoffp >= outsize, false, 3792 outcred); 3793 if (error == 0) { 3794 *inoffp += xfer; 3795 startoff += xfer; 3796 *outoffp += xfer; 3797 copylen -= xfer; 3798 len -= xfer; 3799 if (len < savlen) { 3800 interrupted = sig_intr(); 3801 if (timespecisset(&endts) && 3802 interrupted == 0) { 3803 getnanouptime(&curts); 3804 if (timespeccmp(&curts, 3805 &endts, >=)) 3806 interrupted = 3807 EINTR; 3808 } 3809 } 3810 } 3811 } 3812 xfer = blksize; 3813 } 3814 } 3815 out: 3816 *lenp = savlen - len; 3817 free(dat, M_TEMP); 3818 return (error); 3819 } 3820 3821 static int 3822 vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) 3823 { 3824 struct mount *mp; 3825 struct vnode *vp; 3826 off_t olen, ooffset; 3827 int error; 3828 #ifdef AUDIT 3829 int audited_vnode1 = 0; 3830 #endif 3831 3832 vp = fp->f_vnode; 3833 if (vp->v_type != VREG) 3834 return (ENODEV); 3835 3836 /* Allocating blocks may take a long time, so iterate. */ 3837 for (;;) { 3838 olen = len; 3839 ooffset = offset; 3840 3841 bwillwrite(); 3842 mp = NULL; 3843 error = vn_start_write(vp, &mp, V_WAIT | V_PCATCH); 3844 if (error != 0) 3845 break; 3846 error = vn_lock(vp, LK_EXCLUSIVE); 3847 if (error != 0) { 3848 vn_finished_write(mp); 3849 break; 3850 } 3851 #ifdef AUDIT 3852 if (!audited_vnode1) { 3853 AUDIT_ARG_VNODE1(vp); 3854 audited_vnode1 = 1; 3855 } 3856 #endif 3857 #ifdef MAC 3858 error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); 3859 if (error == 0) 3860 #endif 3861 error = VOP_ALLOCATE(vp, &offset, &len, 0, 3862 td->td_ucred); 3863 VOP_UNLOCK(vp); 3864 vn_finished_write(mp); 3865 3866 if (olen + ooffset != offset + len) { 3867 panic("offset + len changed from %jx/%jx to %jx/%jx", 3868 ooffset, olen, offset, len); 3869 } 3870 if (error != 0 || len == 0) 3871 break; 3872 KASSERT(olen > len, ("Iteration did not make progress?")); 3873 maybe_yield(); 3874 } 3875 3876 return (error); 3877 } 3878 3879 static int 3880 vn_deallocate_impl(struct vnode *vp, off_t *offset, off_t *length, int flags, 3881 int ioflag, struct ucred *cred, struct ucred *active_cred, 3882 struct ucred *file_cred) 3883 { 3884 struct mount *mp; 3885 void *rl_cookie; 3886 off_t off, len; 3887 int error; 3888 #ifdef AUDIT 3889 bool audited_vnode1 = false; 3890 #endif 3891 3892 rl_cookie = NULL; 3893 error = 0; 3894 mp = NULL; 3895 off = *offset; 3896 len = *length; 3897 3898 if ((ioflag & (IO_NODELOCKED | IO_RANGELOCKED)) == 0) 3899 rl_cookie = vn_rangelock_wlock(vp, off, off + len); 3900 while (len > 0 && error == 0) { 3901 /* 3902 * Try to deallocate the longest range in one pass. 3903 * In case a pass takes too long to be executed, it returns 3904 * partial result. The residue will be proceeded in the next 3905 * pass. 3906 */ 3907 3908 if ((ioflag & IO_NODELOCKED) == 0) { 3909 bwillwrite(); 3910 if ((error = vn_start_write(vp, &mp, 3911 V_WAIT | V_PCATCH)) != 0) 3912 goto out; 3913 vn_lock(vp, vn_lktype_write(mp, vp) | LK_RETRY); 3914 } 3915 #ifdef AUDIT 3916 if (!audited_vnode1) { 3917 AUDIT_ARG_VNODE1(vp); 3918 audited_vnode1 = true; 3919 } 3920 #endif 3921 3922 #ifdef MAC 3923 if ((ioflag & IO_NOMACCHECK) == 0) 3924 error = mac_vnode_check_write(active_cred, file_cred, 3925 vp); 3926 #endif 3927 if (error == 0) 3928 error = VOP_DEALLOCATE(vp, &off, &len, flags, ioflag, 3929 cred); 3930 3931 if ((ioflag & IO_NODELOCKED) == 0) { 3932 VOP_UNLOCK(vp); 3933 if (mp != NULL) { 3934 vn_finished_write(mp); 3935 mp = NULL; 3936 } 3937 } 3938 if (error == 0 && len != 0) 3939 maybe_yield(); 3940 } 3941 out: 3942 if (rl_cookie != NULL) 3943 vn_rangelock_unlock(vp, rl_cookie); 3944 *offset = off; 3945 *length = len; 3946 return (error); 3947 } 3948 3949 /* 3950 * This function is supposed to be used in the situations where the deallocation 3951 * is not triggered by a user request. 3952 */ 3953 int 3954 vn_deallocate(struct vnode *vp, off_t *offset, off_t *length, int flags, 3955 int ioflag, struct ucred *active_cred, struct ucred *file_cred) 3956 { 3957 struct ucred *cred; 3958 3959 if (*offset < 0 || *length <= 0 || *length > OFF_MAX - *offset || 3960 flags != 0) 3961 return (EINVAL); 3962 if (vp->v_type != VREG) 3963 return (ENODEV); 3964 3965 cred = file_cred != NOCRED ? file_cred : active_cred; 3966 return (vn_deallocate_impl(vp, offset, length, flags, ioflag, cred, 3967 active_cred, file_cred)); 3968 } 3969 3970 static int 3971 vn_fspacectl(struct file *fp, int cmd, off_t *offset, off_t *length, int flags, 3972 struct ucred *active_cred, struct thread *td) 3973 { 3974 int error; 3975 struct vnode *vp; 3976 int ioflag; 3977 3978 KASSERT(cmd == SPACECTL_DEALLOC, ("vn_fspacectl: Invalid cmd")); 3979 KASSERT((flags & ~SPACECTL_F_SUPPORTED) == 0, 3980 ("vn_fspacectl: non-zero flags")); 3981 KASSERT(*offset >= 0 && *length > 0 && *length <= OFF_MAX - *offset, 3982 ("vn_fspacectl: offset/length overflow or underflow")); 3983 vp = fp->f_vnode; 3984 3985 if (vp->v_type != VREG) 3986 return (ENODEV); 3987 3988 ioflag = get_write_ioflag(fp); 3989 3990 switch (cmd) { 3991 case SPACECTL_DEALLOC: 3992 error = vn_deallocate_impl(vp, offset, length, flags, ioflag, 3993 active_cred, active_cred, fp->f_cred); 3994 break; 3995 default: 3996 panic("vn_fspacectl: unknown cmd %d", cmd); 3997 } 3998 3999 return (error); 4000 } 4001 4002 /* 4003 * Keep this assert as long as sizeof(struct dirent) is used as the maximum 4004 * entry size. 4005 */ 4006 _Static_assert(_GENERIC_MAXDIRSIZ == sizeof(struct dirent), 4007 "'struct dirent' size must be a multiple of its alignment " 4008 "(see _GENERIC_DIRLEN())"); 4009 4010 /* 4011 * Returns successive directory entries through some caller's provided buffer. 4012 * 4013 * This function automatically refills the provided buffer with calls to 4014 * VOP_READDIR() (after MAC permission checks). 4015 * 4016 * 'td' is used for credentials and passed to uiomove(). 'dirbuf' is the 4017 * caller's buffer to fill and 'dirbuflen' its allocated size. 'dirbuf' must 4018 * be properly aligned to access 'struct dirent' structures and 'dirbuflen' 4019 * must be greater than GENERIC_MAXDIRSIZ to avoid VOP_READDIR() returning 4020 * EINVAL (the latter is not a strong guarantee (yet); but EINVAL will always 4021 * be returned if this requirement is not verified). '*dpp' points to the 4022 * current directory entry in the buffer and '*len' contains the remaining 4023 * valid bytes in 'dirbuf' after 'dpp' (including the pointed entry). 4024 * 4025 * At first call (or when restarting the read), '*len' must have been set to 0, 4026 * '*off' to 0 (or any valid start offset) and '*eofflag' to 0. There are no 4027 * more entries as soon as '*len' is 0 after a call that returned 0. Calling 4028 * again this function after such a condition is considered an error and EINVAL 4029 * will be returned. Other possible error codes are those of VOP_READDIR(), 4030 * EINTEGRITY if the returned entries do not pass coherency tests, or EINVAL 4031 * (bad call). All errors are unrecoverable, i.e., the state ('*len', '*off' 4032 * and '*eofflag') must be re-initialized before a subsequent call. On error 4033 * or at end of directory, '*dpp' is reset to NULL. 4034 * 4035 * '*len', '*off' and '*eofflag' are internal state the caller should not 4036 * tamper with except as explained above. '*off' is the next directory offset 4037 * to read from to refill the buffer. '*eofflag' is set to 0 or 1 by the last 4038 * internal call to VOP_READDIR() that returned without error, indicating 4039 * whether it reached the end of the directory, and to 2 by this function after 4040 * all entries have been read. 4041 */ 4042 int 4043 vn_dir_next_dirent(struct vnode *vp, struct thread *td, 4044 char *dirbuf, size_t dirbuflen, 4045 struct dirent **dpp, size_t *len, off_t *off, int *eofflag) 4046 { 4047 struct dirent *dp = NULL; 4048 int reclen; 4049 int error; 4050 struct uio uio; 4051 struct iovec iov; 4052 4053 ASSERT_VOP_LOCKED(vp, "vnode not locked"); 4054 VNASSERT(vp->v_type == VDIR, vp, ("vnode is not a directory")); 4055 MPASS2((uintptr_t)dirbuf < (uintptr_t)dirbuf + dirbuflen, 4056 "Address space overflow"); 4057 4058 if (__predict_false(dirbuflen < GENERIC_MAXDIRSIZ)) { 4059 /* Don't take any chances in this case */ 4060 error = EINVAL; 4061 goto out; 4062 } 4063 4064 if (*len != 0) { 4065 dp = *dpp; 4066 4067 /* 4068 * The caller continued to call us after an error (we set dp to 4069 * NULL in a previous iteration). Bail out right now. 4070 */ 4071 if (__predict_false(dp == NULL)) 4072 return (EINVAL); 4073 4074 MPASS(*len <= dirbuflen); 4075 MPASS2((uintptr_t)dirbuf <= (uintptr_t)dp && 4076 (uintptr_t)dp + *len <= (uintptr_t)dirbuf + dirbuflen, 4077 "Filled range not inside buffer"); 4078 4079 reclen = dp->d_reclen; 4080 if (reclen >= *len) { 4081 /* End of buffer reached */ 4082 *len = 0; 4083 } else { 4084 dp = (struct dirent *)((char *)dp + reclen); 4085 *len -= reclen; 4086 } 4087 } 4088 4089 if (*len == 0) { 4090 dp = NULL; 4091 4092 /* Have to refill. */ 4093 switch (*eofflag) { 4094 case 0: 4095 break; 4096 4097 case 1: 4098 /* Nothing more to read. */ 4099 *eofflag = 2; /* Remember the caller reached EOF. */ 4100 goto success; 4101 4102 default: 4103 /* The caller didn't test for EOF. */ 4104 error = EINVAL; 4105 goto out; 4106 } 4107 4108 iov.iov_base = dirbuf; 4109 iov.iov_len = dirbuflen; 4110 4111 uio.uio_iov = &iov; 4112 uio.uio_iovcnt = 1; 4113 uio.uio_offset = *off; 4114 uio.uio_resid = dirbuflen; 4115 uio.uio_segflg = UIO_SYSSPACE; 4116 uio.uio_rw = UIO_READ; 4117 uio.uio_td = td; 4118 4119 #ifdef MAC 4120 error = mac_vnode_check_readdir(td->td_ucred, vp); 4121 if (error == 0) 4122 #endif 4123 error = VOP_READDIR(vp, &uio, td->td_ucred, eofflag, 4124 NULL, NULL); 4125 if (error != 0) 4126 goto out; 4127 4128 *len = dirbuflen - uio.uio_resid; 4129 *off = uio.uio_offset; 4130 4131 if (*len == 0) { 4132 /* Sanity check on INVARIANTS. */ 4133 MPASS(*eofflag != 0); 4134 *eofflag = 1; 4135 goto success; 4136 } 4137 4138 /* 4139 * Normalize the flag returned by VOP_READDIR(), since we use 2 4140 * as a sentinel value. 4141 */ 4142 if (*eofflag != 0) 4143 *eofflag = 1; 4144 4145 dp = (struct dirent *)dirbuf; 4146 } 4147 4148 if (__predict_false(*len < GENERIC_MINDIRSIZ || 4149 dp->d_reclen < GENERIC_MINDIRSIZ)) { 4150 error = EINTEGRITY; 4151 dp = NULL; 4152 goto out; 4153 } 4154 4155 success: 4156 error = 0; 4157 out: 4158 *dpp = dp; 4159 return (error); 4160 } 4161 4162 /* 4163 * Checks whether a directory is empty or not. 4164 * 4165 * If the directory is empty, returns 0, and if it is not, ENOTEMPTY. Other 4166 * values are genuine errors preventing the check. 4167 */ 4168 int 4169 vn_dir_check_empty(struct vnode *vp) 4170 { 4171 struct thread *const td = curthread; 4172 char *dirbuf; 4173 size_t dirbuflen, len; 4174 off_t off; 4175 int eofflag, error; 4176 struct dirent *dp; 4177 struct vattr va; 4178 4179 ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); 4180 VNPASS(vp->v_type == VDIR, vp); 4181 4182 error = VOP_GETATTR(vp, &va, td->td_ucred); 4183 if (error != 0) 4184 return (error); 4185 4186 dirbuflen = max(DEV_BSIZE, GENERIC_MAXDIRSIZ); 4187 if (dirbuflen < va.va_blocksize) 4188 dirbuflen = va.va_blocksize; 4189 dirbuf = malloc(dirbuflen, M_TEMP, M_WAITOK); 4190 4191 len = 0; 4192 off = 0; 4193 eofflag = 0; 4194 4195 for (;;) { 4196 error = vn_dir_next_dirent(vp, td, dirbuf, dirbuflen, 4197 &dp, &len, &off, &eofflag); 4198 if (error != 0) 4199 goto end; 4200 4201 if (len == 0) { 4202 /* EOF */ 4203 error = 0; 4204 goto end; 4205 } 4206 4207 /* 4208 * Skip whiteouts. Unionfs operates on filesystems only and 4209 * not on hierarchies, so these whiteouts would be shadowed on 4210 * the system hierarchy but not for a union using the 4211 * filesystem of their directories as the upper layer. 4212 * Additionally, unionfs currently transparently exposes 4213 * union-specific metadata of its upper layer, meaning that 4214 * whiteouts can be seen through the union view in empty 4215 * directories. Taking into account these whiteouts would then 4216 * prevent mounting another filesystem on such effectively 4217 * empty directories. 4218 */ 4219 if (dp->d_type == DT_WHT) 4220 continue; 4221 4222 /* 4223 * Any file in the directory which is not '.' or '..' indicates 4224 * the directory is not empty. 4225 */ 4226 switch (dp->d_namlen) { 4227 case 2: 4228 if (dp->d_name[1] != '.') { 4229 /* Can't be '..' (nor '.') */ 4230 error = ENOTEMPTY; 4231 goto end; 4232 } 4233 /* FALLTHROUGH */ 4234 case 1: 4235 if (dp->d_name[0] != '.') { 4236 /* Can't be '..' nor '.' */ 4237 error = ENOTEMPTY; 4238 goto end; 4239 } 4240 break; 4241 4242 default: 4243 error = ENOTEMPTY; 4244 goto end; 4245 } 4246 } 4247 4248 end: 4249 free(dirbuf, M_TEMP); 4250 return (error); 4251 } 4252 4253 4254 static u_long vn_lock_pair_pause_cnt; 4255 SYSCTL_ULONG(_debug, OID_AUTO, vn_lock_pair_pause, CTLFLAG_RD, 4256 &vn_lock_pair_pause_cnt, 0, 4257 "Count of vn_lock_pair deadlocks"); 4258 4259 u_int vn_lock_pair_pause_max; 4260 SYSCTL_UINT(_debug, OID_AUTO, vn_lock_pair_pause_max, CTLFLAG_RW, 4261 &vn_lock_pair_pause_max, 0, 4262 "Max ticks for vn_lock_pair deadlock avoidance sleep"); 4263 4264 static void 4265 vn_lock_pair_pause(const char *wmesg) 4266 { 4267 atomic_add_long(&vn_lock_pair_pause_cnt, 1); 4268 pause(wmesg, prng32_bounded(vn_lock_pair_pause_max)); 4269 } 4270 4271 /* 4272 * Lock pair of (possibly same) vnodes vp1, vp2, avoiding lock order 4273 * reversal. vp1_locked indicates whether vp1 is locked; if not, vp1 4274 * must be unlocked. Same for vp2 and vp2_locked. One of the vnodes 4275 * can be NULL. 4276 * 4277 * The function returns with both vnodes exclusively or shared locked, 4278 * according to corresponding lkflags, and guarantees that it does not 4279 * create lock order reversal with other threads during its execution. 4280 * Both vnodes could be unlocked temporary (and reclaimed). 4281 * 4282 * If requesting shared locking, locked vnode lock must not be recursed. 4283 * 4284 * Only one of LK_SHARED and LK_EXCLUSIVE must be specified. 4285 * LK_NODDLKTREAT can be optionally passed. 4286 * 4287 * If vp1 == vp2, only one, most exclusive, lock is obtained on it. 4288 */ 4289 void 4290 vn_lock_pair(struct vnode *vp1, bool vp1_locked, int lkflags1, 4291 struct vnode *vp2, bool vp2_locked, int lkflags2) 4292 { 4293 int error, locked1; 4294 4295 MPASS((((lkflags1 & LK_SHARED) != 0) ^ ((lkflags1 & LK_EXCLUSIVE) != 0)) || 4296 (vp1 == NULL && lkflags1 == 0)); 4297 MPASS((lkflags1 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0); 4298 MPASS((((lkflags2 & LK_SHARED) != 0) ^ ((lkflags2 & LK_EXCLUSIVE) != 0)) || 4299 (vp2 == NULL && lkflags2 == 0)); 4300 MPASS((lkflags2 & ~(LK_SHARED | LK_EXCLUSIVE | LK_NODDLKTREAT)) == 0); 4301 4302 if (vp1 == NULL && vp2 == NULL) 4303 return; 4304 4305 if (vp1 == vp2) { 4306 MPASS(vp1_locked == vp2_locked); 4307 4308 /* Select the most exclusive mode for lock. */ 4309 if ((lkflags1 & LK_TYPE_MASK) != (lkflags2 & LK_TYPE_MASK)) 4310 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE; 4311 4312 if (vp1_locked) { 4313 ASSERT_VOP_LOCKED(vp1, "vp1"); 4314 4315 /* No need to relock if any lock is exclusive. */ 4316 if ((vp1->v_vnlock->lock_object.lo_flags & 4317 LK_NOSHARE) != 0) 4318 return; 4319 4320 locked1 = VOP_ISLOCKED(vp1); 4321 if (((lkflags1 & LK_SHARED) != 0 && 4322 locked1 != LK_EXCLUSIVE) || 4323 ((lkflags1 & LK_EXCLUSIVE) != 0 && 4324 locked1 == LK_EXCLUSIVE)) 4325 return; 4326 VOP_UNLOCK(vp1); 4327 } 4328 4329 ASSERT_VOP_UNLOCKED(vp1, "vp1"); 4330 vn_lock(vp1, lkflags1 | LK_RETRY); 4331 return; 4332 } 4333 4334 if (vp1 != NULL) { 4335 if ((lkflags1 & LK_SHARED) != 0 && 4336 (vp1->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0) 4337 lkflags1 = (lkflags1 & ~LK_SHARED) | LK_EXCLUSIVE; 4338 if (vp1_locked && VOP_ISLOCKED(vp1) != LK_EXCLUSIVE) { 4339 ASSERT_VOP_LOCKED(vp1, "vp1"); 4340 if ((lkflags1 & LK_EXCLUSIVE) != 0) { 4341 VOP_UNLOCK(vp1); 4342 ASSERT_VOP_UNLOCKED(vp1, 4343 "vp1 shared recursed"); 4344 vp1_locked = false; 4345 } 4346 } else if (!vp1_locked) 4347 ASSERT_VOP_UNLOCKED(vp1, "vp1"); 4348 } else { 4349 vp1_locked = true; 4350 } 4351 4352 if (vp2 != NULL) { 4353 if ((lkflags2 & LK_SHARED) != 0 && 4354 (vp2->v_vnlock->lock_object.lo_flags & LK_NOSHARE) != 0) 4355 lkflags2 = (lkflags2 & ~LK_SHARED) | LK_EXCLUSIVE; 4356 if (vp2_locked && VOP_ISLOCKED(vp2) != LK_EXCLUSIVE) { 4357 ASSERT_VOP_LOCKED(vp2, "vp2"); 4358 if ((lkflags2 & LK_EXCLUSIVE) != 0) { 4359 VOP_UNLOCK(vp2); 4360 ASSERT_VOP_UNLOCKED(vp2, 4361 "vp2 shared recursed"); 4362 vp2_locked = false; 4363 } 4364 } else if (!vp2_locked) 4365 ASSERT_VOP_UNLOCKED(vp2, "vp2"); 4366 } else { 4367 vp2_locked = true; 4368 } 4369 4370 if (!vp1_locked && !vp2_locked) { 4371 vn_lock(vp1, lkflags1 | LK_RETRY); 4372 vp1_locked = true; 4373 } 4374 4375 while (!vp1_locked || !vp2_locked) { 4376 if (vp1_locked && vp2 != NULL) { 4377 if (vp1 != NULL) { 4378 error = VOP_LOCK1(vp2, lkflags2 | LK_NOWAIT, 4379 __FILE__, __LINE__); 4380 if (error == 0) 4381 break; 4382 VOP_UNLOCK(vp1); 4383 vp1_locked = false; 4384 vn_lock_pair_pause("vlp1"); 4385 } 4386 vn_lock(vp2, lkflags2 | LK_RETRY); 4387 vp2_locked = true; 4388 } 4389 if (vp2_locked && vp1 != NULL) { 4390 if (vp2 != NULL) { 4391 error = VOP_LOCK1(vp1, lkflags1 | LK_NOWAIT, 4392 __FILE__, __LINE__); 4393 if (error == 0) 4394 break; 4395 VOP_UNLOCK(vp2); 4396 vp2_locked = false; 4397 vn_lock_pair_pause("vlp2"); 4398 } 4399 vn_lock(vp1, lkflags1 | LK_RETRY); 4400 vp1_locked = true; 4401 } 4402 } 4403 if (vp1 != NULL) { 4404 if (lkflags1 == LK_EXCLUSIVE) 4405 ASSERT_VOP_ELOCKED(vp1, "vp1 ret"); 4406 else 4407 ASSERT_VOP_LOCKED(vp1, "vp1 ret"); 4408 } 4409 if (vp2 != NULL) { 4410 if (lkflags2 == LK_EXCLUSIVE) 4411 ASSERT_VOP_ELOCKED(vp2, "vp2 ret"); 4412 else 4413 ASSERT_VOP_LOCKED(vp2, "vp2 ret"); 4414 } 4415 } 4416 4417 int 4418 vn_lktype_write(struct mount *mp, struct vnode *vp) 4419 { 4420 if (MNT_SHARED_WRITES(mp) || 4421 (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) 4422 return (LK_SHARED); 4423 return (LK_EXCLUSIVE); 4424 } 4425 4426 int 4427 vn_cmp(struct file *fp1, struct file *fp2, struct thread *td) 4428 { 4429 if (fp2->f_type != DTYPE_VNODE) 4430 return (3); 4431 return (kcmp_cmp((uintptr_t)fp1->f_vnode, (uintptr_t)fp2->f_vnode)); 4432 } 4433